mirror of https://github.com/vladmandic/human
9309 lines
1.5 MiB
9309 lines
1.5 MiB
/*
|
|
Human
|
|
homepage: <https://github.com/vladmandic/human>
|
|
author: <https://github.com/vladmandic>'
|
|
*/
|
|
|
|
"use strict";var Human=(()=>{var Oc=Object.defineProperty;var YS=Object.getOwnPropertyDescriptor;var ZS=Object.getOwnPropertyNames;var JS=Object.prototype.hasOwnProperty;var QS=(e,t,a)=>t in e?Oc(e,t,{enumerable:!0,configurable:!0,writable:!0,value:a}):e[t]=a;var yr=(e,t)=>{for(var a in t)Oc(e,a,{get:t[a],enumerable:!0})},eC=(e,t,a,n)=>{if(t&&typeof t=="object"||typeof t=="function")for(let r of ZS(t))!JS.call(e,r)&&r!==a&&Oc(e,r,{get:()=>t[r],enumerable:!(n=YS(t,r))||n.enumerable});return e};var tC=e=>eC(Oc({},"__esModule",{value:!0}),e);var he=(e,t,a)=>(QS(e,typeof t!="symbol"?t+"":t,a),a),C5=(e,t,a)=>{if(!t.has(e))throw TypeError("Cannot "+a)};var Xa=(e,t,a)=>(C5(e,t,"read from private field"),a?a.call(e):t.get(e)),qn=(e,t,a)=>{if(t.has(e))throw TypeError("Cannot add the same private member more than once");t instanceof WeakSet?t.add(e):t.set(e,a)},xr=(e,t,a,n)=>(C5(e,t,"write to private field"),n?n.call(e,a):t.set(e,a),a);var Y3e={};yr(Y3e,{Env:()=>Xp,Human:()=>h5,default:()=>h5,defaults:()=>Qo,draw:()=>v0,empty:()=>pr,env:()=>ne,match:()=>K0,models:()=>c5});var je={};yr(je,{Abs:()=>Zl,Acos:()=>Qs,Acosh:()=>ei,AdadeltaOptimizer:()=>O3,AdagradOptimizer:()=>z3,AdamOptimizer:()=>L3,AdamaxOptimizer:()=>W3,Add:()=>ns,AddN:()=>ti,All:()=>ai,Any:()=>ni,ArgMax:()=>Jl,ArgMin:()=>Ql,Asin:()=>ri,Asinh:()=>si,Atan:()=>ii,Atan2:()=>li,Atanh:()=>oi,AvgPool:()=>ui,AvgPool3D:()=>eu,AvgPool3DGrad:()=>ep,AvgPoolGrad:()=>kh,BackendWasm:()=>jw,BatchMatMul:()=>di,BatchToSpaceND:()=>tu,Bincount:()=>pi,BitwiseAnd:()=>Ih,BroadcastArgs:()=>au,BroadcastTo:()=>jC,Cast:()=>ci,Ceil:()=>hi,ClipByValue:()=>rs,Complex:()=>tp,ComplexAbs:()=>ap,Concat:()=>nu,Conv2D:()=>fi,Conv2DBackpropFilter:()=>np,Conv2DBackpropInput:()=>mi,Conv3D:()=>gi,Conv3DBackpropFilterV2:()=>ru,Conv3DBackpropInputV2:()=>yi,Cos:()=>xi,Cosh:()=>Ai,CropAndResize:()=>wi,Cumprod:()=>bi,Cumsum:()=>vi,DataStorage:()=>Zd,DenseBincount:()=>su,DepthToSpace:()=>ki,DepthwiseConv2dNative:()=>Ii,DepthwiseConv2dNativeBackpropFilter:()=>rp,DepthwiseConv2dNativeBackpropInput:()=>sp,Diag:()=>iu,Dilation2D:()=>Si,Dilation2DBackpropFilter:()=>Pl,Dilation2DBackpropInput:()=>Dl,ENV:()=>W1,Einsum:()=>ip,Elu:()=>Ti,EluGrad:()=>ou,Environment:()=>sA,Equal:()=>Ni,Erf:()=>lu,Exp:()=>Ri,ExpandDims:()=>uu,Expm1:()=>Ei,FFT:()=>op,Fill:()=>du,FlipLeftRight:()=>Mi,Floor:()=>Fi,FloorDiv:()=>_i,FromPixels:()=>Ed,FusedBatchNorm:()=>$i,FusedConv2D:()=>Kr,FusedDepthwiseConv2D:()=>Yr,GPGPUContext:()=>_l,GatherNd:()=>Di,GatherV2:()=>pu,GraphModel:()=>Lp,Greater:()=>Pi,GreaterEqual:()=>Oi,IFFT:()=>lp,Identity:()=>zi,Imag:()=>up,IsFinite:()=>Li,IsInf:()=>Wi,IsNan:()=>Bi,KernelBackend:()=>Kl,LRN:()=>Zi,LRNGrad:()=>cu,LeakyRelu:()=>Vi,Less:()=>Ui,LessEqual:()=>Gi,LinSpace:()=>Hi,Log:()=>ji,Log1p:()=>qi,LogSoftmax:()=>qC,LogicalAnd:()=>Xi,LogicalNot:()=>Ki,LogicalOr:()=>Yi,LogicalXor:()=>oA,LowerBound:()=>XC,MathBackendCPU:()=>Xh,MathBackendWebGL:()=>Bu,MatrixBandPart:()=>KC,Max:()=>Ji,MaxPool:()=>eo,MaxPool3D:()=>hu,MaxPool3DGrad:()=>dp,MaxPoolGrad:()=>Sh,MaxPoolWithArgmax:()=>pp,Maximum:()=>Qi,Mean:()=>to,Min:()=>ao,Minimum:()=>no,MirrorPad:()=>ro,Mod:()=>fu,MomentumOptimizer:()=>B3,Multinomial:()=>so,Multiply:()=>io,Neg:()=>mu,NonMaxSuppressionV3:()=>lo,NonMaxSuppressionV4:()=>gu,NonMaxSuppressionV5:()=>uo,NotEqual:()=>oo,OP_SCOPE_SUFFIX:()=>G1,OneHot:()=>po,OnesLike:()=>yu,Optimizer:()=>ls,OptimizerConstructors:()=>y7,Pack:()=>xu,PadV2:()=>co,Pool:()=>YC,Pow:()=>ho,Prelu:()=>fo,Prod:()=>mo,RMSPropOptimizer:()=>V3,RaggedGather:()=>Ch,RaggedRange:()=>Th,RaggedTensorToTensor:()=>Nh,Range:()=>Au,Rank:()=>j2,Real:()=>cp,RealDiv:()=>Ci,Reciprocal:()=>go,Reduction:()=>ka,Relu:()=>yo,Relu6:()=>bo,Reshape:()=>bu,ResizeBilinear:()=>Ao,ResizeBilinearGrad:()=>wu,ResizeNearestNeighbor:()=>xo,ResizeNearestNeighborGrad:()=>vu,Reverse:()=>vo,RotateWithOffset:()=>Uo,Round:()=>wo,Rsqrt:()=>ko,SGDOptimizer:()=>qh,ScatterNd:()=>Io,SearchSorted:()=>Co,Select:()=>ku,Selu:()=>To,Sigmoid:()=>Eo,Sign:()=>Ro,Sin:()=>No,Sinh:()=>Su,Slice:()=>Iu,Softmax:()=>$o,Softplus:()=>Mo,SpaceToBatchND:()=>Cu,SparseFillEmptyRows:()=>hp,SparseReshape:()=>Nu,SparseSegmentMean:()=>fp,SparseSegmentSum:()=>mp,SparseToDense:()=>Do,SplitV:()=>Tu,Sqrt:()=>Fo,Square:()=>gp,SquaredDifference:()=>Po,StaticRegexReplace:()=>yp,Step:()=>is,StridedSlice:()=>Oo,StringNGrams:()=>Ru,StringSplit:()=>xp,StringToHashBucketFast:()=>Ap,Sub:()=>zo,Sum:()=>_o,Tan:()=>Lo,Tanh:()=>Wo,Tensor:()=>gt,TensorBuffer:()=>Wt,TensorScatterUpdate:()=>So,Tile:()=>ss,TopK:()=>Bo,Transform:()=>Vo,Transpose:()=>wr,Unique:()=>bp,Unpack:()=>Eu,UnsortedSegmentSum:()=>vp,UpperBound:()=>ZC,Variable:()=>_d,WebGPUBackend:()=>r0,ZerosLike:()=>Mu,_FusedMatMul:()=>Xr,abs:()=>Za,acos:()=>MA,acosh:()=>FA,add:()=>ke,addN:()=>Eh,all:()=>_A,any:()=>$A,argMax:()=>rr,argMin:()=>DA,asin:()=>PA,asinh:()=>OA,atan:()=>zA,atan2:()=>LA,atanh:()=>WA,avgPool:()=>Q1,avgPool3d:()=>GA,backend:()=>nr,backend_util:()=>S,basicLSTMCell:()=>HA,batchNorm:()=>Mp,batchNorm2d:()=>jA,batchNorm3d:()=>qA,batchNorm4d:()=>XA,batchToSpaceND:()=>e3,bincount:()=>t3,bitwiseAnd:()=>KA,booleanMaskAsync:()=>$4,broadcastArgs:()=>YA,broadcastTo:()=>Fl,broadcast_util:()=>Ho,browser:()=>Rr,buffer:()=>$e,cast:()=>Ke,ceil:()=>ZA,clipByValue:()=>JA,clone:()=>Sa,complex:()=>Sr,concat:()=>lt,concat1d:()=>QA,concat2d:()=>Fu,concat3d:()=>eb,concat4d:()=>tb,conv1d:()=>ab,conv2d:()=>Fp,conv2dTranspose:()=>rb,conv3d:()=>sb,conv3dTranspose:()=>ib,copyRegisteredKernels:()=>aT,cos:()=>ob,cosh:()=>lb,cosineWindow:()=>Vh,cumprod:()=>ub,cumsum:()=>db,customGrad:()=>tr,denseBincount:()=>pb,deprecationWarn:()=>Y1,depthToSpace:()=>cb,depthwiseConv2d:()=>Mh,deregisterOp:()=>YD,device_util:()=>Sp,diag:()=>hb,dilation2d:()=>fb,disableDeprecationWarnings:()=>cN,dispose:()=>J,disposeVariables:()=>hN,div:()=>ve,divNoNan:()=>gb,dot:()=>yb,dropout:()=>L4,einsum:()=>xb,elu:()=>r3,enableDebugMode:()=>pN,enableProdMode:()=>K1,enclosingPowerOfTwo:()=>$3,engine:()=>It,ensureShape:()=>Ab,env:()=>W,equal:()=>n3,erf:()=>bb,euclideanNorm:()=>kb,exp:()=>Qr,expandDims:()=>Ht,expm1:()=>Ib,eye:()=>i3,fft:()=>zh,fill:()=>sr,findBackend:()=>Z1,findBackendFactory:()=>xN,floor:()=>o3,floorDiv:()=>Rp,forceHalfFloat:()=>Uv,fused:()=>D3,gather:()=>l3,gatherND:()=>z4,gather_util:()=>H3,getBackend:()=>ha,getGradient:()=>G2,getKernel:()=>ah,getKernelsForBackend:()=>Jn,getThreadsCount:()=>Qie,gpgpu_util:()=>vv,grad:()=>xE,grads:()=>AE,greater:()=>Dp,greaterEqual:()=>u3,ifft:()=>Vd,imag:()=>Pp,image:()=>fe,inTopKAsync:()=>W4,io:()=>Kn,irfft:()=>N3,isFinite:()=>Sb,isInf:()=>Cb,isNaN:()=>Tb,keep:()=>zn,kernel_impls:()=>Nn,leakyRelu:()=>d3,less:()=>ih,lessEqual:()=>Fh,linalg:()=>X4,linspace:()=>Nb,loadGraphModel:()=>Y3,loadGraphModelSync:()=>aO,localResponseNormalization:()=>Rb,log:()=>Ll,log1p:()=>p3,logSigmoid:()=>Mb,logSoftmax:()=>Fb,logSumExp:()=>h3,logicalAnd:()=>Ld,logicalNot:()=>f3,logicalOr:()=>m3,logicalXor:()=>_b,losses:()=>K4,lowerBound:()=>$b,matMul:()=>pt,math:()=>i7,max:()=>ga,maxPool:()=>g3,maxPool3d:()=>Db,maxPoolWithArgmax:()=>Pb,maximum:()=>y3,mean:()=>Wd,memory:()=>fN,meshgrid:()=>Ob,min:()=>Jr,minimum:()=>Bd,mirrorPad:()=>zb,mod:()=>_u,moments:()=>Lb,movingAverage:()=>D4,mul:()=>te,multiRNNCell:()=>Wb,multinomial:()=>Bb,neg:()=>Ln,nextFrame:()=>x7,norm:()=>$p,notEqual:()=>x3,oneHot:()=>oh,ones:()=>Ur,onesLike:()=>Vb,op:()=>z,outerProduct:()=>Ub,pad:()=>ir,pad1d:()=>Gb,pad2d:()=>Hb,pad3d:()=>jb,pad4d:()=>qb,pool:()=>Xb,pow:()=>zl,prelu:()=>b3,print:()=>X1,prod:()=>Kb,profile:()=>mN,raggedGather:()=>Yb,raggedRange:()=>Zb,raggedTensorToTensor:()=>Jb,rand:()=>Qb,randomGamma:()=>n4,randomNormal:()=>S3,randomStandardNormal:()=>r4,randomUniform:()=>Dh,randomUniformInt:()=>s4,range:()=>Wl,ready:()=>Np,real:()=>Bl,reciprocal:()=>i4,registerBackend:()=>Go,registerGradient:()=>QC,registerKernel:()=>yn,registerOp:()=>KD,relu:()=>Op,relu6:()=>C3,removeBackend:()=>yN,reshape:()=>Q,reverse:()=>es,reverse1d:()=>o4,reverse2d:()=>l4,reverse3d:()=>u4,reverse4d:()=>d4,rfft:()=>Lh,round:()=>T3,rsqrt:()=>p4,scalar:()=>Ue,scatterND:()=>P4,scatter_util:()=>Wh,searchSorted:()=>$h,selu:()=>c4,separableConv2d:()=>h4,serialization:()=>J4,setBackend:()=>Tp,setPlatform:()=>AN,setThreadsCount:()=>Jie,setWasmPath:()=>Zie,setWasmPaths:()=>n0,setWebGLContext:()=>Zh,setdiff1dAsync:()=>f4,shared:()=>Kh,sigmoid:()=>Wa,sign:()=>m4,signal:()=>q4,sin:()=>g4,sinh:()=>y4,slice:()=>De,slice1d:()=>x4,slice2d:()=>A4,slice3d:()=>zp,slice4d:()=>Ph,slice_util:()=>Nt,softmax:()=>Oh,softplus:()=>c3,spaceToBatchND:()=>A3,sparse:()=>Y4,sparseToDense:()=>O4,spectral:()=>j4,split:()=>Ca,sqrt:()=>er,square:()=>Cn,squaredDifference:()=>R3,squeeze:()=>Oe,stack:()=>ca,step:()=>E3,stridedSlice:()=>b4,string:()=>Z4,sub:()=>ye,sum:()=>ot,sumOutType:()=>Ip,tan:()=>v4,tanh:()=>sh,tensor:()=>Ve,tensor1d:()=>jt,tensor2d:()=>Zn,tensor3d:()=>M3,tensor4d:()=>w4,tensor5d:()=>k4,tensor6d:()=>I4,tensorScatterUpdate:()=>C4,tensor_util:()=>cA,test_util:()=>e4,tidy:()=>Pe,tile:()=>jr,time:()=>gN,topk:()=>T4,train:()=>j$,transpose:()=>js,truncatedNormal:()=>N4,unique:()=>R4,unregisterGradient:()=>tT,unregisterKernel:()=>eT,unsortedSegmentSum:()=>E4,unstack:()=>Ra,upcastType:()=>pa,upperBound:()=>M4,util:()=>v,valueAndGrad:()=>bE,valueAndGrads:()=>vE,variable:()=>F4,variableGrads:()=>Eb,version:()=>qp,version_converter:()=>rO,version_core:()=>j3,version_cpu:()=>yz,version_wasm:()=>eoe,version_webgl:()=>nj,webgl:()=>rj,webgl_util:()=>H6,webgpu_util:()=>Kw,where:()=>kr,whereAsync:()=>_3,zeros:()=>gn,zerosLike:()=>Qa});var aC=Object.create,O1=Object.defineProperty,nC=Object.getOwnPropertyDescriptor,rC=Object.getOwnPropertyNames,sC=Object.getPrototypeOf,iC=Object.prototype.hasOwnProperty,qt=(e,t)=>()=>(t||e((t={exports:{}}).exports,t),t.exports),Ze=(e,t)=>{for(var a in t)O1(e,a,{get:t[a],enumerable:!0})},oC=(e,t,a,n)=>{if(t&&typeof t=="object"||typeof t=="function")for(let r of rC(t))!iC.call(e,r)&&r!==a&&O1(e,r,{get:()=>t[r],enumerable:!(n=nC(t,r))||n.enumerable});return e},Xl=(e,t,a)=>(a=e!=null?aC(sC(e)):{},oC(t||!e||!e.__esModule?O1(a,"default",{value:e,enumerable:!0}):a,e)),lC=qt((e,t)=>{t.exports=n;var a=null;try{a=new WebAssembly.Instance(new WebAssembly.Module(new Uint8Array([0,97,115,109,1,0,0,0,1,13,2,96,0,1,127,96,4,127,127,127,127,1,127,3,7,6,0,1,1,1,1,1,6,6,1,127,1,65,0,11,7,50,6,3,109,117,108,0,1,5,100,105,118,95,115,0,2,5,100,105,118,95,117,0,3,5,114,101,109,95,115,0,4,5,114,101,109,95,117,0,5,8,103,101,116,95,104,105,103,104,0,0,10,191,1,6,4,0,35,0,11,36,1,1,126,32,0,173,32,1,173,66,32,134,132,32,2,173,32,3,173,66,32,134,132,126,34,4,66,32,135,167,36,0,32,4,167,11,36,1,1,126,32,0,173,32,1,173,66,32,134,132,32,2,173,32,3,173,66,32,134,132,127,34,4,66,32,135,167,36,0,32,4,167,11,36,1,1,126,32,0,173,32,1,173,66,32,134,132,32,2,173,32,3,173,66,32,134,132,128,34,4,66,32,135,167,36,0,32,4,167,11,36,1,1,126,32,0,173,32,1,173,66,32,134,132,32,2,173,32,3,173,66,32,134,132,129,34,4,66,32,135,167,36,0,32,4,167,11,36,1,1,126,32,0,173,32,1,173,66,32,134,132,32,2,173,32,3,173,66,32,134,132,130,34,4,66,32,135,167,36,0,32,4,167,11])),{}).exports}catch(I){}function n(I,_,O){this.low=I|0,this.high=_|0,this.unsigned=!!O}n.prototype.__isLong__,Object.defineProperty(n.prototype,"__isLong__",{value:!0});function r(I){return(I&&I.__isLong__)===!0}n.isLong=r;var s={},i={};function o(I,_){var O,B,$;return _?(I>>>=0,($=0<=I&&I<256)&&(B=i[I],B)?B:(O=u(I,(I|0)<0?-1:0,!0),$&&(i[I]=O),O)):(I|=0,($=-128<=I&&I<128)&&(B=s[I],B)?B:(O=u(I,I<0?-1:0,!1),$&&(s[I]=O),O))}n.fromInt=o;function l(I,_){if(isNaN(I))return _?b:A;if(_){if(I<0)return b;if(I>=g)return M}else{if(I<=-y)return F;if(I+1>=y)return N}return I<0?l(-I,_).neg():u(I%m|0,I/m|0,_)}n.fromNumber=l;function u(I,_,O){return new n(I,_,O)}n.fromBits=u;var p=Math.pow;function c(I,_,O){if(I.length===0)throw Error("empty string");if(I==="NaN"||I==="Infinity"||I==="+Infinity"||I==="-Infinity")return A;if(typeof _=="number"?(O=_,_=!1):_=!!_,O=O||10,O<2||36<O)throw RangeError("radix");var B;if((B=I.indexOf("-"))>0)throw Error("interior hyphen");if(B===0)return c(I.substring(1),_,O).neg();for(var $=l(p(O,8)),U=A,G=0;G<I.length;G+=8){var q=Math.min(8,I.length-G),H=parseInt(I.substring(G,G+q),O);if(q<8){var V=l(p(O,q));U=U.mul(V).add(l(H))}else U=U.mul($),U=U.add(l(H))}return U.unsigned=_,U}n.fromString=c;function d(I,_){return typeof I=="number"?l(I,_):typeof I=="string"?c(I,_):u(I.low,I.high,typeof _=="boolean"?_:I.unsigned)}n.fromValue=d;var h=1<<16,f=1<<24,m=h*h,g=m*m,y=g/2,x=o(f),A=o(0);n.ZERO=A;var b=o(0,!0);n.UZERO=b;var w=o(1);n.ONE=w;var C=o(1,!0);n.UONE=C;var T=o(-1);n.NEG_ONE=T;var N=u(-1,2147483647,!1);n.MAX_VALUE=N;var M=u(-1,-1,!0);n.MAX_UNSIGNED_VALUE=M;var F=u(0,-2147483648,!1);n.MIN_VALUE=F;var E=n.prototype;E.toInt=function(){return this.unsigned?this.low>>>0:this.low},E.toNumber=function(){return this.unsigned?(this.high>>>0)*m+(this.low>>>0):this.high*m+(this.low>>>0)},E.toString=function(I){if(I=I||10,I<2||36<I)throw RangeError("radix");if(this.isZero())return"0";if(this.isNegative())if(this.eq(F)){var _=l(I),O=this.div(_),B=O.mul(_).sub(this);return O.toString(I)+B.toInt().toString(I)}else return"-"+this.neg().toString(I);for(var $=l(p(I,6),this.unsigned),U=this,G="";;){var q=U.div($),H=U.sub(q.mul($)).toInt()>>>0,V=H.toString(I);if(U=q,U.isZero())return V+G;for(;V.length<6;)V="0"+V;G=""+V+G}},E.getHighBits=function(){return this.high},E.getHighBitsUnsigned=function(){return this.high>>>0},E.getLowBits=function(){return this.low},E.getLowBitsUnsigned=function(){return this.low>>>0},E.getNumBitsAbs=function(){if(this.isNegative())return this.eq(F)?64:this.neg().getNumBitsAbs();for(var I=this.high!=0?this.high:this.low,_=31;_>0&&!(I&1<<_);_--);return this.high!=0?_+33:_+1},E.isZero=function(){return this.high===0&&this.low===0},E.eqz=E.isZero,E.isNegative=function(){return!this.unsigned&&this.high<0},E.isPositive=function(){return this.unsigned||this.high>=0},E.isOdd=function(){return(this.low&1)===1},E.isEven=function(){return(this.low&1)===0},E.equals=function(I){return r(I)||(I=d(I)),this.unsigned!==I.unsigned&&this.high>>>31===1&&I.high>>>31===1?!1:this.high===I.high&&this.low===I.low},E.eq=E.equals,E.notEquals=function(I){return!this.eq(I)},E.neq=E.notEquals,E.ne=E.notEquals,E.lessThan=function(I){return this.comp(I)<0},E.lt=E.lessThan,E.lessThanOrEqual=function(I){return this.comp(I)<=0},E.lte=E.lessThanOrEqual,E.le=E.lessThanOrEqual,E.greaterThan=function(I){return this.comp(I)>0},E.gt=E.greaterThan,E.greaterThanOrEqual=function(I){return this.comp(I)>=0},E.gte=E.greaterThanOrEqual,E.ge=E.greaterThanOrEqual,E.compare=function(I){if(r(I)||(I=d(I)),this.eq(I))return 0;var _=this.isNegative(),O=I.isNegative();return _&&!O?-1:!_&&O?1:this.unsigned?I.high>>>0>this.high>>>0||I.high===this.high&&I.low>>>0>this.low>>>0?-1:1:this.sub(I).isNegative()?-1:1},E.comp=E.compare,E.negate=function(){return!this.unsigned&&this.eq(F)?F:this.not().add(w)},E.neg=E.negate,E.add=function(I){r(I)||(I=d(I));var _=this.high>>>16,O=this.high&65535,B=this.low>>>16,$=this.low&65535,U=I.high>>>16,G=I.high&65535,q=I.low>>>16,H=I.low&65535,V=0,Z=0,X=0,re=0;return re+=$+H,X+=re>>>16,re&=65535,X+=B+q,Z+=X>>>16,X&=65535,Z+=O+G,V+=Z>>>16,Z&=65535,V+=_+U,V&=65535,u(X<<16|re,V<<16|Z,this.unsigned)},E.subtract=function(I){return r(I)||(I=d(I)),this.add(I.neg())},E.sub=E.subtract,E.multiply=function(I){if(this.isZero())return A;if(r(I)||(I=d(I)),a){var _=a.mul(this.low,this.high,I.low,I.high);return u(_,a.get_high(),this.unsigned)}if(I.isZero())return A;if(this.eq(F))return I.isOdd()?F:A;if(I.eq(F))return this.isOdd()?F:A;if(this.isNegative())return I.isNegative()?this.neg().mul(I.neg()):this.neg().mul(I).neg();if(I.isNegative())return this.mul(I.neg()).neg();if(this.lt(x)&&I.lt(x))return l(this.toNumber()*I.toNumber(),this.unsigned);var O=this.high>>>16,B=this.high&65535,$=this.low>>>16,U=this.low&65535,G=I.high>>>16,q=I.high&65535,H=I.low>>>16,V=I.low&65535,Z=0,X=0,re=0,ee=0;return ee+=U*V,re+=ee>>>16,ee&=65535,re+=$*V,X+=re>>>16,re&=65535,re+=U*H,X+=re>>>16,re&=65535,X+=B*V,Z+=X>>>16,X&=65535,X+=$*H,Z+=X>>>16,X&=65535,X+=U*q,Z+=X>>>16,X&=65535,Z+=O*V+B*H+$*q+U*G,Z&=65535,u(re<<16|ee,Z<<16|X,this.unsigned)},E.mul=E.multiply,E.divide=function(I){if(r(I)||(I=d(I)),I.isZero())throw Error("division by zero");if(a){if(!this.unsigned&&this.high===-2147483648&&I.low===-1&&I.high===-1)return this;var _=(this.unsigned?a.div_u:a.div_s)(this.low,this.high,I.low,I.high);return u(_,a.get_high(),this.unsigned)}if(this.isZero())return this.unsigned?b:A;var O,B,$;if(this.unsigned){if(I.unsigned||(I=I.toUnsigned()),I.gt(this))return b;if(I.gt(this.shru(1)))return C;$=b}else{if(this.eq(F)){if(I.eq(w)||I.eq(T))return F;if(I.eq(F))return w;var U=this.shr(1);return O=U.div(I).shl(1),O.eq(A)?I.isNegative()?w:T:(B=this.sub(I.mul(O)),$=O.add(B.div(I)),$)}else if(I.eq(F))return this.unsigned?b:A;if(this.isNegative())return I.isNegative()?this.neg().div(I.neg()):this.neg().div(I).neg();if(I.isNegative())return this.div(I.neg()).neg();$=A}for(B=this;B.gte(I);){O=Math.max(1,Math.floor(B.toNumber()/I.toNumber()));for(var G=Math.ceil(Math.log(O)/Math.LN2),q=G<=48?1:p(2,G-48),H=l(O),V=H.mul(I);V.isNegative()||V.gt(B);)O-=q,H=l(O,this.unsigned),V=H.mul(I);H.isZero()&&(H=w),$=$.add(H),B=B.sub(V)}return $},E.div=E.divide,E.modulo=function(I){if(r(I)||(I=d(I)),a){var _=(this.unsigned?a.rem_u:a.rem_s)(this.low,this.high,I.low,I.high);return u(_,a.get_high(),this.unsigned)}return this.sub(this.div(I).mul(I))},E.mod=E.modulo,E.rem=E.modulo,E.not=function(){return u(~this.low,~this.high,this.unsigned)},E.and=function(I){return r(I)||(I=d(I)),u(this.low&I.low,this.high&I.high,this.unsigned)},E.or=function(I){return r(I)||(I=d(I)),u(this.low|I.low,this.high|I.high,this.unsigned)},E.xor=function(I){return r(I)||(I=d(I)),u(this.low^I.low,this.high^I.high,this.unsigned)},E.shiftLeft=function(I){return r(I)&&(I=I.toInt()),(I&=63)===0?this:I<32?u(this.low<<I,this.high<<I|this.low>>>32-I,this.unsigned):u(0,this.low<<I-32,this.unsigned)},E.shl=E.shiftLeft,E.shiftRight=function(I){return r(I)&&(I=I.toInt()),(I&=63)===0?this:I<32?u(this.low>>>I|this.high<<32-I,this.high>>I,this.unsigned):u(this.high>>I-32,this.high>=0?0:-1,this.unsigned)},E.shr=E.shiftRight,E.shiftRightUnsigned=function(I){if(r(I)&&(I=I.toInt()),I&=63,I===0)return this;var _=this.high;if(I<32){var O=this.low;return u(O>>>I|_<<32-I,_>>>I,this.unsigned)}else return I===32?u(_,0,this.unsigned):u(_>>>I-32,0,this.unsigned)},E.shru=E.shiftRightUnsigned,E.shr_u=E.shiftRightUnsigned,E.toSigned=function(){return this.unsigned?u(this.low,this.high,!1):this},E.toUnsigned=function(){return this.unsigned?this:u(this.low,this.high,!0)},E.toBytes=function(I){return I?this.toBytesLE():this.toBytesBE()},E.toBytesLE=function(){var I=this.high,_=this.low;return[_&255,_>>>8&255,_>>>16&255,_>>>24,I&255,I>>>8&255,I>>>16&255,I>>>24]},E.toBytesBE=function(){var I=this.high,_=this.low;return[I>>>24,I>>>16&255,I>>>8&255,I&255,_>>>24,_>>>16&255,_>>>8&255,_&255]},n.fromBytes=function(I,_,O){return O?n.fromBytesLE(I,_):n.fromBytesBE(I,_)},n.fromBytesLE=function(I,_){return new n(I[0]|I[1]<<8|I[2]<<16|I[3]<<24,I[4]|I[5]<<8|I[6]<<16|I[7]<<24,_)},n.fromBytesBE=function(I,_){return new n(I[4]<<24|I[5]<<16|I[6]<<8|I[7],I[0]<<24|I[1]<<16|I[2]<<8|I[3],_)}}),uC=qt(()=>{}),dC=qt(()=>{}),pC=qt((e,t)=>{(function(a,n,r){function s(u){var p=this,c=l();p.next=function(){var d=2091639*p.s0+p.c*23283064365386963e-26;return p.s0=p.s1,p.s1=p.s2,p.s2=d-(p.c=d|0)},p.c=1,p.s0=c(" "),p.s1=c(" "),p.s2=c(" "),p.s0-=c(u),p.s0<0&&(p.s0+=1),p.s1-=c(u),p.s1<0&&(p.s1+=1),p.s2-=c(u),p.s2<0&&(p.s2+=1),c=null}function i(u,p){return p.c=u.c,p.s0=u.s0,p.s1=u.s1,p.s2=u.s2,p}function o(u,p){var c=new s(u),d=p&&p.state,h=c.next;return h.int32=function(){return c.next()*4294967296|0},h.double=function(){return h()+(h()*2097152|0)*11102230246251565e-32},h.quick=h,d&&(typeof d=="object"&&i(d,c),h.state=function(){return i(c,{})}),h}function l(){var u=4022871197,p=function(c){c=String(c);for(var d=0;d<c.length;d++){u+=c.charCodeAt(d);var h=.02519603282416938*u;u=h>>>0,h-=u,h*=u,u=h>>>0,h-=u,u+=h*4294967296}return(u>>>0)*23283064365386963e-26};return p}n&&n.exports?n.exports=o:r&&r.amd?r(function(){return o}):this.alea=o})(e,typeof t=="object"&&t,typeof define=="function"&&define)}),cC=qt((e,t)=>{(function(a,n,r){function s(l){var u=this,p="";u.x=0,u.y=0,u.z=0,u.w=0,u.next=function(){var d=u.x^u.x<<11;return u.x=u.y,u.y=u.z,u.z=u.w,u.w^=u.w>>>19^d^d>>>8},l===(l|0)?u.x=l:p+=l;for(var c=0;c<p.length+64;c++)u.x^=p.charCodeAt(c)|0,u.next()}function i(l,u){return u.x=l.x,u.y=l.y,u.z=l.z,u.w=l.w,u}function o(l,u){var p=new s(l),c=u&&u.state,d=function(){return(p.next()>>>0)/4294967296};return d.double=function(){do var h=p.next()>>>11,f=(p.next()>>>0)/4294967296,m=(h+f)/(1<<21);while(m===0);return m},d.int32=p.next,d.quick=d,c&&(typeof c=="object"&&i(c,p),d.state=function(){return i(p,{})}),d}n&&n.exports?n.exports=o:r&&r.amd?r(function(){return o}):this.xor128=o})(e,typeof t=="object"&&t,typeof define=="function"&&define)}),hC=qt((e,t)=>{(function(a,n,r){function s(l){var u=this,p="";u.next=function(){var d=u.x^u.x>>>2;return u.x=u.y,u.y=u.z,u.z=u.w,u.w=u.v,(u.d=u.d+362437|0)+(u.v=u.v^u.v<<4^(d^d<<1))|0},u.x=0,u.y=0,u.z=0,u.w=0,u.v=0,l===(l|0)?u.x=l:p+=l;for(var c=0;c<p.length+64;c++)u.x^=p.charCodeAt(c)|0,c==p.length&&(u.d=u.x<<10^u.x>>>4),u.next()}function i(l,u){return u.x=l.x,u.y=l.y,u.z=l.z,u.w=l.w,u.v=l.v,u.d=l.d,u}function o(l,u){var p=new s(l),c=u&&u.state,d=function(){return(p.next()>>>0)/4294967296};return d.double=function(){do var h=p.next()>>>11,f=(p.next()>>>0)/4294967296,m=(h+f)/(1<<21);while(m===0);return m},d.int32=p.next,d.quick=d,c&&(typeof c=="object"&&i(c,p),d.state=function(){return i(p,{})}),d}n&&n.exports?n.exports=o:r&&r.amd?r(function(){return o}):this.xorwow=o})(e,typeof t=="object"&&t,typeof define=="function"&&define)}),fC=qt((e,t)=>{(function(a,n,r){function s(l){var u=this;u.next=function(){var c=u.x,d=u.i,h,f,m;return h=c[d],h^=h>>>7,f=h^h<<24,h=c[d+1&7],f^=h^h>>>10,h=c[d+3&7],f^=h^h>>>3,h=c[d+4&7],f^=h^h<<7,h=c[d+7&7],h=h^h<<13,f^=h^h<<9,c[d]=f,u.i=d+1&7,f};function p(c,d){var h,f,m=[];if(d===(d|0))f=m[0]=d;else for(d=""+d,h=0;h<d.length;++h)m[h&7]=m[h&7]<<15^d.charCodeAt(h)+m[h+1&7]<<13;for(;m.length<8;)m.push(0);for(h=0;h<8&&m[h]===0;++h);for(h==8?f=m[7]=-1:f=m[h],c.x=m,c.i=0,h=256;h>0;--h)c.next()}p(u,l)}function i(l,u){return u.x=l.x.slice(),u.i=l.i,u}function o(l,u){l==null&&(l=+new Date);var p=new s(l),c=u&&u.state,d=function(){return(p.next()>>>0)/4294967296};return d.double=function(){do var h=p.next()>>>11,f=(p.next()>>>0)/4294967296,m=(h+f)/(1<<21);while(m===0);return m},d.int32=p.next,d.quick=d,c&&(c.x&&i(c,p),d.state=function(){return i(p,{})}),d}n&&n.exports?n.exports=o:r&&r.amd?r(function(){return o}):this.xorshift7=o})(e,typeof t=="object"&&t,typeof define=="function"&&define)}),mC=qt((e,t)=>{(function(a,n,r){function s(l){var u=this;u.next=function(){var c=u.w,d=u.X,h=u.i,f,m;return u.w=c=c+1640531527|0,m=d[h+34&127],f=d[h=h+1&127],m^=m<<13,f^=f<<17,m^=m>>>15,f^=f>>>12,m=d[h]=m^f,u.i=h,m+(c^c>>>16)|0};function p(c,d){var h,f,m,g,y,x=[],A=128;for(d===(d|0)?(f=d,d=null):(d=d+"\0",f=0,A=Math.max(A,d.length)),m=0,g=-32;g<A;++g)d&&(f^=d.charCodeAt((g+32)%d.length)),g===0&&(y=f),f^=f<<10,f^=f>>>15,f^=f<<4,f^=f>>>13,g>=0&&(y=y+1640531527|0,h=x[g&127]^=f+y,m=h==0?m+1:0);for(m>=128&&(x[(d&&d.length||0)&127]=-1),m=127,g=4*128;g>0;--g)f=x[m+34&127],h=x[m=m+1&127],f^=f<<13,h^=h<<17,f^=f>>>15,h^=h>>>12,x[m]=f^h;c.w=y,c.X=x,c.i=m}p(u,l)}function i(l,u){return u.i=l.i,u.w=l.w,u.X=l.X.slice(),u}function o(l,u){l==null&&(l=+new Date);var p=new s(l),c=u&&u.state,d=function(){return(p.next()>>>0)/4294967296};return d.double=function(){do var h=p.next()>>>11,f=(p.next()>>>0)/4294967296,m=(h+f)/(1<<21);while(m===0);return m},d.int32=p.next,d.quick=d,c&&(c.X&&i(c,p),d.state=function(){return i(p,{})}),d}n&&n.exports?n.exports=o:r&&r.amd?r(function(){return o}):this.xor4096=o})(e,typeof t=="object"&&t,typeof define=="function"&&define)}),gC=qt((e,t)=>{(function(a,n,r){function s(l){var u=this,p="";u.next=function(){var d=u.b,h=u.c,f=u.d,m=u.a;return d=d<<25^d>>>7^h,h=h-f|0,f=f<<24^f>>>8^m,m=m-d|0,u.b=d=d<<20^d>>>12^h,u.c=h=h-f|0,u.d=f<<16^h>>>16^m,u.a=m-d|0},u.a=0,u.b=0,u.c=-1640531527,u.d=1367130551,l===Math.floor(l)?(u.a=l/4294967296|0,u.b=l|0):p+=l;for(var c=0;c<p.length+20;c++)u.b^=p.charCodeAt(c)|0,u.next()}function i(l,u){return u.a=l.a,u.b=l.b,u.c=l.c,u.d=l.d,u}function o(l,u){var p=new s(l),c=u&&u.state,d=function(){return(p.next()>>>0)/4294967296};return d.double=function(){do var h=p.next()>>>11,f=(p.next()>>>0)/4294967296,m=(h+f)/(1<<21);while(m===0);return m},d.int32=p.next,d.quick=d,c&&(typeof c=="object"&&i(c,p),d.state=function(){return i(p,{})}),d}n&&n.exports?n.exports=o:r&&r.amd?r(function(){return o}):this.tychei=o})(e,typeof t=="object"&&t,typeof define=="function"&&define)}),yC=qt(()=>{}),xC=qt((e,t)=>{(function(a,n,r){var s=256,i=6,o=52,l="random",u=r.pow(s,i),p=r.pow(2,o),c=p*2,d=s-1,h;function f(w,C,T){var N=[];C=C==!0?{entropy:!0}:C||{};var M=x(y(C.entropy?[w,b(n)]:w==null?A():w,3),N),F=new m(N),E=function(){for(var I=F.g(i),_=u,O=0;I<p;)I=(I+O)*s,_*=s,O=F.g(1);for(;I>=c;)I/=2,_/=2,O>>>=1;return(I+O)/_};return E.int32=function(){return F.g(4)|0},E.quick=function(){return F.g(4)/4294967296},E.double=E,x(b(F.S),n),(C.pass||T||function(I,_,O,B){return B&&(B.S&&g(B,F),I.state=function(){return g(F,{})}),O?(r[l]=I,_):I})(E,M,"global"in C?C.global:this==r,C.state)}function m(w){var C,T=w.length,N=this,M=0,F=N.i=N.j=0,E=N.S=[];for(T||(w=[T++]);M<s;)E[M]=M++;for(M=0;M<s;M++)E[M]=E[F=d&F+w[M%T]+(C=E[M])],E[F]=C;(N.g=function(I){for(var _,O=0,B=N.i,$=N.j,U=N.S;I--;)_=U[B=d&B+1],O=O*s+U[d&(U[B]=U[$=d&$+_])+(U[$]=_)];return N.i=B,N.j=$,O})(s)}function g(w,C){return C.i=w.i,C.j=w.j,C.S=w.S.slice(),C}function y(w,C){var T=[],N=typeof w,M;if(C&&N=="object")for(M in w)try{T.push(y(w[M],C-1))}catch(F){}return T.length?T:N=="string"?w:w+"\0"}function x(w,C){for(var T=w+"",N,M=0;M<T.length;)C[d&M]=d&(N^=C[d&M]*19)+T.charCodeAt(M++);return b(C)}function A(){try{var w;return h&&(w=h.randomBytes)?w=w(s):(w=new Uint8Array(s),(a.crypto||a.msCrypto).getRandomValues(w)),b(w)}catch(N){var C=a.navigator,T=C&&C.plugins;return[+new Date,a,T,a.screen,b(n)]}}function b(w){return String.fromCharCode.apply(0,w)}if(x(r.random(),n),typeof t=="object"&&t.exports){t.exports=f;try{h=yC()}catch(w){}}else typeof define=="function"&&define.amd?define(function(){return f}):r["seed"+l]=f})(typeof self!="undefined"?self:e,[],Math)}),jx=qt((e,t)=>{var a=pC(),n=cC(),r=hC(),s=fC(),i=mC(),o=gC(),l=xC();l.alea=a,l.xor128=n,l.xorwow=r,l.xorshift7=s,l.xor4096=i,l.tychei=o,t.exports=l}),qx=qt(()=>{}),Xx=qt(()=>{}),AC=qt(()=>{}),bC=qt(()=>{}),vC=qt(()=>{}),wC=qt((e,t)=>{var a=(()=>{var n=typeof document!="undefined"&&document.currentScript?document.currentScript.src:void 0;return typeof __filename!="undefined"&&(n=n||__filename),function(r){r=r||{};function s(){return ie.buffer!=Ge&&ht(ie.buffer),yt}function i(){return ie.buffer!=Ge&&ht(ie.buffer),ja}function o(){return ie.buffer!=Ge&&ht(ie.buffer),zt}function l(){return ie.buffer!=Ge&&ht(ie.buffer),la}function u(){return ie.buffer!=Ge&&ht(ie.buffer),Da}function p(){return ie.buffer!=Ge&&ht(ie.buffer),dn}function c(){return ie.buffer!=Ge&&ht(ie.buffer),Pa}var d=typeof r!="undefined"?r:{},h,f;d.ready=new Promise(function(P,j){h=P,f=j});var m;typeof process!="undefined"&&process.listeners&&(m={uncaughtException:process.listeners("uncaughtException"),unhandledRejection:process.listeners("unhandledRejection")});var g=Object.assign({},d),y=[],x="./this.program",A=(P,j)=>{throw j},b=typeof window=="object",w=typeof importScripts=="function",C=typeof process=="object"&&typeof process.versions=="object"&&typeof process.versions.node=="string",T=d.ENVIRONMENT_IS_PTHREAD||!1,N="";function M(P){return d.locateFile?d.locateFile(P,N):N+P}var F,E,I,_;function O(P){P instanceof Es||H("exiting due to exception: "+P)}if(C){var B=qx(),$=Xx();w?N=$.dirname(N)+"/":N=__dirname+"/",F=(j,oe)=>(j=yl(j)?new URL(j):$.normalize(j),B.readFileSync(j,oe?void 0:"utf8")),I=j=>{var oe=F(j,!0);return oe.buffer||(oe=new Uint8Array(oe)),oe},E=(j,oe,Me)=>{j=yl(j)?new URL(j):$.normalize(j),B.readFile(j,function(He,Be){He?Me(He):oe(Be.buffer)})},process.argv.length>1&&(x=process.argv[1].replace(/\\/g,"/")),y=process.argv.slice(2),process.on("uncaughtException",function(j){if(!(j instanceof Es))throw j}),process.on("unhandledRejection",function(j){throw j}),A=(j,oe)=>{if(wn())throw process.exitCode=j,oe;O(oe),process.exit(j)},d.inspect=function(){return"[Emscripten Module object]"};let P;try{P=AC()}catch(j){throw console.error('The "worker_threads" module is not supported in this node.js build - perhaps a newer version is needed?'),j}global.Worker=P.Worker}else(b||w)&&(w?N=self.location.href:typeof document!="undefined"&&document.currentScript&&(N=document.currentScript.src),typeof n!="undefined"&&n&&(N=n),N.indexOf("blob:")!==0?N=N.substr(0,N.replace(/[?#].*/,"").lastIndexOf("/")+1):N="",C||(F=P=>{var j=new XMLHttpRequest;return j.open("GET",P,!1),j.send(null),j.responseText},w&&(I=P=>{var j=new XMLHttpRequest;return j.open("GET",P,!1),j.responseType="arraybuffer",j.send(null),new Uint8Array(j.response)}),E=(P,j,oe)=>{var Me=new XMLHttpRequest;Me.open("GET",P,!0),Me.responseType="arraybuffer",Me.onload=()=>{if(Me.status==200||Me.status==0&&Me.response){j(Me.response);return}oe()},Me.onerror=oe,Me.send(null)}),_=P=>document.title=P);C&&typeof performance=="undefined"&&(global.performance=bC().performance);var U=console.log.bind(console),G=console.warn.bind(console);C&&(U=P=>B.writeSync(1,P+`
|
|
`),G=P=>B.writeSync(2,P+`
|
|
`));var q=d.print||U,H=d.printErr||G;Object.assign(d,g),g=null,d.arguments&&(y=d.arguments),d.thisProgram&&(x=d.thisProgram),d.quit&&(A=d.quit);var V=4,Z=Atomics.load,X=Atomics.store,re=Atomics.compareExchange,ee;d.wasmBinary&&(ee=d.wasmBinary);var me=d.noExitRuntime||!0;typeof WebAssembly!="object"&&Rs("no native wasm support detected");var ie,Ae,Ce=!1,Re;function Le(P,j){P||Rs(j)}var qe=typeof TextDecoder!="undefined"?new TextDecoder("utf8"):void 0;function mt(P,j,oe){j>>>=0;for(var Me=j+oe,He=j;P[He]&&!(He>=Me);)++He;if(He-j>16&&P.buffer&&qe)return qe.decode(P.buffer instanceof SharedArrayBuffer?P.slice(j,He):P.subarray(j,He));for(var Be="";j<He;){var ge=P[j++];if(!(ge&128)){Be+=String.fromCharCode(ge);continue}var Ne=P[j++]&63;if((ge&224)==192){Be+=String.fromCharCode((ge&31)<<6|Ne);continue}var Tt=P[j++]&63;if((ge&240)==224?ge=(ge&15)<<12|Ne<<6|Tt:ge=(ge&7)<<18|Ne<<12|Tt<<6|P[j++]&63,ge<65536)Be+=String.fromCharCode(ge);else{var cn=ge-65536;Be+=String.fromCharCode(55296|cn>>10,56320|cn&1023)}}return Be}function dt(P,j){return P>>>=0,P?mt(i(),P,j):""}function st(P,j,oe,Me){if(oe>>>=0,!(Me>0))return 0;for(var He=oe,Be=oe+Me-1,ge=0;ge<P.length;++ge){var Ne=P.charCodeAt(ge);if(Ne>=55296&&Ne<=57343){var Tt=P.charCodeAt(++ge);Ne=65536+((Ne&1023)<<10)|Tt&1023}if(Ne<=127){if(oe>=Be)break;j[oe++>>>0]=Ne}else if(Ne<=2047){if(oe+1>=Be)break;j[oe++>>>0]=192|Ne>>6,j[oe++>>>0]=128|Ne&63}else if(Ne<=65535){if(oe+2>=Be)break;j[oe++>>>0]=224|Ne>>12,j[oe++>>>0]=128|Ne>>6&63,j[oe++>>>0]=128|Ne&63}else{if(oe+3>=Be)break;j[oe++>>>0]=240|Ne>>18,j[oe++>>>0]=128|Ne>>12&63,j[oe++>>>0]=128|Ne>>6&63,j[oe++>>>0]=128|Ne&63}}return j[oe>>>0]=0,oe-He}function it(P,j,oe){return st(P,i(),j,oe)}var Ge,yt,ja,zt,un,la,Da,dn,Pa;T&&(Ge=d.buffer);function ht(P){Ge=P,d.HEAP8=yt=new Int8Array(P),d.HEAP16=zt=new Int16Array(P),d.HEAP32=la=new Int32Array(P),d.HEAPU8=ja=new Uint8Array(P),d.HEAPU16=un=new Uint16Array(P),d.HEAPU32=Da=new Uint32Array(P),d.HEAPF32=dn=new Float32Array(P),d.HEAPF64=Pa=new Float64Array(P)}var Oa=d.INITIAL_MEMORY||16777216;if(T)ie=d.wasmMemory,Ge=d.buffer;else if(d.wasmMemory)ie=d.wasmMemory;else if(ie=new WebAssembly.Memory({initial:Oa/65536,maximum:65536,shared:!0}),!(ie.buffer instanceof SharedArrayBuffer))throw H("requested a shared WebAssembly.Memory but the returned buffer is not a SharedArrayBuffer, indicating that while the browser has SharedArrayBuffer it does not have WebAssembly threads support - you may need to set a flag"),C&&H("(on node you may need: --experimental-wasm-threads --experimental-wasm-bulk-memory and/or recent version)"),Error("bad memory");ie&&(Ge=ie.buffer),Oa=Ge.byteLength,ht(Ge);var qa,hr=[],ml=[],jn=[],sd=!1;function wn(){return me}function Dr(){if(d.preRun)for(typeof d.preRun=="function"&&(d.preRun=[d.preRun]);d.preRun.length;)of(d.preRun.shift());od(hr)}function Kt(){sd=!0,!T&&od(ml)}function uc(){if(!T){if(d.postRun)for(typeof d.postRun=="function"&&(d.postRun=[d.postRun]);d.postRun.length;)f5(d.postRun.shift());od(jn)}}function of(P){hr.unshift(P)}function lf(P){ml.unshift(P)}function f5(P){jn.unshift(P)}var Pr=0,gl=null,fr=null;function uf(P){Pr++,d.monitorRunDependencies&&d.monitorRunDependencies(Pr)}function dc(P){if(Pr--,d.monitorRunDependencies&&d.monitorRunDependencies(Pr),Pr==0&&(gl!==null&&(clearInterval(gl),gl=null),fr)){var j=fr;fr=null,j()}}function Rs(P){d.onAbort&&d.onAbort(P),P="Aborted("+P+")",H(P),Ce=!0,Re=1,P+=". Build with -sASSERTIONS for more info.";var j=new WebAssembly.RuntimeError(P);throw f(j),j}var df="data:application/octet-stream;base64,";function pc(P){return P.startsWith(df)}function yl(P){return P.startsWith("file://")}var ma;ma="tfjs-backend-wasm-threaded-simd.wasm",pc(ma)||(ma=M(ma));function cc(P){try{if(P==ma&&ee)return new Uint8Array(ee);if(I)return I(P);throw"both async and sync fetching of the wasm failed"}catch(j){Rs(j)}}function pf(){if(!ee&&(b||w)){if(typeof fetch=="function"&&!yl(ma))return fetch(ma,{credentials:"same-origin"}).then(function(P){if(!P.ok)throw"failed to load wasm binary file at '"+ma+"'";return P.arrayBuffer()}).catch(function(){return cc(ma)});if(E)return new Promise(function(P,j){E(ma,function(oe){P(new Uint8Array(oe))},j)})}return Promise.resolve().then(function(){return cc(ma)})}function cf(){var P={env:Sc,wasi_snapshot_preview1:Sc};function j(ge,Ne){var Tt=ge.exports;if(d.asm=Tt,vf(d.asm._emscripten_tls_init),qa=d.asm.__indirect_function_table,lf(d.asm.__wasm_call_ctors),Ae=Ne,!T){var cn=We.unusedWorkers.length;We.unusedWorkers.forEach(function(gr){We.loadWasmModuleToWorker(gr,function(){--cn||dc("wasm-instantiate")})})}}T||uf("wasm-instantiate");function oe(ge){j(ge.instance,ge.module)}function Me(ge){return pf().then(function(Ne){return WebAssembly.instantiate(Ne,P)}).then(function(Ne){return Ne}).then(ge,function(Ne){H("failed to asynchronously prepare wasm: "+Ne),Rs(Ne)})}function He(){return!ee&&typeof WebAssembly.instantiateStreaming=="function"&&!pc(ma)&&!yl(ma)&&!C&&typeof fetch=="function"?fetch(ma,{credentials:"same-origin"}).then(function(ge){var Ne=WebAssembly.instantiateStreaming(ge,P);return Ne.then(oe,function(Tt){return H("wasm streaming compile failed: "+Tt),H("falling back to ArrayBuffer instantiation"),Me(oe)})}):Me(oe)}if(d.instantiateWasm)try{var Be=d.instantiateWasm(P,j);return Be}catch(ge){H("Module.instantiateWasm callback failed with error: "+ge),f(ge)}return He().catch(f),{}}var m5,g5,hc={};function Es(P){this.name="ExitStatus",this.message="Program terminated with exit("+P+")",this.status=P}function hf(P){var j=We.pthreads[P];delete We.pthreads[P],j.terminate(),E2(P),We.runningWorkers.splice(We.runningWorkers.indexOf(j),1),j.pthread_ptr=0}function ff(P){var j=We.pthreads[P];j.postMessage({cmd:"cancel"})}function id(P){var j=We.pthreads[P];Le(j),We.returnWorkerToPool(j)}function mf(P){var j=We.getNewWorker();if(!j)return 6;We.runningWorkers.push(j),We.pthreads[P.pthread_ptr]=j,j.pthread_ptr=P.pthread_ptr;var oe={cmd:"run",start_routine:P.startRoutine,arg:P.arg,pthread_ptr:P.pthread_ptr};return j.runPthread=()=>{C&&j.ref(),j.postMessage(oe,P.transferList),delete j.runPthread},j.loaded&&j.runPthread(),0}var fc={varargs:void 0,get:function(){fc.varargs+=4;var P=l()[fc.varargs-4>>>2];return P},getStr:function(P){var j=dt(P);return j}};function mc(P){if(T)return Or(1,1,P);Re=P,wn()||(We.terminateAllThreads(),d.onExit&&d.onExit(P),Ce=!0),A(P,new Es(P))}function gf(P,j){if(Re=P,!j&&T)throw yc(P),"unwind";mc(P)}var gc=gf;function yf(P){if(P instanceof Es||P=="unwind")return Re;A(1,P)}var We={unusedWorkers:[],runningWorkers:[],tlsInitFunctions:[],pthreads:{},init:function(){T?We.initWorker():We.initMainThread()},initMainThread:function(){for(var P=8;P--;)We.allocateUnusedWorker()},initWorker:function(){me=!1},setExitStatus:function(P){Re=P},terminateAllThreads:function(){for(var P of Object.values(We.pthreads))We.returnWorkerToPool(P);for(var P of We.unusedWorkers)P.terminate();We.unusedWorkers=[]},returnWorkerToPool:function(P){var j=P.pthread_ptr;delete We.pthreads[j],We.unusedWorkers.push(P),We.runningWorkers.splice(We.runningWorkers.indexOf(P),1),P.pthread_ptr=0,C&&P.unref(),E2(j)},receiveObjectTransfer:function(P){},threadInitTLS:function(){We.tlsInitFunctions.forEach(P=>P())},loadWasmModuleToWorker:function(P,j){P.onmessage=Be=>{var ge=Be.data,Ne=ge.cmd;if(P.pthread_ptr&&(We.currentProxiedOperationCallerThread=P.pthread_ptr),ge.targetThread&&ge.targetThread!=Mc()){var Tt=We.pthreads[ge.targetThread];Tt?Tt.postMessage(ge,ge.transferList):H('Internal error! Worker sent a message "'+Ne+'" to target pthread '+ge.targetThread+", but that thread no longer exists!"),We.currentProxiedOperationCallerThread=void 0;return}Ne==="processProxyingQueue"?ld(ge.queue):Ne==="spawnThread"?mf(ge):Ne==="cleanupThread"?id(ge.thread):Ne==="killThread"?hf(ge.thread):Ne==="cancelThread"?ff(ge.thread):Ne==="loaded"?(P.loaded=!0,C&&P.unref(),j&&j(P),P.runPthread&&P.runPthread()):Ne==="print"?q("Thread "+ge.threadId+": "+ge.text):Ne==="printErr"?H("Thread "+ge.threadId+": "+ge.text):Ne==="alert"?alert("Thread "+ge.threadId+": "+ge.text):ge.target==="setimmediate"?P.postMessage(ge):Ne==="callHandler"?d[ge.handler](...ge.args):Ne&&H("worker sent an unknown command "+Ne),We.currentProxiedOperationCallerThread=void 0},P.onerror=Be=>{var ge="worker sent an error!";throw H(ge+" "+Be.filename+":"+Be.lineno+": "+Be.message),Be},C&&(P.on("message",function(Be){P.onmessage({data:Be})}),P.on("error",function(Be){P.onerror(Be)}),P.on("detachedExit",function(){}));var oe=[],Me=["onExit","onAbort","print","printErr"];for(var He of Me)d.hasOwnProperty(He)&&oe.push(He);P.postMessage({cmd:"load",handlers:oe,urlOrBlob:d.mainScriptUrlOrBlob||n,wasmMemory:ie,wasmModule:Ae})},allocateUnusedWorker:function(){var P,j=M("tfjs-backend-wasm-threaded-simd.worker.js");P=new Worker(j),We.unusedWorkers.push(P)},getNewWorker:function(){return We.unusedWorkers.length==0&&(We.allocateUnusedWorker(),We.loadWasmModuleToWorker(We.unusedWorkers[0])),We.unusedWorkers.pop()}};d.PThread=We;function od(P){for(;P.length>0;)P.shift()(d)}function xf(){var P=Mc(),j=l()[P+52>>>2],oe=l()[P+56>>>2],Me=j-oe;w5(j,Me),Fc(j)}d.establishStackSpace=xf;function yc(P){if(T)return Or(2,0,P);try{gc(P)}catch(j){yf(j)}}var xl=[];function Af(P){var j=xl[P];return j||(P>=xl.length&&(xl.length=P+1),xl[P]=j=qa.get(P)),j}function bf(P,j){var oe=Af(P)(j);wn()?We.setExitStatus(oe):v5(oe)}d.invokeEntryPoint=bf;function vf(P){We.tlsInitFunctions.push(P)}function wf(P){x5(P,!w,1,!b),We.threadInitTLS()}function kf(P){T?postMessage({cmd:"cleanupThread",thread:P}):id(P)}function xc(P,j,oe,Me){return T?Or(3,1,P,j,oe,Me):Ac(P,j,oe,Me)}function Ac(P,j,oe,Me){if(typeof SharedArrayBuffer=="undefined")return H("Current environment does not support SharedArrayBuffer, pthreads are not available!"),6;var He=[],Be=0;if(T&&(He.length===0||Be))return xc(P,j,oe,Me);if(Be)return Be;var ge={startRoutine:oe,pthread_ptr:P,arg:Me,transferList:He};return T?(ge.cmd="spawnThread",postMessage(ge,He),0):mf(ge)}function If(){return 65536}var Sf=!0;function Cf(){return Sf}function ld(P){Atomics.store(l(),P>>2,1),Mc()&&b5(P),Atomics.compareExchange(l(),P>>2,1,0)}d.executeNotifiedProxyingQueue=ld;function Tf(P,j,oe,Me){if(P==j)setTimeout(()=>ld(Me));else if(T)postMessage({targetThread:P,cmd:"processProxyingQueue",queue:Me});else{var He=We.pthreads[P];if(!He)return;He.postMessage({cmd:"processProxyingQueue",queue:Me})}return 1}function Nf(P,j,oe){return-1}function Rf(){Rs("")}function Ms(P){Ms.shown||(Ms.shown={}),Ms.shown[P]||(Ms.shown[P]=1,C&&(P="warning: "+P),H(P))}function Ef(){C||w||Ms("Blocking on the main thread is very dangerous, see https://emscripten.org/docs/porting/pthreads.html#blocking-on-the-main-browser-thread")}function Mf(){return Date.now()}function bc(){return 4294901760}function Ff(){return bc()}var ud;C?ud=()=>{var P=process.hrtime();return P[0]*1e3+P[1]/1e6}:ud=()=>performance.timeOrigin+performance.now();function _f(P,j,oe){i().copyWithin(P>>>0,j>>>0,j+oe>>>0)}function $f(){return C?vC().cpus().length:navigator.hardwareConcurrency}function Df(P){var j=M2(),oe=P();return Fc(j),oe}function Or(P,j){var oe=arguments.length-2,Me=arguments;return Df(()=>{for(var He=oe,Be=_c(He*8),ge=Be>>3,Ne=0;Ne<oe;Ne++){var Tt=Me[2+Ne];c()[ge+Ne>>>0]=Tt}return A5(P,He,Be,j)})}var dd=[];function Pf(P,j,oe){dd.length=j;for(var Me=oe>>3,He=0;He<j;He++)dd[He]=c()[Me+He>>>0];var Be=P<0,ge=Be?hc[-P-1]:Hf[P];return ge.apply(null,dd)}function Of(P){try{return ie.grow(P-Ge.byteLength+65535>>>16),ht(ie.buffer),1}catch(j){}}function zf(P){var j=i().length;if(P=P>>>0,P<=j)return!1;var oe=bc();if(P>oe)return!1;let Me=(Tt,cn)=>Tt+(cn-Tt%cn)%cn;for(var He=1;He<=4;He*=2){var Be=j*(1+.2/He);Be=Math.min(Be,P+100663296);var ge=Math.min(oe,Me(Math.max(P,Be),65536)),Ne=Of(ge);if(Ne)return!0}return!1}function Lf(){throw"unwind"}function vc(P){return T?Or(4,1,P):52}function wc(P,j,oe,Me,He){return T?Or(5,1,P,j,oe,Me,He):70}var Wf=[null,[],[]];function Bf(P,j){var oe=Wf[P];j===0||j===10?((P===1?q:H)(mt(oe,0)),oe.length=0):oe.push(j)}function kc(P,j,oe,Me){if(T)return Or(6,1,P,j,oe,Me);for(var He=0,Be=0;Be<oe;Be++){var ge=u()[j>>>2],Ne=u()[j+4>>>2];j+=8;for(var Tt=0;Tt<Ne;Tt++)Bf(P,i()[ge+Tt>>>0]);He+=Ne}return u()[Me>>>2]=He,0}function Ic(P){var j=d["_"+P];return j}function Vf(P,j){s().set(P,j>>>0)}function Uf(P,j,oe,Me,He){var Be={string:hn=>{var wl=0;if(hn!=null&&hn!==0){var S5=(hn.length<<2)+1;wl=_c(S5),it(hn,wl,S5)}return wl},array:hn=>{var wl=_c(hn.length);return Vf(hn,wl),wl}};function ge(hn){return j==="string"?dt(hn):j==="boolean"?!!hn:hn}var Ne=Ic(P),Tt=[],cn=0;if(Me)for(var gr=0;gr<Me.length;gr++){var I5=Be[oe[gr]];I5?(cn===0&&(cn=M2()),Tt[gr]=I5(Me[gr])):Tt[gr]=Me[gr]}var F2=Ne.apply(null,Tt);function KS(hn){return cn!==0&&Fc(cn),ge(hn)}return F2=KS(F2),F2}function Gf(P,j,oe,Me){oe=oe||[];var He=oe.every(ge=>ge==="number"||ge==="boolean"),Be=j!=="string";return Be&&He&&!Me?Ic(P):function(){return Uf(P,j,oe,arguments,Me)}}We.init();var Hf=[null,mc,yc,xc,vc,wc,kc],Sc={__emscripten_init_main_thread_js:wf,__emscripten_thread_cleanup:kf,__pthread_create_js:Ac,_emscripten_default_pthread_stack_size:If,_emscripten_get_now_is_monotonic:Cf,_emscripten_notify_task_queue:Tf,_emscripten_set_offscreencanvas_size:Nf,abort:Rf,emscripten_check_blocking_allowed:Ef,emscripten_date_now:Mf,emscripten_get_heap_max:Ff,emscripten_get_now:ud,emscripten_memcpy_big:_f,emscripten_num_logical_cores:$f,emscripten_receive_on_main_thread_js:Pf,emscripten_resize_heap:zf,emscripten_unwind_to_js_event_loop:Lf,exit:gc,fd_close:vc,fd_seek:wc,fd_write:kc,memory:ie||d.wasmMemory},y5=cf(),jf=d.___wasm_call_ctors=function(){return(jf=d.___wasm_call_ctors=d.asm.__wasm_call_ctors).apply(null,arguments)},qf=d._init=function(){return(qf=d._init=d.asm.init).apply(null,arguments)},Xf=d._init_with_threads_count=function(){return(Xf=d._init_with_threads_count=d.asm.init_with_threads_count).apply(null,arguments)},Kf=d._get_threads_count=function(){return(Kf=d._get_threads_count=d.asm.get_threads_count).apply(null,arguments)},Yf=d._register_tensor=function(){return(Yf=d._register_tensor=d.asm.register_tensor).apply(null,arguments)},Zf=d._dispose_data=function(){return(Zf=d._dispose_data=d.asm.dispose_data).apply(null,arguments)},Jf=d._dispose=function(){return(Jf=d._dispose=d.asm.dispose).apply(null,arguments)},Qf=d._Abs=function(){return(Qf=d._Abs=d.asm.Abs).apply(null,arguments)},em=d._Acos=function(){return(em=d._Acos=d.asm.Acos).apply(null,arguments)},tm=d._Acosh=function(){return(tm=d._Acosh=d.asm.Acosh).apply(null,arguments)},am=d._Add=function(){return(am=d._Add=d.asm.Add).apply(null,arguments)},nm=d._AddN=function(){return(nm=d._AddN=d.asm.AddN).apply(null,arguments)},rm=d._All=function(){return(rm=d._All=d.asm.All).apply(null,arguments)},sm=d._Any=function(){return(sm=d._Any=d.asm.Any).apply(null,arguments)},im=d._ArgMax=function(){return(im=d._ArgMax=d.asm.ArgMax).apply(null,arguments)},om=d._ArgMin=function(){return(om=d._ArgMin=d.asm.ArgMin).apply(null,arguments)},lm=d._Asin=function(){return(lm=d._Asin=d.asm.Asin).apply(null,arguments)},um=d._Asinh=function(){return(um=d._Asinh=d.asm.Asinh).apply(null,arguments)},dm=d._Atan=function(){return(dm=d._Atan=d.asm.Atan).apply(null,arguments)},pm=d._Atan2=function(){return(pm=d._Atan2=d.asm.Atan2).apply(null,arguments)},cm=d._Atanh=function(){return(cm=d._Atanh=d.asm.Atanh).apply(null,arguments)},hm=d._AvgPool=function(){return(hm=d._AvgPool=d.asm.AvgPool).apply(null,arguments)},fm=d._AvgPool3D=function(){return(fm=d._AvgPool3D=d.asm.AvgPool3D).apply(null,arguments)},mm=d._AvgPool3DGrad=function(){return(mm=d._AvgPool3DGrad=d.asm.AvgPool3DGrad).apply(null,arguments)},gm=d._BatchMatMul=function(){return(gm=d._BatchMatMul=d.asm.BatchMatMul).apply(null,arguments)},ym=d._Bincount=function(){return(ym=d._Bincount=d.asm.Bincount).apply(null,arguments)},xm=d._Ceil=function(){return(xm=d._Ceil=d.asm.Ceil).apply(null,arguments)},Am=d._ClipByValue=function(){return(Am=d._ClipByValue=d.asm.ClipByValue).apply(null,arguments)},bm=d._Conv2D=function(){return(bm=d._Conv2D=d.asm.Conv2D).apply(null,arguments)},vm=d._Conv2DBackpropInput=function(){return(vm=d._Conv2DBackpropInput=d.asm.Conv2DBackpropInput).apply(null,arguments)},wm=d._Conv3D=function(){return(wm=d._Conv3D=d.asm.Conv3D).apply(null,arguments)},km=d._Conv3DBackpropFilterV2=function(){return(km=d._Conv3DBackpropFilterV2=d.asm.Conv3DBackpropFilterV2).apply(null,arguments)},Im=d._Conv3DBackpropInputV2=function(){return(Im=d._Conv3DBackpropInputV2=d.asm.Conv3DBackpropInputV2).apply(null,arguments)},Sm=d._Cos=function(){return(Sm=d._Cos=d.asm.Cos).apply(null,arguments)},Cm=d._Cosh=function(){return(Cm=d._Cosh=d.asm.Cosh).apply(null,arguments)},Tm=d._CropAndResize=function(){return(Tm=d._CropAndResize=d.asm.CropAndResize).apply(null,arguments)},Nm=d._Cumprod=function(){return(Nm=d._Cumprod=d.asm.Cumprod).apply(null,arguments)},Rm=d._Cumsum=function(){return(Rm=d._Cumsum=d.asm.Cumsum).apply(null,arguments)},Em=d._DenseBincount=function(){return(Em=d._DenseBincount=d.asm.DenseBincount).apply(null,arguments)},Mm=d._DepthToSpace=function(){return(Mm=d._DepthToSpace=d.asm.DepthToSpace).apply(null,arguments)},Fm=d._DepthwiseConv2dNative=function(){return(Fm=d._DepthwiseConv2dNative=d.asm.DepthwiseConv2dNative).apply(null,arguments)},_m=d._Diag=function(){return(_m=d._Diag=d.asm.Diag).apply(null,arguments)},$m=d._Dilation2D=function(){return($m=d._Dilation2D=d.asm.Dilation2D).apply(null,arguments)},Dm=d._Dilation2DBackpropFilter=function(){return(Dm=d._Dilation2DBackpropFilter=d.asm.Dilation2DBackpropFilter).apply(null,arguments)},Pm=d._Dilation2DBackpropInput=function(){return(Pm=d._Dilation2DBackpropInput=d.asm.Dilation2DBackpropInput).apply(null,arguments)},Om=d._Elu=function(){return(Om=d._Elu=d.asm.Elu).apply(null,arguments)},zm=d._EluGrad=function(){return(zm=d._EluGrad=d.asm.EluGrad).apply(null,arguments)},Lm=d._Equal=function(){return(Lm=d._Equal=d.asm.Equal).apply(null,arguments)},Wm=d._Exp=function(){return(Wm=d._Exp=d.asm.Exp).apply(null,arguments)},Bm=d._Expm1=function(){return(Bm=d._Expm1=d.asm.Expm1).apply(null,arguments)},Vm=d._FlipLeftRight=function(){return(Vm=d._FlipLeftRight=d.asm.FlipLeftRight).apply(null,arguments)},Um=d._Floor=function(){return(Um=d._Floor=d.asm.Floor).apply(null,arguments)},Gm=d._FloorDiv=function(){return(Gm=d._FloorDiv=d.asm.FloorDiv).apply(null,arguments)},Hm=d._FusedBatchNorm=function(){return(Hm=d._FusedBatchNorm=d.asm.FusedBatchNorm).apply(null,arguments)},jm=d._FusedConv2D=function(){return(jm=d._FusedConv2D=d.asm.FusedConv2D).apply(null,arguments)},qm=d._FusedDepthwiseConv2D=function(){return(qm=d._FusedDepthwiseConv2D=d.asm.FusedDepthwiseConv2D).apply(null,arguments)},Xm=d._Gather=function(){return(Xm=d._Gather=d.asm.Gather).apply(null,arguments)},Km=d._GatherNd=function(){return(Km=d._GatherNd=d.asm.GatherNd).apply(null,arguments)},Ym=d._Greater=function(){return(Ym=d._Greater=d.asm.Greater).apply(null,arguments)},Zm=d._GreaterEqual=function(){return(Zm=d._GreaterEqual=d.asm.GreaterEqual).apply(null,arguments)},Jm=d._IsFinite=function(){return(Jm=d._IsFinite=d.asm.IsFinite).apply(null,arguments)},Qm=d._IsInf=function(){return(Qm=d._IsInf=d.asm.IsInf).apply(null,arguments)},e2=d._IsNan=function(){return(e2=d._IsNan=d.asm.IsNan).apply(null,arguments)},t2=d._LRN=function(){return(t2=d._LRN=d.asm.LRN).apply(null,arguments)},a2=d._LRNGrad=function(){return(a2=d._LRNGrad=d.asm.LRNGrad).apply(null,arguments)},n2=d._LeakyRelu=function(){return(n2=d._LeakyRelu=d.asm.LeakyRelu).apply(null,arguments)},r2=d._Less=function(){return(r2=d._Less=d.asm.Less).apply(null,arguments)},s2=d._LessEqual=function(){return(s2=d._LessEqual=d.asm.LessEqual).apply(null,arguments)},i2=d._LinSpace=function(){return(i2=d._LinSpace=d.asm.LinSpace).apply(null,arguments)},o2=d._Log=function(){return(o2=d._Log=d.asm.Log).apply(null,arguments)},l2=d._Log1p=function(){return(l2=d._Log1p=d.asm.Log1p).apply(null,arguments)},u2=d._LogicalAnd=function(){return(u2=d._LogicalAnd=d.asm.LogicalAnd).apply(null,arguments)},d2=d._LogicalNot=function(){return(d2=d._LogicalNot=d.asm.LogicalNot).apply(null,arguments)},p2=d._LogicalOr=function(){return(p2=d._LogicalOr=d.asm.LogicalOr).apply(null,arguments)},c2=d._LogicalXor=function(){return(c2=d._LogicalXor=d.asm.LogicalXor).apply(null,arguments)},h2=d._Max=function(){return(h2=d._Max=d.asm.Max).apply(null,arguments)},f2=d._MaxPool=function(){return(f2=d._MaxPool=d.asm.MaxPool).apply(null,arguments)},m2=d._MaxPool3D=function(){return(m2=d._MaxPool3D=d.asm.MaxPool3D).apply(null,arguments)},g2=d._MaxPool3DGrad=function(){return(g2=d._MaxPool3DGrad=d.asm.MaxPool3DGrad).apply(null,arguments)},y2=d._Maximum=function(){return(y2=d._Maximum=d.asm.Maximum).apply(null,arguments)},x2=d._Mean=function(){return(x2=d._Mean=d.asm.Mean).apply(null,arguments)},A2=d._Min=function(){return(A2=d._Min=d.asm.Min).apply(null,arguments)},b2=d._Minimum=function(){return(b2=d._Minimum=d.asm.Minimum).apply(null,arguments)},v2=d._MirrorPad=function(){return(v2=d._MirrorPad=d.asm.MirrorPad).apply(null,arguments)},w2=d._Multinomial=function(){return(w2=d._Multinomial=d.asm.Multinomial).apply(null,arguments)},k2=d._Multiply=function(){return(k2=d._Multiply=d.asm.Multiply).apply(null,arguments)},I2=d._Neg=function(){return(I2=d._Neg=d.asm.Neg).apply(null,arguments)},S2=d._NonMaxSuppressionV3=function(){return(S2=d._NonMaxSuppressionV3=d.asm.NonMaxSuppressionV3).apply(null,arguments)},Cc=d._NonMaxSuppressionV4=function(){return(Cc=d._NonMaxSuppressionV4=d.asm.NonMaxSuppressionV4).apply(null,arguments)},Tc=d._NonMaxSuppressionV5=function(){return(Tc=d._NonMaxSuppressionV5=d.asm.NonMaxSuppressionV5).apply(null,arguments)},pd=d._NotEqual=function(){return(pd=d._NotEqual=d.asm.NotEqual).apply(null,arguments)},C2=d._OneHot=function(){return(C2=d._OneHot=d.asm.OneHot).apply(null,arguments)},T2=d._PadV2=function(){return(T2=d._PadV2=d.asm.PadV2).apply(null,arguments)},Al=d._Pow=function(){return(Al=d._Pow=d.asm.Pow).apply(null,arguments)},Nc=d._Prelu=function(){return(Nc=d._Prelu=d.asm.Prelu).apply(null,arguments)},bl=d._Prod=function(){return(bl=d._Prod=d.asm.Prod).apply(null,arguments)},vl=d._RealDiv=function(){return(vl=d._RealDiv=d.asm.RealDiv).apply(null,arguments)},N2=d._Reciprocal=function(){return(N2=d._Reciprocal=d.asm.Reciprocal).apply(null,arguments)},Y=d._Relu=function(){return(Y=d._Relu=d.asm.Relu).apply(null,arguments)},se=d._Relu6=function(){return(se=d._Relu6=d.asm.Relu6).apply(null,arguments)},Ee=d._ResizeBilinear=function(){return(Ee=d._ResizeBilinear=d.asm.ResizeBilinear).apply(null,arguments)},Qe=d._ResizeBilinearGrad=function(){return(Qe=d._ResizeBilinearGrad=d.asm.ResizeBilinearGrad).apply(null,arguments)},wt=d._ResizeNearestNeighbor=function(){return(wt=d._ResizeNearestNeighbor=d.asm.ResizeNearestNeighbor).apply(null,arguments)},kt=d._ResizeNearestNeighborGrad=function(){return(kt=d._ResizeNearestNeighborGrad=d.asm.ResizeNearestNeighborGrad).apply(null,arguments)},Je=d._Reverse=function(){return(Je=d._Reverse=d.asm.Reverse).apply(null,arguments)},Ye=d._RotateWithOffset=function(){return(Ye=d._RotateWithOffset=d.asm.RotateWithOffset).apply(null,arguments)},Lt=d._Round=function(){return(Lt=d._Round=d.asm.Round).apply(null,arguments)},pn=d._Rsqrt=function(){return(pn=d._Rsqrt=d.asm.Rsqrt).apply(null,arguments)},mr=d._ScatterNd=function(){return(mr=d._ScatterNd=d.asm.ScatterNd).apply(null,arguments)},Rc=d._SearchSorted=function(){return(Rc=d._SearchSorted=d.asm.SearchSorted).apply(null,arguments)},cd=d._SelectV2=function(){return(cd=d._SelectV2=d.asm.SelectV2).apply(null,arguments)},R2=d._Selu=function(){return(R2=d._Selu=d.asm.Selu).apply(null,arguments)},za=d._Sigmoid=function(){return(za=d._Sigmoid=d.asm.Sigmoid).apply(null,arguments)},zr=d._Sign=function(){return(zr=d._Sign=d.asm.Sign).apply(null,arguments)},Ec=d._Sin=function(){return(Ec=d._Sin=d.asm.Sin).apply(null,arguments)},yS=d._Softmax=function(){return(yS=d._Softmax=d.asm.Softmax).apply(null,arguments)},xS=d._Softplus=function(){return(xS=d._Softplus=d.asm.Softplus).apply(null,arguments)},AS=d._SparseFillEmptyRows=function(){return(AS=d._SparseFillEmptyRows=d.asm.SparseFillEmptyRows).apply(null,arguments)},bS=d._SparseReshape=function(){return(bS=d._SparseReshape=d.asm.SparseReshape).apply(null,arguments)},vS=d._SparseSegmentReduction=function(){return(vS=d._SparseSegmentReduction=d.asm.SparseSegmentReduction).apply(null,arguments)},wS=d._SparseToDense=function(){return(wS=d._SparseToDense=d.asm.SparseToDense).apply(null,arguments)},kS=d._Sqrt=function(){return(kS=d._Sqrt=d.asm.Sqrt).apply(null,arguments)},IS=d._Square=function(){return(IS=d._Square=d.asm.Square).apply(null,arguments)},SS=d._SquaredDifference=function(){return(SS=d._SquaredDifference=d.asm.SquaredDifference).apply(null,arguments)},CS=d._Step=function(){return(CS=d._Step=d.asm.Step).apply(null,arguments)},TS=d._StridedSlice=function(){return(TS=d._StridedSlice=d.asm.StridedSlice).apply(null,arguments)},NS=d._Sub=function(){return(NS=d._Sub=d.asm.Sub).apply(null,arguments)},RS=d._Sum=function(){return(RS=d._Sum=d.asm.Sum).apply(null,arguments)},ES=d._Tan=function(){return(ES=d._Tan=d.asm.Tan).apply(null,arguments)},MS=d._Tanh=function(){return(MS=d._Tanh=d.asm.Tanh).apply(null,arguments)},FS=d._TensorScatterUpdate=function(){return(FS=d._TensorScatterUpdate=d.asm.TensorScatterUpdate).apply(null,arguments)},_S=d._Tile=function(){return(_S=d._Tile=d.asm.Tile).apply(null,arguments)},$S=d._TopK=function(){return($S=d._TopK=d.asm.TopK).apply(null,arguments)},DS=d._Transform=function(){return(DS=d._Transform=d.asm.Transform).apply(null,arguments)},PS=d._Transpose=function(){return(PS=d._Transpose=d.asm.Transpose).apply(null,arguments)},OS=d.__FusedMatMul=function(){return(OS=d.__FusedMatMul=d.asm._FusedMatMul).apply(null,arguments)},zS=d._malloc=function(){return(zS=d._malloc=d.asm.malloc).apply(null,arguments)},LS=d._free=function(){return(LS=d._free=d.asm.free).apply(null,arguments)},WS=d.__emscripten_tls_init=function(){return(WS=d.__emscripten_tls_init=d.asm._emscripten_tls_init).apply(null,arguments)},Mc=d._pthread_self=function(){return(Mc=d._pthread_self=d.asm.pthread_self).apply(null,arguments)},BS=d.___errno_location=function(){return(BS=d.___errno_location=d.asm.__errno_location).apply(null,arguments)},x5=d.__emscripten_thread_init=function(){return(x5=d.__emscripten_thread_init=d.asm._emscripten_thread_init).apply(null,arguments)},VS=d.__emscripten_thread_crashed=function(){return(VS=d.__emscripten_thread_crashed=d.asm._emscripten_thread_crashed).apply(null,arguments)},US=d._emscripten_main_thread_process_queued_calls=function(){return(US=d._emscripten_main_thread_process_queued_calls=d.asm.emscripten_main_thread_process_queued_calls).apply(null,arguments)},GS=d._emscripten_main_browser_thread_id=function(){return(GS=d._emscripten_main_browser_thread_id=d.asm.emscripten_main_browser_thread_id).apply(null,arguments)},A5=d._emscripten_run_in_main_runtime_thread_js=function(){return(A5=d._emscripten_run_in_main_runtime_thread_js=d.asm.emscripten_run_in_main_runtime_thread_js).apply(null,arguments)},HS=d._emscripten_dispatch_to_thread_=function(){return(HS=d._emscripten_dispatch_to_thread_=d.asm.emscripten_dispatch_to_thread_).apply(null,arguments)},b5=d.__emscripten_proxy_execute_task_queue=function(){return(b5=d.__emscripten_proxy_execute_task_queue=d.asm._emscripten_proxy_execute_task_queue).apply(null,arguments)},E2=d.__emscripten_thread_free_data=function(){return(E2=d.__emscripten_thread_free_data=d.asm._emscripten_thread_free_data).apply(null,arguments)},v5=d.__emscripten_thread_exit=function(){return(v5=d.__emscripten_thread_exit=d.asm._emscripten_thread_exit).apply(null,arguments)},w5=d._emscripten_stack_set_limits=function(){return(w5=d._emscripten_stack_set_limits=d.asm.emscripten_stack_set_limits).apply(null,arguments)},M2=d.stackSave=function(){return(M2=d.stackSave=d.asm.stackSave).apply(null,arguments)},Fc=d.stackRestore=function(){return(Fc=d.stackRestore=d.asm.stackRestore).apply(null,arguments)},_c=d.stackAlloc=function(){return(_c=d.stackAlloc=d.asm.stackAlloc).apply(null,arguments)},jS=d.dynCall_iijjiiii=function(){return(jS=d.dynCall_iijjiiii=d.asm.dynCall_iijjiiii).apply(null,arguments)},qS=d.dynCall_jiji=function(){return(qS=d.dynCall_jiji=d.asm.dynCall_jiji).apply(null,arguments)};d.keepRuntimeAlive=wn,d.wasmMemory=ie,d.cwrap=Gf,d.ExitStatus=Es,d.PThread=We;var $c;fr=function P(){$c||k5(),$c||(fr=P)};function k5(P){if(P=P||y,Pr>0)return;if(T){h(d),Kt(),startWorker(d);return}if(Dr(),Pr>0)return;function j(){$c||($c=!0,d.calledRun=!0,!Ce&&(Kt(),h(d),d.onRuntimeInitialized&&d.onRuntimeInitialized(),uc()))}d.setStatus?(d.setStatus("Running..."),setTimeout(function(){setTimeout(function(){d.setStatus("")},1),j()},1)):j()}if(d.preInit)for(typeof d.preInit=="function"&&(d.preInit=[d.preInit]);d.preInit.length>0;)d.preInit.pop()();k5();var Dc;m&&(Dc={uncaughtException:process.listeners("uncaughtException").filter(function(P){return!m.uncaughtException.indexOf(P)>-1}),unhandledRejection:process.listeners("unhandledRejection").filter(function(P){return!m.unhandledRejection.indexOf(P)>-1})});var Pc;if(typeof WasmBackendModule!="undefined")Pc=WasmBackendModule;else if(typeof r!="undefined")Pc=r;else throw new Error("Could not find wasm module in post.js");if(Dc){var XS=Pc._dispose;Pc._dispose=function(){XS(),Dc.uncaughtException.forEach(function(P){process.removeListener("uncaughtException",P)}),Dc.unhandledRejection.forEach(function(P){process.removeListener("unhandledRejection",P)})}}return r.ready}})();typeof e=="object"&&typeof t=="object"?t.exports=a:typeof define=="function"&&define.amd?define([],function(){return a}):typeof e=="object"&&(e.WasmBackendModuleThreadedSimd=a)}),kC=qt((e,t)=>{t.exports.wasmWorkerContents=`"use strict";var Module={};var ENVIRONMENT_IS_NODE=typeof process=="object"&&typeof process.versions=="object"&&typeof process.versions.node=="string";if(ENVIRONMENT_IS_NODE){var nodeWorkerThreads=require("worker_threads");var parentPort=nodeWorkerThreads.parentPort;parentPort.on("message",data=>onmessage({data:data}));var fs=require("fs");Object.assign(global,{self:global,require:require,Module:Module,location:{href:__filename},Worker:nodeWorkerThreads.Worker,importScripts:function(f){(0,eval)(fs.readFileSync(f,"utf8")+"//# sourceURL="+f)},postMessage:function(msg){parentPort.postMessage(msg)},performance:global.performance||{now:function(){return Date.now()}}})}var initializedJS=false;var pendingNotifiedProxyingQueues=[];function threadPrintErr(){var text=Array.prototype.slice.call(arguments).join(" ");if(ENVIRONMENT_IS_NODE){fs.writeSync(2,text+"
|
|
");return}console.error(text)}function threadAlert(){var text=Array.prototype.slice.call(arguments).join(" ");postMessage({cmd:"alert",text:text,threadId:Module["_pthread_self"]()})}var err=threadPrintErr;self.alert=threadAlert;Module["instantiateWasm"]=(info,receiveInstance)=>{var instance=new WebAssembly.Instance(Module["wasmModule"],info);receiveInstance(instance);Module["wasmModule"]=null;return instance.exports};self.onunhandledrejection=e=>{throw e.reason??e};self.startWorker=instance=>{Module=instance;postMessage({"cmd":"loaded"})};self.onmessage=e=>{try{if(e.data.cmd==="load"){Module["wasmModule"]=e.data.wasmModule;for(const handler of e.data.handlers){Module[handler]=function(){postMessage({cmd:"callHandler",handler:handler,args:[...arguments]})}}Module["wasmMemory"]=e.data.wasmMemory;Module["buffer"]=Module["wasmMemory"].buffer;Module["ENVIRONMENT_IS_PTHREAD"]=true;if(typeof e.data.urlOrBlob=="string"){importScripts(e.data.urlOrBlob)}else{var objectUrl=URL.createObjectURL(e.data.urlOrBlob);importScripts(objectUrl);URL.revokeObjectURL(objectUrl)}WasmBackendModuleThreadedSimd(Module)}else if(e.data.cmd==="run"){Module["__emscripten_thread_init"](e.data.pthread_ptr,0,0,1);Module["establishStackSpace"]();Module["PThread"].receiveObjectTransfer(e.data);Module["PThread"].threadInitTLS();if(!initializedJS){pendingNotifiedProxyingQueues.forEach(queue=>{Module["executeNotifiedProxyingQueue"](queue)});pendingNotifiedProxyingQueues=[];initializedJS=true}try{Module["invokeEntryPoint"](e.data.start_routine,e.data.arg)}catch(ex){if(ex!="unwind"){if(ex instanceof Module["ExitStatus"]){if(Module["keepRuntimeAlive"]()){}else{Module["__emscripten_thread_exit"](ex.status)}}else{throw ex}}}}else if(e.data.cmd==="cancel"){if(Module["_pthread_self"]()){Module["__emscripten_thread_exit"](-1)}}else if(e.data.target==="setimmediate"){}else if(e.data.cmd==="processProxyingQueue"){if(initializedJS){Module["executeNotifiedProxyingQueue"](e.data.queue)}else{pendingNotifiedProxyingQueues.push(e.data.queue)}}else if(e.data.cmd){err("worker.js received unknown command "+e.data.cmd);err(e.data)}}catch(ex){if(Module["__emscripten_thread_crashed"]){Module["__emscripten_thread_crashed"]()}throw ex}};`}),IC=qt((e,t)=>{var a=(()=>{var n=typeof document!="undefined"&&document.currentScript?document.currentScript.src:void 0;return typeof __filename!="undefined"&&(n=n||__filename),function(r){r=r||{};var s=typeof r!="undefined"?r:{},i,o;s.ready=new Promise(function(Y,se){i=Y,o=se});var l;typeof process!="undefined"&&process.listeners&&(l={uncaughtException:process.listeners("uncaughtException"),unhandledRejection:process.listeners("unhandledRejection")});var u=Object.assign({},s),p=[],c="./this.program",d=(Y,se)=>{throw se},h=typeof window=="object",f=typeof importScripts=="function",m=typeof process=="object"&&typeof process.versions=="object"&&typeof process.versions.node=="string",g="";function y(Y){return s.locateFile?s.locateFile(Y,g):g+Y}var x,A,b,w;function C(Y){Y instanceof gl||F("exiting due to exception: "+Y)}if(m){var T=qx(),N=Xx();f?g=N.dirname(g)+"/":g=__dirname+"/",x=(Y,se)=>(Y=Dr(Y)?new URL(Y):N.normalize(Y),T.readFileSync(Y,se?void 0:"utf8")),b=Y=>{var se=x(Y,!0);return se.buffer||(se=new Uint8Array(se)),se},A=(Y,se,Ee)=>{Y=Dr(Y)?new URL(Y):N.normalize(Y),T.readFile(Y,function(Qe,wt){Qe?Ee(Qe):se(wt.buffer)})},process.argv.length>1&&(c=process.argv[1].replace(/\\/g,"/")),p=process.argv.slice(2),process.on("uncaughtException",function(Y){if(!(Y instanceof gl))throw Y}),process.on("unhandledRejection",function(Y){throw Y}),d=(Y,se)=>{if(ja())throw process.exitCode=Y,se;C(se),process.exit(Y)},s.inspect=function(){return"[Emscripten Module object]"}}else(h||f)&&(f?g=self.location.href:typeof document!="undefined"&&document.currentScript&&(g=document.currentScript.src),n&&(g=n),g.indexOf("blob:")!==0?g=g.substr(0,g.replace(/[?#].*/,"").lastIndexOf("/")+1):g="",x=Y=>{var se=new XMLHttpRequest;return se.open("GET",Y,!1),se.send(null),se.responseText},f&&(b=Y=>{var se=new XMLHttpRequest;return se.open("GET",Y,!1),se.responseType="arraybuffer",se.send(null),new Uint8Array(se.response)}),A=(Y,se,Ee)=>{var Qe=new XMLHttpRequest;Qe.open("GET",Y,!0),Qe.responseType="arraybuffer",Qe.onload=()=>{if(Qe.status==200||Qe.status==0&&Qe.response){se(Qe.response);return}Ee()},Qe.onerror=Ee,Qe.send(null)},w=Y=>document.title=Y);var M=s.print||console.log.bind(console),F=s.printErr||console.warn.bind(console);Object.assign(s,u),u=null,s.arguments&&(p=s.arguments),s.thisProgram&&(c=s.thisProgram),s.quit&&(d=s.quit);var E=4,I;s.wasmBinary&&(I=s.wasmBinary);var _=s.noExitRuntime||!0;typeof WebAssembly!="object"&&jn("no native wasm support detected");var O,B=!1,$;function U(Y,se){Y||jn(se)}var G=typeof TextDecoder!="undefined"?new TextDecoder("utf8"):void 0;function q(Y,se,Ee){se>>>=0;for(var Qe=se+Ee,wt=se;Y[wt]&&!(wt>=Qe);)++wt;if(wt-se>16&&Y.buffer&&G)return G.decode(Y.subarray(se,wt));for(var kt="";se<wt;){var Je=Y[se++];if(!(Je&128)){kt+=String.fromCharCode(Je);continue}var Ye=Y[se++]&63;if((Je&224)==192){kt+=String.fromCharCode((Je&31)<<6|Ye);continue}var Lt=Y[se++]&63;if((Je&240)==224?Je=(Je&15)<<12|Ye<<6|Lt:Je=(Je&7)<<18|Ye<<12|Lt<<6|Y[se++]&63,Je<65536)kt+=String.fromCharCode(Je);else{var pn=Je-65536;kt+=String.fromCharCode(55296|pn>>10,56320|pn&1023)}}return kt}function H(Y,se){return Y>>>=0,Y?q(ee,Y,se):""}function V(Y,se,Ee,Qe){if(Ee>>>=0,!(Qe>0))return 0;for(var wt=Ee,kt=Ee+Qe-1,Je=0;Je<Y.length;++Je){var Ye=Y.charCodeAt(Je);if(Ye>=55296&&Ye<=57343){var Lt=Y.charCodeAt(++Je);Ye=65536+((Ye&1023)<<10)|Lt&1023}if(Ye<=127){if(Ee>=kt)break;se[Ee++>>>0]=Ye}else if(Ye<=2047){if(Ee+1>=kt)break;se[Ee++>>>0]=192|Ye>>6,se[Ee++>>>0]=128|Ye&63}else if(Ye<=65535){if(Ee+2>=kt)break;se[Ee++>>>0]=224|Ye>>12,se[Ee++>>>0]=128|Ye>>6&63,se[Ee++>>>0]=128|Ye&63}else{if(Ee+3>=kt)break;se[Ee++>>>0]=240|Ye>>18,se[Ee++>>>0]=128|Ye>>12&63,se[Ee++>>>0]=128|Ye>>6&63,se[Ee++>>>0]=128|Ye&63}}return se[Ee>>>0]=0,Ee-wt}function Z(Y,se,Ee){return V(Y,ee,se,Ee)}var X,re,ee,me,ie,Ae,Ce,Re,Le;function qe(Y){X=Y,s.HEAP8=re=new Int8Array(Y),s.HEAP16=me=new Int16Array(Y),s.HEAP32=Ae=new Int32Array(Y),s.HEAPU8=ee=new Uint8Array(Y),s.HEAPU16=ie=new Uint16Array(Y),s.HEAPU32=Ce=new Uint32Array(Y),s.HEAPF32=Re=new Float32Array(Y),s.HEAPF64=Le=new Float64Array(Y)}var mt=s.INITIAL_MEMORY||16777216,dt,st=[],it=[],Ge=[],yt=!1;function ja(){return _}function zt(){if(s.preRun)for(typeof s.preRun=="function"&&(s.preRun=[s.preRun]);s.preRun.length;)Da(s.preRun.shift());fr(st)}function un(){yt=!0,fr(it)}function la(){if(s.postRun)for(typeof s.postRun=="function"&&(s.postRun=[s.postRun]);s.postRun.length;)Pa(s.postRun.shift());fr(Ge)}function Da(Y){st.unshift(Y)}function dn(Y){it.unshift(Y)}function Pa(Y){Ge.unshift(Y)}var ht=0,Oa=null,qa=null;function hr(Y){ht++,s.monitorRunDependencies&&s.monitorRunDependencies(ht)}function ml(Y){if(ht--,s.monitorRunDependencies&&s.monitorRunDependencies(ht),ht==0&&(Oa!==null&&(clearInterval(Oa),Oa=null),qa)){var se=qa;qa=null,se()}}function jn(Y){s.onAbort&&s.onAbort(Y),Y="Aborted("+Y+")",F(Y),B=!0,$=1,Y+=". Build with -sASSERTIONS for more info.";var se=new WebAssembly.RuntimeError(Y);throw o(se),se}var sd="data:application/octet-stream;base64,";function wn(Y){return Y.startsWith(sd)}function Dr(Y){return Y.startsWith("file://")}var Kt;Kt="tfjs-backend-wasm.wasm",wn(Kt)||(Kt=y(Kt));function uc(Y){try{if(Y==Kt&&I)return new Uint8Array(I);if(b)return b(Y);throw"both async and sync fetching of the wasm failed"}catch(se){jn(se)}}function of(){if(!I&&(h||f)){if(typeof fetch=="function"&&!Dr(Kt))return fetch(Kt,{credentials:"same-origin"}).then(function(Y){if(!Y.ok)throw"failed to load wasm binary file at '"+Kt+"'";return Y.arrayBuffer()}).catch(function(){return uc(Kt)});if(A)return new Promise(function(Y,se){A(Kt,function(Ee){Y(new Uint8Array(Ee))},se)})}return Promise.resolve().then(function(){return uc(Kt)})}function lf(){var Y={env:id,wasi_snapshot_preview1:id};function se(Je,Ye){var Lt=Je.exports;s.asm=Lt,O=s.asm.memory,qe(O.buffer),dt=s.asm.__indirect_function_table,dn(s.asm.__wasm_call_ctors),ml("wasm-instantiate")}hr("wasm-instantiate");function Ee(Je){se(Je.instance)}function Qe(Je){return of().then(function(Ye){return WebAssembly.instantiate(Ye,Y)}).then(function(Ye){return Ye}).then(Je,function(Ye){F("failed to asynchronously prepare wasm: "+Ye),jn(Ye)})}function wt(){return!I&&typeof WebAssembly.instantiateStreaming=="function"&&!wn(Kt)&&!Dr(Kt)&&!m&&typeof fetch=="function"?fetch(Kt,{credentials:"same-origin"}).then(function(Je){var Ye=WebAssembly.instantiateStreaming(Je,Y);return Ye.then(Ee,function(Lt){return F("wasm streaming compile failed: "+Lt),F("falling back to ArrayBuffer instantiation"),Qe(Ee)})}):Qe(Ee)}if(s.instantiateWasm)try{var kt=s.instantiateWasm(Y,se);return kt}catch(Je){F("Module.instantiateWasm callback failed with error: "+Je),o(Je)}return wt().catch(o),{}}var f5,Pr;function gl(Y){this.name="ExitStatus",this.message="Program terminated with exit("+Y+")",this.status=Y}function fr(Y){for(;Y.length>0;)Y.shift()(s)}function uf(){jn("")}function dc(){return 4294901760}function Rs(){return dc()}function df(Y,se,Ee){ee.copyWithin(Y>>>0,se>>>0,se+Ee>>>0)}function pc(Y){try{return O.grow(Y-X.byteLength+65535>>>16),qe(O.buffer),1}catch(se){}}function yl(Y){var se=ee.length;Y=Y>>>0;var Ee=dc();if(Y>Ee)return!1;let Qe=(Lt,pn)=>Lt+(pn-Lt%pn)%pn;for(var wt=1;wt<=4;wt*=2){var kt=se*(1+.2/wt);kt=Math.min(kt,Y+100663296);var Je=Math.min(Ee,Qe(Math.max(Y,kt),65536)),Ye=pc(Je);if(Ye)return!0}return!1}var ma={varargs:void 0,get:function(){ma.varargs+=4;var Y=Ae[ma.varargs-4>>>2];return Y},getStr:function(Y){var se=H(Y);return se}};function cc(Y){return 52}function pf(Y,se,Ee,Qe,wt){return 70}var cf=[null,[],[]];function m5(Y,se){var Ee=cf[Y];se===0||se===10?((Y===1?M:F)(q(Ee,0)),Ee.length=0):Ee.push(se)}function g5(Y,se,Ee,Qe){for(var wt=0,kt=0;kt<Ee;kt++){var Je=Ce[se>>>2],Ye=Ce[se+4>>>2];se+=8;for(var Lt=0;Lt<Ye;Lt++)m5(Y,ee[Je+Lt>>>0]);wt+=Ye}return Ce[Qe>>>2]=wt,0}function hc(Y){var se=s["_"+Y];return se}function Es(Y,se){re.set(Y,se>>>0)}function hf(Y,se,Ee,Qe,wt){var kt={string:za=>{var zr=0;if(za!=null&&za!==0){var Ec=(za.length<<2)+1;zr=pd(Ec),Z(za,zr,Ec)}return zr},array:za=>{var zr=pd(za.length);return Es(za,zr),zr}};function Je(za){return se==="string"?H(za):se==="boolean"?!!za:za}var Ye=hc(Y),Lt=[],pn=0;if(Qe)for(var mr=0;mr<Qe.length;mr++){var Rc=kt[Ee[mr]];Rc?(pn===0&&(pn=Cc()),Lt[mr]=Rc(Qe[mr])):Lt[mr]=Qe[mr]}var cd=Ye.apply(null,Lt);function R2(za){return pn!==0&&Tc(pn),Je(za)}return cd=R2(cd),cd}function ff(Y,se,Ee,Qe){Ee=Ee||[];var wt=Ee.every(Je=>Je==="number"||Je==="boolean"),kt=se!=="string";return kt&&wt&&!Qe?hc(Y):function(){return hf(Y,se,Ee,arguments,Qe)}}var id={abort:uf,emscripten_get_heap_max:Rs,emscripten_memcpy_big:df,emscripten_resize_heap:yl,fd_close:cc,fd_seek:pf,fd_write:g5},mf=lf(),fc=s.___wasm_call_ctors=function(){return(fc=s.___wasm_call_ctors=s.asm.__wasm_call_ctors).apply(null,arguments)},mc=s._init=function(){return(mc=s._init=s.asm.init).apply(null,arguments)},gf=s._init_with_threads_count=function(){return(gf=s._init_with_threads_count=s.asm.init_with_threads_count).apply(null,arguments)},gc=s._get_threads_count=function(){return(gc=s._get_threads_count=s.asm.get_threads_count).apply(null,arguments)},yf=s._register_tensor=function(){return(yf=s._register_tensor=s.asm.register_tensor).apply(null,arguments)},We=s._dispose_data=function(){return(We=s._dispose_data=s.asm.dispose_data).apply(null,arguments)},od=s._dispose=function(){return(od=s._dispose=s.asm.dispose).apply(null,arguments)},xf=s._Abs=function(){return(xf=s._Abs=s.asm.Abs).apply(null,arguments)},yc=s._Acos=function(){return(yc=s._Acos=s.asm.Acos).apply(null,arguments)},xl=s._Acosh=function(){return(xl=s._Acosh=s.asm.Acosh).apply(null,arguments)},Af=s._Add=function(){return(Af=s._Add=s.asm.Add).apply(null,arguments)},bf=s._AddN=function(){return(bf=s._AddN=s.asm.AddN).apply(null,arguments)},vf=s._All=function(){return(vf=s._All=s.asm.All).apply(null,arguments)},wf=s._Any=function(){return(wf=s._Any=s.asm.Any).apply(null,arguments)},kf=s._ArgMax=function(){return(kf=s._ArgMax=s.asm.ArgMax).apply(null,arguments)},xc=s._ArgMin=function(){return(xc=s._ArgMin=s.asm.ArgMin).apply(null,arguments)},Ac=s._Asin=function(){return(Ac=s._Asin=s.asm.Asin).apply(null,arguments)},If=s._Asinh=function(){return(If=s._Asinh=s.asm.Asinh).apply(null,arguments)},Sf=s._Atan=function(){return(Sf=s._Atan=s.asm.Atan).apply(null,arguments)},Cf=s._Atan2=function(){return(Cf=s._Atan2=s.asm.Atan2).apply(null,arguments)},ld=s._Atanh=function(){return(ld=s._Atanh=s.asm.Atanh).apply(null,arguments)},Tf=s._AvgPool=function(){return(Tf=s._AvgPool=s.asm.AvgPool).apply(null,arguments)},Nf=s._AvgPool3D=function(){return(Nf=s._AvgPool3D=s.asm.AvgPool3D).apply(null,arguments)},Rf=s._AvgPool3DGrad=function(){return(Rf=s._AvgPool3DGrad=s.asm.AvgPool3DGrad).apply(null,arguments)},Ms=s._BatchMatMul=function(){return(Ms=s._BatchMatMul=s.asm.BatchMatMul).apply(null,arguments)},Ef=s._Bincount=function(){return(Ef=s._Bincount=s.asm.Bincount).apply(null,arguments)},Mf=s._Ceil=function(){return(Mf=s._Ceil=s.asm.Ceil).apply(null,arguments)},bc=s._ClipByValue=function(){return(bc=s._ClipByValue=s.asm.ClipByValue).apply(null,arguments)},Ff=s._Conv2D=function(){return(Ff=s._Conv2D=s.asm.Conv2D).apply(null,arguments)},ud=s._Conv2DBackpropInput=function(){return(ud=s._Conv2DBackpropInput=s.asm.Conv2DBackpropInput).apply(null,arguments)},_f=s._Conv3D=function(){return(_f=s._Conv3D=s.asm.Conv3D).apply(null,arguments)},$f=s._Conv3DBackpropFilterV2=function(){return($f=s._Conv3DBackpropFilterV2=s.asm.Conv3DBackpropFilterV2).apply(null,arguments)},Df=s._Conv3DBackpropInputV2=function(){return(Df=s._Conv3DBackpropInputV2=s.asm.Conv3DBackpropInputV2).apply(null,arguments)},Or=s._Cos=function(){return(Or=s._Cos=s.asm.Cos).apply(null,arguments)},dd=s._Cosh=function(){return(dd=s._Cosh=s.asm.Cosh).apply(null,arguments)},Pf=s._CropAndResize=function(){return(Pf=s._CropAndResize=s.asm.CropAndResize).apply(null,arguments)},Of=s._Cumprod=function(){return(Of=s._Cumprod=s.asm.Cumprod).apply(null,arguments)},zf=s._Cumsum=function(){return(zf=s._Cumsum=s.asm.Cumsum).apply(null,arguments)},Lf=s._DenseBincount=function(){return(Lf=s._DenseBincount=s.asm.DenseBincount).apply(null,arguments)},vc=s._DepthToSpace=function(){return(vc=s._DepthToSpace=s.asm.DepthToSpace).apply(null,arguments)},wc=s._DepthwiseConv2dNative=function(){return(wc=s._DepthwiseConv2dNative=s.asm.DepthwiseConv2dNative).apply(null,arguments)},Wf=s._Diag=function(){return(Wf=s._Diag=s.asm.Diag).apply(null,arguments)},Bf=s._Dilation2D=function(){return(Bf=s._Dilation2D=s.asm.Dilation2D).apply(null,arguments)},kc=s._Dilation2DBackpropFilter=function(){return(kc=s._Dilation2DBackpropFilter=s.asm.Dilation2DBackpropFilter).apply(null,arguments)},Ic=s._Dilation2DBackpropInput=function(){return(Ic=s._Dilation2DBackpropInput=s.asm.Dilation2DBackpropInput).apply(null,arguments)},Vf=s._Elu=function(){return(Vf=s._Elu=s.asm.Elu).apply(null,arguments)},Uf=s._EluGrad=function(){return(Uf=s._EluGrad=s.asm.EluGrad).apply(null,arguments)},Gf=s._Equal=function(){return(Gf=s._Equal=s.asm.Equal).apply(null,arguments)},Hf=s._Exp=function(){return(Hf=s._Exp=s.asm.Exp).apply(null,arguments)},Sc=s._Expm1=function(){return(Sc=s._Expm1=s.asm.Expm1).apply(null,arguments)},y5=s._FlipLeftRight=function(){return(y5=s._FlipLeftRight=s.asm.FlipLeftRight).apply(null,arguments)},jf=s._Floor=function(){return(jf=s._Floor=s.asm.Floor).apply(null,arguments)},qf=s._FloorDiv=function(){return(qf=s._FloorDiv=s.asm.FloorDiv).apply(null,arguments)},Xf=s._FusedBatchNorm=function(){return(Xf=s._FusedBatchNorm=s.asm.FusedBatchNorm).apply(null,arguments)},Kf=s._FusedConv2D=function(){return(Kf=s._FusedConv2D=s.asm.FusedConv2D).apply(null,arguments)},Yf=s._FusedDepthwiseConv2D=function(){return(Yf=s._FusedDepthwiseConv2D=s.asm.FusedDepthwiseConv2D).apply(null,arguments)},Zf=s._Gather=function(){return(Zf=s._Gather=s.asm.Gather).apply(null,arguments)},Jf=s._GatherNd=function(){return(Jf=s._GatherNd=s.asm.GatherNd).apply(null,arguments)},Qf=s._Greater=function(){return(Qf=s._Greater=s.asm.Greater).apply(null,arguments)},em=s._GreaterEqual=function(){return(em=s._GreaterEqual=s.asm.GreaterEqual).apply(null,arguments)},tm=s._IsFinite=function(){return(tm=s._IsFinite=s.asm.IsFinite).apply(null,arguments)},am=s._IsInf=function(){return(am=s._IsInf=s.asm.IsInf).apply(null,arguments)},nm=s._IsNan=function(){return(nm=s._IsNan=s.asm.IsNan).apply(null,arguments)},rm=s._LRN=function(){return(rm=s._LRN=s.asm.LRN).apply(null,arguments)},sm=s._LRNGrad=function(){return(sm=s._LRNGrad=s.asm.LRNGrad).apply(null,arguments)},im=s._LeakyRelu=function(){return(im=s._LeakyRelu=s.asm.LeakyRelu).apply(null,arguments)},om=s._Less=function(){return(om=s._Less=s.asm.Less).apply(null,arguments)},lm=s._LessEqual=function(){return(lm=s._LessEqual=s.asm.LessEqual).apply(null,arguments)},um=s._LinSpace=function(){return(um=s._LinSpace=s.asm.LinSpace).apply(null,arguments)},dm=s._Log=function(){return(dm=s._Log=s.asm.Log).apply(null,arguments)},pm=s._Log1p=function(){return(pm=s._Log1p=s.asm.Log1p).apply(null,arguments)},cm=s._LogicalAnd=function(){return(cm=s._LogicalAnd=s.asm.LogicalAnd).apply(null,arguments)},hm=s._LogicalNot=function(){return(hm=s._LogicalNot=s.asm.LogicalNot).apply(null,arguments)},fm=s._LogicalOr=function(){return(fm=s._LogicalOr=s.asm.LogicalOr).apply(null,arguments)},mm=s._LogicalXor=function(){return(mm=s._LogicalXor=s.asm.LogicalXor).apply(null,arguments)},gm=s._Max=function(){return(gm=s._Max=s.asm.Max).apply(null,arguments)},ym=s._MaxPool=function(){return(ym=s._MaxPool=s.asm.MaxPool).apply(null,arguments)},xm=s._MaxPool3D=function(){return(xm=s._MaxPool3D=s.asm.MaxPool3D).apply(null,arguments)},Am=s._MaxPool3DGrad=function(){return(Am=s._MaxPool3DGrad=s.asm.MaxPool3DGrad).apply(null,arguments)},bm=s._Maximum=function(){return(bm=s._Maximum=s.asm.Maximum).apply(null,arguments)},vm=s._Mean=function(){return(vm=s._Mean=s.asm.Mean).apply(null,arguments)},wm=s._Min=function(){return(wm=s._Min=s.asm.Min).apply(null,arguments)},km=s._Minimum=function(){return(km=s._Minimum=s.asm.Minimum).apply(null,arguments)},Im=s._MirrorPad=function(){return(Im=s._MirrorPad=s.asm.MirrorPad).apply(null,arguments)},Sm=s._Multinomial=function(){return(Sm=s._Multinomial=s.asm.Multinomial).apply(null,arguments)},Cm=s._Multiply=function(){return(Cm=s._Multiply=s.asm.Multiply).apply(null,arguments)},Tm=s._Neg=function(){return(Tm=s._Neg=s.asm.Neg).apply(null,arguments)},Nm=s._NonMaxSuppressionV3=function(){return(Nm=s._NonMaxSuppressionV3=s.asm.NonMaxSuppressionV3).apply(null,arguments)},Rm=s._NonMaxSuppressionV4=function(){return(Rm=s._NonMaxSuppressionV4=s.asm.NonMaxSuppressionV4).apply(null,arguments)},Em=s._NonMaxSuppressionV5=function(){return(Em=s._NonMaxSuppressionV5=s.asm.NonMaxSuppressionV5).apply(null,arguments)},Mm=s._NotEqual=function(){return(Mm=s._NotEqual=s.asm.NotEqual).apply(null,arguments)},Fm=s._OneHot=function(){return(Fm=s._OneHot=s.asm.OneHot).apply(null,arguments)},_m=s._PadV2=function(){return(_m=s._PadV2=s.asm.PadV2).apply(null,arguments)},$m=s._Pow=function(){return($m=s._Pow=s.asm.Pow).apply(null,arguments)},Dm=s._Prelu=function(){return(Dm=s._Prelu=s.asm.Prelu).apply(null,arguments)},Pm=s._Prod=function(){return(Pm=s._Prod=s.asm.Prod).apply(null,arguments)},Om=s._RealDiv=function(){return(Om=s._RealDiv=s.asm.RealDiv).apply(null,arguments)},zm=s._Reciprocal=function(){return(zm=s._Reciprocal=s.asm.Reciprocal).apply(null,arguments)},Lm=s._Relu=function(){return(Lm=s._Relu=s.asm.Relu).apply(null,arguments)},Wm=s._Relu6=function(){return(Wm=s._Relu6=s.asm.Relu6).apply(null,arguments)},Bm=s._ResizeBilinear=function(){return(Bm=s._ResizeBilinear=s.asm.ResizeBilinear).apply(null,arguments)},Vm=s._ResizeBilinearGrad=function(){return(Vm=s._ResizeBilinearGrad=s.asm.ResizeBilinearGrad).apply(null,arguments)},Um=s._ResizeNearestNeighbor=function(){return(Um=s._ResizeNearestNeighbor=s.asm.ResizeNearestNeighbor).apply(null,arguments)},Gm=s._ResizeNearestNeighborGrad=function(){return(Gm=s._ResizeNearestNeighborGrad=s.asm.ResizeNearestNeighborGrad).apply(null,arguments)},Hm=s._Reverse=function(){return(Hm=s._Reverse=s.asm.Reverse).apply(null,arguments)},jm=s._RotateWithOffset=function(){return(jm=s._RotateWithOffset=s.asm.RotateWithOffset).apply(null,arguments)},qm=s._Round=function(){return(qm=s._Round=s.asm.Round).apply(null,arguments)},Xm=s._Rsqrt=function(){return(Xm=s._Rsqrt=s.asm.Rsqrt).apply(null,arguments)},Km=s._ScatterNd=function(){return(Km=s._ScatterNd=s.asm.ScatterNd).apply(null,arguments)},Ym=s._SearchSorted=function(){return(Ym=s._SearchSorted=s.asm.SearchSorted).apply(null,arguments)},Zm=s._SelectV2=function(){return(Zm=s._SelectV2=s.asm.SelectV2).apply(null,arguments)},Jm=s._Selu=function(){return(Jm=s._Selu=s.asm.Selu).apply(null,arguments)},Qm=s._Sigmoid=function(){return(Qm=s._Sigmoid=s.asm.Sigmoid).apply(null,arguments)},e2=s._Sign=function(){return(e2=s._Sign=s.asm.Sign).apply(null,arguments)},t2=s._Sin=function(){return(t2=s._Sin=s.asm.Sin).apply(null,arguments)},a2=s._Softmax=function(){return(a2=s._Softmax=s.asm.Softmax).apply(null,arguments)},n2=s._Softplus=function(){return(n2=s._Softplus=s.asm.Softplus).apply(null,arguments)},r2=s._SparseFillEmptyRows=function(){return(r2=s._SparseFillEmptyRows=s.asm.SparseFillEmptyRows).apply(null,arguments)},s2=s._SparseReshape=function(){return(s2=s._SparseReshape=s.asm.SparseReshape).apply(null,arguments)},i2=s._SparseSegmentReduction=function(){return(i2=s._SparseSegmentReduction=s.asm.SparseSegmentReduction).apply(null,arguments)},o2=s._SparseToDense=function(){return(o2=s._SparseToDense=s.asm.SparseToDense).apply(null,arguments)},l2=s._Sqrt=function(){return(l2=s._Sqrt=s.asm.Sqrt).apply(null,arguments)},u2=s._Square=function(){return(u2=s._Square=s.asm.Square).apply(null,arguments)},d2=s._SquaredDifference=function(){return(d2=s._SquaredDifference=s.asm.SquaredDifference).apply(null,arguments)},p2=s._Step=function(){return(p2=s._Step=s.asm.Step).apply(null,arguments)},c2=s._StridedSlice=function(){return(c2=s._StridedSlice=s.asm.StridedSlice).apply(null,arguments)},h2=s._Sub=function(){return(h2=s._Sub=s.asm.Sub).apply(null,arguments)},f2=s._Sum=function(){return(f2=s._Sum=s.asm.Sum).apply(null,arguments)},m2=s._Tan=function(){return(m2=s._Tan=s.asm.Tan).apply(null,arguments)},g2=s._Tanh=function(){return(g2=s._Tanh=s.asm.Tanh).apply(null,arguments)},y2=s._TensorScatterUpdate=function(){return(y2=s._TensorScatterUpdate=s.asm.TensorScatterUpdate).apply(null,arguments)},x2=s._Tile=function(){return(x2=s._Tile=s.asm.Tile).apply(null,arguments)},A2=s._TopK=function(){return(A2=s._TopK=s.asm.TopK).apply(null,arguments)},b2=s._Transform=function(){return(b2=s._Transform=s.asm.Transform).apply(null,arguments)},v2=s._Transpose=function(){return(v2=s._Transpose=s.asm.Transpose).apply(null,arguments)},w2=s.__FusedMatMul=function(){return(w2=s.__FusedMatMul=s.asm._FusedMatMul).apply(null,arguments)},k2=s._malloc=function(){return(k2=s._malloc=s.asm.malloc).apply(null,arguments)},I2=s._free=function(){return(I2=s._free=s.asm.free).apply(null,arguments)},S2=s.___errno_location=function(){return(S2=s.___errno_location=s.asm.__errno_location).apply(null,arguments)},Cc=s.stackSave=function(){return(Cc=s.stackSave=s.asm.stackSave).apply(null,arguments)},Tc=s.stackRestore=function(){return(Tc=s.stackRestore=s.asm.stackRestore).apply(null,arguments)},pd=s.stackAlloc=function(){return(pd=s.stackAlloc=s.asm.stackAlloc).apply(null,arguments)},C2=s.dynCall_iijjiiii=function(){return(C2=s.dynCall_iijjiiii=s.asm.dynCall_iijjiiii).apply(null,arguments)},T2=s.dynCall_jiji=function(){return(T2=s.dynCall_jiji=s.asm.dynCall_jiji).apply(null,arguments)};s.cwrap=ff;var Al;qa=function Y(){Al||Nc(),Al||(qa=Y)};function Nc(Y){if(Y=Y||p,ht>0||(zt(),ht>0))return;function se(){Al||(Al=!0,s.calledRun=!0,!B&&(un(),i(s),s.onRuntimeInitialized&&s.onRuntimeInitialized(),la()))}s.setStatus?(s.setStatus("Running..."),setTimeout(function(){setTimeout(function(){s.setStatus("")},1),se()},1)):se()}if(s.preInit)for(typeof s.preInit=="function"&&(s.preInit=[s.preInit]);s.preInit.length>0;)s.preInit.pop()();Nc();var bl;l&&(bl={uncaughtException:process.listeners("uncaughtException").filter(function(Y){return!l.uncaughtException.indexOf(Y)>-1}),unhandledRejection:process.listeners("unhandledRejection").filter(function(Y){return!l.unhandledRejection.indexOf(Y)>-1})});var vl;if(typeof r!="undefined")vl=r;else if(typeof WasmBackendModuleThreadedSimd!="undefined")vl=WasmBackendModuleThreadedSimd;else throw new Error("Could not find wasm module in post.js");if(bl){var N2=vl._dispose;vl._dispose=function(){N2(),bl.uncaughtException.forEach(function(Y){process.removeListener("uncaughtException",Y)}),bl.unhandledRejection.forEach(function(Y){process.removeListener("unhandledRejection",Y)})}}return r.ready}})();typeof e=="object"&&typeof t=="object"?t.exports=a:typeof define=="function"&&define.amd?define([],function(){return a}):typeof e=="object"&&(e.WasmBackendModule=a)}),Zd=class{constructor(e,t){this.backend=e,this.dataMover=t,this.data=new WeakMap,this.dataIdsCount=0}get(e){return this.data.has(e)||this.dataMover.moveData(this.backend,e),this.data.get(e)}set(e,t){this.dataIdsCount++,this.data.set(e,t)}has(e){return this.data.has(e)}delete(e){return this.dataIdsCount--,this.data.delete(e)}numDataIds(){return this.dataIdsCount}},Kl=class{refCount(e){return Ka("refCount")}incRef(e){return Ka("incRef")}timerAvailable(){return!0}time(e){return Ka("time")}read(e){return Ka("read")}readSync(e){return Ka("readSync")}readToGPU(e,t){return Ka("readToGPU")}numDataIds(){return Ka("numDataIds")}disposeData(e,t){return Ka("disposeData")}write(e,t,a){return Ka("write")}move(e,t,a,n,r){return Ka("move")}createTensorFromGPUData(e,t,a){return Ka("createTensorFromGPUData")}memory(){return Ka("memory")}floatPrecision(){return Ka("floatPrecision")}epsilon(){return this.floatPrecision()===32?1e-7:1e-4}dispose(){return Ka("dispose")}};function Ka(e){throw new Error(`'${e}' not yet implemented or not found in the registry. This kernel may not be supported by the tfjs backend you have chosen`)}function Kx(e){let t=e.length,a=0;for(;t>0;)a=Math.random()*t|0,t--,Qc(e,t,a)}function SC(e,t){if(e.length!==t.length)throw new Error(`Array sizes must match to be shuffled together First array length was ${e.length}Second array length was ${t.length}`);let a=e.length,n=0;for(;a>0;)n=Math.random()*a|0,a--,Qc(e,a,n),Qc(t,a,n)}function Rd(e,t,a){return Math.max(e,Math.min(t,a))}function CC(e){return e%2===0?e:e+1}function Qc(e,t,a){let n=e[t];e[t]=e[a],e[a]=n}function TC(e){let t=0;for(let a=0;a<e.length;a++)t+=e[a];return t}function NC(e,t){let a=Math.random();return t*a+(1-a)*e}function RC(e,t){let a=0;for(let n=0;n<e.length;n++){let r=Number(e[n])-Number(t[n]);a+=r*r}return a}function D(e,t){if(!e)throw new Error(typeof t=="string"?t:t())}function Na(e,t,a=""){D(Cr(e,t),()=>a+` Shapes ${e} and ${t} must match`)}function Js(e){D(e!=null,()=>"The input to the tensor constructor must be a non-null value.")}function vt(e){if(e.length===0)return 1;let t=e[0];for(let a=1;a<e.length;a++)t*=e[a];return t}function EC(e){return e.length===0}function Yx(e,t){if(e===t)return!0;if(e==null||t==null||e.length!==t.length)return!1;for(let a=0;a<e.length;a++)if(e[a]!==null&&t[a]!==null&&e[a]!==t[a])return!1;return!0}function Cr(e,t){if(e===t)return!0;if(e==null||t==null||e.length!==t.length)return!1;for(let a=0;a<e.length;a++)if(e[a]!==t[a])return!1;return!0}function $l(e){return e%1===0}function MC(e){if(Math.tanh!=null)return Math.tanh(e);if(e===1/0)return 1;if(e===-1/0)return-1;{let t=Math.exp(2*e);return(t-1)/(t+1)}}function FC(e){let t=Math.ceil(Math.sqrt(e));return[t,Math.ceil(e/t)]}function _C(e){let t=new Uint32Array(e);for(let a=0;a<e;++a)t[a]=a;return Kx(t),t}function Sd(e,t){return t<=e.length?e:e+" ".repeat(t-e.length)}function $C(e,t=r=>0,a,n){return new Promise((r,s)=>{let i=0,o=()=>{if(e()){r();return}i++;let l=t(i);if(a!=null&&i>=a){s();return}n!=null?n(o,l):setTimeout(o,l)};o()})}function DC(e,t){let a=1,n=-1;for(let s=0;s<e.length;++s)if(e[s]>=0)a*=e[s];else if(e[s]===-1){if(n!==-1)throw Error(`Shapes can only have 1 implicit size. Found -1 at dim ${n} and dim ${s}`);n=s}else if(e[s]<0)throw Error(`Shapes can not be < 0. Found ${e[s]} at dim ${s}`);if(n===-1){if(t>0&&t!==a)throw Error(`Size(${t}) must match the product of shape ${e}`);return e}if(a===0)throw Error(`Cannot infer the missing size in [${e}] when there are 0 elements`);if(t%a!==0)throw Error(`The implicit shape can't be a fractional number. Got ${t} / ${a}`);let r=e.slice();return r[n]=t/a,r}function Jd(e,t){let a=t.length;return e=e==null?t.map((n,r)=>r):[].concat(e),D(e.every(n=>n>=-a&&n<a),()=>`All values in axis param must be in range [-${a}, ${a}) but got axis ${e}`),D(e.every(n=>$l(n)),()=>`All values in axis param must be integers but got axis ${e}`),e.map(n=>n<0?a+n:n)}function Zx(e,t){let a=[],n=[],r=t!=null&&Array.isArray(t)&&t.length===0,s=t==null||r?null:Jd(t,e).sort(),i=0;for(let o=0;o<e.length;++o){if(s!=null){if(s[i]===o&&e[o]!==1)throw new Error(`Can't squeeze axis ${o} since its dim '${e[o]}' is not 1`);(s[i]==null||s[i]>o)&&e[o]===1&&(a.push(e[o]),n.push(o)),s[i]<=o&&i++}e[o]!==1&&(a.push(e[o]),n.push(o))}return{newShape:a,keptDims:n}}function Jx(e,t){return z1(e,t)}function z1(e,t){let a=null;if(e==null||e==="float32")a=new Float32Array(t);else if(e==="int32")a=new Int32Array(t);else if(e==="bool")a=new Uint8Array(t);else if(e==="string")a=new Array(t);else throw new Error(`Unknown data type ${e}`);return a}function Qx(e,t){for(let a=0;a<e.length;a++){let n=e[a];if(isNaN(n)||!isFinite(n))throw Error(`A tensor of type ${t} being uploaded contains ${n}.`)}}function eA(e){return e==="bool"||e==="complex64"||e==="float32"||e==="int32"||e==="string"}function PC(e,t){return!(t==="complex64"||t==="float32"&&e!=="complex64"||t==="int32"&&e!=="float32"&&e!=="complex64"||t==="bool"&&e==="bool")}function eh(e){if(e==="float32"||e==="int32")return 4;if(e==="complex64")return 8;if(e==="bool")return 1;throw new Error(`Unknown dtype ${e}`)}function tA(e){if(e==null)return 0;let t=0;return e.forEach(a=>t+=a.length),t}function Br(e){return typeof e=="string"||e instanceof String}function aA(e){return typeof e=="boolean"}function nA(e){return typeof e=="number"}function Qd(e){return Array.isArray(e)?Qd(e[0]):e instanceof Float32Array?"float32":e instanceof Int32Array||e instanceof Uint8Array||e instanceof Uint8ClampedArray?"int32":nA(e)?"float32":Br(e)?"string":aA(e)?"bool":"float32"}function qr(e){return!!(e&&e.constructor&&e.call&&e.apply)}function th(e,t){for(let a=t;a<e;++a)if(e%a===0)return a;return e}function Yl(e){let t=e.length;if(t<2)return[];let a=new Array(t-1);a[t-2]=e[t-1];for(let n=t-3;n>=0;--n)a[n]=a[n+1]*e[n+1];return a}function rA(e,t,a,n=!1){let r=new Array;if(t.length===1){let s=t[0]*(n?2:1);for(let i=0;i<s;i++)r[i]=a[e+i]}else{let s=t[0],i=t.slice(1),o=i.reduce((l,u)=>l*u)*(n?2:1);for(let l=0;l<s;l++)r[l]=rA(e+l*o,i,a,n)}return r}function Rl(e,t,a=!1){if(e.length===0)return t[0];let n=e.reduce((r,s)=>r*s)*(a?2:1);if(n===0)return[];if(n!==t.length)throw new Error(`[${e}] does not match the input size ${t.length}${a?" for a complex tensor":""}.`);return rA(0,e,t,a)}function OC(e,t){if(Array.isArray(e))return e;if(t==="float32")return e instanceof Float32Array?e:new Float32Array(e);if(t==="int32")return e instanceof Int32Array?e:new Int32Array(e);if(t==="bool"||t==="string")return Uint8Array.from(new Int32Array(e));throw new Error(`Unknown dtype ${t}`)}function L1(e,t){let a=vh(e,t);for(let n=0;n<a.length;n++)a[n]=1;return a}function vh(e,t){if(t==null||t==="float32"||t==="complex64")return new Float32Array(e);if(t==="int32")return new Int32Array(e);if(t==="bool")return new Uint8Array(e);throw new Error(`Unknown data type ${t}`)}function zC(e,t){let a=e.reduce((n,r)=>n*r,1);if(t==null||t==="float32")return Rl(e,new Float32Array(a));if(t==="int32")return Rl(e,new Int32Array(a));if(t==="bool")return Rl(e,new Uint8Array(a));throw new Error(`Unknown data type ${t}`)}function an(e){e.forEach(t=>{D(Number.isInteger(t)&&t>=0,()=>`Tensor must have a shape comprised of positive integers but got shape [${e}].`)})}function LC(e,t,a){if(t===0)return 0;if(t===1)return e[0];let n=e[e.length-1];for(let r=0;r<e.length-1;++r)n+=a[r]*e[r];return n}function WC(e,t,a){if(t===0)return[];if(t===1)return[e];let n=new Array(t);for(let r=0;r<n.length-1;++r)n[r]=Math.floor(e/a[r]),e-=n[r]*a[r];return n[n.length-1]=e,n}function wh(e){return e&&e.then&&typeof e.then=="function"}var T5="tfjsflags",sA=class{constructor(e){this.global=e,this.flags={},this.flagRegistry={},this.urlFlags={},this.getQueryParams=BC,this.populateURLFlags()}setPlatform(e,t){this.platform!=null&&(W().getBool("IS_TEST")||W().getBool("PROD")||console.warn(`Platform ${this.platformName} has already been set. Overwriting the platform with ${e}.`)),this.platformName=e,this.platform=t}registerFlag(e,t,a){if(this.flagRegistry[e]={evaluationFn:t,setHook:a},this.urlFlags[e]!=null){let n=this.urlFlags[e];W().getBool("IS_TEST")||W().getBool("PROD")||console.warn(`Setting feature override from URL ${e}: ${n}.`),this.set(e,n)}}async getAsync(e){return e in this.flags?this.flags[e]:(this.flags[e]=await this.evaluateFlag(e),this.flags[e])}get(e){if(e in this.flags)return this.flags[e];let t=this.evaluateFlag(e);if(wh(t))throw new Error(`Flag ${e} cannot be synchronously evaluated. Please use getAsync() instead.`);return this.flags[e]=t,this.flags[e]}getNumber(e){return this.get(e)}getBool(e){return this.get(e)}getString(e){return this.get(e)}getFlags(){return this.flags}get features(){return this.flags}set(e,t){if(this.flagRegistry[e]==null)throw new Error(`Cannot set flag ${e} as it has not been registered.`);this.flags[e]=t,this.flagRegistry[e].setHook!=null&&this.flagRegistry[e].setHook(t)}evaluateFlag(e){if(this.flagRegistry[e]==null)throw new Error(`Cannot evaluate flag '${e}': no evaluation function found.`);return this.flagRegistry[e].evaluationFn()}setFlags(e){this.flags=Object.assign({},e)}reset(){this.flags={},this.urlFlags={},this.populateURLFlags()}populateURLFlags(){if(typeof this.global=="undefined"||typeof this.global.location=="undefined"||typeof this.global.location.search=="undefined")return;let e=this.getQueryParams(this.global.location.search);T5 in e&&e[T5].split(",").forEach(t=>{let[a,n]=t.split(":");this.urlFlags[a]=UC(a,n)})}};function BC(e){let t={};return e.replace(/[?&]([^=?&]+)(?:=([^&]*))?/g,(a,...n)=>(VC(t,n[0],n[1]),n.join("="))),t}function VC(e,t,a){e[decodeURIComponent(t)]=decodeURIComponent(a||"")}function UC(e,t){let a=t.toLowerCase();return a==="true"||a==="false"?a==="true":`${+a}`===a?+a:t}function W(){return W1}var W1=null;function GC(e){W1=e}var _2;function iA(){if(_2==null){let e;if(typeof window!="undefined")e=window;else if(typeof global!="undefined")e=global;else if(typeof process!="undefined")e=process;else if(typeof self!="undefined")e=self;else throw new Error("Could not find a global object");_2=e}return _2}function HC(){let e=iA();return e._tfGlobals==null&&(e._tfGlobals=new Map),e._tfGlobals}function B1(e,t){let a=HC();if(a.has(e))return a.get(e);{let n=t();return a.set(e,n),a.get(e)}}var Zl="Abs",Qs="Acos",ei="Acosh",ns="Add",ti="AddN",ai="All",ni="Any",Jl="ArgMax",Ql="ArgMin",ri="Asin",si="Asinh",ii="Atan",oi="Atanh",li="Atan2",ui="AvgPool",kh="AvgPoolGrad",eu="AvgPool3D",ep="AvgPool3DGrad",di="BatchMatMul",tu="BatchToSpaceND",pi="Bincount",Ih="BitwiseAnd",jC="BroadcastTo",au="BroadcastArgs",ci="Cast",hi="Ceil",rs="ClipByValue",tp="Complex",ap="ComplexAbs",nu="Concat",fi="Conv2D",np="Conv2DBackpropFilter",mi="Conv2DBackpropInput",gi="Conv3D",ru="Conv3DBackpropFilterV2",yi="Conv3DBackpropInputV2",xi="Cos",Ai="Cosh",bi="Cumprod",vi="Cumsum",wi="CropAndResize",su="DenseBincount",ki="DepthToSpace",Ii="DepthwiseConv2dNative",rp="DepthwiseConv2dNativeBackpropFilter",sp="DepthwiseConv2dNativeBackpropInput",iu="Diag",Si="Dilation2D",Dl="Dilation2DBackpropInput",Pl="Dilation2DBackpropFilter",Ci="RealDiv",ip="Einsum",Ti="Elu",ou="EluGrad",lu="Erf",Ni="Equal",Ri="Exp",uu="ExpandDims",Ei="Expm1",op="FFT",du="Fill",Mi="FlipLeftRight",Fi="Floor",_i="FloorDiv",$i="FusedBatchNorm",pu="GatherV2",Di="GatherNd",Pi="Greater",Oi="GreaterEqual",zi="Identity",lp="IFFT",up="Imag",Li="IsFinite",Wi="IsInf",Bi="IsNan",Vi="LeakyRelu",Ui="Less",Gi="LessEqual",Hi="LinSpace",ji="Log",qi="Log1p",Xi="LogicalAnd",Ki="LogicalNot",Yi="LogicalOr",oA="LogicalXor",qC="LogSoftmax",XC="LowerBound",Zi="LRN",cu="LRNGrad",KC="MatrixBandPart",Ji="Max",Qi="Maximum",eo="MaxPool",Sh="MaxPoolGrad",hu="MaxPool3D",dp="MaxPool3DGrad",pp="MaxPoolWithArgmax",to="Mean",ao="Min",no="Minimum",ro="MirrorPad",fu="Mod",so="Multinomial",io="Multiply",mu="Neg",oo="NotEqual",lo="NonMaxSuppressionV3",gu="NonMaxSuppressionV4",uo="NonMaxSuppressionV5",yu="OnesLike",po="OneHot",xu="Pack",co="PadV2",YC="Pool",ho="Pow",fo="Prelu",mo="Prod",Ch="RaggedGather",Th="RaggedRange",Nh="RaggedTensorToTensor",Au="Range",cp="Real",go="Reciprocal",yo="Relu",bu="Reshape",xo="ResizeNearestNeighbor",vu="ResizeNearestNeighborGrad",Ao="ResizeBilinear",wu="ResizeBilinearGrad",bo="Relu6",vo="Reverse",wo="Round",ko="Rsqrt",Io="ScatterNd",So="TensorScatterUpdate",Co="SearchSorted",ku="Select",To="Selu",Iu="Slice",No="Sin",Su="Sinh",Ro="Sign",Eo="Sigmoid",Mo="Softplus",Fo="Sqrt",_o="Sum",Cu="SpaceToBatchND",Tu="SplitV",$o="Softmax",hp="SparseFillEmptyRows",Nu="SparseReshape",fp="SparseSegmentMean",mp="SparseSegmentSum",Do="SparseToDense",Po="SquaredDifference",gp="Square",yp="StaticRegexReplace",Oo="StridedSlice",Ru="StringNGrams",xp="StringSplit",Ap="StringToHashBucketFast",zo="Sub",Lo="Tan",Wo="Tanh",ss="Tile",Bo="TopK",Vo="Transform",wr="Transpose",bp="Unique",Eu="Unpack",vp="UnsortedSegmentSum",ZC="UpperBound",Mu="ZerosLike",is="Step",Ed="FromPixels",Uo="RotateWithOffset",Xr="_FusedMatMul",Kr="FusedConv2D",Yr="FusedDepthwiseConv2D";function Wr(...e){W().getBool("IS_TEST")||W().getBool("PROD")||console.warn(...e)}function JC(...e){W().getBool("IS_TEST")||W().getBool("PROD")||console.log(...e)}var Ol=B1("kernelRegistry",()=>new Map),Md=B1("gradRegistry",()=>new Map);function ah(e,t){let a=V1(e,t);return Ol.get(a)}function G2(e){return Md.get(e)}function Jn(e){let t=Ol.entries(),a=[];for(;;){let{done:n,value:r}=t.next();if(n)break;let[s,i]=r,[o]=s.split("_");o===e&&a.push(i)}return a}function yn(e){let{kernelName:t,backendName:a}=e,n=V1(t,a);Ol.has(n)&&Wr(`The kernel '${t}' for backend '${a}' is already registered`),Ol.set(n,e)}function QC(e){let{kernelName:t}=e;Md.has(t)&&W().getBool("DEBUG")&&Wr(`Overriding the gradient for '${t}'`),Md.set(t,e)}function eT(e,t){let a=V1(e,t);if(!Ol.has(a))throw new Error(`The kernel '${e}' for backend '${t}' is not registered`);Ol.delete(a)}function tT(e){if(!Md.has(e))throw new Error(`The gradient '${e}' for backend is not registered`);Md.delete(e)}function aT(e,t){Jn(e).forEach(a=>{let n=Object.assign({},a,{backendName:t});yn(n)})}function V1(e,t){return`${t}_${e}`}var v={};Ze(v,{arraysEqual:()=>Cr,arraysEqualWithNull:()=>Yx,assert:()=>D,assertNonNegativeIntegerDimensions:()=>an,assertNonNull:()=>Js,assertShapesMatch:()=>Na,bytesFromStringArray:()=>tA,bytesPerElement:()=>eh,checkConversionForErrors:()=>Qx,clamp:()=>Rd,computeStrides:()=>Yl,convertBackendValuesAndArrayBuffer:()=>OC,createScalarValue:()=>lT,createShuffledIndices:()=>_C,decodeString:()=>nh,distSquared:()=>RC,encodeString:()=>kp,fetch:()=>dT,fingerPrint64:()=>oT,flatten:()=>Zr,getArrayFromDType:()=>z1,getTypedArrayFromDType:()=>Jx,hasEncodingLoss:()=>PC,hexToLong:()=>wp,indexToLoc:()=>WC,inferDtype:()=>Qd,inferFromImplicitShape:()=>DC,isBoolean:()=>aA,isFunction:()=>qr,isInt:()=>$l,isNumber:()=>nA,isPromise:()=>wh,isScalarShape:()=>EC,isString:()=>Br,isTypedArray:()=>Zt,isValidDtype:()=>eA,locToIndex:()=>LC,makeOnesTypedArray:()=>L1,makeZerosNestedTypedArray:()=>zC,makeZerosTypedArray:()=>vh,nearestDivisor:()=>th,nearestLargerEven:()=>CC,now:()=>Fd,parseAxisParam:()=>Jd,randUniform:()=>NC,repeatedTry:()=>$C,rightPad:()=>Sd,shuffle:()=>Kx,shuffleCombo:()=>SC,sizeFromShape:()=>vt,sizeToSquarishShape:()=>FC,squeezeShape:()=>Zx,sum:()=>TC,swap:()=>Qc,tanh:()=>MC,toNestedArray:()=>Rl,toTypedArray:()=>Rh});function lA(e){return e instanceof Float32Array||e instanceof Int32Array||e instanceof Uint8Array||e instanceof Uint8ClampedArray}var N5=Xl(lC()),Ps=N5.default||N5;function wp(e){return Ps.fromString(e,!0,16)}var uA=wp("c3a5c85c97cb3127"),$s=wp("b492b66fbe98f273"),wa=wp("9ae16a3b2f90404f");function H2(e){return e.xor(e.shru(47))}function dA(e,t,a){let n=e.slice(t,t+a);return Ps.fromBytes(Array.from(n),!0,!0)}function xt(e,t){return dA(e,t,8)}function R5(e,t){return dA(e,t,4)}function Yt(e,t){return t===0?e:e.shru(t).or(e.shl(64-t))}function Hr(e,t,a=wp("9ddfea08eb382d69")){let n=e.xor(t).mul(a);n=n.xor(n.shru(47));let r=t.xor(n).mul(a);return r=r.xor(r.shru(47)),r=r.mul(a),r}function nT(e,t,a,n,r,s){r=r.add(e),s=Yt(s.add(r).add(n),21);let i=r;return r=r.add(t),r=r.add(a),s=s.add(Yt(r,44)),[r.add(n),s.add(i)]}function zc(e,t,a,n){return nT(xt(e,t),xt(e,t+8),xt(e,t+16),xt(e,t+24),a,n)}function rT(e,t=e.length){if(t>=8){let a=wa.add(t*2),n=xt(e,0).add(wa),r=xt(e,t-8),s=Yt(r,37).mul(a).add(n),i=Yt(n,25).add(r).mul(a);return Hr(s,i,a)}if(t>=4){let a=wa.add(t*2),n=R5(e,0);return Hr(n.shl(3).add(t),R5(e,t-4),a)}if(t>0){let a=e[0],n=e[t>>1],r=e[t-1],s=a+(n<<8),i=t+(r<<2);return H2(wa.mul(s).xor(uA.mul(i))).mul(wa)}return wa}function sT(e,t=e.length){let a=wa.add(t*2),n=xt(e,0).mul($s),r=xt(e,8),s=xt(e,t-8).mul(a),i=xt(e,t-16).mul(wa);return Hr(Yt(n.add(r),43).add(Yt(s,30)).add(i),n.add(Yt(r.add(wa),18)).add(s),a)}function iT(e,t=e.length){let a=wa.add(t*2),n=xt(e,0).mul(wa),r=xt(e,8),s=xt(e,t-8).mul(a),i=xt(e,t-16).mul(wa),o=Yt(n.add(r),43).add(Yt(s,30)).add(i),l=Hr(o,n.add(Yt(r.add(wa),18)).add(s),a),u=xt(e,16).mul(a),p=xt(e,24),c=o.add(xt(e,t-32)).mul(a),d=l.add(xt(e,t-24)).mul(a);return Hr(Yt(u.add(p),43).add(Yt(c,30)).add(d),u.add(Yt(p.add(n),18)).add(c),a)}function oT(e,t=e.length){let a=Ps.fromNumber(81,!0);if(t<=32)return t<=16?rT(e,t):sT(e,t);if(t<=64)return iT(e,t);let n=a,r=a.mul($s).add(113),s=H2(r.mul(wa).add(113)).mul(wa),i=[Ps.UZERO,Ps.UZERO],o=[Ps.UZERO,Ps.UZERO];n=n.mul(wa).add(xt(e,0));let l=0,u=(t-1>>6)*64,p=u+(t-1&63)-63;do n=Yt(n.add(r).add(i[0]).add(xt(e,l+8)),37).mul($s),r=Yt(r.add(i[1]).add(xt(e,l+48)),42).mul($s),n=n.xor(o[1]),r=r.add(i[0]).add(xt(e,l+40)),s=Yt(s.add(o[0]),33).mul($s),i=zc(e,l,i[1].mul($s),n.add(o[0])),o=zc(e,l+32,s.add(o[1]),r.add(xt(e,l+16))),[s,n]=[n,s],l+=64;while(l!==u);let c=$s.add(s.and(255).shl(1));return l=p,o[0]=o[0].add(t-1&63),i[0]=i[0].add(o[0]),o[0]=o[0].add(i[0]),n=Yt(n.add(r).add(i[0]).add(xt(e,l+8)),37).mul(c),r=Yt(r.add(i[1]).add(xt(e,l+48)),42).mul(c),n=n.xor(o[1].mul(9)),r=r.add(i[0].mul(9).add(xt(e,l+40))),s=Yt(s.add(o[0]),33).mul(c),i=zc(e,l,i[1].mul(c),n.add(o[0])),o=zc(e,l+32,s.add(o[1]),r.add(xt(e,l+16))),[s,n]=[n,s],Hr(Hr(i[0],o[0],c).add(H2(r).mul(uA)).add(s),Hr(i[1],o[1],c).add(n),c)}function lT(e,t){return t==="string"?kp(e):Rh([e],t)}function uT(e,t){return e instanceof Float32Array&&t==="float32"||e instanceof Int32Array&&t==="int32"||e instanceof Uint8Array&&t==="bool"}function Rh(e,t){if(t==="string")throw new Error("Cannot convert a string[] to a TypedArray");if(Array.isArray(e)&&(e=Zr(e)),W().getBool("DEBUG")&&Qx(e,t),uT(e,t))return e;if(t==null||t==="float32"||t==="complex64")return new Float32Array(e);if(t==="int32")return new Int32Array(e);if(t==="bool"){let a=new Uint8Array(e.length);for(let n=0;n<a.length;++n)Math.round(e[n])!==0&&(a[n]=1);return a}else throw new Error(`Unknown data type ${t}`)}function Fd(){return W().platform.now()}function dT(e,t){return W().platform.fetch(e,t)}function kp(e,t="utf-8"){return t=t||"utf-8",W().platform.encode(e,t)}function nh(e,t="utf-8"){return t=t||"utf-8",W().platform.decode(e,t)}function Zt(e){return W().platform.isTypedArray!=null?W().platform.isTypedArray(e):lA(e)}function Zr(e,t=[],a=!1){if(t==null&&(t=[]),typeof e=="boolean"||typeof e=="number"||typeof e=="string"||wh(e)||e==null||Zt(e)&&a)t.push(e);else if(Array.isArray(e)||Zt(e))for(let n=0;n<e.length;++n)Zr(e[n],t,a);else{let n=-1;for(let r of Object.keys(e))/^([1-9]+[0-9]*|0)$/.test(r)&&(n=Math.max(n,Number(r)));for(let r=0;r<=n;r++)Zr(e[r],t,a)}return t}var pT=class{constructor(e,t){this.backendTimer=e,this.logger=t,t==null&&(this.logger=new hT)}profileKernel(e,t,a){let n,r=()=>{n=a()},s,i=Fd();if(this.backendTimer.timerAvailable())s=this.backendTimer.time(r);else{r();for(let o of n)o.dataSync();s=Promise.resolve({kernelMs:Fd()-i})}if(W().getBool("CHECK_COMPUTATION_FOR_ERRORS"))for(let o=0;o<n.length;o++){let l=n[o];l.data().then(u=>{cT(u,l.dtype,e)})}return{kernelName:e,outputs:n,inputs:t,timeMs:s.then(o=>o.kernelMs),extraInfo:s.then(o=>o.getExtraProfileInfo!=null?o.getExtraProfileInfo():"")}}logKernelProfile(e){let{kernelName:t,outputs:a,timeMs:n,inputs:r,extraInfo:s}=e;a.forEach(i=>{Promise.all([i.data(),n,s]).then(o=>{this.logger.logKernelProfile(t,i,o[0],o[1],r,o[2])})})}};function cT(e,t,a){if(t!=="float32")return!1;for(let n=0;n<e.length;n++){let r=e[n];if(isNaN(r)||!isFinite(r))return console.warn(`Found ${r} in the result of '${a}'`),!0}return!1}var hT=class{logKernelProfile(e,t,a,n,r,s){let i=typeof n=="number"?Sd(`${n}ms`,9):n.error,o=Sd(e,25),l=t.rank,u=t.size,p=Sd(t.shape.toString(),14),c="";for(let d in r){let h=r[d];if(h!=null){let f=h.shape||t.shape,m=f.length;c+=`${d}: ${m}D ${m>0?f:""} `}}console.log(`%c${o} %c${i} %c${l}D ${p} %c${u} %c${c} %c${s}`,"font-weight:bold","color:red","color:blue","color: orange","color: green","color: steelblue")}};function fT(e,t,a){let n={},r={};for(let l=0;l<t.length;l++)n[t[l].id]=!0;for(let l=0;l<e.length;l++){let u=e[l],p=u.inputs;for(let c in p){let d=p[c],h=!1;for(let f=0;f<t.length;f++)if(n[d.id]){u.outputs.forEach(m=>n[m.id]=!0),h=!0,r[u.id]=!0;break}if(h)break}}let s={};s[a.id]=!0;let i={};for(let l=e.length-1;l>=0;l--){let u=e[l],p=u.inputs;for(let c=0;c<u.outputs.length;c++)if(s[u.outputs[c].id]){for(let d in p)s[p[d].id]=!0,i[u.id]=!0;break}}let o=[];for(let l=0;l<e.length;l++){let u=e[l];if(r[u.id]&&i[u.id]){let p={};for(let d in u.inputs){let h=u.inputs[d];n[h.id]&&(p[d]=h)}let c=Object.assign({},u);c.inputs=p,c.outputs=u.outputs,o.push(c)}}return o}function mT(e,t,a,n){for(let r=t.length-1;r>=0;r--){let s=t[r],i=[];if(s.outputs.forEach(l=>{let u=e[l.id];u!=null?i.push(u):i.push(null)}),s.gradient==null)throw new Error(`Cannot compute gradient: gradient function not found for ${s.kernelName}.`);let o=s.gradient(i);for(let l in s.inputs){if(!(l in o))throw new Error(`Cannot backprop through input ${l}. Available gradients found: ${Object.keys(o)}.`);let u=a(()=>o[l]());if(u.dtype!=="float32")throw new Error(`Error in gradient for op ${s.kernelName}. The gradient of input ${l} must have 'float32' dtype, but has '${u.dtype}'`);let p=s.inputs[l];if(!Cr(u.shape,p.shape))throw new Error(`Error in gradient for op ${s.kernelName}. The gradient of input '${l}' has shape '${u.shape}', which does not match the shape of the input '${p.shape}'`);if(e[p.id]==null)e[p.id]=u;else{let c=e[p.id];e[p.id]=n(c,u),c.dispose()}}}}var E5=20,hd=3,$2=7;function gT(e,t,a,n){let r=Yl(t),s=yT(e,t,a,r),i=t.length,o=Hc(e,t,a,r,s),l=["Tensor"];return n&&(l.push(` dtype: ${a}`),l.push(` rank: ${i}`),l.push(` shape: [${t}]`),l.push(" values:")),l.push(o.map(u=>" "+u).join(`
|
|
`)),l.join(`
|
|
`)}function yT(e,t,a,n){let r=vt(t),s=n[n.length-1],i=new Array(s).fill(0),o=t.length,l=a==="complex64"?gd(e):e;if(o>1)for(let u=0;u<r/s;u++){let p=u*s;for(let c=0;c<s;c++)i[c]=Math.max(i[c],md(l[p+c],0,a).length)}return i}function md(e,t,a){let n;return Array.isArray(e)?n=`${parseFloat(e[0].toFixed($2))} + ${parseFloat(e[1].toFixed($2))}j`:Br(e)?n=`'${e}'`:a==="bool"?n=pA(e):n=parseFloat(e.toFixed($2)).toString(),Sd(n,t)}function pA(e){return e===0?"false":"true"}function Hc(e,t,a,n,r,s=!0){let i=a==="complex64"?2:1,o=t[0],l=t.length;if(l===0){if(a==="complex64"){let m=gd(e);return[md(m[0],0,a)]}return a==="bool"?[pA(e[0])]:[e[0].toString()]}if(l===1){if(o>E5){let m=hd*i,g=Array.from(e.slice(0,m)),y=Array.from(e.slice((o-hd)*i,o*i));return a==="complex64"&&(g=gd(g),y=gd(y)),["["+g.map((x,A)=>md(x,r[A],a)).join(", ")+", ..., "+y.map((x,A)=>md(x,r[o-hd+A],a)).join(", ")+"]"]}return["["+(a==="complex64"?gd(e):Array.from(e)).map((m,g)=>md(m,r[g],a)).join(", ")+"]"]}let u=t.slice(1),p=n.slice(1),c=n[0]*i,d=[];if(o>E5){for(let m=0;m<hd;m++){let g=m*c,y=g+c;d.push(...Hc(e.slice(g,y),u,a,p,r,!1))}d.push("...");for(let m=o-hd;m<o;m++){let g=m*c,y=g+c;d.push(...Hc(e.slice(g,y),u,a,p,r,m===o-1))}}else for(let m=0;m<o;m++){let g=m*c,y=g+c;d.push(...Hc(e.slice(g,y),u,a,p,r,m===o-1))}let h=l===2?",":"";d[0]="["+(o>0?d[0]+h:"");for(let m=1;m<d.length-1;m++)d[m]=" "+d[m]+h;let f=`,
|
|
`;for(let m=2;m<l;m++)f+=`
|
|
`;return d[d.length-1]=" "+d[d.length-1]+"]"+(s?"":f),d}function gd(e){let t=[];for(let a=0;a<e.length;a+=2)t.push([e[a],e[a+1]]);return t}var Wt=class{constructor(e,t,a){if(this.dtype=t,this.shape=e.slice(),this.size=vt(e),a!=null){let n=a.length;D(n===this.size,()=>`Length of values '${n}' does not match the size inferred by the shape '${this.size}'.`)}if(t==="complex64")throw new Error("complex64 dtype TensorBuffers are not supported. Please create a TensorBuffer for the real and imaginary parts separately and call tf.complex(real, imag).");this.values=a||z1(t,this.size),this.strides=Yl(e)}set(e,...t){t.length===0&&(t=[0]),D(t.length===this.rank,()=>`The number of provided coordinates (${t.length}) must match the rank (${this.rank})`);let a=this.locToIndex(t);this.values[a]=e}get(...e){e.length===0&&(e=[0]);let t=0;for(let n of e){if(n<0||n>=this.shape[t]){let r=`Requested out of range element at ${e}. Buffer shape=${this.shape}`;throw new Error(r)}t++}let a=e[e.length-1];for(let n=0;n<e.length-1;++n)a+=this.strides[n]*e[n];return this.values[a]}locToIndex(e){if(this.rank===0)return 0;if(this.rank===1)return e[0];let t=e[e.length-1];for(let a=0;a<e.length-1;++a)t+=this.strides[a]*e[a];return t}indexToLoc(e){if(this.rank===0)return[];if(this.rank===1)return[e];let t=new Array(this.shape.length);for(let a=0;a<t.length-1;++a)t[a]=Math.floor(e/this.strides[a]),e-=t[a]*this.strides[a];return t[t.length-1]=e,t}get rank(){return this.shape.length}toTensor(){return On().makeTensor(this.values,this.shape,this.dtype)}},On=null,Cl=null,xT=null;function AT(e){On=e}function bT(e){Cl=e}function vT(e){xT=e}var gt=class{constructor(e,t,a,n){this.kept=!1,this.isDisposedInternal=!1,this.shape=e.slice(),this.dtype=t||"float32",this.size=vt(e),this.strides=Yl(e),this.dataId=a,this.id=n,this.rankType=this.rank<5?this.rank.toString():"higher"}get rank(){return this.shape.length}async buffer(){let e=await this.data();return Cl.buffer(this.shape,this.dtype,e)}bufferSync(){return Cl.buffer(this.shape,this.dtype,this.dataSync())}async array(){let e=await this.data();return Rl(this.shape,e,this.dtype==="complex64")}arraySync(){return Rl(this.shape,this.dataSync(),this.dtype==="complex64")}async data(){this.throwIfDisposed();let e=On().read(this.dataId);if(this.dtype==="string"){let t=await e;try{return t.map(a=>nh(a))}catch(a){throw new Error("Failed to decode the string bytes into utf-8. To get the original bytes, call tensor.bytes().")}}return e}dataToGPU(e){return this.throwIfDisposed(),On().readToGPU(this.dataId,e)}dataSync(){this.throwIfDisposed();let e=On().readSync(this.dataId);if(this.dtype==="string")try{return e.map(t=>nh(t))}catch(t){throw new Error("Failed to decode the string bytes into utf-8. To get the original bytes, call tensor.bytes().")}return e}async bytes(){this.throwIfDisposed();let e=await On().read(this.dataId);return this.dtype==="string"?e:new Uint8Array(e.buffer)}dispose(){this.isDisposed||(On().disposeTensor(this),this.isDisposedInternal=!0)}get isDisposed(){return this.isDisposedInternal}throwIfDisposed(){if(this.isDisposed)throw new Error("Tensor is disposed.")}print(e=!1){return Cl.print(this,e)}clone(){return this.throwIfDisposed(),Cl.clone(this)}toString(e=!1){let t=this.dataSync();return gT(t,this.shape,this.dtype,e)}cast(e){return this.throwIfDisposed(),Cl.cast(this,e)}variable(e=!0,t,a){return this.throwIfDisposed(),On().makeVariable(this,e,t,a)}};Object.defineProperty(gt,Symbol.hasInstance,{value:e=>!!e&&e.data!=null&&e.dataSync!=null&&e.throwIfDisposed!=null});function wT(){return B1("Tensor",()=>gt)}wT();var _d=class extends gt{constructor(e,t,a,n){super(e.shape,e.dtype,e.dataId,n),this.trainable=t,this.name=a}assign(e){if(e.dtype!==this.dtype)throw new Error(`dtype of the new value (${e.dtype}) and previous value (${this.dtype}) must match`);if(!Cr(e.shape,this.shape))throw new Error(`shape of the new value (${e.shape}) and previous value (${this.shape}) must match`);On().disposeTensor(this),this.dataId=e.dataId,On().incRef(this,null)}dispose(){On().disposeVariable(this),this.isDisposedInternal=!0}};Object.defineProperty(_d,Symbol.hasInstance,{value:e=>e instanceof gt&&e.assign!=null&&e.assign instanceof Function});var cA={};Ze(cA,{assertTypesMatch:()=>mA,getTensorsInContainer:()=>U1,isTensorInList:()=>IT,makeTypesMatch:()=>Rt});var j2;(function(e){e.R0="R0",e.R1="R1",e.R2="R2",e.R3="R3",e.R4="R4",e.R5="R5",e.R6="R6"})(j2||(j2={}));var q2;(function(e){e.float32="float32",e.int32="int32",e.bool="int32",e.complex64="complex64"})(q2||(q2={}));var X2;(function(e){e.float32="float32",e.int32="int32",e.bool="bool",e.complex64="complex64"})(X2||(X2={}));var K2;(function(e){e.float32="float32",e.int32="float32",e.bool="float32",e.complex64="complex64"})(K2||(K2={}));var Y2;(function(e){e.float32="complex64",e.int32="complex64",e.bool="complex64",e.complex64="complex64"})(Y2||(Y2={}));var kT={float32:K2,int32:q2,bool:X2,complex64:Y2};function pa(e,t){if(e==="string"||t==="string"){if(e==="string"&&t==="string")return"string";throw new Error(`Can not upcast ${e} with ${t}`)}return kT[e][t]}function Ip(e){return pa(e,"int32")}function hA(e){return e!=null&&typeof e=="object"&&"texture"in e&&e.texture instanceof WebGLTexture}function fA(e){return typeof GPUBuffer!="undefined"&&e!=null&&typeof e=="object"&&"buffer"in e&&e.buffer instanceof GPUBuffer}function Rt(e,t){if(e.dtype===t.dtype)return[e,t];let a=pa(e.dtype,t.dtype);return[e.cast(a),t.cast(a)]}function mA(e,t){D(e.dtype===t.dtype,()=>`The dtypes of the first(${e.dtype}) and second(${t.dtype}) input must match`)}function IT(e,t){return t.some(a=>a.id===e.id)}function U1(e){let t=[];return gA(e,t,new Set),t}function gA(e,t,a){if(e==null)return;if(e instanceof gt){t.push(e);return}if(!ST(e))return;let n=e;for(let r in n){let s=n[r];a.has(s)||(a.add(s),gA(s,t,a))}}function ST(e){return Array.isArray(e)||typeof e=="object"}function D2(e){return e.kernelName!=null}var M5=class{constructor(){this.registeredVariables={},this.nextTapeNodeId=0,this.numBytes=0,this.numTensors=0,this.numStringTensors=0,this.numDataBuffers=0,this.gradientDepth=0,this.kernelDepth=0,this.scopeStack=[],this.numDataMovesStack=[],this.nextScopeId=0,this.tensorInfo=new WeakMap,this.profiling=!1,this.activeProfile={newBytes:0,newTensors:0,peakBytes:0,kernels:[],result:null,get kernelNames(){return Array.from(new Set(this.kernels.map(e=>e.name)))}}}dispose(){for(let e in this.registeredVariables)this.registeredVariables[e].dispose()}},$d=class{constructor(e){this.ENV=e,this.registry={},this.registryFactory={},this.pendingBackendInitId=0,this.state=new M5}async ready(){if(this.pendingBackendInit!=null)return this.pendingBackendInit.then(()=>{});if(this.backendInstance!=null)return;let e=this.getSortedBackends();for(let t=0;t<e.length;t++){let a=e[t];if(await this.initializeBackend(a).success){await this.setBackend(a);return}}throw new Error("Could not initialize any backends, all backend initializations failed.")}get backend(){if(this.pendingBackendInit!=null)throw new Error(`Backend '${this.backendName}' has not yet been initialized. Make sure to await tf.ready() or await tf.setBackend() before calling other methods`);if(this.backendInstance==null){let{name:e,asyncInit:t}=this.initializeBackendsAndReturnBest();if(t)throw new Error(`The highest priority backend '${e}' has not yet been initialized. Make sure to await tf.ready() or await tf.setBackend() before calling other methods`);this.setBackend(e)}return this.backendInstance}backendNames(){return Object.keys(this.registryFactory)}findBackend(e){if(!(e in this.registry))if(e in this.registryFactory){let{asyncInit:t}=this.initializeBackend(e);if(t)return null}else return null;return this.registry[e]}findBackendFactory(e){return e in this.registryFactory?this.registryFactory[e].factory:null}registerBackend(e,t,a=1){return e in this.registryFactory?(Wr(`${e} backend was already registered. Reusing existing backend factory.`),!1):(this.registryFactory[e]={factory:t,priority:a},!0)}async setBackend(e){if(this.registryFactory[e]==null)throw new Error(`Backend name '${e}' not found in registry`);if(this.backendName=e,this.registry[e]==null){this.backendInstance=null;let{success:t,asyncInit:a}=this.initializeBackend(e);if(!(a?await t:t))return!1}return this.backendInstance=this.registry[e],this.setupRegisteredKernels(),this.profiler=new pT(this.backendInstance),!0}setupRegisteredKernels(){Jn(this.backendName).forEach(e=>{e.setupFunc!=null&&e.setupFunc(this.backendInstance)})}disposeRegisteredKernels(e){Jn(e).forEach(t=>{t.disposeFunc!=null&&t.disposeFunc(this.registry[e])})}initializeBackend(e){let t=this.registryFactory[e];if(t==null)throw new Error(`Cannot initialize backend ${e}, no registration found.`);try{let a=t.factory();if(a&&!(a instanceof Kl)&&typeof a.then=="function"){let n=++this.pendingBackendInitId,r=a.then(s=>n<this.pendingBackendInitId?!1:(this.registry[e]=s,this.pendingBackendInit=null,!0)).catch(s=>(n<this.pendingBackendInitId||(this.pendingBackendInit=null,Wr(`Initialization of backend ${e} failed`),Wr(s.stack||s.message)),!1));return this.pendingBackendInit=r,{success:r,asyncInit:!0}}else return this.registry[e]=a,{success:!0,asyncInit:!1}}catch(a){return Wr(`Initialization of backend ${e} failed`),Wr(a.stack||a.message),{success:!1,asyncInit:!1}}}removeBackend(e){if(!(e in this.registryFactory))throw new Error(`${e} backend not found in registry`);this.backendName===e&&this.pendingBackendInit!=null&&this.pendingBackendInitId++,e in this.registry&&(this.disposeRegisteredKernels(e),this.registry[e].dispose(),delete this.registry[e]),delete this.registryFactory[e],this.backendName===e&&(this.pendingBackendInit=null,this.backendName=null,this.backendInstance=null)}getSortedBackends(){if(Object.keys(this.registryFactory).length===0)throw new Error("No backend found in registry.");return Object.keys(this.registryFactory).sort((e,t)=>this.registryFactory[t].priority-this.registryFactory[e].priority)}initializeBackendsAndReturnBest(){let e=this.getSortedBackends();for(let t=0;t<e.length;t++){let a=e[t],{success:n,asyncInit:r}=this.initializeBackend(a);if(r||n)return{name:a,asyncInit:r}}throw new Error("Could not initialize any backends, all backend initializations failed.")}moveData(e,t){let a=this.state.tensorInfo.get(t),n=a.backend,r=this.readSync(t),s=n.refCount(t);n.disposeData(t,!0),a.backend=e,e.move(t,r,a.shape,a.dtype,s),this.shouldCheckForMemLeaks()&&this.state.numDataMovesStack[this.state.numDataMovesStack.length-1]++}tidy(e,t){let a=null;if(t==null){if(typeof e!="function")throw new Error("Please provide a function to tidy()");t=e}else{if(typeof e!="string"&&!(e instanceof String))throw new Error("When calling with two arguments, the first argument to tidy() must be a string");if(typeof t!="function")throw new Error("When calling with two arguments, the 2nd argument to tidy() must be a function");a=e}let n;return this.scopedRun(()=>this.startScope(a),()=>this.endScope(n),()=>(n=t(),n instanceof Promise&&console.error("Cannot return a Promise inside of tidy."),n))}scopedRun(e,t,a){e();try{let n=a();return t(),n}catch(n){throw t(),n}}nextTensorId(){return $d.nextTensorId++}nextVariableId(){return $d.nextVariableId++}clone(e){let t=L.runKernel(zi,{x:e}),a={x:e},n=s=>({x:()=>{let i="float32",o={x:s},l={dtype:i};return L.runKernel(ci,o,l)}}),r=[];return this.addTapeNode(this.state.activeScope.name,a,[t],n,r,{}),t}runKernel(e,t,a){if(this.backendName==null&&this.backend,ah(e,this.backendName)==null)throw new Error(`Kernel '${e}' not registered for backend '${this.backendName}'`);return this.runKernelFunc({kernelName:e,inputs:t,attrs:a})}shouldCheckForMemLeaks(){return this.ENV.getBool("IS_TEST")}checkKernelForMemLeak(e,t,a){let n=this.backend.numDataIds(),r=0;a.forEach(o=>{r+=o.dtype==="complex64"?3:1});let s=this.state.numDataMovesStack[this.state.numDataMovesStack.length-1],i=n-t-r-s;if(i>0)throw new Error(`Backend '${this.backendName}' has an internal memory leak (${i} data ids) after running '${e}'`)}runKernelFunc(e){let t,a=[],n=this.isTapeOn(),r=this.state.numBytes,s=this.state.numTensors;this.shouldCheckForMemLeaks()&&this.state.numDataMovesStack.push(0);let i;this.backendName==null&&this.backend;let o,l=D2(e)?e.kernelName:this.state.activeScope!=null?this.state.activeScope.name:"";if(D2(e)){let{kernelName:h,inputs:f,attrs:m}=e;this.backendName==null&&this.backend;let g=ah(h,this.backendName);D(g!=null,()=>`Cannot find registered kernel '${h}' for backend '${this.backendName}'`),i=()=>{let y=this.backend.numDataIds();o=g.kernelFunc({inputs:f,attrs:m,backend:this.backend});let x=Array.isArray(o)?o:[o];this.shouldCheckForMemLeaks()&&this.checkKernelForMemLeak(h,y,x);let A=x.map(b=>b.rank!=null?b:this.makeTensorFromTensorInfo(b));if(n){let b=this.getTensorsForGradient(h,f,A);a=this.saveTensorsForBackwardMode(b)}return A}}else{let{forwardFunc:h}=e,f=m=>{n&&(a=m.map(g=>this.keep(this.clone(g))))};i=()=>{let m=this.backend.numDataIds();o=this.tidy(()=>h(this.backend,f));let g=Array.isArray(o)?o:[o];return this.shouldCheckForMemLeaks()&&this.checkKernelForMemLeak(l,m,g),g}}let{inputs:u,attrs:p}=e,c=D2(e)?null:e.backwardsFunc,d;return this.scopedRun(()=>this.state.kernelDepth++,()=>this.state.kernelDepth--,()=>{!this.ENV.getBool("DEBUG")&&!this.state.profiling?t=i():(d=this.profiler.profileKernel(l,u,()=>i()),this.ENV.getBool("DEBUG")&&this.profiler.logKernelProfile(d),t=d.outputs)}),n&&this.addTapeNode(l,u,t,c,a,p),this.state.profiling&&this.state.activeProfile.kernels.push({name:l,bytesAdded:this.state.numBytes-r,totalBytesSnapshot:this.state.numBytes,tensorsAdded:this.state.numTensors-s,totalTensorsSnapshot:this.state.numTensors,inputShapes:Object.keys(u).map(h=>u[h]!=null?u[h].shape:null),outputShapes:t.map(h=>h.shape),kernelTimeMs:d.timeMs,extraInfo:d.extraInfo}),Array.isArray(o)?t:t[0]}saveTensorsForBackwardMode(e){return e.map(t=>this.keep(this.clone(t)))}getTensorsForGradient(e,t,a){let n=G2(e);if(n!=null){let r=n.inputsToSave||[],s=n.outputsToSave||[],i;n.saveAllInputs?(D(Array.isArray(t),()=>"saveAllInputs is true, expected inputs to be an array."),i=Object.keys(t).map(l=>t[l])):i=r.map(l=>t[l]);let o=a.filter((l,u)=>s[u]);return i.concat(o)}return[]}makeTensor(e,t,a,n){if(e==null)throw new Error("Values passed to engine.makeTensor() are null");a=a||"float32",n=n||this.backend;let r=e;a==="string"&&Br(e[0])&&(r=e.map(o=>kp(o)));let s=n.write(r,t,a),i=new gt(t,a,s,this.nextTensorId());if(this.trackTensor(i,n),a==="string"){let o=this.state.tensorInfo.get(s),l=tA(r);this.state.numBytes+=l-o.bytes,o.bytes=l}return i}makeTensorFromDataId(e,t,a,n){a=a||"float32";let r={dataId:e,shape:t,dtype:a};return this.makeTensorFromTensorInfo(r,n)}makeTensorFromTensorInfo(e,t){let{dataId:a,shape:n,dtype:r}=e,s=new gt(n,r,a,this.nextTensorId());return this.trackTensor(s,t),s}makeVariable(e,t=!0,a,n){a=a||this.nextVariableId().toString(),n!=null&&n!==e.dtype&&(e=e.cast(n));let r=new _d(e,t,a,this.nextTensorId());if(this.state.registeredVariables[r.name]!=null)throw new Error(`Variable with name ${r.name} was already registered`);return this.state.registeredVariables[r.name]=r,this.incRef(r,this.backend),r}trackTensor(e,t){this.state.numTensors++,e.dtype==="string"&&this.state.numStringTensors++;let a=0;e.dtype!=="complex64"&&e.dtype!=="string"&&(a=e.size*eh(e.dtype)),this.state.numBytes+=a,this.state.tensorInfo.has(e.dataId)||(this.state.numDataBuffers++,this.state.tensorInfo.set(e.dataId,{backend:t||this.backend,dtype:e.dtype,shape:e.shape,bytes:a})),e instanceof _d||this.track(e)}incRef(e,t){this.trackTensor(e,t),this.backend.incRef(e.dataId)}removeDataId(e,t){this.state.tensorInfo.has(e)&&this.state.tensorInfo.get(e).backend===t&&(this.state.tensorInfo.delete(e),this.state.numDataBuffers--)}disposeTensor(e){if(!this.state.tensorInfo.has(e.dataId))return;let t=this.state.tensorInfo.get(e.dataId);if(this.state.numTensors--,e.dtype==="string"&&(this.state.numStringTensors--,this.state.numBytes-=t.bytes),e.dtype!=="complex64"&&e.dtype!=="string"){let a=e.size*eh(e.dtype);this.state.numBytes-=a}t.backend.disposeData(e.dataId)&&this.removeDataId(e.dataId,t.backend)}disposeVariables(){for(let e in this.state.registeredVariables){let t=this.state.registeredVariables[e];this.disposeVariable(t)}}disposeVariable(e){this.disposeTensor(e),this.state.registeredVariables[e.name]!=null&&delete this.state.registeredVariables[e.name]}memory(){let e=this.backend.memory();return e.numTensors=this.state.numTensors,e.numDataBuffers=this.state.numDataBuffers,e.numBytes=this.state.numBytes,this.state.numStringTensors>0&&(e.unreliable=!0,e.reasons==null&&(e.reasons=[]),e.reasons.push("Memory usage by string tensors is approximate (2 bytes per character)")),e}async profile(e){this.state.profiling=!0;let t=this.state.numBytes,a=this.state.numTensors;this.state.activeProfile.kernels=[],this.state.activeProfile.result=await e(),this.state.profiling=!1,this.state.activeProfile.peakBytes=Math.max(...this.state.activeProfile.kernels.map(n=>n.totalBytesSnapshot)),this.state.activeProfile.newBytes=this.state.numBytes-t,this.state.activeProfile.newTensors=this.state.numTensors-a;for(let n of this.state.activeProfile.kernels)n.kernelTimeMs=await n.kernelTimeMs,n.extraInfo=await n.extraInfo;return this.state.activeProfile}isTapeOn(){return this.state.gradientDepth>0&&this.state.kernelDepth===0}addTapeNode(e,t,a,n,r,s){let i={id:this.state.nextTapeNodeId++,kernelName:e,inputs:t,outputs:a,saved:r},o=G2(e);o!=null&&(n=o.gradFunc),n!=null&&(i.gradient=l=>(l=l.map((u,p)=>{if(u==null){let c=a[p],d=vh(c.size,c.dtype);return this.makeTensor(d,c.shape,c.dtype)}return u}),n(l.length>1?l:l[0],r,s))),this.state.activeTape.push(i)}keep(e){return e.kept=!0,e}startTape(){this.state.gradientDepth===0&&(this.state.activeTape=[]),this.state.gradientDepth++}endTape(){this.state.gradientDepth--}startScope(e){let t={track:[],name:"unnamed scope",id:this.state.nextScopeId++};e&&(t.name=e),this.state.scopeStack.push(t),this.state.activeScope=t}endScope(e){let t=U1(e),a=new Set(t.map(r=>r.id));for(let r=0;r<this.state.activeScope.track.length;r++){let s=this.state.activeScope.track[r];!s.kept&&!a.has(s.id)&&s.dispose()}let n=this.state.scopeStack.pop();this.state.activeScope=this.state.scopeStack.length===0?null:this.state.scopeStack[this.state.scopeStack.length-1],t.forEach(r=>{!r.kept&&r.scopeId===n.id&&this.track(r)})}gradients(e,t,a,n=!1){if(D(t.length>0,()=>"gradients() received an empty list of xs."),a!=null&&a.dtype!=="float32")throw new Error(`dy must have 'float32' dtype, but has '${a.dtype}'`);let r=this.scopedRun(()=>this.startTape(),()=>this.endTape(),()=>this.tidy("forward",e));D(r instanceof gt,()=>"The result y returned by f() must be a tensor.");let s=fT(this.state.activeTape,t,r);if(!n&&s.length===0&&t.length>0)throw new Error("Cannot compute gradient of y=f(x) with respect to x. Make sure that the f you passed encloses all operations that lead from x to y.");return this.tidy("backward",()=>{let i={};i[r.id]=a==null?CT(r.shape):a,mT(i,s,l=>this.tidy(l),TT);let o=t.map(l=>i[l.id]);return this.state.gradientDepth===0&&(this.state.activeTape.forEach(l=>{for(let u of l.saved)u.dispose()}),this.state.activeTape=null),{value:r,grads:o}})}customGrad(e){return D(qr(e),()=>"The f passed in customGrad(f) must be a function."),(...t)=>{D(t.every(i=>i instanceof gt),()=>"The args passed in customGrad(f)(x1, x2,...) must all be tensors");let a,n={};t.forEach((i,o)=>{n[o]=i});let r=(i,o)=>(a=e(...t,o),D(a.value instanceof gt,()=>"The function f passed in customGrad(f) must return an object where `obj.value` is a tensor"),D(qr(a.gradFunc),()=>"The function f passed in customGrad(f) must return an object where `obj.gradFunc` is a function."),a.value),s=(i,o)=>{let l=a.gradFunc(i,o),u=Array.isArray(l)?l:[l];D(u.length===t.length,()=>"The function f passed in customGrad(f) must return an object where `obj.gradFunc` is a function that returns the same number of tensors as inputs passed to f(...)."),D(u.every(c=>c instanceof gt),()=>"The function f passed in customGrad(f) must return an object where `obj.gradFunc` is a function that returns a list of only tensors.");let p={};return u.forEach((c,d)=>{p[d]=()=>c}),p};return this.runKernelFunc({forwardFunc:r,backwardsFunc:s,inputs:n})}}readSync(e){return this.state.tensorInfo.get(e).backend.readSync(e)}read(e){return this.state.tensorInfo.get(e).backend.read(e)}readToGPU(e,t){return this.state.tensorInfo.get(e).backend.readToGPU(e,t)}async time(e){let t=Fd(),a=await this.backend.time(e);return a.wallMs=Fd()-t,a}track(e){return this.state.activeScope!=null&&(e.scopeId=this.state.activeScope.id,this.state.activeScope.track.push(e)),e}get registeredVariables(){return this.state.registeredVariables}reset(){this.pendingBackendInitId++,this.state.dispose(),this.ENV.reset(),this.state=new M5;for(let e in this.registry)this.disposeRegisteredKernels(e),this.registry[e].dispose(),delete this.registry[e];this.backendName=null,this.backendInstance=null,this.pendingBackendInit=null}};$d.nextTensorId=0;$d.nextVariableId=0;function CT(e){let t=L1(vt(e),"float32");return L.makeTensor(t,e,"float32")}function yA(){let e=iA();if(e._tfengine==null){let t=new sA(e);e._tfengine=new $d(t)}return GC(e._tfengine.ENV),AT(()=>e._tfengine),e._tfengine}var L=yA();function TT(e,t){let a={a:e,b:t};return L.runKernel(ns,a)}var Sp={};Ze(Sp,{isBrowser:()=>xA,isMobile:()=>ET,mockIsMobile:()=>RT});function NT(){return typeof navigator!="undefined"&&navigator!=null}var Z2;function RT(e){Z2=e}function ET(e){if(Z2!==void 0)return Z2;if(e||NT()){if(e||(e=navigator),e.product==="ReactNative")return!0;let t=e.userAgent||e.vendor||(typeof window!="undefined"?window.opera:"");if(!t){let a=e;return a.userAgentData&&a.userAgentData.mobile}return/(android|bb\d+|meego).+mobile|avantgo|bada\/|blackberry|blazer|compal|elaine|fennec|hiptop|iemobile|ip(hone|od)|iris|kindle|lge |maemo|midp|mmp|mobile.+firefox|netfront|opera m(ob|in)i|palm( os)?|phone|p(ixi|re)\/|plucker|pocket|psp|series(4|6)0|symbian|treo|up\.(browser|link)|vodafone|wap|windows ce|xda|xiino/i.test(t)||/1207|6310|6590|3gso|4thp|50[1-6]i|770s|802s|a wa|abac|ac(er|oo|s\-)|ai(ko|rn)|al(av|ca|co)|amoi|an(ex|ny|yw)|aptu|ar(ch|go)|as(te|us)|attw|au(di|\-m|r |s )|avan|be(ck|ll|nq)|bi(lb|rd)|bl(ac|az)|br(e|v)w|bumb|bw\-(n|u)|c55\/|capi|ccwa|cdm\-|cell|chtm|cldc|cmd\-|co(mp|nd)|craw|da(it|ll|ng)|dbte|dc\-s|devi|dica|dmob|do(c|p)o|ds(12|\-d)|el(49|ai)|em(l2|ul)|er(ic|k0)|esl8|ez([4-7]0|os|wa|ze)|fetc|fly(\-|_)|g1 u|g560|gene|gf\-5|g\-mo|go(\.w|od)|gr(ad|un)|haie|hcit|hd\-(m|p|t)|hei\-|hi(pt|ta)|hp( i|ip)|hs\-c|ht(c(\-| |_|a|g|p|s|t)|tp)|hu(aw|tc)|i\-(20|go|ma)|i230|iac( |\-|\/)|ibro|idea|ig01|ikom|im1k|inno|ipaq|iris|ja(t|v)a|jbro|jemu|jigs|kddi|keji|kgt( |\/)|klon|kpt |kwc\-|kyo(c|k)|le(no|xi)|lg( g|\/(k|l|u)|50|54|\-[a-w])|libw|lynx|m1\-w|m3ga|m50\/|ma(te|ui|xo)|mc(01|21|ca)|m\-cr|me(rc|ri)|mi(o8|oa|ts)|mmef|mo(01|02|bi|de|do|t(\-| |o|v)|zz)|mt(50|p1|v )|mwbp|mywa|n10[0-2]|n20[2-3]|n30(0|2)|n50(0|2|5)|n7(0(0|1)|10)|ne((c|m)\-|on|tf|wf|wg|wt)|nok(6|i)|nzph|o2im|op(ti|wv)|oran|owg1|p800|pan(a|d|t)|pdxg|pg(13|\-([1-8]|c))|phil|pire|pl(ay|uc)|pn\-2|po(ck|rt|se)|prox|psio|pt\-g|qa\-a|qc(07|12|21|32|60|\-[2-7]|i\-)|qtek|r380|r600|raks|rim9|ro(ve|zo)|s55\/|sa(ge|ma|mm|ms|ny|va)|sc(01|h\-|oo|p\-)|sdk\/|se(c(\-|0|1)|47|mc|nd|ri)|sgh\-|shar|sie(\-|m)|sk\-0|sl(45|id)|sm(al|ar|b3|it|t5)|so(ft|ny)|sp(01|h\-|v\-|v )|sy(01|mb)|t2(18|50)|t6(00|10|18)|ta(gt|lk)|tcl\-|tdg\-|tel(i|m)|tim\-|t\-mo|to(pl|sh)|ts(70|m\-|m3|m5)|tx\-9|up(\.b|g1|si)|utst|v400|v750|veri|vi(rg|te)|vk(40|5[0-3]|\-v)|vm40|voda|vulc|vx(52|53|60|61|70|80|81|83|85|98)|w3c(\-| )|webc|whit|wi(g |nc|nw)|wmlb|wonu|x700|yas\-|your|zeto|zte\-/i.test(t.substr(0,4))}return!1}function xA(){return typeof window!="undefined"&&window.document!=null||typeof WorkerGlobalScope!="undefined"}var Va=W();Va.registerFlag("DEBUG",()=>!1,e=>{e&&console.warn("Debugging mode is ON. The output of every math call will be downloaded to CPU and checked for NaNs. This significantly impacts performance.")});Va.registerFlag("IS_BROWSER",()=>xA());Va.registerFlag("IS_NODE",()=>typeof process!="undefined"&&typeof process.versions!="undefined"&&typeof process.versions.node!="undefined");Va.registerFlag("IS_CHROME",()=>typeof navigator!="undefined"&&navigator!=null&&navigator.userAgent!=null&&/Chrome/.test(navigator.userAgent)&&/Google Inc/.test(navigator.vendor));Va.registerFlag("IS_SAFARI",()=>typeof navigator!="undefined"&&navigator!=null&&navigator.userAgent!=null&&/Safari/.test(navigator.userAgent)&&/Apple/.test(navigator.vendor));Va.registerFlag("PROD",()=>!1);Va.registerFlag("TENSORLIKE_CHECK_SHAPE_CONSISTENCY",()=>Va.getBool("DEBUG"));Va.registerFlag("DEPRECATION_WARNINGS_ENABLED",()=>!0);Va.registerFlag("IS_TEST",()=>!1);Va.registerFlag("CHECK_COMPUTATION_FOR_ERRORS",()=>Va.getBool("DEBUG"));Va.registerFlag("WRAP_TO_IMAGEBITMAP",()=>!1);Va.registerFlag("CANVAS2D_WILL_READ_FREQUENTLY_FOR_GPU",()=>!1);Va.registerFlag("USE_SETTIMEOUTCUSTOM",()=>!1);function Qn(e,t){let a=e;if(Zt(e))return t==="string"?[]:[e.length];if(hA(e)){let r=e.channels||"RGBA";return[e.height,e.width*r.length]}else if(fA(e))return[e.buffer.size/(t==null?4:eh(t))];if(!Array.isArray(e))return[];let n=[];for(;Array.isArray(a)||Zt(a)&&t!=="string";)n.push(a.length),a=a[0];return Array.isArray(e)&&W().getBool("TENSORLIKE_CHECK_SHAPE_CONSISTENCY")&&AA(e,n,[]),n}function AA(e,t,a){if(a=a||[],!Array.isArray(e)&&!Zt(e)){D(t.length===0,()=>`Element arr[${a.join("][")}] is a primitive, but should be an array/TypedArray of ${t[0]} elements`);return}D(t.length>0,()=>`Element arr[${a.join("][")}] should be a primitive, but is an array of ${e.length} elements`),D(e.length===t[0],()=>`Element arr[${a.join("][")}] should have ${t[0]} elements, but has ${e.length} elements`);let n=t.slice(1);for(let r=0;r<e.length;++r)AA(e[r],n,a.concat(r))}function F5(e,t,a,n){if(e!=="string_or_numeric"){if(e==null)throw new Error("Expected dtype cannot be null.");if(e!=="numeric"&&e!==t||e==="numeric"&&t==="string")throw new Error(`Argument '${a}' passed to '${n}' must be ${e} tensor, but got ${t} tensor`)}}function R(e,t,a,n="numeric"){if(e instanceof gt)return F5(n,e.dtype,t,a),e;let r=Qd(e);if(r!=="string"&&["bool","int32","float32"].indexOf(n)>=0&&(r=n),F5(n,r,t,a),e==null||!Zt(e)&&!Array.isArray(e)&&typeof e!="number"&&typeof e!="boolean"&&typeof e!="string"){let o=e==null?"null":e.constructor.name;throw new Error(`Argument '${t}' passed to '${a}' must be a Tensor or TensorLike, but got '${o}'`)}let s=Qn(e,r);!Zt(e)&&!Array.isArray(e)&&(e=[e]);let i=r!=="string"?Rh(e,r):Zr(e,[],!0);return L.makeTensor(i,s,r)}function Dd(e,t,a,n="numeric"){if(!Array.isArray(e))throw new Error(`Argument ${t} passed to ${a} must be a \`Tensor[]\` or \`TensorLike[]\``);return e.map((r,s)=>R(r,`${t}[${s}]`,a,n))}var G1="__op";function z(e){let t=Object.keys(e);if(t.length!==1)throw new Error(`Please provide an object with a single key (operation name) mapping to a function. Got an object with ${t.length} keys.`);let a=t[0],n=e[a];a.endsWith("_")&&(a=a.substring(0,a.length-1)),a=a+G1;let r=(...s)=>{L.startScope(a);try{let i=n(...s);return wh(i)&&console.error("Cannot return a Promise inside of tidy."),L.endScope(i),i}catch(i){throw L.endScope(null),i}};return Object.defineProperty(r,"name",{value:a,configurable:!0}),r}function MT(e,t){let a=R(e,"real","complex"),n=R(t,"imag","complex");Na(a.shape,n.shape,`real and imag shapes, ${a.shape} and ${n.shape}, must match in call to tf.complex().`);let r={real:a,imag:n};return L.runKernel(tp,r)}var Sr=z({complex_:MT});function os(e,t,a,n){if(n==null)n=Qd(e);else if(n==="complex64")throw new Error("Cannot construct a complex64 tensor directly. Please use tf.complex(real, imag).");if(fA(e)||hA(e)){if(n!=="float32"&&n!=="int32")throw new Error(`Creating tensor from GPU data only supports 'float32'|'int32' dtype, while the dtype is ${n}.`);return L.backend.createTensorFromGPUData(e,t||a,n)}if(!Zt(e)&&!Array.isArray(e)&&typeof e!="number"&&typeof e!="boolean"&&typeof e!="string")throw new Error("values passed to tensor(values) must be a number/boolean/string or an array of numbers/booleans/strings, or a TypedArray");if(t!=null){an(t);let r=vt(t),s=vt(a);D(r===s,()=>`Based on the provided shape, [${t}], the tensor should have ${r} values but has ${s}`);for(let i=0;i<a.length;++i){let o=a[i],l=i===a.length-1?o!==vt(t.slice(i)):!0;D(a[i]===t[i]||!l,()=>`Error creating a new Tensor. Inferred shape (${a}) does not match the provided shape (${t}). `)}}return!Zt(e)&&!Array.isArray(e)&&(e=[e]),t=t||a,e=n!=="string"?Rh(e,n):Zr(e,[],!0),L.makeTensor(e,t,n)}function Ve(e,t,a){let n=Qn(e,a);return os(e,t,n,a)}var J2={float32:4,float16:2,int32:4,uint16:2,uint8:1,bool:1,complex64:8},rh=4;async function FT(e,t){let a=[],n=[],r=Array.isArray(e)?e.map(i=>i.name):Object.keys(e);for(let i=0;i<r.length;++i){let o=r[i],l=Array.isArray(e)?e[i].tensor:e[o];if(l.dtype!=="float32"&&l.dtype!=="int32"&&l.dtype!=="bool"&&l.dtype!=="string"&&l.dtype!=="complex64")throw new Error(`Unsupported dtype in weight '${o}': ${l.dtype}`);let u={name:o,shape:l.shape,dtype:l.dtype};if(l.dtype==="string"){let p=new Promise(async c=>{let d=await l.bytes(),h=d.reduce((g,y)=>g+y.length,0)+rh*d.length,f=new Uint8Array(h),m=0;for(let g=0;g<d.length;g++){let y=d[g],x=new Uint8Array(new Uint32Array([y.length]).buffer);f.set(x,m),m+=rh,f.set(y,m),m+=y.length}c(f)});n.push(p)}else n.push(l.data());t!=null&&(u.group=t),a.push(u)}let s=await Promise.all(n);return{data:_T(s),specs:a}}function bA(e,t){let a={},n,r=0;for(let s of t){let i=s.name,o=s.dtype,l=s.shape,u=vt(l),p;if("quantization"in s){let c=s.quantization;if(c.dtype==="uint8"||c.dtype==="uint16"){if(!("min"in c&&"scale"in c))throw new Error(`Weight ${s.name} with quantization ${c.dtype} doesn't have corresponding metadata min and scale.`)}else if(c.dtype==="float16"){if(o!=="float32")throw new Error(`Weight ${s.name} is quantized with ${c.dtype} which only supports weights of type float32 not ${o}.`)}else throw new Error(`Weight ${s.name} has unknown quantization dtype ${c.dtype}. Supported quantization dtypes are: 'uint8', 'uint16', and 'float16'.`);let d=J2[c.dtype],h=e.slice(r,r+u*d),f=c.dtype==="uint8"?new Uint8Array(h):new Uint16Array(h);if(o==="float32")if(c.dtype==="uint8"||c.dtype==="uint16"){p=new Float32Array(f.length);for(let m=0;m<f.length;m++){let g=f[m];p[m]=g*c.scale+c.min}}else if(c.dtype==="float16")n===void 0&&(n=LT()),p=n(f);else throw new Error(`Unsupported quantization type ${c.dtype} for weight type float32.`);else if(o==="int32"){if(c.dtype!=="uint8"&&c.dtype!=="uint16")throw new Error(`Unsupported quantization type ${c.dtype} for weight type int32.`);p=new Int32Array(f.length);for(let m=0;m<f.length;m++){let g=f[m];p[m]=Math.round(g*c.scale+c.min)}}else throw new Error(`Unsupported dtype in weight '${i}': ${o}`);r+=u*d}else if(o==="string"){let c=vt(s.shape);p=[];for(let d=0;d<c;d++){let h=new Uint32Array(e.slice(r,r+rh))[0];r+=rh;let f=new Uint8Array(e.slice(r,r+h));p.push(f),r+=h}}else{let c=J2[o],d=e.slice(r,r+u*c);if(o==="float32")p=new Float32Array(d);else if(o==="int32")p=new Int32Array(d);else if(o==="bool")p=new Uint8Array(d);else if(o==="complex64"){p=new Float32Array(d);let h=new Float32Array(p.length/2),f=new Float32Array(p.length/2);for(let y=0;y<h.length;y++)h[y]=p[y*2],f[y]=p[y*2+1];let m=Ve(h,l,"float32"),g=Ve(f,l,"float32");a[i]=Sr(m,g),m.dispose(),g.dispose()}else throw new Error(`Unsupported dtype in weight '${i}': ${o}`);r+=u*c}o!=="complex64"&&(a[i]=Ve(p,l,o))}return a}function _T(e){if(e===null)throw new Error(`Invalid input value: ${JSON.stringify(e)}`);let t=0,a=[];e.forEach(s=>{if(t+=s.byteLength,a.push(s.byteLength===s.buffer.byteLength?s:new s.constructor(s)),!(s instanceof Float32Array||s instanceof Int32Array||s instanceof Uint8Array))throw new Error(`Unsupported TypedArray subtype: ${s.constructor.name}`)});let n=new Uint8Array(t),r=0;return a.forEach(s=>{n.set(new Uint8Array(s.buffer),r),r+=s.byteLength}),n.buffer}var H1=typeof Buffer!="undefined"&&(typeof Blob=="undefined"||typeof atob=="undefined"||typeof btoa=="undefined");function _5(e){return H1?Buffer.byteLength(e):new Blob([e]).size}function $T(e){if(H1)return Buffer.from(e).toString("base64");let t=new Uint8Array(e),a="";for(let n=0,r=t.length;n<r;n++)a+=String.fromCharCode(t[n]);return btoa(a)}function DT(e){if(H1){let n=Buffer.from(e,"base64");return n.buffer.slice(n.byteOffset,n.byteOffset+n.byteLength)}let t=atob(e),a=new Uint8Array(t.length);for(let n=0;n<t.length;++n)a.set([t.charCodeAt(n)],n);return a.buffer}function j1(e){if(e.length===1)return e[0];let t=0;e.forEach(r=>{t+=r.byteLength});let a=new Uint8Array(t),n=0;return e.forEach(r=>{a.set(new Uint8Array(r),n),n+=r.byteLength}),a.buffer}function $5(e){let t="/";for(e=e.trim();e.endsWith(t);)e=e.slice(0,e.length-1);let a=e.split(t);return a[a.length-1]}function vA(e,t){let a={modelTopology:e.modelTopology,format:e.format,generatedBy:e.generatedBy,convertedBy:e.convertedBy,weightsManifest:t};return e.signature!=null&&(a.signature=e.signature),e.userDefinedMetadata!=null&&(a.userDefinedMetadata=e.userDefinedMetadata),e.modelInitializer!=null&&(a.modelInitializer=e.modelInitializer),e.initializerSignature!=null&&(a.initializerSignature=e.initializerSignature),e.trainingConfig!=null&&(a.trainingConfig=e.trainingConfig),a}function wA(e,t,a){let n={modelTopology:e.modelTopology,format:e.format,generatedBy:e.generatedBy,convertedBy:e.convertedBy};if(e.trainingConfig!=null&&(n.trainingConfig=e.trainingConfig),e.weightsManifest!=null){if(!t)throw new Error("modelJSON has weightsManifest but weightSpecs is null");if(!a)throw new Error("modelJSON has weightsManifest but weightData is null");n.weightSpecs=t,n.weightData=a}return e.signature!=null&&(n.signature=e.signature),e.userDefinedMetadata!=null&&(n.userDefinedMetadata=e.userDefinedMetadata),e.modelInitializer!=null&&(n.modelInitializer=e.modelInitializer),e.initializerSignature!=null&&(n.initializerSignature=e.initializerSignature),n}async function q1(e,t){let a,n;return e.weightsManifest!=null&&([a,n]=await t(e.weightsManifest)),wA(e,a,n)}function Cp(e){if(e.modelTopology instanceof ArrayBuffer)throw new Error("Expected JSON model topology, received ArrayBuffer.");return{dateSaved:new Date,modelTopologyType:"JSON",modelTopologyBytes:e.modelTopology==null?0:_5(JSON.stringify(e.modelTopology)),weightSpecsBytes:e.weightSpecs==null?0:_5(JSON.stringify(e.weightSpecs)),weightDataBytes:e.weightData==null?0:e.weightData.byteLength}}function kA(e){let t=[];for(let a of e)t.push(...a.weights);return t}function PT(){let e=a=>{let n=a<<13,r=0;for(;!(n&8388608);)r-=8388608,n<<=1;return n&=-8388609,r+=947912704,n|r},t=new Uint32Array(2048);t[0]=0;for(let a=1;a<1024;a++)t[a]=e(a);for(let a=1024;a<2048;a++)t[a]=939524096+(a-1024<<13);return t}function OT(){let e=new Uint32Array(64);e[0]=0,e[31]=1199570944,e[32]=2147483648,e[63]=3347054592;for(let t=1;t<31;t++)e[t]=t<<23;for(let t=33;t<63;t++)e[t]=2147483648+(t-32<<23);return e}function zT(){let e=new Uint32Array(64);for(let t=0;t<64;t++)e[t]=1024;return e[0]=e[32]=0,e}function LT(){let e=PT(),t=OT(),a=zT();return n=>{let r=new ArrayBuffer(4*n.length),s=new Uint32Array(r);for(let i=0;i<n.length;i++){let o=n[i],l=e[a[o>>10]+(o&1023)]+t[o>>10];s[i]=l}return new Float32Array(r)}}var _t=class{constructor(){this.saveRouters=[],this.loadRouters=[]}static getInstance(){return _t.instance==null&&(_t.instance=new _t),_t.instance}static registerSaveRouter(e){_t.getInstance().saveRouters.push(e)}static registerLoadRouter(e){_t.getInstance().loadRouters.push(e)}static getSaveHandlers(e){return _t.getHandlers(e,"save")}static getLoadHandlers(e,t){return _t.getHandlers(e,"load",t)}static getHandlers(e,t,a){let n=[];return(t==="load"?_t.getInstance().loadRouters:_t.getInstance().saveRouters).forEach(r=>{let s=r(e,a);s!==null&&n.push(s)}),n}},WT=e=>_t.registerSaveRouter(e),BT=e=>_t.registerLoadRouter(e),VT=e=>_t.getSaveHandlers(e),UT=(e,t)=>_t.getLoadHandlers(e,t),Q2="tensorflowjs",e1=1,Ws="models_store",Vr="model_info_store";function IA(){if(!W().getBool("IS_BROWSER"))throw new Error("Failed to obtain IndexedDB factory because the current environmentis not a web browser.");let e=typeof window=="undefined"?self:window,t=e.indexedDB||e.mozIndexedDB||e.webkitIndexedDB||e.msIndexedDB||e.shimIndexedDB;if(t==null)throw new Error("The current browser does not appear to support IndexedDB.");return t}function t1(e){let t=e.result;t.createObjectStore(Ws,{keyPath:"modelPath"}),t.createObjectStore(Vr,{keyPath:"modelPath"})}var Us=class{constructor(e){if(this.indexedDB=IA(),e==null||!e)throw new Error("For IndexedDB, modelPath must not be null, undefined or empty.");this.modelPath=e}async save(e){if(e.modelTopology instanceof ArrayBuffer)throw new Error("BrowserLocalStorage.save() does not support saving model topology in binary formats yet.");return this.databaseAction(this.modelPath,e)}async load(){return this.databaseAction(this.modelPath)}databaseAction(e,t){return new Promise((a,n)=>{let r=this.indexedDB.open(Q2,e1);r.onupgradeneeded=()=>t1(r),r.onsuccess=()=>{let s=r.result;if(t==null){let i=s.transaction(Ws,"readonly"),o=i.objectStore(Ws).get(this.modelPath);o.onsuccess=()=>{if(o.result==null)return s.close(),n(new Error(`Cannot find model with path '${this.modelPath}' in IndexedDB.`));a(o.result.modelArtifacts)},o.onerror=l=>(s.close(),n(o.error)),i.oncomplete=()=>s.close()}else{let i=Cp(t),o=s.transaction(Vr,"readwrite"),l=o.objectStore(Vr),u;try{u=l.put({modelPath:this.modelPath,modelArtifactsInfo:i})}catch(c){return n(c)}let p;u.onsuccess=()=>{p=s.transaction(Ws,"readwrite");let c=p.objectStore(Ws),d;try{d=c.put({modelPath:this.modelPath,modelArtifacts:t,modelArtifactsInfo:i})}catch(h){return n(h)}d.onsuccess=()=>a({modelArtifactsInfo:i}),d.onerror=h=>{l=o.objectStore(Vr);let f=l.delete(this.modelPath);f.onsuccess=()=>(s.close(),n(d.error)),f.onerror=m=>(s.close(),n(d.error))}},u.onerror=c=>(s.close(),n(u.error)),o.oncomplete=()=>{p==null?s.close():p.oncomplete=()=>s.close()}}},r.onerror=s=>n(r.error)})}};Us.URL_SCHEME="indexeddb://";var SA=e=>W().getBool("IS_BROWSER")&&!Array.isArray(e)&&e.startsWith(Us.URL_SCHEME)?GT(e.slice(Us.URL_SCHEME.length)):null;_t.registerSaveRouter(SA);_t.registerLoadRouter(SA);function GT(e){return new Us(e)}function HT(e){return e.startsWith(Us.URL_SCHEME)?e.slice(Us.URL_SCHEME.length):e}var jT=class{constructor(){this.indexedDB=IA()}async listModels(){return new Promise((e,t)=>{let a=this.indexedDB.open(Q2,e1);a.onupgradeneeded=()=>t1(a),a.onsuccess=()=>{let n=a.result,r=n.transaction(Vr,"readonly"),s=r.objectStore(Vr).getAll();s.onsuccess=()=>{let i={};for(let o of s.result)i[o.modelPath]=o.modelArtifactsInfo;e(i)},s.onerror=i=>(n.close(),t(s.error)),r.oncomplete=()=>n.close()},a.onerror=n=>t(a.error)})}async removeModel(e){return e=HT(e),new Promise((t,a)=>{let n=this.indexedDB.open(Q2,e1);n.onupgradeneeded=()=>t1(n),n.onsuccess=()=>{let r=n.result,s=r.transaction(Vr,"readwrite"),i=s.objectStore(Vr),o=i.get(e),l;o.onsuccess=()=>{if(o.result==null)return r.close(),a(new Error(`Cannot find model with path '${e}' in IndexedDB.`));{let u=i.delete(e),p=()=>{l=r.transaction(Ws,"readwrite");let c=l.objectStore(Ws).delete(e);c.onsuccess=()=>t(o.result.modelArtifactsInfo),c.onerror=d=>a(o.error)};u.onsuccess=p,u.onerror=c=>(p(),r.close(),a(o.error))}},o.onerror=u=>(r.close(),a(o.error)),s.oncomplete=()=>{l==null?r.close():l.oncomplete=()=>r.close()}},n.onerror=r=>a(n.error)})}},vr="/",Tl="tensorflowjs_models",CA="info",qT="model_topology",XT="weight_specs",KT="weight_data",YT="model_metadata";function TA(e){return{info:[Tl,e,CA].join(vr),topology:[Tl,e,qT].join(vr),weightSpecs:[Tl,e,XT].join(vr),weightData:[Tl,e,KT].join(vr),modelMetadata:[Tl,e,YT].join(vr)}}function NA(e){for(let t of Object.values(e))window.localStorage.removeItem(t)}function ZT(e){let t=e.split(vr);if(t.length<3)throw new Error(`Invalid key format: ${e}`);return t.slice(1,t.length-1).join(vr)}function JT(e){return e.startsWith(Gs.URL_SCHEME)?e.slice(Gs.URL_SCHEME.length):e}var Gs=class{constructor(e){if(!W().getBool("IS_BROWSER")||typeof window=="undefined"||typeof window.localStorage=="undefined")throw new Error("The current environment does not support local storage.");if(this.LS=window.localStorage,e==null||!e)throw new Error("For local storage, modelPath must not be null, undefined or empty.");this.modelPath=e,this.keys=TA(this.modelPath)}async save(e){if(e.modelTopology instanceof ArrayBuffer)throw new Error("BrowserLocalStorage.save() does not support saving model topology in binary formats yet.");{let t=JSON.stringify(e.modelTopology),a=JSON.stringify(e.weightSpecs),n=Cp(e);try{this.LS.setItem(this.keys.info,JSON.stringify(n)),this.LS.setItem(this.keys.topology,t),this.LS.setItem(this.keys.weightSpecs,a),this.LS.setItem(this.keys.weightData,$T(e.weightData));let r={format:e.format,generatedBy:e.generatedBy,convertedBy:e.convertedBy,signature:e.signature!=null?e.signature:void 0,userDefinedMetadata:e.userDefinedMetadata!=null?e.userDefinedMetadata:void 0,modelInitializer:e.modelInitializer!=null?e.modelInitializer:void 0,initializerSignature:e.initializerSignature!=null?e.initializerSignature:void 0,trainingConfig:e.trainingConfig!=null?e.trainingConfig:void 0};return this.LS.setItem(this.keys.modelMetadata,JSON.stringify(r)),{modelArtifactsInfo:n}}catch(r){throw NA(this.keys),new Error(`Failed to save model '${this.modelPath}' to local storage: size quota being exceeded is a possible cause of this failure: modelTopologyBytes=${n.modelTopologyBytes}, weightSpecsBytes=${n.weightSpecsBytes}, weightDataBytes=${n.weightDataBytes}.`)}}}async load(){let e=JSON.parse(this.LS.getItem(this.keys.info));if(e==null)throw new Error(`In local storage, there is no model with name '${this.modelPath}'`);if(e.modelTopologyType!=="JSON")throw new Error("BrowserLocalStorage does not support loading non-JSON model topology yet.");let t={},a=JSON.parse(this.LS.getItem(this.keys.topology));if(a==null)throw new Error(`In local storage, the topology of model '${this.modelPath}' is missing.`);t.modelTopology=a;let n=JSON.parse(this.LS.getItem(this.keys.weightSpecs));if(n==null)throw new Error(`In local storage, the weight specs of model '${this.modelPath}' are missing.`);t.weightSpecs=n;let r=this.LS.getItem(this.keys.modelMetadata);if(r!=null){let i=JSON.parse(r);t.format=i.format,t.generatedBy=i.generatedBy,t.convertedBy=i.convertedBy,i.signature!=null&&(t.signature=i.signature),i.userDefinedMetadata!=null&&(t.userDefinedMetadata=i.userDefinedMetadata),i.modelInitializer!=null&&(t.modelInitializer=i.modelInitializer),i.initializerSignature!=null&&(t.initializerSignature=i.initializerSignature),i.trainingConfig!=null&&(t.trainingConfig=i.trainingConfig)}let s=this.LS.getItem(this.keys.weightData);if(s==null)throw new Error(`In local storage, the binary weight values of model '${this.modelPath}' are missing.`);return t.weightData=DT(s),t}};Gs.URL_SCHEME="localstorage://";var RA=e=>W().getBool("IS_BROWSER")&&!Array.isArray(e)&&e.startsWith(Gs.URL_SCHEME)?QT(e.slice(Gs.URL_SCHEME.length)):null;_t.registerSaveRouter(RA);_t.registerLoadRouter(RA);function QT(e){return new Gs(e)}var eN=class{constructor(){D(W().getBool("IS_BROWSER"),()=>"Current environment is not a web browser"),D(typeof window=="undefined"||typeof window.localStorage!="undefined",()=>"Current browser does not appear to support localStorage"),this.LS=window.localStorage}async listModels(){let e={},t=Tl+vr,a=vr+CA;for(let n=0;n<this.LS.length;++n){let r=this.LS.key(n);if(r.startsWith(t)&&r.endsWith(a)){let s=ZT(r);e[s]=JSON.parse(this.LS.getItem(r))}}return e}async removeModel(e){e=JT(e);let t=TA(e);if(this.LS.getItem(t.info)==null)throw new Error(`Cannot find model at path '${e}'`);let a=JSON.parse(this.LS.getItem(t.info));return NA(t),a}},El="://",La=class{constructor(){this.managers={}}static getInstance(){return La.instance==null&&(La.instance=new La),La.instance}static registerManager(e,t){D(e!=null,()=>"scheme must not be undefined or null."),e.endsWith(El)&&(e=e.slice(0,e.indexOf(El))),D(e.length>0,()=>"scheme must not be an empty string.");let a=La.getInstance();D(a.managers[e]==null,()=>`A model store manager is already registered for scheme '${e}'.`),a.managers[e]=t}static getManager(e){let t=La.getInstance().managers[e];if(t==null)throw new Error(`Cannot find model manager for scheme '${e}'`);return t}static getSchemes(){return Object.keys(La.getInstance().managers)}};function jc(e){if(e.indexOf(El)===-1)throw new Error(`The url string provided does not contain a scheme. Supported schemes are: ${La.getSchemes().join(",")}`);return{scheme:e.split(El)[0],path:e.split(El)[1]}}async function EA(e,t,a=!1){D(e!==t,()=>`Old path and new path are the same: '${e}'`);let n=_t.getLoadHandlers(e);D(n.length>0,()=>`Copying failed because no load handler is found for source URL ${e}.`),D(n.length<2,()=>`Copying failed because more than one (${n.length}) load handlers for source URL ${e}.`);let r=n[0],s=_t.getSaveHandlers(t);D(s.length>0,()=>`Copying failed because no save handler is found for destination URL ${t}.`),D(s.length<2,()=>`Copying failed because more than one (${n.length}) save handlers for destination URL ${t}.`);let i=s[0],o=jc(e).scheme,l=jc(e).path,u=o===jc(e).scheme,p=await r.load();a&&u&&await La.getManager(o).removeModel(l);let c=await i.save(p);return a&&!u&&await La.getManager(o).removeModel(l),c.modelArtifactsInfo}async function tN(){let e=La.getSchemes(),t={};for(let a of e){let n=await La.getManager(a).listModels();for(let r in n){let s=a+El+r;t[s]=n[r]}}return t}async function aN(e){let t=jc(e);return La.getManager(t.scheme).removeModel(t.path)}async function nN(e,t){return EA(e,t,!1)}async function rN(e,t){return EA(e,t,!0)}var sN=class{constructor(){this.messageName="setTimeoutCustom",this.functionRefs=[],this.handledMessageCount=0,this.hasEventListener=!1}fetch(e,t){return fetch(e,t)}now(){return performance.now()}encode(e,t){if(t!=="utf-8"&&t!=="utf8")throw new Error(`Browser's encoder only supports utf-8, but got ${t}`);return this.textEncoder==null&&(this.textEncoder=new TextEncoder),this.textEncoder.encode(e)}decode(e,t){return new TextDecoder(t).decode(e)}setTimeoutCustom(e,t){if(typeof window=="undefined"||!W().getBool("USE_SETTIMEOUTCUSTOM")){setTimeout(e,t);return}this.functionRefs.push(e),setTimeout(()=>{window.postMessage({name:this.messageName,index:this.functionRefs.length-1},"*")},t),this.hasEventListener||(this.hasEventListener=!0,window.addEventListener("message",a=>{if(a.source===window&&a.data.name===this.messageName){a.stopPropagation();let n=this.functionRefs[a.data.index];n(),this.handledMessageCount++,this.handledMessageCount===this.functionRefs.length&&(this.functionRefs=[],this.handledMessageCount=0)}},!0))}isTypedArray(e){return lA(e)}};if(W().get("IS_BROWSER")){W().setPlatform("browser",new sN);try{La.registerManager(Gs.URL_SCHEME,new eN)}catch(e){}try{La.registerManager(Us.URL_SCHEME,new jT)}catch(e){}}var iN={importFetch:()=>uC()},P2,oN=class{constructor(){this.util=dC(),this.textEncoder=new this.util.TextEncoder}fetch(e,t){return W().global.fetch!=null?W().global.fetch(e,t):(P2==null&&(P2=iN.importFetch()),P2(e,t))}now(){let e=process.hrtime();return e[0]*1e3+e[1]/1e6}encode(e,t){if(t!=="utf-8"&&t!=="utf8")throw new Error(`Node built-in encoder only supports utf-8, but got ${t}`);return this.textEncoder.encode(e)}decode(e,t){return e.length===0?"":new this.util.TextDecoder(t).decode(e)}isTypedArray(e){return this.util.types.isFloat32Array(e)||this.util.types.isInt32Array(e)||this.util.types.isUint8Array(e)||this.util.types.isUint8ClampedArray(e)}};W().get("IS_NODE")&&!W().get("IS_BROWSER")&&W().setPlatform("node",new oN);function $e(e,t="float32",a){return t=t||"float32",an(e),new Wt(e,t,a)}function lN(e,t){let a=R(e,"x","cast");if(!eA(t))throw new Error(`Failed to cast to unknown dtype ${t}`);if(t==="string"&&a.dtype!=="string"||t!=="string"&&a.dtype==="string")throw new Error("Only strings can be casted to strings");let n={x:a},r={dtype:t};return L.runKernel(ci,n,r)}var Ke=z({cast_:lN});function uN(e){let t={x:R(e,"x","clone","string_or_numeric")};return L.runKernel(zi,t)}var Sa=z({clone_:uN});function X1(e,t=!1){console.log(e.toString(t))}yA();var dN={buffer:$e,cast:Ke,clone:Sa,print:X1};bT(dN);function K1(){W().set("PROD",!0)}function pN(){W().set("DEBUG",!0)}function cN(){W().set("DEPRECATION_WARNINGS_ENABLED",!1),console.warn("TensorFlow.js deprecation warnings have been disabled.")}function Y1(e){W().getBool("DEPRECATION_WARNINGS_ENABLED")&&console.warn(e+" You can disable deprecation warnings with tf.disableDeprecationWarnings().")}vT(Y1);function hN(){L.disposeVariables()}function It(){return L}function fN(){return L.memory()}function mN(e){return L.profile(e)}function Pe(e,t){return L.tidy(e,t)}function J(e){U1(e).forEach(t=>t.dispose())}function zn(e){return L.keep(e)}function gN(e){return L.time(e)}function Tp(e){return L.setBackend(e)}function Np(){return L.ready()}function ha(){return L.backendName}function yN(e){L.removeBackend(e)}function Z1(e){return L.findBackend(e)}function xN(e){return L.findBackendFactory(e)}function Go(e,t,a=1){return L.registerBackend(e,t,a)}function nr(){return L.backend}function AN(e,t){W().setPlatform(e,t)}function bN(e,t){let a=R(e,"a","add"),n=R(t,"b","add");[a,n]=Rt(a,n);let r={a,b:n};return L.runKernel(ns,r)}var ke=z({add_:bN});function vN(e,t){let a=R(e,"a","floorDiv"),n=R(t,"b","floorDiv");[a,n]=Rt(a,n);let r={a,b:n};return L.runKernel(_i,r)}var Rp=z({floorDiv_:vN});function wN(e,t){let a=R(e,"a","div"),n=R(t,"b","div");if([a,n]=Rt(a,n),a.dtype==="int32"&&n.dtype==="int32")return Rp(a,n);let r={a,b:n},s={};return L.runKernel(Ci,r,s)}var ve=z({div_:wN});function kN(e,t){let a=R(e,"a","mul"),n=R(t,"b","mul");[a,n]=Rt(a,n);let r={a,b:n};return L.runKernel(io,r)}var te=z({mul_:kN});function IN(e){let t=R(e,"x","abs");if(t.dtype==="complex64"){let a={x:t};return L.runKernel(ap,a)}else{let a={x:t};return L.runKernel(Zl,a)}}var Za=z({abs_:IN});function SN(e){let t={x:R(e,"x","acos")};return L.runKernel(Qs,t)}var MA=z({acos_:SN});function CN(e){let t={x:R(e,"x","acosh")};return L.runKernel(ei,t)}var FA=z({acosh_:CN});function TN(e){D(Array.isArray(e),()=>"The argument passed to tf.addN() must be a list of tensors"),D(e.length>=1,()=>`Must pass at least one tensor to tf.addN(), but got ${e.length}`);let t=e.map((r,s)=>R(r,`tensors${s}`,"addN")),a=t[0];t.forEach(r=>{if(r.dtype!==a.dtype)throw new Error("All tensors passed to tf.addN() must have the same dtype")}),t.forEach(r=>{if(!Cr(r.shape,a.shape))throw new Error("All tensors passed to tf.addN() must have the same shape")});let n=t;return L.runKernel(ti,n)}var Eh=z({addN_:TN});function NN(e,t=null,a=!1){let n={x:R(e,"x","all","bool")},r={axis:t,keepDims:a};return L.runKernel(ai,n,r)}var _A=z({all_:NN});function RN(e,t=null,a=!1){let n={x:R(e,"x","any","bool")},r={axis:t,keepDims:a};return L.runKernel(ni,n,r)}var $A=z({any_:RN});function EN(e,t=0){let a={x:R(e,"x","argMax")},n={axis:t};return L.runKernel(Jl,a,n)}var rr=z({argMax_:EN});function MN(e,t=0){let a={x:R(e,"x","argMin")},n={axis:t};return L.runKernel(Ql,a,n)}var DA=z({argMin_:MN});function FN(e){let t={x:R(e,"x","asin")};return L.runKernel(ri,t)}var PA=z({asin_:FN});function _N(e){let t={x:R(e,"x","asinh")};return L.runKernel(si,t)}var OA=z({asinh_:_N});function $N(e){let t={x:R(e,"x","atan")};return L.runKernel(ii,t)}var zA=z({atan_:$N});function DN(e,t){let a=R(e,"a","atan2"),n=R(t,"b","atan2");[a,n]=Rt(a,n);let r={a,b:n};return L.runKernel(li,r)}var LA=z({atan2_:DN});function PN(e){let t={x:R(e,"x","atanh")};return L.runKernel(oi,t)}var WA=z({atanh_:PN});function ON(e,t,a,n,r="NHWC",s){let i=e[3],o=[...t,i],l=UA(r);return Ep(e,o,a,s,n,null,null,l)}function BA(e,t,a,n,r,s,i="channelsLast"){let[o,l]=Pd(t),u;if(i==="channelsLast")u=[o,l,e[3],e[3]];else if(i==="channelsFirst")u=[o,l,e[1],e[1]];else throw new Error(`Unknown dataFormat ${i}`);return Ep(e,u,a,n,r,s,!1,i)}function zN(e,t,a,n,r,s,i="NDHWC"){let[o,l,u]=a1(t),p,c;if(i==="NDHWC")c="channelsLast",p=[o,l,u,e[4],e[4]];else if(i==="NCDHW")c="channelsFirst",p=[o,l,u,e[1],e[1]];else throw new Error(`Unknown dataFormat ${i}`);return VA(e,p,a,n,r,!1,c,s)}function Ep(e,t,a,n,r,s,i=!1,o="channelsLast"){let[l,u,p,c]=[-1,-1,-1,-1];if(o==="channelsLast")[l,u,p,c]=e;else if(o==="channelsFirst")[l,c,u,p]=e;else throw new Error(`Unknown dataFormat ${o}`);let[d,h,,f]=t,[m,g]=Pd(a),[y,x]=Pd(n),A=Ml(d,y),b=Ml(h,x),{padInfo:w,outHeight:C,outWidth:T}=BN(r,u,p,m,g,A,b,s,o),N=i?f*c:f,M;return o==="channelsFirst"?M=[l,N,C,T]:o==="channelsLast"&&(M=[l,C,T,N]),{batchSize:l,dataFormat:o,inHeight:u,inWidth:p,inChannels:c,outHeight:C,outWidth:T,outChannels:N,padInfo:w,strideHeight:m,strideWidth:g,filterHeight:d,filterWidth:h,effectiveFilterHeight:A,effectiveFilterWidth:b,dilationHeight:y,dilationWidth:x,inShape:e,outShape:M,filterShape:t}}function VA(e,t,a,n,r,s=!1,i="channelsLast",o){let[l,u,p,c,d]=[-1,-1,-1,-1,-1];if(i==="channelsLast")[l,u,p,c,d]=e;else if(i==="channelsFirst")[l,d,u,p,c]=e;else throw new Error(`Unknown dataFormat ${i}`);let[h,f,m,,g]=t,[y,x,A]=a1(a),[b,w,C]=a1(n),T=Ml(h,b),N=Ml(f,w),M=Ml(m,C),{padInfo:F,outDepth:E,outHeight:I,outWidth:_}=VN(r,u,p,c,y,x,A,T,N,M,o),O=s?g*d:g,B;return i==="channelsFirst"?B=[l,O,E,I,_]:i==="channelsLast"&&(B=[l,E,I,_,O]),{batchSize:l,dataFormat:i,inDepth:u,inHeight:p,inWidth:c,inChannels:d,outDepth:E,outHeight:I,outWidth:_,outChannels:O,padInfo:F,strideDepth:y,strideHeight:x,strideWidth:A,filterDepth:h,filterHeight:f,filterWidth:m,effectiveFilterDepth:T,effectiveFilterHeight:N,effectiveFilterWidth:M,dilationDepth:b,dilationHeight:w,dilationWidth:C,inShape:e,outShape:B,filterShape:t}}function LN(e,t,a,n,r){n==null&&(n=J1(e,t,a));let s=e[0],i=e[1],o=Od((s-t+2*n)/a+1,r),l=Od((i-t+2*n)/a+1,r);return[o,l]}function WN(e,t,a,n,r,s){r==null&&(r=J1(e,t[0],n[0]));let i=[0,0,0,a];for(let o=0;o<3;o++)e[o]+2*r>=t[o]&&(i[o]=Od((e[o]-t[o]+2*r)/n[o]+1,s));return i}function J1(e,t,a,n=1){let r=Ml(t,n);return Math.floor((e[0]*(a-1)-a+r)/2)}function Pd(e){return typeof e=="number"?[e,e,e]:e.length===2?[e[0],e[1],1]:e}function a1(e){return typeof e=="number"?[e,e,e]:e}function Ml(e,t){return t<=1?e:e+(e-1)*(t-1)}function BN(e,t,a,n,r,s,i,o,l){let u,p,c;if(typeof e=="number"){u={top:e,bottom:e,left:e,right:e,type:e===0?"VALID":"NUMBER"};let d=LN([t,a],s,n,e,o);p=d[0],c=d[1]}else if(e==="same"){p=Math.ceil(t/n),c=Math.ceil(a/r);let d=Math.max(0,(p-1)*n+s-t),h=Math.max(0,(c-1)*r+i-a),f=Math.floor(d/2),m=d-f,g=Math.floor(h/2),y=h-g;u={top:f,bottom:m,left:g,right:y,type:"SAME"}}else if(e==="valid")u={top:0,bottom:0,left:0,right:0,type:"VALID"},p=Math.ceil((t-s+1)/n),c=Math.ceil((a-i+1)/r);else if(typeof e=="object"){let d=l==="channelsLast"?e[1][0]:e[2][0],h=l==="channelsLast"?e[1][1]:e[2][1],f=l==="channelsLast"?e[2][0]:e[3][0],m=l==="channelsLast"?e[2][1]:e[3][1];u={top:d,bottom:h,left:f,right:m,type:d===0&&h===0&&f===0&&m===0?"VALID":"EXPLICIT"},p=Od((t-s+d+h)/n+1,o),c=Od((a-i+f+m)/r+1,o)}else throw Error(`Unknown padding parameter: ${e}`);return{padInfo:u,outHeight:p,outWidth:c}}function VN(e,t,a,n,r,s,i,o,l,u,p){let c,d,h,f;if(e==="valid"&&(e=0),typeof e=="number"){c={top:e,bottom:e,left:e,right:e,front:e,back:e,type:e===0?"VALID":"NUMBER"};let m=WN([t,a,n,1],[o,l,u],1,[r,s,i],e,p);d=m[0],h=m[1],f=m[2]}else if(e==="same"){d=Math.ceil(t/r),h=Math.ceil(a/s),f=Math.ceil(n/i);let m=(d-1)*r+o-t,g=(h-1)*s+l-a,y=(f-1)*i+u-n,x=Math.floor(m/2),A=m-x,b=Math.floor(g/2),w=g-b,C=Math.floor(y/2),T=y-C;c={top:b,bottom:w,left:C,right:T,front:x,back:A,type:"SAME"}}else throw Error(`Unknown padding parameter: ${e}`);return{padInfo:c,outDepth:d,outHeight:h,outWidth:f}}function Od(e,t){if(!t)return Math.trunc(e);switch(t){case"round":return Math.round(e);case"ceil":return Math.ceil(e);case"floor":return Math.floor(e);default:throw new Error(`Unknown roundingMode ${t}`)}}function zd(e){let[t,a,n]=Pd(e);return t===1&&a===1&&n===1}function Tr(e,t){return zd(e)||zd(t)}function Hs(e){return Pd(e).every(t=>t>0)}function UA(e){if(e==="NHWC")return"channelsLast";if(e==="NCHW")return"channelsFirst";throw new Error(`Unknown dataFormat ${e}`)}function Tn(e,t,a){if(a!=null){if(typeof t=="string")throw Error(`Error in ${e}: pad must be an integer when using dimRoundingMode ${a} but got pad ${t}.`);if(typeof t=="number")D($l(t),()=>`Error in ${e}: pad must be an integer when using dimRoundingMode ${a} but got pad ${t}.`);else if(typeof t=="object")t.forEach(n=>{n.forEach(r=>{D($l(r),()=>`Error in ${e}: pad must be an integer when using dimRoundingMode ${a} but got pad ${r}.`)})});else throw Error(`Error in ${e}: Unknown padding parameter: ${t}`)}}function UN(e,t){let a={x:R(e,"x","reshape","string_or_numeric")},n={shape:t};return L.runKernel(bu,a,n)}var Q=z({reshape_:UN});function GN(e,t,a,n,r){let s=R(e,"x","avgPool","float32"),i=1;D(Tr(a,i),()=>`Error in avgPool: Either strides or dilations must be 1. Got strides ${a} and dilations '${i}'`);let o=s,l=!1;s.rank===3&&(l=!0,o=Q(s,[1,s.shape[0],s.shape[1],s.shape[2]])),D(o.rank===4,()=>`Error in avgPool: x must be rank 4 but got rank ${o.rank}.`),Tn("avgPool",n,r);let u={x:o},p={filterSize:t,strides:a,pad:n,dimRoundingMode:r},c=L.runKernel(ui,u,p);return c=Ke(c,s.dtype),l?Q(c,[c.shape[1],c.shape[2],c.shape[3]]):c}var Q1=z({avgPool_:GN});function HN(e,t,a,n,r,s="NDHWC"){let i=R(e,"x","avgPool3d","float32"),o=i,l=!1;i.rank===4&&(l=!0,o=Q(i,[1,i.shape[0],i.shape[1],i.shape[2],i.shape[3]])),D(o.rank===5,()=>`Error in avgPool3d: x must be rank 5 but got rank ${o.rank}.`),D(s==="NDHWC",()=>`Error in avgPool3d: Only NDHWC is currently supported, but got dataFormat of ${s}`),D(typeof a=="number"&&a>0||Array.isArray(a)&&a[0]>0&&a[1]>0&&a[2]>0,()=>`Error in avgPool3d: Stride must be > 0, but got '${a}'`),Tn("avgPool3d",n,r);let u={x:o},p={filterSize:t,strides:a,pad:n,dimRoundingMode:r,dataFormat:s},c=L.runKernel(eu,u,p);return c=Ke(c,o.dtype),l?Q(c,[c.shape[1],c.shape[2],c.shape[3],c.shape[4]]):c}var GA=z({avgPool3d_:HN});function jN(e,t=0){D(e.length>=1,()=>"Pass at least one tensor to concat");let a=Dd(e,"tensors","concat","string_or_numeric");if(a[0].dtype==="complex64"&&a.forEach(s=>{if(s.dtype!=="complex64")throw new Error(`Cannot concatenate complex64 tensors with a tensor
|
|
with dtype ${s.dtype}. `)}),a.length===1)return Sa(a[0]);let n=a,r={axis:t};return L.runKernel(nu,n,r)}var lt=z({concat_:jN});function qN(e,t,a=!1,n=!1){let r=R(e,"a","matMul"),s=R(t,"b","matMul");[r,s]=Rt(r,s);let i={a:r,b:s},o={transposeA:a,transposeB:n};return L.runKernel(di,i,o)}var pt=z({matMul_:qN});function XN(e){let t={x:R(e,"x","sigmoid","float32")};return L.runKernel(Eo,t)}var Wa=z({sigmoid_:XN});function KN(e,t,a){let n=R(e,"x","slice","string_or_numeric");if(n.rank===0)throw new Error("Slicing scalar is not possible");let r={x:n},s={begin:t,size:a};return L.runKernel(Iu,r,s)}var De=z({slice_:KN});function YN(e){let t={x:R(e,"x","tanh","float32")};return L.runKernel(Wo,t)}var sh=z({tanh_:YN});function ZN(e,t,a,n,r,s){let i=R(e,"forgetBias","basicLSTMCell"),o=R(t,"lstmKernel","basicLSTMCell"),l=R(a,"lstmBias","basicLSTMCell"),u=R(n,"data","basicLSTMCell"),p=R(r,"c","basicLSTMCell"),c=R(s,"h","basicLSTMCell"),d=lt([u,c],1),h=pt(d,o),f=ke(h,l),m=f.shape[0],g=f.shape[1]/4,y=[m,g],x=De(f,[0,0],y),A=De(f,[0,g],y),b=De(f,[0,g*2],y),w=De(f,[0,g*3],y),C=ke(te(Wa(x),sh(A)),te(p,Wa(ke(i,b)))),T=te(sh(C),Wa(w));return[C,T]}var HA=z({basicLSTMCell_:ZN});function JN(e,t,a){let n=R(e,"x","batchToSpaceND"),r=t.reduce((o,l)=>o*l);D(n.rank>=1+t.length,()=>`input rank is ${n.rank} but should be > than blockShape.length ${t.length}`),D(a.length===t.length,()=>`crops.length is ${a.length} but should be equal to blockShape.length ${t.length}`),D(n.shape[0]%r===0,()=>`input tensor batch is ${n.shape[0]} but is not divisible by the product of the elements of blockShape ${t.join(" * ")} === ${r}`);let s={x:n},i={blockShape:t,crops:a};return L.runKernel(tu,s,i)}var e3=z({batchToSpaceND_:JN});function QN(e){let t;return e.rank===0||e.rank===1?t=Q(e,[1,1,1,e.size]):e.rank===2?t=Q(e,[1,1,e.shape[0],e.shape[1]]):e.rank===3?t=Q(e,[1,e.shape[0],e.shape[1],e.shape[2]]):t=e,t}function eR(e,t,a,n,r,s){s==null&&(s=.001);let i=R(e,"x","batchNorm"),o=R(t,"mean","batchNorm"),l=R(a,"variance","batchNorm"),u;r!=null&&(u=R(r,"scale","batchNorm"));let p;n!=null&&(p=R(n,"offset","batchNorm")),D(o.rank===l.rank,()=>"Batch normalization gradient requires mean and variance to have equal ranks."),D(p==null||o.rank===p.rank,()=>"Batch normalization gradient requires mean and offset to have equal ranks."),D(u==null||o.rank===u.rank,()=>"Batch normalization gradient requires mean and scale to have equal ranks.");let c={x:QN(i),scale:u,offset:p,mean:o,variance:l},d={varianceEpsilon:s},h=L.runKernel($i,c,d);return Q(h,i.shape)}var Mp=z({batchNorm_:eR});function tR(e,t,a,n,r,s){let i=R(e,"x","batchNorm"),o=R(t,"mean","batchNorm"),l=R(a,"variance","batchNorm"),u;r!=null&&(u=R(r,"scale","batchNorm"));let p;return n!=null&&(p=R(n,"offset","batchNorm")),D(i.rank===2,()=>`Error in batchNorm2D: x must be rank 2 but got rank ${i.rank}.`),D(o.rank===2||o.rank===1,()=>`Error in batchNorm2D: mean must be rank 2 or rank 1 but got rank ${o.rank}.`),D(l.rank===2||l.rank===1,()=>`Error in batchNorm2D: variance must be rank 2 or rank 1 but got rank ${l.rank}.`),u!=null&&D(u.rank===2||u.rank===1,()=>`Error in batchNorm2D: scale must be rank 2 or rank 1 but got rank ${u.rank}.`),p!=null&&D(p.rank===2||p.rank===1,()=>`Error in batchNorm2D: offset must be rank 2 or rank 1 but got rank ${p.rank}.`),Mp(i,o,l,p,u,s)}var jA=z({batchNorm2d_:tR});function aR(e,t,a,n,r,s){let i=R(e,"x","batchNorm"),o=R(t,"mean","batchNorm"),l=R(a,"variance","batchNorm"),u;r!=null&&(u=R(r,"scale","batchNorm"));let p;return n!=null&&(p=R(n,"offset","batchNorm")),D(i.rank===3,()=>`Error in batchNorm3D: x must be rank 3 but got rank ${i.rank}.`),D(o.rank===3||o.rank===1,()=>`Error in batchNorm3D: mean must be rank 3 or rank 1 but got rank ${o.rank}.`),D(l.rank===3||l.rank===1,()=>`Error in batchNorm3D: variance must be rank 3 or rank 1 but got rank ${l.rank}.`),u!=null&&D(u.rank===3||u.rank===1,()=>`Error in batchNorm3D: scale must be rank 3 or rank 1 but got rank ${u.rank}.`),p!=null&&D(p.rank===3||p.rank===1,()=>`Error in batchNorm3D: offset must be rank 3 or rank 1 but got rank ${p.rank}.`),Mp(i,o,l,p,u,s)}var qA=z({batchNorm3d_:aR});function nR(e,t,a,n,r,s){let i=R(e,"x","batchNorm"),o=R(t,"mean","batchNorm"),l=R(a,"variance","batchNorm"),u;r!=null&&(u=R(r,"scale","batchNorm"));let p;return n!=null&&(p=R(n,"offset","batchNorm")),D(i.rank===4,()=>`Error in batchNorm4D: x must be rank 4 but got rank ${i.rank}.`),D(o.rank===4||o.rank===1,()=>`Error in batchNorm4D: mean must be rank 4 or rank 1 but got rank ${o.rank}.`),D(l.rank===4||l.rank===1,()=>`Error in batchNorm4D: variance must be rank 4 or rank 1 but got rank ${l.rank}.`),u!=null&&D(u.rank===4||u.rank===1,()=>`Error in batchNorm4D: scale must be rank 4 or rank 1 but got rank ${u.rank}.`),p!=null&&D(p.rank===4||p.rank===1,()=>`Error in batchNorm4D: offset must be rank 4 or rank 1 but got rank ${p.rank}.`),Mp(i,o,l,p,u,s)}var XA=z({batchNorm4d_:nR});function rR(e,t,a){let n=R(e,"x","bincount"),r=R(t,"weights","bincount");D(n.dtype==="int32",()=>`Error in bincount: input dtype must be int32, but got ${n.dtype}`),D(a>=0,()=>`size must be non-negative, but got ${a}.`),D(r.size===n.size||r.size===0,()=>`Error in bincount: weights must have the same size as input or0-length, but got input shape: ${n.shape}, weights shape: ${r.shape}.`);let s={x:n,weights:r},i={size:a};return L.runKernel(pi,s,i)}var t3=z({bincount_:rR});function sR(e,t){let a=R(e,"x","bitwiseAnd"),n=R(t,"y","bitwiseAnd");if(!Cr(a.shape,n.shape))throw new Error(`BitwiseAnd: Tensors must have the same shape. x: ${a.shape}, y: ${n.shape}`);if(a.dtype!=="int32"||n.dtype!=="int32")throw new Error(`BitwiseAnd: Only supports 'int32' values in tensor, found type of x: ${a.dtype} and type of y: ${n.dtype}`);let r={a,b:n};return L.runKernel(Ih,r)}var KA=z({bitwiseAnd_:sR});function iR(e,t){let a=R(e,"s0","broadcastArgs","int32"),n=R(t,"s1","broadcastArgs","int32");if(a.rank!==1)throw new Error(`broadcastArgs(): first input must be a vector (rank=1). Has rank ${a.rank}`);if(n.rank!==1)throw new Error(`broadcastArgs(): second input must be a vector (rank=1). Has rank ${n.rank}`);let r={s0:a,s1:n};return L.runKernel(au,r)}var YA=z({broadcastArgs_:iR});function oR(e,t){let a=R(e,"broadcastTo","x"),n=a.shape;if(an(t),t.length<a.rank)throw new Error(`broadcastTo(): shape.length=${t.length} < input.rank=${a.rank}.`);if(t.length>a.rank){let l=a.shape.slice();for(;l.length<t.length;)l.unshift(1);a=Q(a,l)}let r=a.shape,s=Array.from(t);for(let l=t.length-1;l>=0;l--)if(r[l]===t[l])s[l]=1;else if(a.shape[l]!==1)throw new Error(`broadcastTo(): [${n}] cannot be broadcast to [${t}].`);if(s.map((l,u)=>l>1?u:-1).filter(l=>l>=0).length===0)return Sa(a);let i={x:a},o={reps:s};return L.runKernel(ss,i,o)}var Fl=z({broadcastTo_:oR});function lR(e){let t={x:R(e,"x","ceil","float32")};return L.runKernel(hi,t)}var ZA=z({ceil_:lR});function sr(e,t,a){an(e),a=a||Qd(t);let n={shape:e,value:t,dtype:a};return L.runKernel(du,{},n)}function uR(e,t,a){let n=R(e,"x","clipByValue");if(D(t<=a,()=>`Error in clip: min (${t}) must be less than or equal to max (${a}).`),t===a)return sr(n.shape,t,n.dtype);let r={x:n},s={clipValueMin:t,clipValueMax:a};return L.runKernel(rs,r,s)}var JA=z({clipByValue_:uR});function dR(e){return lt(e,0)}var QA=z({concat1d_:dR});function pR(e,t){return lt(e,t)}var Fu=z({concat2d_:pR});function cR(e,t){return lt(e,t)}var eb=z({concat3d_:cR});function hR(e,t){return lt(e,t)}var tb=z({concat4d_:hR});function fR(e,t,a,n,r="NHWC",s=[1,1],i){let o=R(e,"x","conv2d","float32"),l=R(t,"filter","conv2d","float32"),u=o,p=!1;o.rank===3&&(p=!0,u=Q(o,[1,o.shape[0],o.shape[1],o.shape[2]])),D(u.rank===4,()=>`Error in conv2d: input must be rank 4, but got rank ${u.rank}.`),D(l.rank===4,()=>`Error in conv2d: filter must be rank 4, but got rank ${l.rank}.`),Tn("conv2d",n,i);let c=r==="NHWC"?u.shape[3]:u.shape[1];D(c===l.shape[2],()=>`Error in conv2d: depth of input (${c}) must match input depth for filter ${l.shape[2]}.`),D(Tr(a,s),()=>`Error in conv2D: Either strides or dilations must be 1. Got strides ${a} and dilations '${s}'`),D(Hs(s),()=>"Error in conv2D: Dilated rates should be larger than 0."),D(Hs(a),()=>"Error in conv2D: Strides should be larger than 0.");let d={x:u,filter:l},h={strides:a,pad:n,dataFormat:r,dilations:s,dimRoundingMode:i},f=L.runKernel(fi,d,h);return p?Q(f,[f.shape[1],f.shape[2],f.shape[3]]):f}var Fp=z({conv2d_:fR});function mR(e,t,a,n,r="NWC",s=1,i){let o=R(e,"x","conv1d"),l=R(t,"filter","conv1d"),u=o,p=!1;o.rank===2&&(p=!0,u=Q(o,[1,o.shape[0],o.shape[1]])),D(u.rank===3,()=>`Error in conv1d: input must be rank 3, but got rank ${u.rank}.`),D(l.rank===3,()=>`Error in conv1d: filter must be rank 3, but got rank ${l.rank}.`),Tn("conv1d",n,i),D(u.shape[2]===l.shape[1],()=>`Error in conv1d: depth of input (${u.shape[2]}) must match input depth for filter ${l.shape[1]}.`),D(Tr(a,s),()=>`Error in conv1D: Either stride or dilation must be 1. Got stride ${a} and dilation '${s}'`),D(Hs(s),()=>"Error in conv1D: Dilated rates should be larger than 0."),D(Hs(a),()=>"Error in conv1D: Stride should be larger than 0."),D(r==="NWC",()=>`Error in conv1d: got dataFormat of ${r} but only NWC is currently supported.`);let c=Q(l,[1,l.shape[0],l.shape[1],l.shape[2]]),d=Q(u,[u.shape[0],1,u.shape[1],u.shape[2]]),h=Fp(d,c,[1,a],n,"NHWC",[1,s],i);return p?Q(h,[h.shape[2],h.shape[3]]):Q(h,[h.shape[0],h.shape[2],h.shape[3]])}var ab=z({conv1d_:mR});function gR(e,t,a,n,r,s="NHWC",i){D(e.length===t.rank,()=>`Length of inShape (${e.length}) and rank of dy (${t.rank}) must match`);let o=e,l=t,u=!1;t.rank===3&&(u=!0,l=Q(t,[1,t.shape[0],t.shape[1],t.shape[2]]),o=[1,e[0],e[1],e[2]]),D(o.length===4,()=>`Error in conv2dDerInput: inShape must be length 4, but got length ${o.length}.`),D(l.rank===4,()=>`Error in conv2dDerInput: dy must be rank 4, but got rank ${l.rank}`),D(a.rank===4,()=>`Error in conv2dDerInput: filter must be rank 4, but got rank ${a.rank}`);let p=s==="NHWC"?o[3]:o[1],c=s==="NHWC"?l.shape[3]:l.shape[1];D(p===a.shape[2],()=>`Error in conv2dDerInput: depth of input (${p}) must match input depth for filter ${a.shape[2]}.`),D(c===a.shape[3],()=>`Error in conv2dDerInput: depth of output (${c}) must match output depth for filter ${a.shape[3]}.`),Tn("conv2dDerInput",r,i);let d={dy:l,filter:a},h={strides:n,pad:r,dataFormat:s,dimRoundingMode:i,inputShape:o},f=L.runKernel(mi,d,h);return u?Q(f,[f.shape[1],f.shape[2],f.shape[3]]):f}var nb=z({conv2DBackpropInput_:gR});function yR(e,t,a,n,r,s){let i=R(e,"x","conv2dTranspose"),o=R(t,"filter","conv2dTranspose");return nb(a,i,o,n,r,"NHWC",s)}var rb=z({conv2dTranspose_:yR});function xR(e,t,a,n,r="NDHWC",s=[1,1,1]){let i=R(e,"x","conv3d"),o=R(t,"filter","conv3d"),l=i,u=!1;i.rank===4&&(u=!0,l=Q(i,[1,i.shape[0],i.shape[1],i.shape[2],i.shape[3]])),D(l.rank===5,()=>`Error in conv3d: input must be rank 5, but got rank ${l.rank}.`),D(o.rank===5,()=>`Error in conv3d: filter must be rank 5, but got rank ${o.rank}.`),D(l.shape[4]===o.shape[3],()=>`Error in conv3d: depth of input (${l.shape[4]}) must match input depth for filter ${o.shape[3]}.`),D(Tr(a,s),()=>`Error in conv3D: Either strides or dilations must be 1. Got strides ${a} and dilations '${s}'`),D(r==="NDHWC",()=>`Error in conv3d: got dataFormat of ${r} but only NDHWC is currently supported.`),D(Hs(s),()=>"Error in conv3D: Dilated rates should be larger than 0."),D(Hs(a),()=>"Error in conv3D: Strides should be larger than 0.");let p={x:l,filter:o},c={strides:a,pad:n,dataFormat:r,dilations:s},d=L.runKernel(gi,p,c);return u?Q(d,[d.shape[1],d.shape[2],d.shape[3],d.shape[4]]):d}var sb=z({conv3d_:xR});function AR(e,t,a,n,r){D(e.length===t.rank,()=>`Length of inShape (${e.length}) and rank of dy (${t.rank}) must match`);let s=e,i=t,o=!1;t.rank===4&&(o=!0,i=Q(t,[1,t.shape[0],t.shape[1],t.shape[2],t.shape[3]]),s=[1,e[0],e[1],e[2],e[3]]);let l=s[4],u=i.shape[4];D(s.length===5,()=>`Error in conv3dDerInput: inShape must be length 5, but got length ${s.length}.`),D(i.rank===5,()=>`Error in conv3dDerInput: dy must be rank 5, but got rank ${i.rank}`),D(a.rank===5,()=>`Error in conv3dDerInput: filter must be rank 5, but got rank ${a.rank}`),D(l===a.shape[3],()=>`Error in conv3dDerInput: depth of input (${l}) must match input depth for filter ${a.shape[3]}.`),D(u===a.shape[4],()=>`Error in conv3dDerInput: depth of output (${u}) must match output depth for filter ${a.shape[4]}.`);let p={dy:i,filter:a},c={pad:r,strides:n,inputShape:s},d=L.runKernel(yi,p,c);return o?Q(d,[d.shape[1],d.shape[2],d.shape[3],d.shape[4]]):d}var bR=z({conv3DBackpropInput_:AR});function vR(e,t,a,n,r){let s=R(e,"x","conv3dTranspose"),i=R(t,"filter","conv3dTranspose");return bR(a,s,i,n,r)}var ib=z({conv3dTranspose_:vR});function wR(e){let t={x:R(e,"x","cos","float32")};return L.runKernel(xi,t)}var ob=z({cos_:wR});function kR(e){let t={x:R(e,"x","cosh","float32")};return L.runKernel(Ai,t)}var lb=z({cosh_:kR});function IR(e,t=0,a=!1,n=!1){let r={x:R(e,"x","cumprod")},s={axis:t,exclusive:a,reverse:n};return L.runKernel(bi,r,s)}var ub=z({cumprod_:IR});function SR(e,t=0,a=!1,n=!1){let r={x:R(e,"x","cumsum")},s={axis:t,exclusive:a,reverse:n};return L.runKernel(vi,r,s)}var db=z({cumsum_:SR});function CR(e,t,a,n=!1){let r=R(e,"x","denseBincount"),s=R(t,"weights","denseBincount");D(r.dtype==="int32",()=>`Error in denseBincount: input dtype must be int32, but got ${r.dtype}`),D(r.rank<=2,()=>`Error in denseBincount: input must be at most rank 2, but got rank ${r.rank}.`),D(a>=0,()=>`size must be non-negative, but got ${a}.`),D(s.size===r.size||s.size===0,()=>`Error in denseBincount: weights must have the same shape as x or 0-length, but got x shape: ${r.shape}, weights shape: ${s.shape}.`);let i={x:r,weights:s},o={size:a,binaryOutput:n};return L.runKernel(su,i,o)}var pb=z({denseBincount_:CR});function TR(e,t,a="NHWC"){let n=R(e,"x","depthToSpace","float32"),r=a==="NHWC"?n.shape[1]:n.shape[2],s=a==="NHWC"?n.shape[2]:n.shape[3],i=a==="NHWC"?n.shape[3]:n.shape[1];D(t>1,()=>`blockSize should be > 1 for depthToSpace, but was: ${t}`),D(r*t>=0,()=>`Negative dimension size caused by overflow when multiplying
|
|
${r} and ${t} for depthToSpace with input shape
|
|
${n.shape}`),D(s*t>=0,()=>`Negative dimension size caused by overflow when multiplying
|
|
${s} and ${t} for depthToSpace with input shape
|
|
${n.shape}`),D(i%(t*t)===0,()=>`Dimension size must be evenly divisible by ${t*t} but is ${i} for depthToSpace with input shape ${n.shape}`);let o={x:n},l={blockSize:t,dataFormat:a};return L.runKernel(ki,o,l)}var cb=z({depthToSpace_:TR});function NR(e,t,a,n,r="NHWC",s=[1,1],i){let o=R(e,"x","depthwiseConv2d","float32"),l=R(t,"filter","depthwiseConv2d","float32"),u=o,p=!1;o.rank===3&&(p=!0,u=Q(o,[1,o.shape[0],o.shape[1],o.shape[2]])),D(u.rank===4,()=>`Error in depthwiseConv2d: input must be rank 4, but got rank ${u.rank}.`),D(l.rank===4,()=>`Error in depthwiseConv2d: filter must be rank 4, but got rank ${l.rank}.`);let c=r==="NHWC"?u.shape[3]:u.shape[1];D(c===l.shape[2],()=>`Error in depthwiseConv2d: number of input channels (${c}) must match the inChannels dimension in filter ${l.shape[2]}.`),Tn("depthwiseConv2d",n,i);let d={x:u,filter:l},h={strides:a,pad:n,dataFormat:r,dilations:s,dimRoundingMode:i},f=L.runKernel(Ii,d,h);return p?Q(f,[f.shape[1],f.shape[2],f.shape[3]]):f}var Mh=z({depthwiseConv2d_:NR});function RR(e){let t={x:R(e,"x","diag")};return L.runKernel(iu,t)}var hb=z({diag_:RR});function ER(e,t,a,n,r=[1,1],s="NHWC"){let i=R(e,"x","dilation2d"),o=R(t,"filter","dilation2d");D(i.rank===3||i.rank===4,()=>`Error in dilation2d: input must be rank 3 or 4, but got rank ${i.rank}.`),D(o.rank===3,()=>`Error in dilation2d: filter must be rank 3, but got rank ${o.rank}.`),D(s==="NHWC",()=>`Error in dilation2d: Only NHWC is currently supported, but got dataFormat of ${s}`);let l=i,u=!1;i.rank===3&&(l=Q(i,[1,i.shape[0],i.shape[1],i.shape[2]]),u=!0),D(l.shape[3]===o.shape[2],()=>`Error in dilation2d: input and filter must have the same depth: ${l.shape[3]} vs ${o.shape[2]}`);let p={x:l,filter:o},c={strides:a,pad:n,dilations:r},d=L.runKernel(Si,p,c);return u?Q(d,[d.shape[1],d.shape[2],d.shape[3]]):d}var fb=z({dilation2d_:ER}),Ho={};Ze(Ho,{assertAndGetBroadcastShape:()=>Bt,getBroadcastDims:()=>mb,getReductionAxes:()=>a3});function mb(e,t){let a=e.length,n=[];for(let r=0;r<a;r++){let s=a-1-r,i=e[s]||1;(t[t.length-1-r]||1)>1&&i===1&&n.unshift(s)}return n}function a3(e,t){let a=[];for(let n=0;n<t.length;n++){let r=e[e.length-n-1],s=t.length-n-1,i=t[s];(r==null||r===1&&i>1)&&a.unshift(s)}return a}function Bt(e,t){let a=Math.max(e.length,t.length),n=new Array(a);for(let r=0;r<a;r++){let s=e[e.length-r-1];s==null&&(s=1);let i=t[t.length-r-1];if(i==null&&(i=1),s===1)n[a-r-1]=i;else if(i===1)n[a-r-1]=s;else if(s!==i){let o=`Operands could not be broadcast together with shapes ${e} and ${t}.`;throw Error(o)}else n[a-r-1]=s}return n}function MR(e,t){let a=R(e,"a","equal","string_or_numeric"),n=R(t,"b","equal","string_or_numeric");[a,n]=Rt(a,n),Bt(a.shape,n.shape);let r={a,b:n};return L.runKernel(Ni,r)}var n3=z({equal_:MR});function FR(e,t,a){let n=R(t,"a","where"),r=R(a,"b","where"),s=R(e,"condition","where","bool"),i=Bt(Bt(s.shape,n.shape),r.shape),o=Fl(s,i),l=Fl(n,i),u=Fl(r,i),p={condition:o,t:l,e:u};return L.runKernel(ku,p)}var kr=z({where_:FR});function _R(e){let t={x:R(e,"x","zerosLike")};return L.runKernel(Mu,t)}var Qa=z({zerosLike_:_R});function $R(e,t){let a=R(e,"a","div"),n=R(t,"b","div");[a,n]=Rt(a,n);let r=ve(a,n),s=Qa(r),i=n3(n,s);return kr(i,s,r)}var gb=z({divNoNan_:$R});function DR(e,t){let a=R(e,"t1","dot"),n=R(t,"t2","dot");D((a.rank===1||a.rank===2)&&(n.rank===1||n.rank===2),()=>`Error in dot: inputs must all be rank 1 or 2, but got ranks ${a.rank} and ${n.rank}.`);let r=a.rank===1?a.size:a.shape[1],s=n.rank===1?n.size:n.shape[0];if(D(r===s,()=>`Error in dot: inner dimensions of inputs must match, but got ${r} and ${s}.`),a.rank===1&&n.rank===1){let i=Q(a,[1,-1]),o=Q(n,[-1,1]),l=pt(i,o);return Q(l,[])}else if(a.rank===1&&n.rank===2){let i=Q(a,[1,-1]),o=Q(n,[n.shape[0],n.shape[1]]),l=pt(i,o);return Q(l,[l.size])}else if(a.rank===2&&n.rank===1){let i=Q(n,[-1,1]),o=pt(a,i);return Q(o,[o.size])}else{let i=Q(n,[n.shape[0],n.shape[1]]);return pt(a,i)}}var yb=z({dot_:DR});function PR(e,...t){let a=t.map((r,s)=>R(r,`tensors${s}`,"einsum")),n={equation:e};return L.runKernel(ip,a,n)}var xb=z({einsum_:PR});function OR(e){let t={x:R(e,"x","elu","float32")};return L.runKernel(Ti,t)}var r3=z({elu_:OR});function zR(e,t){let a=R(e,"x","ensureShape","string_or_numeric");if(!Yx(a.shape,t))throw new Error(`EnsureShape: Shape of tensor ${a.shape} is not compatible with expected shape ${t}`);return e}var Ab=z({ensureShape_:zR});function LR(e){let t=R(e,"x","erf");D(t.dtype==="int32"||t.dtype==="float32",()=>"Input dtype must be `int32` or `float32`."),t.dtype==="int32"&&(t=Ke(t,"float32"));let a={x:t};return L.runKernel(lu,a)}var bb=z({erf_:LR});function s3(e,t){for(let a=0;a<e.length;++a)if(e[e.length-a-1]!==t-1-a)return!1;return!0}function vb(e,t,a){let n=e.length+t.length,r=[],s=0,i=0;for(let o=0;o<n;o++)a.indexOf(o)===-1?r.push(e[s++]):r.push(t[i++]);return r}function WR(e,t){let a=[],n=e.length;for(let s=0;s<n;s++)t.indexOf(s)===-1&&a.push(e[s]);let r=t.map(s=>e[s]);return[a,r]}function _p(e,t){let a=t.map(n=>1);return vb(e,a,t)}function BR(e,t,a){D(s3(t,a),()=>`${e} supports only inner-most axes for now. Got axes ${t} and rank-${a} input.`)}function VR(e,t){if(s3(e,t))return null;let a=[];for(let n=0;n<t;++n)e.indexOf(n)===-1&&a.push(n);return e.forEach(n=>a.push(n)),a}function UR(e){return e.map((t,a)=>[a,t]).sort((t,a)=>t[1]-a[1]).map(t=>t[0])}function GR(e,t){let a=[];for(let n=t-e;n<t;++n)a.push(n);return a}function HR(e,t=null,a=!1){let n={x:R(e,"x","max")},r={reductionIndices:t,keepDims:a};return L.runKernel(Ji,n,r)}var ga=z({max_:HR});function jR(e,t=null,a=!1){let n={x:R(e,"x","min")},r={axis:t,keepDims:a};return L.runKernel(ao,n,r)}var Jr=z({min_:jR});function qR(e,t){let a=R(e,"base","pow"),n=R(t,"exp","pow");[a,n]=Rt(a,n);let r={a,b:n};return L.runKernel(ho,r)}var zl=z({pow_:qR});function Ue(e,t){if((Zt(e)&&t!=="string"||Array.isArray(e))&&t!=="complex64")throw new Error("Error creating a new Scalar: value must be a primitive (number|boolean|string)");if(t==="string"&&Zt(e)&&!(e instanceof Uint8Array))throw new Error("When making a scalar from encoded string, the value must be `Uint8Array`.");return os(e,[],[],t)}function XR(e){let t={x:R(e,"x","sqrt","float32")};return L.runKernel(Fo,t)}var er=z({sqrt_:XR});function KR(e){let t=R(e,"x","square"),a={};return L.runKernel("Square",{x:t},a)}var Cn=z({square_:KR});function YR(e,t=null,a=!1){let n=R(e,"x","sum");n.dtype==="bool"&&(n=Ke(n,"int32"));let r={x:n},s={axis:t,keepDims:a};return L.runKernel(_o,r,s)}var ot=z({sum_:YR});function ZR(e,t="euclidean",a=null,n=!1){e=R(e,"x","norm");let r=wb(e,t,a),s=r.shape;if(n){let i=Jd(a,e.shape);s=_p(r.shape,i)}return Q(r,s)}function wb(e,t,a=null){if(e.rank===0)return Za(e);if(e.rank!==1&&a===null)return wb(Q(e,[-1]),t,a);if(e.rank===1||typeof a=="number"||Array.isArray(a)&&a.length===1){if(t===1)return ot(Za(e),a);if(t===1/0)return ga(Za(e),a);if(t===-1/0)return Jr(Za(e),a);if(t==="euclidean"||t===2)return er(ot(zl(Za(e),Ue(2,"int32")),a));throw new Error(`Error in norm: invalid ord value: ${t}`)}if(Array.isArray(a)&&a.length===2){if(t===1)return ga(ot(Za(e),a[0]),a[1]-1);if(t===1/0)return ga(ot(Za(e),a[1]),a[0]);if(t===-1/0)return Jr(ot(Za(e),a[1]),a[0]);if(t==="fro"||t==="euclidean")return er(ot(Cn(e),a));throw new Error(`Error in norm: invalid ord value: ${t}`)}throw new Error(`Error in norm: invalid axis: ${a}`)}var $p=z({norm_:ZR});function JR(e,t=null,a=!1){return $p(e,"euclidean",t,a)}var kb=z({euclideanNorm_:JR});function QR(e){let t={x:R(e,"x","exp")};return L.runKernel(Ri,t)}var Qr=z({exp_:QR});function eE(e,t=0){let a=R(e,"x","expandDims","string_or_numeric");D(t<=a.rank,()=>"Axis must be <= rank of the tensor");let n={input:a},r={dim:t};return L.runKernel(uu,n,r)}var Ht=z({expandDims_:eE});function tE(e){let t={x:R(e,"x","expm1")};return L.runKernel(Ei,t)}var Ib=z({expm1_:tE});function aE(e,t){let a=R(e,"x","tile","string_or_numeric");D(a.rank===t.length,()=>`Error in transpose: rank of input ${a.rank} must match length of reps ${t}.`);let n={x:a},r={reps:t};return L.runKernel(ss,n,r)}var jr=z({tile_:aE});function nE(e,t,a,n="float32"){t==null&&(t=e);let r=$e([e,t],n),s=e<=t?e:t;for(let o=0;o<s;++o)r.set(1,o,o);let i=Q(r.toTensor(),[e,t]);if(a==null)return i;if(a.length===1)return jr(Ht(i,0),[a[0],1,1]);if(a.length===2)return jr(Ht(Ht(i,0),0),[a[0],a[1],1,1]);if(a.length===3)return jr(Ht(Ht(Ht(i,0),0),0),[a[0],a[1],a[2],1,1]);throw new Error(`eye() currently supports only 1D and 2D batchShapes, but received ${a.length}D.`)}var i3=z({eye_:nE});function rE(e){let t={x:R(e,"x","floor","float32")};return L.runKernel(Fi,t)}var o3=z({floor_:rE});function sE(e,t,a=0,n=0){let r=R(e,"x","gather"),s=R(t,"indices","gather","int32"),i={x:r,indices:s},o={axis:a,batchDims:n};return L.runKernel(pu,i,o)}var l3=z({gather_:sE});function iE(e,t){let a=R(e,"a","greater","string_or_numeric"),n=R(t,"b","greater","string_or_numeric");[a,n]=Rt(a,n),Bt(a.shape,n.shape);let r={a,b:n};return L.runKernel(Pi,r)}var Dp=z({greater_:iE});function oE(e,t){let a=R(e,"a","greaterEqual","string_or_numeric"),n=R(t,"b","greaterEqual","string_or_numeric");[a,n]=Rt(a,n),Bt(a.shape,n.shape);let r={a,b:n};return L.runKernel(Oi,r)}var u3=z({greaterEqual_:oE});function lE(e){let t={input:R(e,"input","imag")};return L.runKernel(up,t)}var Pp=z({imag_:lE});function uE(e){let t={x:R(e,"x","isFinite")};return L.runKernel(Li,t)}var Sb=z({isFinite_:uE});function dE(e){let t={x:R(e,"x","isInf")};return L.runKernel(Wi,t)}var Cb=z({isInf_:dE});function pE(e){let t={x:R(e,"x","isNaN")};return L.runKernel(Bi,t)}var Tb=z({isNaN_:pE});function cE(e,t=.2){let a={x:R(e,"x","leakyRelu")},n={alpha:t};return L.runKernel(Vi,a,n)}var d3=z({leakyRelu_:cE});function hE(e,t){let a=R(e,"a","less","string_or_numeric"),n=R(t,"b","less","string_or_numeric");[a,n]=Rt(a,n),Bt(a.shape,n.shape);let r={a,b:n};return L.runKernel(Ui,r)}var ih=z({less_:hE});function fE(e,t){let a=R(e,"a","lessEqual","string_or_numeric"),n=R(t,"b","lessEqual","string_or_numeric");[a,n]=Rt(a,n),Bt(a.shape,n.shape);let r={a,b:n};return L.runKernel(Gi,r)}var Fh=z({lessEqual_:fE});function Nb(e,t,a){if(a<=0)throw new Error("The number of values should be positive.");let n={start:e,stop:t,num:a};return L.runKernel(Hi,{},n)}function mE(e,t=5,a=1,n=1,r=.5){let s=R(e,"x","localResponseNormalization");D(s.rank===4||s.rank===3,()=>`Error in localResponseNormalization: x must be rank 3 or 4 but got
|
|
rank ${s.rank}.`),D($l(t),()=>`Error in localResponseNormalization: depthRadius must be an integer but got depthRadius ${t}.`);let i=s,o=!1;s.rank===3&&(o=!0,i=Q(s,[1,s.shape[0],s.shape[1],s.shape[2]]));let l={x:i},u={depthRadius:t,bias:a,alpha:n,beta:r},p=L.runKernel(Zi,l,u);return o?Q(p,[p.shape[1],p.shape[2],p.shape[3]]):p}var Rb=z({localResponseNormalization_:mE});function gE(e){let t={x:R(e,"x","log","float32")};return L.runKernel(ji,t)}var Ll=z({log_:gE});function yE(e){let t={x:R(e,"x","log1p")};return L.runKernel(qi,t)}var p3=z({log1p_:yE});function xE(e){return D(qr(e),()=>"The f passed in grad(f) must be a function"),(t,a)=>{let n=R(t,"x","tf.grad","string_or_numeric"),r=a!=null?R(a,"dy","tf.grad"):null;return L.tidy(()=>{let{value:s,grads:i}=L.gradients(()=>e(n),[n],r);return r!=null&&Na(s.shape,r.shape,"The shape of dy passed in grad(f)(x, dy) must match the shape returned by f(x)"),_h(i),i[0]})}}function AE(e){return D(qr(e),()=>"The f passed in grads(f) must be a function"),(t,a)=>{D(Array.isArray(t),()=>"The args passed in grads(f)(args) must be an array of `Tensor`s or `TensorLike`s");let n=Dd(t,"args","tf.grads","string_or_numeric"),r=a!=null?R(a,"dy","tf.grads"):null;return L.tidy(()=>{let{value:s,grads:i}=L.gradients(()=>e(...n),n,r);return r!=null&&Na(s.shape,r.shape,"The shape of dy passed in grads(f)([x1,...], dy) must match the shape returned by f([x1,...])"),_h(i),i})}}function bE(e){return D(qr(e),()=>"The f passed in valueAndGrad(f) must be a function"),(t,a)=>{D(t instanceof gt,()=>"The x passed in valueAndGrad(f)(x) must be a tensor"),D(a==null||a instanceof gt,()=>"The dy passed in valueAndGrad(f)(x, dy) must be a tensor");let{grads:n,value:r}=L.gradients(()=>e(t),[t],a);return _h(n),{grad:n[0],value:r}}}function vE(e){return D(qr(e),()=>"The f passed in valueAndGrads(f) must be a function"),(t,a)=>{D(Array.isArray(t)&&t.every(r=>r instanceof gt),()=>"The args passed in valueAndGrads(f)(args) must be array of tensors"),D(a==null||a instanceof gt,()=>"The dy passed in valueAndGrads(f)(args, dy) must be a tensor");let n=L.gradients(()=>e(...t),t,a);return a!=null&&Na(n.value.shape,a.shape,"The shape of dy passed in valueAndGrads(f)([x1,...], dy) must match the shape returned by f([x1,...])"),_h(n.grads),n}}function Eb(e,t){D(qr(e),()=>"The f passed in variableGrads(f) must be a function"),D(t==null||Array.isArray(t)&&t.every(u=>u instanceof _d),()=>"The varList passed in variableGrads(f, varList) must be an array of variables");let a=t!=null;if(!a){t=[];for(let u in L.registeredVariables)t.push(L.registeredVariables[u])}let n=a?t.filter(u=>!u.trainable):null,r=t.length;t=t.filter(u=>u.trainable),D(t.length>0,()=>`variableGrads() expects at least one of the input variables to be trainable, but none of the ${r} variables is trainable.`);let s=!0,{value:i,grads:o}=L.gradients(e,t,null,s);D(o.some(u=>u!=null),()=>"Cannot find a connection between any variable and the result of the loss function y=f(x). Please make sure the operations that use variables are inside the function f passed to minimize()."),D(i.rank===0,()=>`The f passed in variableGrads(f) must return a scalar, but it returned a rank-${i.rank} tensor`);let l={};return t.forEach((u,p)=>{o[p]!=null&&(l[u.name]=o[p])}),n!=null&&n.forEach(u=>l[u.name]=null),{value:i,grads:l}}function tr(e){return L.customGrad(e)}function _h(e){if(e.filter(t=>t==null).length>0)throw new Error(`Cannot compute gradient of y=f(x) with respect to x. Make sure that
|
|
the f you passed encloses all operations that lead from x to y.`)}function wE(e){let t={x:R(e,"x","neg")};return L.runKernel(mu,t)}var Ln=z({neg_:wE});function kE(e){let t={x:R(e,"x","softplus")};return L.runKernel(Mo,t)}var c3=z({softplus_:kE});function IE(e){let t=R(e,"x","logSigmoid");return tr(a=>({value:Ln(c3(Ln(a))),gradFunc:n=>te(n,Wa(Ln(a)))}))(t)}var Mb=z({logSigmoid_:IE});function SE(e,t){let a=R(e,"a","sub"),n=R(t,"b","sub");[a,n]=Rt(a,n);let r={a,b:n};return L.runKernel(zo,r)}var ye=z({sub_:SE});function CE(e,t=-1){let a=R(e,"logits","logSoftmax");if(t===-1&&(t=a.rank-1),t!==a.rank-1)throw Error(`Log Softmax along a non-last dimension is not yet supported. Logits was rank ${a.rank} and axis was ${t}`);return tr((n,r)=>{let s=ga(n,t,!0),i=ye(n,s),o=ye(Ke(i,"float32"),Ll(ot(Qr(i),t,!0)));return r([o]),{value:o,gradFunc:(l,u)=>{let[p]=u,c=!0,d=Qr(p);return ye(l,te(ot(l,t,c),d))}}})(a)}var Fb=z({logSoftmax_:CE});function TE(e,t=null,a=!1){let n=R(e,"x","logSumExp"),r=Jd(t,n.shape),s=ga(n,r,!0),i=ye(n,s),o=Qr(i),l=ot(o,r),u=Ll(l),p=ke(Q(s,u.shape),u);if(a){let c=_p(p.shape,r);return Q(p,c)}return p}var h3=z({logSumExp_:TE});function NE(e,t){let a=R(e,"a","logicalAnd","bool"),n=R(t,"b","logicalAnd","bool");Bt(a.shape,n.shape);let r={a,b:n};return L.runKernel(Xi,r)}var Ld=z({logicalAnd_:NE});function RE(e){let t={x:R(e,"x","logicalNot","bool")};return L.runKernel(Ki,t)}var f3=z({logicalNot_:RE});function EE(e,t){let a=R(e,"a","logicalOr","bool"),n=R(t,"b","logicalOr","bool");Bt(a.shape,n.shape);let r={a,b:n};return L.runKernel(Yi,r)}var m3=z({logicalOr_:EE});function ME(e,t){let a=R(e,"a","logicalXor","bool"),n=R(t,"b","logicalXor","bool");return Bt(a.shape,n.shape),Ld(m3(e,t),f3(Ld(e,t)))}var _b=z({logicalXor_:ME}),Lc=2147483648;function FE(e,t,a="left"){let n=R(e,"sortedSequence","searchSorted"),r=R(t,"values","searchSorted"),s=n.shape[n.shape.length-1],i=r.shape[r.shape.length-1],o=Q(n,[-1,s]),l=Q(r,[-1,i]);if(o.rank<2)throw new Error("Sorted input argument must be at least 2-dimensional");if(o.shape[0]!==l.shape[0])throw new Error("Leading dimension of 'sortedSequence' and 'values' must match.");if(vt(l.shape)>=Lc)throw new Error(`values tensor size must less than ${Lc}`);if(o.shape[1]>=Lc)throw new Error(`trailing dim_size must less than ${Lc} for int32 output type, was ${o.shape[1]}`);let u={sortedSequence:o,values:l},p={side:a};return L.runKernel(Co,u,p)}var $h=z({searchSorted_:FE});function $b(e,t){return $h(e,t,"left")}function _E(e,t,a,n,r){let s=R(e,"x","maxPool"),i=1,o=s,l=!1;s.rank===3&&(l=!0,o=Q(s,[1,s.shape[0],s.shape[1],s.shape[2]])),D(o.rank===4,()=>`Error in maxPool: input must be rank 4 but got rank ${o.rank}.`),D(Tr(a,i),()=>`Error in maxPool: Either strides or dilations must be 1. Got strides ${a} and dilations '${i}'`),Tn("maxPool",n,r);let u={x:o},p={filterSize:t,strides:a,pad:n,dimRoundingMode:r},c=L.runKernel(eo,u,p);return l?Q(c,[c.shape[1],c.shape[2],c.shape[3]]):c}var g3=z({maxPool_:_E});function $E(e,t=[1,1,1],a,n,r,s="NDHWC"){let i=R(e,"x","maxPool3d"),o=i,l=!1;i.rank===4&&(l=!0,o=Q(i,[1,i.shape[0],i.shape[1],i.shape[2],i.shape[3]])),D(o.rank===5,()=>`Error in maxPool3d: x must be rank 5 but got rank ${o.rank}.`),D(s==="NDHWC",()=>`Error in maxPool3d: Only NDHWC is currently supported, but got dataFormat of ${s}`),Tn("maxPool3d",n,r);let u={x:o},p={filterSize:t,strides:a,pad:n,dimRoundingMode:r,dataFormat:s},c=L.runKernel(hu,u,p);return l?Q(c,[c.shape[1],c.shape[2],c.shape[3],c.shape[4]]):c}var Db=z({maxPool3d_:$E});function DE(e,t,a,n,r=!1){let s={x:R(e,"x","maxPoolWithArgmax")},i={filterSize:t,strides:a,pad:n,includeBatchInIndex:r},o=L.runKernel(pp,s,i);return{result:o[0],indexes:o[1]}}var Pb=z({maxPoolWithArgmax_:DE});function PE(e,t){let a=R(e,"a","maximum"),n=R(t,"b","maximum");[a,n]=Rt(a,n),a.dtype==="bool"&&(a=Ke(a,"int32"),n=Ke(n,"int32")),Bt(a.shape,n.shape);let r={a,b:n};return L.runKernel(Qi,r)}var y3=z({maximum_:PE});function OE(e,t=null,a=!1){let n={x:R(e,"x","mean")},r={axis:t,keepDims:a};return L.runKernel(to,n,r)}var Wd=z({mean_:OE});function gn(e,t="float32"){if(an(e),t==="complex64"){let n=gn(e,"float32"),r=gn(e,"float32");return Sr(n,r)}let a=vh(vt(e),t);return L.makeTensor(a,e,t)}function Ur(e,t="float32"){if(an(e),t==="complex64"){let n=Ur(e,"float32"),r=gn(e,"float32");return Sr(n,r)}let a=L1(vt(e),t);return L.makeTensor(a,e,t)}function Ob(e,t,{indexing:a="xy"}={}){if(a!=="xy"&&a!=="ij")throw new TypeError(`${a} is not a valid third argument to meshgrid`);if(e===void 0)return[];let n=R(e,"x","meshgrid",e instanceof gt?e.dtype:"float32");if(t===void 0)return[n];let r=R(t,"y","meshgrid",t instanceof gt?t.dtype:"float32"),s=vt(n.shape),i=vt(r.shape);return a==="xy"?(n=Q(n,[1,-1]),r=Q(r,[-1,1]),[pt(Ur([i,1],n.dtype),n),pt(r,Ur([1,s],r.dtype))]):(n=Q(n,[-1,1]),r=Q(r,[1,-1]),[pt(n,Ur([1,i],n.dtype)),pt(Ur([s,1],r.dtype),r)])}function zE(e,t){let a=R(e,"a","minimum"),n=R(t,"b","minimum");[a,n]=Rt(a,n),a.dtype==="bool"&&(a=Ke(a,"int32"),n=Ke(n,"int32")),Bt(a.shape,n.shape);let r={a,b:n};return L.runKernel(no,r)}var Bd=z({minimum_:zE});function LE(e,t,a){D(a==="reflect"||a==="symmetric",()=>`Invalid mode. Mode must be either reflect or symmetric. Got ${a}.`);let n=R(e,"x","mirrorPad");if(n.rank===0)throw new Error("mirrorPad(scalar) is not defined. Pass non-scalar to mirrorPad");D(t.length===n.rank,()=>`Padding doesn't match input. Must be ${n.rank}. Got ${t.length}.`);let r=a==="reflect"?1:0;for(let o=0;o<n.rank;o++)D(t[o].length===2,()=>"Invalid number of paddings. Must be length of 2 each."),D(t[o][0]>=0&&t[o][0]<=n.shape[o]-r&&t[o][1]>=0&&t[o][1]<=n.shape[o]-r,()=>`Padding in dimension ${o} cannot be greater than or equal to ${n.shape[o]-r} or less than 0 for input of shape ${n.shape}`);let s={paddings:t,mode:a},i={x:n};return L.runKernel(ro,i,s)}var zb=z({mirrorPad_:LE});function WE(e,t){let a=R(e,"a","mod"),n=R(t,"b","mod");[a,n]=Rt(a,n);let r={a,b:n};return L.runKernel(fu,r)}var _u=z({mod_:WE});function BE(e,t=null,a=!1){e=R(e,"x","moments");let n=Jd(t,e.shape),r=Wd(e,n,a),s=r.shape;a||(s=_p(r.shape,n));let i=Cn(ye(Ke(e,"float32"),Q(r,s))),o=Wd(i,n,a);return{mean:r,variance:o}}var Lb=z({moments_:BE});function VE(e,t,a,n){let r=R(t,"data","multiRNNCell"),s=Dd(a,"c","multiRNNCell"),i=Dd(n,"h","multiRNNCell"),o=r,l=[];for(let c=0;c<e.length;c++){let d=e[c](o,s[c],i[c]);l.push(d[0]),l.push(d[1]),o=d[1]}let u=[],p=[];for(let c=0;c<l.length;c+=2)u.push(l[c]),p.push(l[c+1]);return[u,p]}var Wb=z({multiRNNCell_:VE});function UE(e,t,a,n=!1){let r=R(e,"logits","multinomial"),s=r.size,i=r.rank;if(s<2)throw new Error(`Error in multinomial: you need at least 2 outcomes, but got ${s}.`);if(i>2)throw new Error(`Rank of probabilities must be 1 or 2, but is ${i}`);a=a||Math.random();let o={logits:i===1?Q(r,[1,-1]):r},l={numSamples:t,seed:a,normalized:n},u=L.runKernel(so,o,l);return i===1?Q(u,[u.size]):u}var Bb=z({multinomial_:UE});function GE(e,t){let a=R(e,"a","notEqual","string_or_numeric"),n=R(t,"b","notEqual","string_or_numeric");[a,n]=Rt(a,n),Bt(a.shape,n.shape);let r={a,b:n};return L.runKernel(oo,r)}var x3=z({notEqual_:GE});function HE(e,t,a=1,n=0,r="int32"){if(t<2)throw new Error(`Error in oneHot: depth must be >=2, but it is ${t}`);let s={indices:R(e,"indices","oneHot","int32")},i={dtype:r,depth:t,onValue:a,offValue:n};return L.runKernel(po,s,i)}var oh=z({oneHot_:HE});function jE(e){let t={x:R(e,"x","onesLike")};return L.runKernel(yu,t)}var Vb=z({onesLike_:jE});function qE(e,t){let a=R(e,"v1","outerProduct"),n=R(t,"v2","outerProduct");D(a.rank===1&&n.rank===1,()=>`Error in outerProduct: inputs must be rank 1, but got ranks ${a.rank} and ${n.rank}.`);let r=Q(a,[-1,1]),s=Q(n,[1,-1]);return pt(r,s)}var Ub=z({outerProduct_:qE});function XE(e,t,a=0){let n=R(e,"x","pad");if(n.rank===0)throw new Error("pad(scalar) is not defined. Pass non-scalar to pad");let r={paddings:t,constantValue:a},s={x:n};return L.runKernel(co,s,r)}var ir=z({pad_:XE});function KE(e,t,a=0){return D(t.length===2,()=>"Invalid number of paddings. Must be length of 2."),ir(e,[t],a)}var Gb=z({pad1d_:KE});function YE(e,t,a=0){return D(t.length===2&&t[0].length===2&&t[1].length===2,()=>"Invalid number of paddings. Must be length of 2 each."),ir(e,t,a)}var Hb=z({pad2d_:YE});function ZE(e,t,a=0){return D(t.length===3&&t[0].length===2&&t[1].length===2&&t[2].length===2,()=>"Invalid number of paddings. Must be length of 2 each."),ir(e,t,a)}var jb=z({pad3d_:ZE});function JE(e,t,a=0){return D(t.length===4&&t[0].length===2&&t[1].length===2&&t[2].length===2&&t[3].length===2,()=>"Invalid number of paddings. Must be length of 2 each."),ir(e,t,a)}var qb=z({pad4d_:JE});function QE(e,t,a){let n=R(e,"x","spaceToBatchND");D(n.rank>=1+t.length,()=>`input rank ${n.rank} should be > than [blockShape] ${t.length}`),D(a.length===t.length,()=>`paddings.shape[0] ${a.length} must be equal to [blockShape] ${t.length}`),D(n.shape.reduce((i,o,l)=>l>0&&l<=t.length?i&&(o+a[l-1][0]+a[l-1][1])%t[l-1]===0:i,!0),()=>`input spatial dimensions ${n.shape.slice(1)} with paddings ${a.toString()} must be divisible by blockShapes ${t.toString()}`);let r={x:n},s={blockShape:t,paddings:a};return L.runKernel(Cu,r,s)}var A3=z({spaceToBatchND_:QE});function eM(e,t,a,n,r,s,i){r==null&&(r=[1,1]),s==null&&(s=1),n===0&&(n="valid");let o=R(e,"x","maxPool"),l=o,u=!1;o.rank===3&&(u=!0,l=Q(o,[1,o.shape[0],o.shape[1],o.shape[2]])),D(Tr(s,r),()=>`Error in pool: Either strides or dilations must be 1. Got strides ${s} and dilations '${r}'`);let p=BA(l.shape,t,s,r,n),c=[p.dilationHeight,p.dilationWidth],d;n==="same"?d=aM([p.filterHeight,p.filterWidth],c):d=[[0,0],[0,0]];let h=c[0]===1&&c[1]===1,[f,m]=tM([p.inHeight,p.inWidth],c,d),g=h?n:"valid",y=h?l:A3(l,c,f),x=(a==="avg"?()=>Q1(y,t,s,g,i):()=>g3(y,t,s,g,i))(),A=h?x:e3(x,c,m);return u?Q(A,[A.shape[1],A.shape[2],A.shape[3]]):A}function tM(e,t,a){let n=a.map(p=>p[0]),r=a.map(p=>p[1]),s=e.concat(n,r),i=t.map((p,c)=>(p-s[c]%p)%p),o=r.map((p,c)=>p+i[c]),l=t.map((p,c)=>[n[c],o[c]]),u=t.map((p,c)=>[0,i[c]]);return[l,u]}function aM(e,t){let a=e.map((s,i)=>s+(s-1)*(t[i]-1)).map(s=>s-1),n=a.map(s=>Math.floor(s/2)),r=a.map((s,i)=>s-n[i]);return a.map((s,i)=>[n[i],r[i]])}var Xb=z({pool_:eM});function nM(e,t){let a=R(e,"x","prelu"),n=R(t,"alpha","prelu"),r={x:a,alpha:n};return L.runKernel(fo,r)}var b3=z({prelu_:nM});function rM(e,t=null,a=!1){let n=R(e,"x","prod");n.dtype==="bool"&&(n=Ke(n,"int32"));let r={x:n},s={axis:t,keepDims:a};return L.runKernel(mo,r,s)}var Kb=z({prod_:rM});function sM(e,t,a,n){let r=e.map((p,c)=>R(p,`tensors${c}`,"raggedGather","int32")),s=R(t,"paramsDenseValues","raggedGather"),i=R(a,"indices","raggedGather","int32"),o={paramsNestedSplits:r,paramsDenseValues:s,indices:i},l={outputRaggedRank:n},u=L.runKernel(Ch,o,l);return{outputNestedSplits:u.slice(0,u.length-1),outputDenseValues:u[u.length-1]}}var Yb=z({raggedGather_:sM});function iM(e,t,a){let n=R(e,"starts","raggedRange"),r=R(t,"limits","raggedRange",n.dtype),s=R(a,"deltas","raggedRange",n.dtype),i={starts:n,limits:r,deltas:s},o=L.runKernel(Th,i);return{rtNestedSplits:o[0],rtDenseValues:o[1]}}var Zb=z({raggedRange_:iM});function oM(e,t,a,n,r){let s=R(e,"shape","raggedTensorToTensor","int32"),i=R(t,"values","raggedTensorToTensor"),o=R(a,"defaultValue","raggedTensorToTensor",i.dtype),l=n.map((c,d)=>R(c,`tensors${d}`,"raggedTensorToTensor","int32")),u={shape:s,values:i,defaultValue:o,rowPartitionTensors:l},p={rowPartitionTypes:r};return L.runKernel(Nh,u,p)}var Jb=z({raggedTensorToTensor_:oM});function lM(e,t,a){an(e);let n=vt(e),r=null;if(a==null||a==="float32")r=new Float32Array(n);else if(a==="int32")r=new Int32Array(n);else if(a==="bool")r=new Uint8Array(n);else throw new Error(`Unknown data type ${a}`);for(let s=0;s<n;s++)r[s]=t();return L.makeTensor(r,e,a)}var Qb=z({rand_:lM}),v3=Xl(jx()),e4={};Ze(e4,{TEST_EPSILON_FLOAT16:()=>t4,createVideoElement:()=>gM,encodeStrings:()=>a4,expectArrayBuffersEqual:()=>mM,expectArraysClose:()=>dM,expectArraysEqual:()=>cM,expectNumbersClose:()=>hM,expectPromiseToFail:()=>pM,expectValuesInRange:()=>fM,play:()=>yM,testEpsilon:()=>w3});var uM=.001,t4=.1;function dM(e,t,a){return a==null&&(a=w3()),n1(e,t,(n,r)=>k3(n,r,a))}function w3(){return L.backend.floatPrecision()===32?uM:t4}function n1(e,t,a){let n=!0;if((Zt(e)||Zt(t))&&(n=!1),Zt(e)&&Zt(t)&&(n=!0),n){let i=e.constructor.name,o=t.constructor.name;if(i!==o)throw new Error(`Arrays are of different type. Actual: ${i}. Expected: ${o}`)}if(Array.isArray(e)&&Array.isArray(t)){let i=Qn(e),o=Qn(t);if(!Cr(i,o))throw new Error(`Arrays have different shapes. Actual: [${i}]. Expected: [${o}]`)}let r=Zt(e)?e:Zr(e),s=Zt(t)?t:Zr(t);if(r.length!==s.length)throw new Error(`Arrays have different lengths actual: ${r.length} vs expected: ${s.length}.
|
|
Actual: ${r}.
|
|
Expected: ${s}.`);for(let i=0;i<s.length;++i){let o=r[i],l=s[i];if(!a(o,l))throw new Error(`Arrays differ: actual[${i}] = ${o}, expected[${i}] = ${l}.
|
|
Actual: ${r}.
|
|
Expected: ${s}.`)}typeof expect!="undefined"&&expect().nothing()}function pM(e,t){e().then(()=>t.fail(),()=>t()),typeof expect!="undefined"&&expect().nothing()}function cM(e,t){let a=typeof t=="string"||typeof t=="number"||typeof t=="boolean"?[t]:t;return Br(e)||Br(e[0])||Br(t)||Br(t[0])?n1(e,a,(n,r)=>n==r):n1(e,t,(n,r)=>k3(n,r,0))}function hM(e,t,a){if(a==null&&(a=w3()),!k3(e,t,a))throw new Error(`Numbers differ: actual === ${e}, expected === ${t}`);typeof expect!="undefined"&&expect().nothing()}function k3(e,t,a){return!isFinite(e)&&!isFinite(t)?!0:!(isNaN(e)||isNaN(t)||Math.abs(e-t)>a)}function fM(e,t,a){for(let n=0;n<e.length;n++)if(e[n]<t||e[n]>a)throw new Error(`Value out of range:${e[n]} low: ${t}, high: ${a}`)}function mM(e,t){let a=new Float32Array(e),n=new Float32Array(t);if(a.length!==n.length)throw new Error(`Expected ArrayBuffer to be of length ${n.length}, but it was ${a.length}`);for(let r=0;r<n.length;r++)if(a[r]!==n[r])throw new Error(`Expected ArrayBuffer value at ${r} to be ${n[r]} but got ${a[r]} instead`)}function a4(e){for(let t=0;t<e.length;t++){let a=e[t];Array.isArray(a)?a4(a):e[t]=kp(a)}return e}function gM(e){let t=document.createElement("video");return"playsInline"in t&&(t.playsInline=!0),t.muted=!0,t.loop=!0,t.style.position="fixed",t.style.left="0px",t.style.top="0px",t.preload="auto",t.appendChild(e),new Promise(a=>{t.addEventListener("loadeddata",n=>a(t)),t.load()})}async function yM(e){await e.play(),"requestVideoFrameCallback"in e&&await new Promise(t=>{e.requestVideoFrameCallback(t)})}var I3=class{constructor(e,t,a,n,r){this.mean=e,this.stdDev=t,this.dtype=a,this.nextVal=NaN,this.truncated=n,this.truncated&&(this.upper=this.mean+this.stdDev*2,this.lower=this.mean-this.stdDev*2);let s=r||Math.random();this.random=v3.alea(s.toString())}nextValue(){if(!isNaN(this.nextVal)){let n=this.nextVal;return this.nextVal=NaN,n}let e,t,a=!1;for(;!a;){let n,r,s;do n=2*this.random()-1,r=2*this.random()-1,s=n*n+r*r;while(s>=1||s===0);let i=Math.sqrt(-2*Math.log(s)/s);e=this.mean+this.stdDev*n*i,t=this.mean+this.stdDev*r*i,(!this.truncated||this.isValidTruncated(e))&&(a=!0)}return(!this.truncated||this.isValidTruncated(t))&&(this.nextVal=this.convertValue(t)),this.convertValue(e)}convertValue(e){return this.dtype==null||this.dtype==="float32"?e:Math.round(e)}isValidTruncated(e){return e<=this.upper&&e>=this.lower}},xM=class{constructor(e,t,a,n){this.alpha=e,this.beta=1/t,this.dtype=a;let r=n||Math.random();this.randu=v3.alea(r.toString()),this.randn=new I3(0,1,a,!1,this.randu()),e<1?this.d=e+2/3:this.d=e-1/3,this.c=1/Math.sqrt(9*this.d)}nextValue(){let e,t,a,n,r,s;for(;;){do n=this.randn.nextValue(),s=1+this.c*n;while(s<=0);if(s*=s*s,e=n*n,t=1-.331*e*e,a=.5*e+this.d*(1-s+Math.log(s)),r=this.randu(),r<t||Math.log(r)<a)break}return s=1/this.beta*this.d*s,this.alpha<1&&(s*=Math.pow(this.randu(),1/this.alpha)),this.convertValue(s)}convertValue(e){return this.dtype==="float32"?e:Math.round(e)}},AM=class{constructor(e=0,t=1,a,n){if(this.canReturnFloat=()=>this.dtype==null||this.dtype==="float32",this.min=e,this.range=t-e,this.dtype=a,n==null&&(n=Math.random()),typeof n=="number"&&(n=n.toString()),!this.canReturnFloat()&&this.range<=1)throw new Error(`The difference between ${e} - ${t} <= 1 and dtype is not float`);this.random=v3.alea(n)}convertValue(e){return this.canReturnFloat()?e:Math.round(e)}nextValue(){return this.convertValue(this.min+this.range*this.random())}};function bM(e,t,a=1,n="float32",r){if(an(e),a==null&&(a=1),n==null&&(n="float32"),n!=="float32"&&n!=="int32")throw new Error(`Unsupported data type ${n}`);let s=new xM(t,a,n,r),i=$e(e,n);for(let o=0;o<i.values.length;o++)i.values[o]=s.nextValue();return i.toTensor()}var n4=z({randomGamma_:bM});function vM(e,t=0,a=1,n,r){if(an(e),n!=null&&n==="bool")throw new Error(`Unsupported data type ${n}`);let s=new I3(t,a,n,!1,r),i=$e(e,n);for(let o=0;o<i.values.length;o++)i.values[o]=s.nextValue();return i.toTensor()}var S3=z({randomNormal_:vM});function wM(e,t,a){if(t!=null&&t==="bool")throw new Error(`Unsupported data type ${t}`);return S3(e,0,1,t,a)}var r4=z({randomStandardNormal_:wM});function kM(e,t=0,a=1,n="float32",r){an(e);let s=$e(e,n),i=new AM(t,a,null,r);for(let o=0;o<s.values.length;o++)s.values[o]=i.nextValue();return s.toTensor()}var Dh=z({randomUniform_:kM});function IM(e,t,a,n){return Dh(e,t,a,"int32",n)}var s4=z({randomUniformInt_:IM});function Wl(e,t,a=1,n="float32"){if(a===0)throw new Error("Cannot have a step of zero");let r={start:e,stop:t,step:a,dtype:n};return L.runKernel(Au,{},r)}function SM(e){let t={input:R(e,"input","real")};return L.runKernel(cp,t)}var Bl=z({real_:SM});function CM(e){let t={x:R(e,"x","reciprocal")};return L.runKernel(go,t)}var i4=z({reciprocal_:CM});function TM(e){let t={x:R(e,"x","relu")};return L.runKernel(yo,t)}var Op=z({relu_:TM});function NM(e){let t={x:R(e,"x","relu6")};return L.runKernel(bo,t)}var C3=z({relu6_:NM});function RM(e,t){let a={x:R(e,"x","reverse")},n={dims:t};return L.runKernel(vo,a,n)}var es=z({reverse_:RM});function EM(e){let t=R(e,"x","reverse");return D(t.rank===1,()=>`Error in reverse1D: x must be rank 1 but got rank ${t.rank}.`),es(t,0)}var o4=z({reverse1d_:EM});function MM(e,t){let a=R(e,"x","reverse");return D(a.rank===2,()=>`Error in reverse2D: x must be rank 2 but got rank ${a.rank}.`),es(a,t)}var l4=z({reverse2d_:MM});function FM(e,t){let a=R(e,"x","reverse");return D(a.rank===3,()=>`Error in reverse3D: x must be rank 3 but got rank ${a.rank}.`),es(a,t)}var u4=z({reverse3d_:FM});function _M(e,t){let a=R(e,"x","reverse");return D(a.rank===4,()=>`Error in reverse4D: x must be rank 4 but got rank ${a.rank}.`),es(a,t)}var d4=z({reverse4d_:_M});function $M(e){let t={x:R(e,"x","round")};return L.runKernel(wo,t)}var T3=z({round_:$M});function DM(e){let t={x:R(e,"x","rsqrt","float32")};return L.runKernel(ko,t)}var p4=z({rsqrt_:DM});function PM(e){let t={x:R(e,"x","selu")};return L.runKernel(To,t)}var c4=z({selu_:PM});function OM(e,t,a,n,r,s=[1,1],i="NHWC"){let o=R(e,"x","separableConv2d"),l=R(t,"depthwiseFilter","separableConv2d"),u=R(a,"pointwiseFilter","separableConv2d"),p=o,c=!1;if(o.rank===3&&(c=!0,p=Q(o,[1,o.shape[0],o.shape[1],o.shape[2]])),i==="NCHW")throw new Error("separableConv2d currently does not support dataFormat NCHW; only NHWC is supported");D(p.rank===4,()=>`Error in separableConv2d: input must be rank 4, but got rank ${p.rank}.`),D(l.rank===4,()=>`Error in separableConv2d: depthwise filter must be rank 4, but got rank ${l.rank}.`),D(u.rank===4,()=>`Error in separableConv2d: pointwise filter must be rank 4, but got rank ${l.rank}.`),D(u.shape[0]===1,()=>`Error in separableConv2d: the first dimension of pointwise filter must be 1, but got ${u.shape[0]}.`),D(u.shape[1]===1,()=>`Error in separableConv2d: the second dimension of pointwise filter must be 1, but got ${u.shape[1]}.`);let d=l.shape[2],h=l.shape[3];D(u.shape[2]===d*h,()=>`Error in separableConv2d: the third dimension of pointwise filter must be ${d*h}, but got ${u.shape[2]}.`);let f=Mh(p,l,n,r,i,s),m=Fp(f,u,1,"valid",i);return c?Q(m,[m.shape[1],m.shape[2],m.shape[3]]):m}var h4=z({separableConv2d_:OM});async function zM(e,t){let a=R(e,"x","setdiff1d"),n=R(t,"y","setdiff1d");D(a.dtype===n.dtype,()=>`x and y should have the same dtype, but got x (${a.dtype}) and y (${n.dtype}).`),D(a.rank===1,()=>`x should be 1D tensor, but got x (${a.shape}).`),D(n.rank===1,()=>`y should be 1D tensor, but got y (${n.shape}).`);let r=await a.data(),s=await n.data(),i=new Set(s),o=0;for(let p=0;p<r.length;p++)i.has(r[p])||o++;let l=new Wt([o],a.dtype),u=new Wt([o],"int32");for(let p=0,c=0;p<r.length;p++)i.has(r[p])||(l.values[c]=r[p],u.values[c]=p,c++);return[l.toTensor(),u.toTensor()]}var f4=zM;function LM(e){let t={x:R(e,"x","sign")};return L.runKernel(Ro,t)}var m4=z({sign_:LM});function WM(e){let t={x:R(e,"x","sin","float32")};return L.runKernel(No,t)}var g4=z({sin_:WM});function BM(e){let t={x:R(e,"x","sinh")};return L.runKernel(Su,t)}var y4=z({sinh_:BM});function VM(e,t,a){let n=R(e,"x","slice1d");return D(n.rank===1,()=>`slice1d expects a rank-1 tensor, but got a rank-${n.rank} tensor`),De(n,[t],[a])}var x4=z({slice1d_:VM});function UM(e,t,a){let n=R(e,"x","slice2d");return D(n.rank===2,()=>`slice2d expects a rank-2 tensor, but got a rank-${n.rank} tensor`),De(n,t,a)}var A4=z({slice2d_:UM});function GM(e,t,a){let n=R(e,"x","slice3d");return D(n.rank===3,()=>`slice3d expects a rank-3 tensor, but got a rank-${n.rank} tensor`),De(n,t,a)}var zp=z({slice3d_:GM});function HM(e,t,a){let n=R(e,"x","slice4d");return D(n.rank===4,()=>`slice4d expects a rank-4 tensor, but got a rank-${n.rank} tensor`),De(n,t,a)}var Ph=z({slice4d_:HM});function jM(e,t=-1){let a=R(e,"logits","softmax","float32");if(t===-1&&(t=a.rank-1),t!==a.rank-1)throw Error(`Softmax along a non-last dimension is not yet supported. Logits was rank ${a.rank} and dim was ${t}`);let n={logits:a},r={dim:t};return L.runKernel($o,n,r)}var Oh=z({softmax_:jM});function qM(e){D(e.dtype==="complex64",()=>`The dtype for tf.spectral.fft() must be complex64 but got ${e.dtype}.`);let t={input:e};return L.runKernel(op,t)}var zh=z({fft_:qM});function XM(e){D(e.dtype==="complex64",()=>`The dtype for tf.spectral.ifft() must be complex64 but got ${e.dtype}.`);let t={input:e};return L.runKernel(lp,t)}var Vd=z({ifft_:XM});function KM(e){let t=e.shape[e.shape.length-1],a=e.size/t,n;if(t<=2){let r=Q(e,[a,t]);n=Vd(r)}else{let r=[a,2*(t-1)],s=Q(Bl(e),[a,t]),i=Q(Pp(e),[a,t]),o=es(De(s,[0,1],[a,t-2]),1),l=te(es(De(i,[0,1],[a,t-2]),1),Ue(-1)),u=lt([s,o],1),p=lt([i,l],1),c=Q(Sr(u,p),[r[0],r[1]]);n=Vd(c)}if(n=Bl(n),e.rank===3&&e.shape[0]!==0){let r=n,s=e.shape[0];n=Q(n,[s,n.shape[0]/s,n.shape[1]]),r.dispose()}return n}var N3=z({irfft_:KM});function YM(e,t,a=0){let n={x:R(e,"x","split")},r={numOrSizeSplits:t,axis:a};return L.runKernel(Tu,n,r)}var Ca=z({split_:YM});function ZM(e,t){D(e.dtype==="float32",()=>`The dtype for rfft() must be real value but got ${e.dtype}`);let a=e.shape[e.shape.length-1],n=e.size/a,r;if(t!=null&&t<a){let f=e.shape.map(g=>0),m=e.shape.map(g=>g);m[e.shape.length-1]=t,r=De(e,f,m),a=t}else if(t!=null&&t>a){let f=e.shape.map(m=>m);f[e.shape.length-1]=t-a,r=lt([e,gn(f)],e.shape.length-1),a=t}else r=e;let s=Qa(r),i=Q(Sr(r,s),[n,a]),o=zh(i),l=Math.floor(a/2)+1,u=Bl(o),p=Pp(o),c=Ca(u,[l,a-l],u.shape.length-1),d=Ca(p,[l,a-l],p.shape.length-1),h=r.shape.slice();return h[r.shape.length-1]=l,Q(Sr(c[0],d[0]),h)}var Lh=z({rfft_:ZM});function JM(e,t){let a=R(e,"a","squaredDifference"),n=R(t,"b","squaredDifference");[a,n]=Rt(a,n),Bt(a.shape,n.shape);let r={a,b:n},s={};return L.runKernel(Po,r,s)}var R3=z({squaredDifference_:JM});function QM(e,t){let a=R(e,"x","squeeze","string_or_numeric");return Q(a,Zx(a.shape,t).newShape)}var Oe=z({squeeze_:QM});function eF(e,t=0){let a=Dd(e,"tensors","stack","string_or_numeric");D(a.length>=1,()=>"Pass at least one tensor to tf.stack"),a.length>0&&D(t<=a[0].rank,()=>"Axis must be <= rank of the tensor");let n=a,r={axis:t};return L.runKernel(xu,n,r)}var ca=z({stack_:eF});function tF(e,t=0){let a={x:R(e,"x","step")},n={alpha:t};return L.runKernel(is,a,n)}var E3=z({step_:tF});function aF(e,t,a,n,r=0,s=0,i=0,o=0,l=0){let u={x:R(e,"x","stridedSlice","string_or_numeric")},p={begin:t,end:a,strides:n,beginMask:r,endMask:s,ellipsisMask:i,newAxisMask:o,shrinkAxisMask:l};return L.runKernel(Oo,u,p)}var b4=z({stridedSlice_:aF});function nF(e){let t={x:R(e,"x","tan","float32")};return L.runKernel(Lo,t)}var v4=z({tan_:nF});function jt(e,t){Js(e);let a=Qn(e,t);if(a.length!==1)throw new Error("tensor1d() requires values to be a flat/TypedArray");return os(e,null,a,t)}function Zn(e,t,a){if(Js(e),t!=null&&t.length!==2)throw new Error("tensor2d() requires shape to have two numbers");let n=Qn(e,a);if(n.length!==2&&n.length!==1)throw new Error("tensor2d() requires values to be number[][] or flat/TypedArray");if(n.length===1&&t==null)throw new Error("tensor2d() requires shape to be provided when `values` are a flat/TypedArray");return os(e,t,n,a)}function M3(e,t,a){if(Js(e),t!=null&&t.length!==3)throw new Error("tensor3d() requires shape to have three numbers");let n=Qn(e,a);if(n.length!==3&&n.length!==1)throw new Error("tensor3d() requires values to be number[][][] or flat/TypedArray");if(n.length===1&&t==null)throw new Error("tensor3d() requires shape to be provided when `values` are a flat array");return os(e,t,n,a)}function w4(e,t,a){if(Js(e),t!=null&&t.length!==4)throw new Error("tensor4d() requires shape to have four numbers");let n=Qn(e,a);if(n.length!==4&&n.length!==1)throw new Error("tensor4d() requires values to be number[][][][] or flat/TypedArray");if(n.length===1&&t==null)throw new Error("tensor4d() requires shape to be provided when `values` are a flat array");return os(e,t,n,a)}function k4(e,t,a){if(Js(e),t!=null&&t.length!==5)throw new Error("tensor5d() requires shape to have five numbers");let n=Qn(e,a);if(n.length!==5&&n.length!==1)throw new Error("tensor5d() requires values to be number[][][][][] or flat/TypedArray");if(n.length===1&&t==null)throw new Error("tensor5d() requires shape to be provided when `values` are a flat array");return os(e,t,n,a)}function I4(e,t,a){if(Js(e),t!=null&&t.length!==6)throw new Error("tensor6d() requires shape to have six numbers");let n=Qn(e,a);if(n.length!==6&&n.length!==1)throw new Error("tensor6d() requires values to be number[][][][][][] or flat/TypedArray");if(n.length===1&&t==null)throw new Error("tensor6d() requires shape to be provided when `values` are a flat array");return t=t||n,os(e,t,n,a)}var Wh={};Ze(Wh,{calculateShapes:()=>S4,validateInput:()=>Bh,validateUpdateShape:()=>F3});function F3(e,t,a){let n=t.rank>1?t.shape[t.rank-1]:1,r=t.rank>1?t.rank-1:1,s=`Must have updates.shape = indices.shape[:batchDim] + shape[sliceDim:], got updates.shape: ${a.shape}, indices.shape: ${t.shape}, shape: ${e}, sliceDim: ${n}, and batchDim: ${r}.`;if(a.rank<r)throw new Error(s+` update.rank < ${r}. `);if(e.length<n+(a.rank-r))throw new Error(s+` Output shape length < ${n+(a.rank-r)}`);if(a.rank!==r+e.length-n)throw new Error(s+` update.rank != ${r+e.length-n}`);for(let i=0;i<r;++i)if(a.shape[i]!==t.shape[i])throw new Error(s+` updates.shape[${i}] (${a.shape[i]}) != indices.shape[${i}] (${t.shape[i]}).`);for(let i=0;i<a.rank-r;++i)if(a.shape[i+r]!==e[i+n])throw new Error(s+` updates.shape[${i+r}] (${a.shape[i+r]}) != shape[${i+r}] (${e[i+r]})`)}function Bh(e,t,a){if(t.rank<1)throw new Error(`tf.scatterND() expects the indices to be rank 1 or higher, but the rank was ${t.rank}.`);if(e.rank<1)throw new Error(`tf.scatterND() expects the updates to be rank 1 or higher, but the rank was ${e.rank}.`);if(t.dtype!=="int32")throw new Error(`The dtype of 'indices' should be int32, but got dtype: ${t.dtype}`);if(a.length<1)throw new Error(`Output rank must be greater or equal to 1, but got shape: ${a}`);if(a.length===0){if(t.size===0)throw new Error(`Indices specified for empty output. indices shape: ${t.shape}`);if(e.size===0)throw new Error(`Updates specified for empty output. updates shape: ${e.shape}`)}F3(a,t,e)}function S4(e,t,a){let n=t.shape.length,r=n>1?t.shape[n-1]:1,s=a.length,i=1;for(let c=r;c<s;++c)i*=a[c];let o=r<1?1:r,l=vt(t.shape)/o,u=[...Yl(a.slice(0,r)),1],p=vt(a);return{sliceRank:r,numUpdates:l,sliceSize:i,strides:u,outputSize:p}}function rF(e,t,a){let n=R(e,"tensor","tensorScatterupdate"),r=R(t,"indices","tensorScatterupdate","int32"),s=R(a,"updates","tensorScatterupdate");if(Bh(s,r,n.shape),n.dtype!==s.dtype)throw new Error(`tensor and updates must have the same dtype, instead they are ${n.dtype} and ${s.dtype}.`);let i={tensor:n,indices:r,updates:s},o={};return L.runKernel(So,i,o)}var C4=z({tensorScatterUpdate_:rF});function sF(e,t=1,a=!0){let n=R(e,"x","topk");if(n.rank===0)throw new Error("topk() expects the input to be of rank 1 or higher");let r=n.shape[n.shape.length-1];if(t<0)throw new Error(`'k' passed to topk() must be >= 0 but got ${t}`);if(t>r)throw new Error(`'k' passed to topk() must be <= the last dimension (${r}) but got ${t}`);let s={x:n},i={k:t,sorted:a},[o,l]=L.runKernel(Bo,s,i);return{values:o,indices:l}}var T4=z({topk_:sF});function iF(e,t=0,a=1,n,r){if(an(e),n!=null&&n==="bool")throw new Error("Unsupported data type $ { dtype }");let s=new I3(t,a,n,!0,r),i=$e(e,n);for(let o=0;o<i.values.length;o++)i.values[o]=s.nextValue();return i.toTensor()}var N4=z({truncatedNormal_:iF});function oF(e,t=0){let a=R(e,"x","unique","string_or_numeric");D(a.rank>0,()=>"The input tensor must be at least 1D");let n={x:a},r={axis:t},[s,i]=L.runKernel(bp,n,r);return{values:s,indices:i}}var R4=z({unique_:oF});function lF(e,t,a){let n=R(e,"x","unsortedSegmentSum"),r=R(t,"segmentIds","unsortedSegmentSum","int32");D($l(a),()=>"numSegments must be of dtype int");let s={x:n,segmentIds:r},i={numSegments:a};return L.runKernel(vp,s,i)}var E4=z({unsortedSegmentSum_:lF});function uF(e,t=0){let a=R(e,"x","unstack","string_or_numeric");D(t>=-a.shape.length&&t<a.shape.length,()=>`Axis = ${t} is not in [-${a.shape.length}, ${a.shape.length})`);let n={value:a},r={axis:t};return L.runKernel(Eu,n,r)}var Ra=z({unstack_:uF});function M4(e,t){return $h(e,t,"right")}function F4(e,t=!0,a,n){return L.makeVariable(e,t,a,n)}function _4(e,t){let a=[];for(let s=0;s<t.length;s++)t[s]&&a.push(s);let n=$e(e,"int32"),r=$e([a.length,e.length],"int32");for(let s=0;s<a.length;s++){let i=n.indexToLoc(a[s]),o=s*e.length;r.values.set(i,o)}return r.toTensor()}async function dF(e){let t=R(e,"condition","whereAsync","bool"),a=await t.data(),n=_4(t.shape,a);return e!==t&&t.dispose(),n}var _3=dF;async function pF(e,t,a){let n=R(e,"tensor","boolMask"),r=R(t,"mask","boolMask","bool"),s=a==null?0:a,i=r.rank,o=n.shape;D(i>0,()=>"mask cannot be scalar"),Na(o.slice(s,s+i),r.shape,"mask's shape must match the first K dimensions of tensor's shape,");let l=1;for(let m=s;m<s+i;m++)l*=o[m];let u=o.slice(0,s).concat([l],o.slice(s+i)),p=Q(n,u),c=Q(r,[-1]),d=await _3(c),h=Oe(d,[1]),f=l3(p,h,s);return e!==n&&n.dispose(),t!==r&&r.dispose(),h.dispose(),p.dispose(),c.dispose(),d.dispose(),f}var $4=pF;function cF(e,t,a){let n=R(e,"x","transpose");if(t==null&&(t=n.shape.map((i,o)=>o).reverse()),D(n.rank===t.length,()=>`Error in transpose: rank of input ${n.rank} must match length of perm ${t}.`),t.forEach(i=>{D(i>=0&&i<n.rank,()=>`All entries in 'perm' must be between 0 and ${n.rank-1} but got ${t}`)}),n.rank<=1)return n.clone();let r={x:n},s={perm:t};return n.dtype==="complex64"?Pe(()=>{let i=Bl(n),o=Pp(n);return i=L.runKernel(wr,{x:i},s),o=L.runKernel(wr,{x:o},s),a&&(o=Ln(o)),Sr(i,o)}):L.runKernel(wr,r,s)}var js=z({transpose_:cF});function hF(e,t,a,n,r=!0){let s=R(e,"v","movingAverage"),i=R(t,"x","movingAverage"),o=R(a,"decay","movingAverage");mA(s,i),D(Cr(s.shape,i.shape),()=>"Shape mismatch in v and x");let l=Ue(1),u=ye(l,o),p=te(ye(i,s),u);if(r){D(n!=null,()=>"When using zeroDebias: true, step is required.");let c=R(n,"step","movingAverage");p=ve(p,ye(l,zl(o,c)))}return ke(s,p)}var D4=z({movingAverage_:hF});function fF(e,t,a){an(a);let n=R(e,"indices","scatterND","int32"),r=R(t,"updates","scatterND");Bh(r,n,a);let s={indices:n,updates:r},i={shape:a};return L.runKernel(Io,s,i)}var P4=z({scatterND_:fF});function mF(e,t,a,n){if(e.dtype!=="int32")throw new Error(`tf.sparseToDense() expects the indices to be int32 type, but the dtype was ${e.dtype}.`);if(e.rank>2)throw new Error(`sparseIndices should be a scalar, vector, or matrix, but got shape ${e.shape}.`);let r=e.rank>0?e.shape[0]:1,s=e.rank>1?e.shape[1]:1;if(a.length!==s)throw new Error(`outputShape has incorrect number of elements:, ${a.length}, should be: ${s}.`);let i=t.size;if(!(t.rank===0||t.rank===1&&i===r))throw new Error(`sparseValues has incorrect shape ${t.shape}, should be [] or [${r}]`);if(t.dtype!==n.dtype)throw new Error("sparseValues.dtype must match defaultValues.dtype")}function gF(e,t,a,n=0){an(a);let r=R(e,"sparseIndices","sparseToDense","int32"),s=R(t,"sparseValues","sparseToDense","string_or_numeric"),i=R(n,"defaultValue","sparseToDense",s.dtype);mF(r,s,a,i);let o={sparseIndices:r,sparseValues:s,defaultValue:i},l={outputShape:a};return L.runKernel(Do,o,l)}var O4=z({sparseToDense_:gF});function yF(e,t){let a=R(t,"indices","gatherND","int32"),n={params:R(e,"x","gatherND","string_or_numeric"),indices:a};return L.runKernel(Di,n)}var z4=z({gatherND_:yF});function xF(e,t){if(t==null)return e.shape.slice();if(Cr(e.shape,t))return t;if(e.shape.length===t.length){let a=[];for(let n=0;n<e.shape.length;n++)t[n]==null&&e.shape[n]!=null?a.push(e.shape[n]):a.push(t[n]);return a}return t}function AF(e,t,a,n){let r=R(e,"x","dropout");if(D(r.dtype==="float32",()=>`x has to be a floating point tensor since it's going to be scaled, but got a ${r.dtype} tensor instead.`),D(t>=0&&t<1,()=>`rate must be a float in the range [0, 1), but got ${t}.`),t===0)return e instanceof gt?r.clone():r;let s=xF(r,a),i=1-t,o=ve(o3(ke(Dh(s,0,1,"float32",n),i)),i);return te(r,o)}var L4=z({dropout_:AF});function $3(e){return Math.floor(Math.pow(2,Math.ceil(Math.log(e)/Math.log(2))))}function Vh(e,t,a){let n=1-e%2,r=new Float32Array(e);for(let s=0;s<e;++s){let i=2*Math.PI*s/(e+n-1);r[s]=t-a*Math.cos(i)}return jt(r,"float32")}async function bF(e,t,a=1){let n=R(e,"predictions","inTopK"),r=R(t,"targets","inTopK");D(n.rank>1,()=>`inTopK() expects the predictions to be of rank 2 or higher, but got ${n.rank}`),D(n.rank-1===r.rank,()=>`predictions rank should be 1 larger than targets rank, but got predictions rank ${n.rank} and targets rank ${r.rank}`),Na(n.shape.slice(0,n.shape.length-1),r.shape,"predictions's shape should be align with the targets' shape, except the last dimension.");let s=n.shape[n.shape.length-1];D(a>0&&a<=s,()=>`'k' passed to inTopK() must be > 0 && <= the predictions last dimension (${s}), but got ${a}`);let i=await n.data(),o=await r.data(),[l,u]=[i.length/s,s],p=Jx("bool",l);for(let c=0;c<l;c++){let d=c*u,h=i.subarray(d,d+u),f=[];for(let m=0;m<h.length;m++)f.push({value:h[m],index:m});f.sort((m,g)=>g.value-m.value),p[c]=0;for(let m=0;m<a;m++)if(f[m].index===o[c]){p[c]=1;break}}return e!==n&&n.dispose(),t!==r&&r.dispose(),Ve(p,r.shape,"bool")}var W4=bF,D3={};Ze(D3,{conv2d:()=>IF,depthwiseConv2d:()=>EF,matMul:()=>FF});function vF(e,t,a,n,r,s="NHWC",i){let o=e;e.rank===3&&(o=Q(e,[1,e.shape[0],e.shape[1],e.shape[2]]));let l=t;l.rank===3&&(l=Q(t,[1,t.shape[0],t.shape[1],t.shape[2]])),D(o.rank===4,()=>`Error in conv2dDerFilter: input must be rank 4, but got shape ${o.shape}.`),D(l.rank===4,()=>`Error in conv2dDerFilter: dy must be rank 4, but got shape ${l.shape}.`),D(a.length===4,()=>`Error in conv2dDerFilter: filterShape must be length 4, but got ${a}.`);let u=s==="NHWC"?o.shape[3]:o.shape[1],p=s==="NHWC"?l.shape[3]:l.shape[1];D(u===a[2],()=>`Error in conv2dDerFilter: depth of input ${u}) must match input depth in filter (${a[2]}.`),D(p===a[3],()=>`Error in conv2dDerFilter: depth of dy (${p}) must match output depth for filter (${a[3]}).`),Tn("conv2dDerFilter",r,i);let c={x:o,dy:l},d={strides:n,pad:r,dataFormat:s,dimRoundingMode:i,filterShape:a};return L.runKernel(np,c,d)}var wF=z({conv2DBackpropFilter_:vF});function Uh(e,t,a){if(a==null||a==="linear")return e;if(a==="relu")return te(e,E3(t));throw new Error(`Cannot compute gradient for fused activation ${a}.`)}function Gh(e,t){let a=t,n=a3(e.shape,t.shape);return n.length>0&&(a=ot(a,n)),Q(a,e.shape)}function Hh(e,t,a,n){if(t==="linear")return e;if(t==="relu")return Op(e);if(t==="elu")return r3(e);if(t==="relu6")return C3(e);if(t==="prelu")return b3(e,a);if(t==="leakyrelu")return d3(e,n);if(t==="sigmoid")return Wa(e);throw new Error(`Unknown fused activation ${t}.`)}var jh=(e,t)=>!(e>0)||t==="linear";function kF({x:e,filter:t,strides:a,pad:n,dataFormat:r="NHWC",dilations:s=[1,1],dimRoundingMode:i,bias:o,activation:l="linear",preluActivationWeights:u,leakyreluAlpha:p}){if(l=l||"linear",jh(L.state.gradientDepth,l)===!1){D(r==="NHWC",()=>`Error in fused conv2d: got dataFormat of ${r} but only NHWC is currently supported for the case of gradient depth is 0 and the activation is not linear.`);let C=Fp(e,t,a,n,r,s,i);return o!=null&&(C=ke(C,o)),Hh(C,l,u,p)}let c=R(e,"x","conv2d","float32"),d=R(t,"filter","conv2d","float32"),h=c,f=!1;c.rank===3&&(f=!0,h=Q(c,[1,c.shape[0],c.shape[1],c.shape[2]])),D(h.rank===4,()=>`Error in fused conv2d: input must be rank 4, but got rank ${h.rank}.`),D(d.rank===4,()=>`Error in fused conv2d: filter must be rank 4, but got rank ${d.rank}.`),Tn("fused conv2d",n,i);let m=r==="NHWC"?h.shape[3]:h.shape[1];D(d.shape[2]===m,()=>`Error in conv2d: depth of input (${m}) must match input depth for filter ${d.shape[2]}.`),D(Tr(a,s),()=>`Error in conv2D: Either strides or dilations must be 1. Got strides ${a} and dilations '${s}'`);let g=Ep(h.shape,d.shape,a,s,n,i),y;o!=null&&(y=R(o,"bias","fused conv2d"),[y]=Rt(y,c),r==="NHWC"?Bt(g.outShape,y.shape):(D(y.shape.length<=1,()=>`Error in fused conv2d: only supports scalar or 1-D Tensor bias for NCHW format but got the bias of rank-${y.shape.length}.`),D(y.shape.length===0||y.shape[0]===g.outChannels||y.shape[0]===1,()=>`Error in fused conv2d: bias shape (${y.shape}) is not compatible with the number of output channels (${g.outChannels})`)));let x;if(u!=null){let C=u.shape;if(D(C.length<=1||C.length===3,()=>`Error in fused conv2d: only supports scalar, 1-D Tensor or 3-D Tensor PReLU activation weights but got a tensor of rank-${C.length}.`),C.length===1)D(C[0]===1||C[0]===g.outChannels,()=>`Error in fused conv2d: PReLU activation weights (${C}) is not compatible with the number of output channels (${g.outChannels}).`);else if(C.length===3)try{Bt(C,g.outShape)}catch(T){let N=`Error in fused conv2d: PReLU activation weights (${C}) is not compatible with the output shape of the conv2d (${g.outShape}).`;throw Error(N)}x=R(u,"prelu weights","fused conv2d")}let A=(C,T)=>{D(r==="NHWC",()=>`Error in gradient of fused conv2D: got dataFormat of ${r} but only NHWC is currently supported.`);let[N,M,F,E]=T,I=Uh(C,F,l);D(zd(s),()=>`Error in gradient of fused conv2D: dilation rates greater than 1 are not yet supported in gradients. Got dilations '${s}'`);let _=nb(M.shape,I,N,a,n),O=wF(M,I,N.shape,a,n),B=[_,O];if(E!=null){let $=Gh(E,I);B.push($)}return B},b={x:h,filter:d,bias:y,preluActivationWeights:x},w={strides:a,pad:n,dataFormat:r,dilations:s,dimRoundingMode:i,activation:l,leakyreluAlpha:p};return o==null?tr((C,T,N)=>{let M=L.runKernel(Kr,b,w);return N([T,C,M]),f&&(M=Q(M,[M.shape[1],M.shape[2],M.shape[3]])),{value:M,gradFunc:A}})(h,d):tr((C,T,N,M)=>{let F=L.runKernel(Kr,b,w);return M([T,C,F,N]),f&&(F=Q(F,[F.shape[1],F.shape[2],F.shape[3]])),{value:F,gradFunc:A}})(h,d,y)}var IF=z({fusedConv2d_:kF});function SF(e,t,a,n,r,s=[1,1],i){let o=e;e.rank===3&&(o=Q(e,[1,e.shape[0],e.shape[1],e.shape[2]]));let l=t;l.rank===3&&(l=Q(t,[1,t.shape[0],t.shape[1],t.shape[2]]));let u={x:o,dy:l},p={strides:n,pad:r,dimRoundingMode:i,dilations:s,filterShape:a};return L.runKernel(rp,u,p)}var CF=z({depthwiseConv2dNativeBackpropFilter_:SF});function TF(e,t,a,n,r,s=[1,1],i){let o=t,l=!1;t.rank===3&&(l=!0,o=Q(t,[1,t.shape[0],t.shape[1],t.shape[2]]));let u={dy:o,filter:a},p={strides:n,pad:r,dimRoundingMode:i,dilations:s,inputShape:e},c=L.runKernel(sp,u,p);return l?Q(c,[c.shape[1],c.shape[2],c.shape[3]]):c}var NF=z({depthwiseConv2dNativeBackpropInput_:TF});function RF({x:e,filter:t,strides:a,pad:n,dataFormat:r="NHWC",dilations:s=[1,1],dimRoundingMode:i,bias:o,activation:l="linear",preluActivationWeights:u,leakyreluAlpha:p}){if(jh(L.state.gradientDepth,l)===!1){let w=Mh(e,t,a,n,r,s,i);return o!=null&&(w=ke(w,o)),Hh(w,l,u,p)}let c=R(e,"x","depthwiseConv2d","float32"),d=R(t,"filter","depthwiseConv2d","float32"),h=c,f=!1;c.rank===3&&(f=!0,h=Q(c,[1,c.shape[0],c.shape[1],c.shape[2]])),D(h.rank===4,()=>`Error in fused depthwiseConv2d: input must be rank 4, but got rank ${h.rank}.`),D(d.rank===4,()=>`Error in fused depthwiseConv2d: filter must be rank 4, but got rank ${d.rank}.`),D(h.shape[3]===d.shape[2],()=>`Error in fused depthwiseConv2d: number of input channels (${h.shape[3]}) must match the inChannels dimension in filter ${d.shape[2]}.`),s==null&&(s=[1,1]),D(Tr(a,s),()=>`Error in fused depthwiseConv2d: Either strides or dilations must be 1. Got strides ${a} and dilations '${s}'`),Tn("fused depthwiseConv2d",n,i);let m=Ep(h.shape,d.shape,a,s,n,i,!0),g;o!=null&&(g=R(o,"bias","fused conv2d"),[g]=Rt(g,c),Bt(m.outShape,g.shape));let y;u!=null&&(y=R(u,"prelu weights","fused depthwiseConv2d"));let x=(w,C)=>{D(zd(s),()=>`Error in gradient of fused depthwiseConv2d: dilation rates greater than 1 are not yet supported. Got dilations '${s}'`);let[T,N,M,F]=C,E=Uh(w,M,l),I=NF(N.shape,E,T,a,n,s,i),_=CF(N,E,T.shape,a,n,s,i);if(F!=null){let O=Gh(g,E);return[I,_,O]}return[I,_]},A={x:h,filter:d,bias:g,preluActivationWeights:y},b={strides:a,pad:n,dataFormat:r,dilations:s,dimRoundingMode:i,activation:l,leakyreluAlpha:p};return o==null?tr((w,C,T)=>{let N=L.runKernel(Yr,A,b);return T([C,w,N]),f&&(N=Q(N,[N.shape[1],N.shape[2],N.shape[3]])),{value:N,gradFunc:x}})(h,d):tr((w,C,T,N)=>{let M=L.runKernel(Yr,A,b);return N([C,w,M,T]),f&&(M=Q(M,[M.shape[1],M.shape[2],M.shape[3]])),{value:M,gradFunc:x}})(h,d,g)}var EF=z({fusedDepthwiseConv2d_:RF});function MF({a:e,b:t,transposeA:a=!1,transposeB:n=!1,bias:r,activation:s="linear",preluActivationWeights:i,leakyreluAlpha:o=.2}){if(jh(L.state.gradientDepth,s)===!1){let F=pt(e,t,a,n);return r!=null&&(F=ke(F,r)),Hh(F,s,i,o)}let l=R(e,"a","fused matMul"),u=R(t,"b","fused matMul");[l,u]=Rt(l,u);let p=a?l.shape[l.rank-2]:l.shape[l.rank-1],c=n?u.shape[u.rank-1]:u.shape[u.rank-2],d=a?l.shape[l.rank-1]:l.shape[l.rank-2],h=n?u.shape[u.rank-2]:u.shape[u.rank-1],f=l.shape.slice(0,-2),m=u.shape.slice(0,-2),g=vt(f),y=vt(m);D(p===c,()=>`Error in fused matMul: inner shapes (${p}) and (${c}) of Tensors with shapes ${l.shape} and ${u.shape} and transposeA=${a} and transposeB=${n} must match.`);let x=Bt(l.shape.slice(0,-2),u.shape.slice(0,-2)).concat([d,h]),A=a?Q(l,[g,p,d]):Q(l,[g,d,p]),b=n?Q(u,[y,h,c]):Q(u,[y,c,h]),w;r!=null&&(w=R(r,"bias","fused matMul"),[w]=Rt(w,l),Bt(x,w.shape));let C;i!=null&&(C=R(i,"prelu weights","fused matMul"));let T=(F,E)=>{let[I,_,O,B]=E,$=Uh(Q(F,O.shape),O,s),U,G;if(!a&&!n?(U=pt($,_,!1,!0),G=pt(I,$,!0,!1)):!a&&n?(U=pt($,_,!1,!1),G=pt($,I,!0,!1)):a&&!n?(U=pt(_,$,!1,!0),G=pt(I,$,!1,!1)):(U=pt(_,$,!0,!0),G=pt($,I,!0,!0)),r!=null){let q=Gh(B,$);return[U,G,q]}else return[U,G]},N={a:A,b,bias:w,preluActivationWeights:C},M={transposeA:a,transposeB:n,activation:s,leakyreluAlpha:o};return r==null?tr((F,E,I)=>{let _=L.runKernel(Xr,N,M);return I([F,E,_]),{value:Q(_,x),gradFunc:T}})(A,b):tr((F,E,I,_)=>{let O=L.runKernel(Xr,N,M);return _([F,E,O,I]),{value:Q(O,x),gradFunc:T}})(A,b,w)}var FF=z({fusedMatMul_:MF});function _F(e){return Vh(e,.54,.46)}var $F=z({hammingWindow_:_F});function DF(e){return Vh(e,.5,.5)}var B4=z({hannWindow_:DF});function PF(e,t,a,n=!1,r=0){let s=0,i=[];for(;s+t<=e.size;)i.push(De(e,s,t)),s+=a;if(n)for(;s<e.size;){let o=s+t-e.size,l=lt([De(e,s,t-o),sr([o],r)]);i.push(l),s+=a}return i.length===0?Zn([],[0,t]):Q(lt(i),[i.length,t])}var V4=z({frame_:PF});function OF(e,t,a,n,r=B4){n==null&&(n=$3(t));let s=V4(e,t,a),i=te(s,r(t));return Lh(i,n)}var zF=z({stft_:OF});function LF(e,t,a,n,r="bilinear",s=0){let i=R(e,"image","cropAndResize"),o=R(t,"boxes","cropAndResize","float32"),l=R(a,"boxInd","cropAndResize","int32"),u=o.shape[0];D(i.rank===4,()=>`Error in cropAndResize: image must be rank 4,but got rank ${i.rank}.`),D(o.rank===2&&o.shape[1]===4,()=>`Error in cropAndResize: boxes must be have size [${u},4] but had shape ${o.shape}.`),D(l.rank===1&&l.shape[0]===u,()=>`Error in cropAndResize: boxInd must be have size [${u}] but had shape ${o.shape}.`),D(n.length===2,()=>`Error in cropAndResize: cropSize must be of length 2, but got length ${n.length}.`),D(n[0]>=1&&n[1]>=1,()=>`cropSize must be atleast [1,1], but was ${n}`),D(r==="bilinear"||r==="nearest",()=>`method must be bilinear or nearest, but was ${r}`);let p={image:i,boxes:o,boxInd:l},c={method:r,extrapolationValue:s,cropSize:n};return L.runKernel(wi,p,c)}var WF=z({cropAndResize_:LF});function BF(e){let t=R(e,"image","flipLeftRight","float32");D(t.rank===4,()=>`Error in flipLeftRight: image must be rank 4,but got rank ${t.rank}.`);let a={image:t};return L.runKernel(Mi,a,{})}var VF=z({flipLeftRight_:BF});function UF(e){let t=R(e,"image","grayscaleToRGB"),a=t.rank-1,n=t.shape[a];D(t.rank>=2,()=>`Error in grayscaleToRGB: images must be at least rank 2, but got rank ${t.rank}.`),D(n===1,()=>`Error in grayscaleToRGB: last dimension of a grayscale image should be size 1, but got size ${n}.`);let r=new Array(t.rank);return r.fill(1,0,a),r[a]=3,jr(t,r)}var GF=z({grayscaleToRGB_:UF});function HF(e,t,a=0,n=.5){let r=R(e,"image","rotateWithOffset","float32");D(r.rank===4,()=>`Error in rotateWithOffset: image must be rank 4,but got rank ${r.rank}.`);let s={image:r},i={radians:t,fillValue:a,center:n};return L.runKernel(Uo,s,i)}var jF=z({rotateWithOffset_:HF});function $u(e,t,a,n,r,s){n==null&&(n=.5),r==null&&(r=Number.NEGATIVE_INFINITY),s==null&&(s=0);let i=e.shape[0];return a=Math.min(a,i),D(0<=n&&n<=1,()=>`iouThreshold must be in [0, 1], but was '${n}'`),D(e.rank===2,()=>`boxes must be a 2D tensor, but was of rank '${e.rank}'`),D(e.shape[1]===4,()=>`boxes must have 4 columns, but 2nd dimension was ${e.shape[1]}`),D(t.rank===1,()=>"scores must be a 1D tensor"),D(t.shape[0]===i,()=>`scores has incompatible shape with boxes. Expected ${i}, but was ${t.shape[0]}`),D(0<=s&&s<=1,()=>`softNmsSigma must be in [0, 1], but was '${s}'`),{maxOutputSize:a,iouThreshold:n,scoreThreshold:r,softNmsSigma:s}}function qF(e,t,a,n=.5,r=Number.NEGATIVE_INFINITY){let s=R(e,"boxes","nonMaxSuppression","float32"),i=R(t,"scores","nonMaxSuppression","float32"),o=$u(s,i,a,n,r);a=o.maxOutputSize,n=o.iouThreshold,r=o.scoreThreshold;let l={maxOutputSize:a,iouThreshold:n,scoreThreshold:r};return L.runKernel(lo,{boxes:s,scores:i},l)}var XF=z({nonMaxSuppression_:qF});function KF(e,t,a){let n=YF(e,t,a),r=n<0?-(n+1):n;e.splice(r,0,t)}function YF(e,t,a){return JF(e,t,a||ZF)}function ZF(e,t){return e>t?1:e<t?-1:0}function JF(e,t,a){let n=0,r=e.length,s=0,i=!1;for(;n<r;){s=n+(r-n>>>1);let o=a(t,e[s]);o>0?n=s+1:(r=s,i=!o)}return i?n:-n-1}function U4(e,t,a,n,r){return P3(e,t,a,n,r,0)}function G4(e,t,a,n,r,s){return P3(e,t,a,n,r,0,!1,s,!0)}function H4(e,t,a,n,r,s){return P3(e,t,a,n,r,s,!0)}function P3(e,t,a,n,r,s,i=!1,o=!1,l=!1){let u=[];for(let g=0;g<t.length;g++)t[g]>r&&u.push({score:t[g],boxIndex:g,suppressBeginIndex:0});u.sort(D5);let p=s>0?-.5/s:0,c=[],d=[];for(;c.length<a&&u.length>0;){let g=u.pop(),{score:y,boxIndex:x,suppressBeginIndex:A}=g;if(y<r)break;let b=!1;for(let w=c.length-1;w>=A;--w){let C=QF(e,x,c[w]);if(C>=n){b=!0;break}if(g.score=g.score*e_(n,p,C),g.score<=r)break}g.suppressBeginIndex=c.length,b||(g.score===y?(c.push(x),d.push(g.score)):g.score>r&&KF(u,g,D5))}let h=c.length,f=a-h;o&&f>0&&(c.push(...new Array(f).fill(0)),d.push(...new Array(f).fill(0)));let m={selectedIndices:c};return i&&(m.selectedScores=d),l&&(m.validOutputs=h),m}function QF(e,t,a){let n=e.subarray(t*4,t*4+4),r=e.subarray(a*4,a*4+4),s=Math.min(n[0],n[2]),i=Math.min(n[1],n[3]),o=Math.max(n[0],n[2]),l=Math.max(n[1],n[3]),u=Math.min(r[0],r[2]),p=Math.min(r[1],r[3]),c=Math.max(r[0],r[2]),d=Math.max(r[1],r[3]),h=(o-s)*(l-i),f=(c-u)*(d-p);if(h<=0||f<=0)return 0;let m=Math.max(s,u),g=Math.max(i,p),y=Math.min(o,c),x=Math.min(l,d),A=Math.max(y-m,0)*Math.max(x-g,0);return A/(h+f-A)}function e_(e,t,a){let n=Math.exp(t*a*a);return a<=e?n:0}function D5(e,t){return e.score-t.score||e.score===t.score&&t.boxIndex-e.boxIndex}async function t_(e,t,a,n=.5,r=Number.NEGATIVE_INFINITY){let s=R(e,"boxes","nonMaxSuppressionAsync"),i=R(t,"scores","nonMaxSuppressionAsync"),o=$u(s,i,a,n,r);a=o.maxOutputSize,n=o.iouThreshold,r=o.scoreThreshold;let l=await Promise.all([s.data(),i.data()]),u=l[0],p=l[1],{selectedIndices:c}=U4(u,p,a,n,r);return s!==e&&s.dispose(),i!==t&&i.dispose(),jt(c,"int32")}var a_=t_;function n_(e,t,a,n=.5,r=Number.NEGATIVE_INFINITY,s=0){let i=R(e,"boxes","nonMaxSuppression"),o=R(t,"scores","nonMaxSuppression"),l=$u(i,o,a,n,r,s);a=l.maxOutputSize,n=l.iouThreshold,r=l.scoreThreshold,s=l.softNmsSigma;let u={boxes:i,scores:o},p={maxOutputSize:a,iouThreshold:n,scoreThreshold:r,softNmsSigma:s},c=L.runKernel(uo,u,p);return{selectedIndices:c[0],selectedScores:c[1]}}var r_=z({nonMaxSuppressionWithScore_:n_});async function s_(e,t,a,n=.5,r=Number.NEGATIVE_INFINITY,s=0){let i=R(e,"boxes","nonMaxSuppressionAsync"),o=R(t,"scores","nonMaxSuppressionAsync"),l=$u(i,o,a,n,r,s);a=l.maxOutputSize,n=l.iouThreshold,r=l.scoreThreshold,s=l.softNmsSigma;let u=await Promise.all([i.data(),o.data()]),p=u[0],c=u[1],{selectedIndices:d,selectedScores:h}=H4(p,c,a,n,r,s);return i!==e&&i.dispose(),o!==t&&o.dispose(),{selectedIndices:jt(d,"int32"),selectedScores:jt(h)}}var i_=s_;function o_(e,t,a,n=.5,r=Number.NEGATIVE_INFINITY,s=!1){let i=R(e,"boxes","nonMaxSuppression"),o=R(t,"scores","nonMaxSuppression"),l=$u(i,o,a,n,r,null),u=l.maxOutputSize,p=l.iouThreshold,c=l.scoreThreshold,d={boxes:i,scores:o},h={maxOutputSize:u,iouThreshold:p,scoreThreshold:c,padToMaxOutputSize:s},f=L.runKernel(gu,d,h);return{selectedIndices:f[0],validOutputs:f[1]}}var l_=z({nonMaxSuppressionPadded_:o_});async function u_(e,t,a,n=.5,r=Number.NEGATIVE_INFINITY,s=!1){let i=R(e,"boxes","nonMaxSuppressionAsync"),o=R(t,"scores","nonMaxSuppressionAsync"),l=$u(i,o,a,n,r,null),u=l.maxOutputSize,p=l.iouThreshold,c=l.scoreThreshold,[d,h]=await Promise.all([i.data(),o.data()]),{selectedIndices:f,validOutputs:m}=G4(d,h,u,p,c,s);return i!==e&&i.dispose(),o!==t&&o.dispose(),{selectedIndices:jt(f,"int32"),validOutputs:Ue(m,"int32")}}var d_=u_;function p_(e,t,a=!1,n=!1){let r=R(e,"images","resizeBilinear");D(r.rank===3||r.rank===4,()=>`Error in resizeBilinear: x must be rank 3 or 4, but got rank ${r.rank}.`),D(t.length===2,()=>`Error in resizeBilinear: new shape must 2D, but got shape ${t}.`),D(n===!1||a===!1,()=>"Error in resizeBilinear: If halfPixelCenters is true, alignCorners must be false.");let s=r,i=!1;r.rank===3&&(i=!0,s=Q(r,[1,r.shape[0],r.shape[1],r.shape[2]]));let[]=t,o={images:s},l={alignCorners:a,halfPixelCenters:n,size:t},u=L.runKernel(Ao,o,l);return i?Q(u,[u.shape[1],u.shape[2],u.shape[3]]):u}var c_=z({resizeBilinear_:p_});function h_(e,t,a=!1,n=!1){let r=R(e,"images","resizeNearestNeighbor");D(r.rank===3||r.rank===4,()=>`Error in resizeNearestNeighbor: x must be rank 3 or 4, but got rank ${r.rank}.`),D(t.length===2,()=>`Error in resizeNearestNeighbor: new shape must 2D, but got shape ${t}.`),D(r.dtype==="float32"||r.dtype==="int32",()=>"`images` must have `int32` or `float32` as dtype"),D(n===!1||a===!1,()=>"Error in resizeNearestNeighbor: If halfPixelCenters is true, alignCorners must be false.");let s=r,i=!1;r.rank===3&&(i=!0,s=Q(r,[1,r.shape[0],r.shape[1],r.shape[2]]));let[]=t,o={images:s},l={alignCorners:a,halfPixelCenters:n,size:t},u=L.runKernel(xo,o,l);return i?Q(u,[u.shape[1],u.shape[2],u.shape[3]]):u}var f_=z({resizeNearestNeighbor_:h_});function m_(e,t="binary",a=!1,n=.5){let r=R(e,"image","threshold"),s=.2989,i=.587,o=.114,l=r.shape[0]*r.shape[1],u=te(jt([n]),255),p,c,d,h;if(D(r.rank===3,()=>`Error in threshold: image must be rank 3,but got rank ${r.rank}.`),D(r.shape[2]===3||r.shape[2]===1,()=>`Error in threshold: image color channel must be equal to 3 or 1but got ${r.shape[2]}.`),D(r.dtype==="int32"||r.dtype==="float32",()=>`Error in dtype: image dtype must be int32 or float32,but got dtype ${r.dtype}.`),D(t==="otsu"||t==="binary",()=>`Method must be binary or otsu, but was ${t}`),r.shape[2]===3){[p,c,d]=Ca(r,[1,1,1],-1);let m=te(p,s),g=te(c,i),y=te(d,o);h=ke(ke(m,g),y)}else h=e;if(t==="otsu"){let m=t3(Ke(T3(h),"int32"),Ve([]),256);u=g_(m,l)}let f=a?Fh(h,u):Dp(h,u);return Ke(te(f,255),"int32")}function g_(e,t){let a=jt([-1]),n=jt([0]),r=jt([0]),s,i,o,l,u,p;for(let c=0;c<e.size-1;c++){s=De(e,0,c+1),i=De(e,c+1),u=ve(ot(s),t),p=ve(ot(i),t);let d=ot(te(s,Wl(0,s.size)));o=ve(d,ot(s));let h=sr(i.shape,s.size),f=ke(Wl(0,i.size),h),m=te(i,f);l=ve(ot(m),ot(i));let g=ye(o,l),y=ye(o,l),x=te(u,p);r=te(te(x,g),y);let A=Dp(r,n);n=kr(A,r,n),a=kr(A,jt([c]),a)}return a}var y_=z({threshold_:m_});function x_(e,t,a="nearest",n="constant",r=0,s){let i=R(e,"image","transform","float32"),o=R(t,"transforms","transform","float32");D(i.rank===4,()=>`Error in transform: image must be rank 4,but got rank ${i.rank}.`),D(o.rank===2&&(o.shape[0]===i.shape[0]||o.shape[0]===1)&&o.shape[1]===8,()=>"Error in transform: Input transform should be batch x 8 or 1 x 8"),D(s==null||s.length===2,()=>`Error in transform: outputShape must be [height, width] or null, but got ${s}.`);let l={image:i,transforms:o},u={interpolation:a,fillMode:n,fillValue:r,outputShape:s};return L.runKernel(Vo,l,u)}var A_=z({transform_:x_});function b_(e,t,a){let n=R(e,"a","bandPart");D(n.rank>=2,()=>`bandPart(): Rank must be at least 2, got ${n.rank}.`);let r=n.shape,[s,i]=n.shape.slice(-2),o,l;typeof t=="number"?(D(t%1===0,()=>`bandPart(): numLower must be an integer, got ${t}.`),D(t<=s,()=>`bandPart(): numLower (${t}) must not be greater than the number of rows (${s}).`),o=R(t<0?s:t,"numLower","bandPart")):(D(t.dtype==="int32",()=>"bandPart(): numLower's dtype must be an int32."),o=kr(ih(t,0),s,Bd(t,s))),typeof a=="number"?(D(a%1===0,()=>`bandPart(): numUpper must be an integer, got ${a}.`),D(a<=i,()=>`bandPart(): numUpper (${a}) must not be greater than the number of columns (${i}).`),l=R(a<0?i:a,"numUpper","bandPart")):(D(a.dtype==="int32",()=>"bandPart(): numUpper's dtype must be an int32."),l=kr(ih(a,0),i,Bd(a,i)));let u=Q(Wl(0,s,1,"int32"),[-1,1]),p=Wl(0,i,1,"int32"),c=ye(u,p),d=Ld(Fh(c,o),u3(c,Ln(l))),h=gn([s,i],n.dtype);return Q(ca(Ra(Q(n,[-1,s,i])).map(f=>kr(d,f,h))),r)}var v_=z({bandPart_:b_});function w_(e){let t;if(Array.isArray(e)){t=!1,D(e!=null&&e.length>0,()=>"Gram-Schmidt process: input must not be null, undefined, or empty");let r=e[0].shape[0];for(let s=1;s<e.length;++s)D(e[s].shape[0]===r,()=>`Gram-Schmidt: Non-unique lengths found in the input vectors: (${e[s].shape[0]} vs. ${r})`)}else t=!0,e=Ca(e,e.shape[0],0).map(r=>Oe(r,[0]));D(e.length<=e[0].shape[0],()=>`Gram-Schmidt: Number of vectors (${e.length}) exceeds number of dimensions (${e[0].shape[0]}).`);let a=[],n=e;for(let r=0;r<e.length;++r)a.push(L.tidy(()=>{let s=n[r];if(r>0)for(let i=0;i<r;++i){let o=te(ot(te(a[i],s)),a[i]);s=ye(s,o)}return ve(s,$p(s,"euclidean"))}));return t?ca(a,0):a}var k_=z({gramSchmidt_:w_});function I_(e,t=!1){if(D(e.rank>=2,()=>`qr() requires input tensor to have a rank >= 2, but got rank ${e.rank}`),e.rank===2)return P5(e,t);{let a=e.shape.slice(0,e.shape.length-2).reduce((l,u)=>l*u),n=Ra(Q(e,[a,e.shape[e.shape.length-2],e.shape[e.shape.length-1]]),0),r=[],s=[];n.forEach(l=>{let[u,p]=P5(l,t);r.push(u),s.push(p)});let i=Q(ca(r,0),e.shape),o=Q(ca(s,0),e.shape);return[i,o]}}function P5(e,t=!1){return L.tidy(()=>{D(e.shape.length===2,()=>`qr2d() requires a 2D Tensor, but got a ${e.shape.length}D Tensor.`);let a=e.shape[0],n=e.shape[1],r=i3(a),s=Sa(e),i=Zn([[1]],[1,1]),o=Sa(i),l=a>=n?n:a;for(let u=0;u<l;++u){let p=s,c=o,d=r;[o,s,r]=L.tidy(()=>{let h=De(s,[u,u],[a-u,1]),f=$p(h),m=De(s,[u,u],[1,1]),g=kr(Dp(m,0),Zn([[-1]]),Zn([[1]])),y=ye(m,te(g,f)),x=ve(h,y);x.shape[0]===1?o=Sa(i):o=lt([i,De(x,[1,0],[x.shape[0]-1,x.shape[1]])],0);let A=Ln(ve(pt(g,y),f)),b=De(s,[u,0],[a-u,n]),w=te(A,o),C=js(o);if(u===0)s=ye(b,pt(w,pt(C,b)));else{let M=ye(b,pt(w,pt(C,b)));s=lt([De(s,[0,0],[u,n]),M],0)}let T=js(w),N=De(r,[0,u],[a,r.shape[1]-u]);if(u===0)r=ye(N,pt(pt(N,o),T));else{let M=ye(N,pt(pt(N,o),T));r=lt([De(r,[0,0],[a,u]),M],1)}return[o,s,r]}),J([p,c,d])}return!t&&a>n&&(r=De(r,[0,0],[a,n]),s=De(s,[0,0],[n,n])),[r,s]})}var S_=z({qr_:I_}),ka;(function(e){e[e.NONE=0]="NONE",e[e.MEAN=1]="MEAN",e[e.SUM=2]="SUM",e[e.SUM_BY_NONZERO_WEIGHTS=3]="SUM_BY_NONZERO_WEIGHTS"})(ka||(ka={}));function C_(e,t,a=ka.SUM_BY_NONZERO_WEIGHTS){let n=R(e,"losses","computeWeightedLoss"),r=null;t!=null&&(r=R(t,"weights","computeWeightedLoss"));let s=r==null?n:te(n,r);if(a===ka.NONE)return s;if(a===ka.SUM)return ot(s);if(a===ka.MEAN){if(r==null)return Wd(s);{let i=n.size/r.size,o=ve(ot(s),ot(r));return i>1?ve(o,Ue(i)):o}}if(a===ka.SUM_BY_NONZERO_WEIGHTS){if(r==null)return ve(ot(s),Ue(n.size));{let i=te(r,Ur(n.shape)),o=Ke(ot(x3(i,Ue(0))),"float32");return ve(ot(s),o)}}throw Error(`Unknown reduction: ${a}`)}var Nr=z({computeWeightedLoss_:C_});function T_(e,t,a,n=ka.SUM_BY_NONZERO_WEIGHTS){let r=R(e,"labels","absoluteDifference"),s=R(t,"predictions","absoluteDifference"),i=null;a!=null&&(i=R(a,"weights","absoluteDifference")),Na(r.shape,s.shape,"Error in absoluteDifference: ");let o=Za(ye(r,s));return Nr(o,i,n)}var N_=z({absoluteDifference_:T_});function R_(e,t,a,n,r=ka.SUM_BY_NONZERO_WEIGHTS){let s=R(e,"labels","cosineDistance"),i=R(t,"predictions","cosineDistance"),o=null;n!=null&&(o=R(n,"weights","cosineDistance")),Na(s.shape,i.shape,"Error in cosineDistance: ");let l=Ue(1),u=ye(l,ot(te(s,i),a,!0));return Nr(u,o,r)}var E_=z({cosineDistance_:R_});function M_(e,t,a,n=ka.SUM_BY_NONZERO_WEIGHTS){let r=R(e,"labels","hingeLoss"),s=R(t,"predictions","hingeLoss"),i=null;a!=null&&(i=R(a,"weights","hingeLoss")),Na(r.shape,s.shape,"Error in hingeLoss: ");let o=Ue(1);r=ye(te(Ue(2),r),o);let l=Op(ye(o,te(r,s)));return Nr(l,i,n)}var F_=z({hingeLoss_:M_});function __(e,t,a,n=1,r=ka.SUM_BY_NONZERO_WEIGHTS){let s=R(e,"labels","huberLoss"),i=R(t,"predictions","huberLoss"),o=null;a!=null&&(o=R(a,"weights","huberLoss")),Na(s.shape,i.shape,"Error in huberLoss: ");let l=Ue(n),u=Za(ye(i,s)),p=Bd(u,l),c=ye(u,p),d=ke(te(Ue(.5),Cn(p)),te(l,c));return Nr(d,o,r)}var $_=z({huberLoss_:__});function D_(e,t,a,n=1e-7,r=ka.SUM_BY_NONZERO_WEIGHTS){let s=R(e,"labels","logLoss"),i=R(t,"predictions","logLoss"),o=null;a!=null&&(o=R(a,"weights","logLoss")),Na(s.shape,i.shape,"Error in logLoss: ");let l=Ue(1),u=Ue(n),p=Ln(te(s,Ll(ke(i,u)))),c=te(ye(l,s),Ll(ke(ye(l,i),u))),d=ye(p,c);return Nr(d,o,r)}var P_=z({logLoss_:D_});function O_(e,t,a,n=ka.SUM_BY_NONZERO_WEIGHTS){let r=R(e,"labels","meanSquaredError"),s=R(t,"predictions","meanSquaredError"),i=null;a!=null&&(i=R(a,"weights","meanSquaredError")),Na(r.shape,s.shape,"Error in meanSquaredError: ");let o=R3(r,s);return Nr(o,i,n)}var z_=z({meanSquaredError_:O_});function L_(e,t){let a=R(e,"labels","sigmoidCrossEntropyWithLogits"),n=R(t,"logits","sigmoidCrossEntropyWithLogits");Na(a.shape,n.shape,"Error in sigmoidCrossEntropyWithLogits: ");let r=Op(n),s=te(n,a),i=p3(Qr(Ln(Za(n))));return ke(ye(r,s),i)}function W_(e,t,a,n=0,r=ka.SUM_BY_NONZERO_WEIGHTS){let s=R(e,"multiClassLabels","sigmoidCrossEntropy"),i=R(t,"logits","sigmoidCrossEntropy"),o=null;if(a!=null&&(o=R(a,"weights","sigmoidCrossEntropy")),Na(s.shape,i.shape,"Error in sigmoidCrossEntropy: "),n>0){let u=Ue(n),p=Ue(1),c=Ue(.5);s=ke(te(s,ye(p,u)),te(c,u))}let l=L_(s,i);return Nr(l,o,r)}var B_=z({sigmoidCrossEntropy_:W_});function V_(e,t,a=-1){if(a===-1&&(a=t.rank-1),a!==t.rank-1)throw Error(`Softmax cross entropy along a non-last dimension is not yet supported. Labels / logits was rank ${t.rank} and dim was ${a}`);return tr((n,r,s)=>{let i=h3(r,[a],!0),o=ye(Ke(r,"float32"),i);s([n,o]);let l=Ln(te(o,n));return{value:ot(l,[a]),gradFunc:(u,p)=>{let[c,d]=p,h=_p(u.shape,[a]);return[te(Q(u,h),ye(Ke(c,"float32"),Qr(d))),te(Q(u,h),ye(Qr(d),Ke(c,"float32")))]}}})(e,t)}function U_(e,t,a,n=0,r=ka.SUM_BY_NONZERO_WEIGHTS){let s=R(e,"onehotLabels","softmaxCrossEntropy"),i=R(t,"logits","softmaxCrossEntropy"),o=null;if(a!=null&&(o=R(a,"weights","softmaxCrossEntropy")),Na(s.shape,i.shape,"Error in softmaxCrossEntropy: "),n>0){let u=Ue(n),p=Ue(1),c=Ue(s.shape[1]);s=ke(te(s,ye(p,u)),ve(u,c))}let l=V_(s,i);return Nr(l,o,r)}var G_=z({softmaxCrossEntropy_:U_});function H_(e,t,a,n){let r=R(e,"indices","sparseFillEmptyRows","int32"),s=R(t,"values","sparseFillEmptyRows"),i=R(a,"denseShape","sparseFillEmptyRows","int32"),o=R(n,"defaultValue","sparseFillEmptyRows",s.dtype);if(r.rank!==2)throw new Error(`Indices should be Tensor2D but received shape
|
|
${r.shape}`);if(s.rank!==1)throw new Error(`Values should be Tensor1D but received shape ${s.shape}`);if(i.rank!==1)throw new Error(`Dense shape should be Tensor1D but received shape ${i.shape}`);if(o.rank!==0)throw new Error(`Default value should be a scalar but received shape ${o.shape}`);let l={indices:r,values:s,denseShape:i,defaultValue:o},u=L.runKernel(hp,l);return{outputIndices:u[0],outputValues:u[1],emptyRowIndicator:u[2],reverseIndexMap:u[3]}}var j_=z({sparseFillEmptyRows_:H_});function q_(e,t,a){let n=R(e,"inputIndices","sparseReshape","int32"),r=R(t,"inputShape","sparseReshape","int32"),s=R(a,"newShape","sparseReshape","int32");if(n.rank!==2)throw new Error(`Input indices should be Tensor2D but received shape
|
|
${n.shape}`);if(r.rank!==1)throw new Error(`Input shape should be Tensor1D but received shape ${r.shape}`);if(s.rank!==1)throw new Error(`New shape should be Tensor1D but received shape ${s.shape}`);let i={inputIndices:n,inputShape:r,newShape:s},o=L.runKernel(Nu,i);return{outputIndices:o[0],outputShape:o[1]}}var X_=z({sparseReshape_:q_});function K_(e,t,a){let n=R(e,"data","sparseSegmentMean"),r=R(t,"indices","sparseSegmentMean","int32"),s=R(a,"segmentIds","sparseSegmentMean","int32");if(n.rank<1)throw new Error("Data should be at least 1 dimensional but received scalar");if(r.rank!==1)throw new Error(`Indices should be Tensor1D but received shape
|
|
${r.shape}`);if(s.rank!==1)throw new Error(`Segment ids should be Tensor1D but received shape
|
|
${s.shape}`);let i={data:n,indices:r,segmentIds:s};return L.runKernel(fp,i)}var Y_=z({sparseSegmentMean_:K_});function Z_(e,t,a){let n=R(e,"data","sparseSegmentSum"),r=R(t,"indices","sparseSegmentSum","int32"),s=R(a,"segmentIds","sparseSegmentSum","int32");if(n.rank<1)throw new Error("Data should be at least 1 dimensional but received scalar");if(r.rank!==1)throw new Error(`Indices should be Tensor1D but received shape
|
|
${r.shape}`);if(s.rank!==1)throw new Error(`Segment ids should be Tensor1D but received shape
|
|
${s.shape}`);let i={data:n,indices:r,segmentIds:s};return L.runKernel(mp,i)}var J_=z({sparseSegmentSum_:Z_});function Q_(e,t,a,n,r,s,i,o){let l=R(e,"data","stringNGrams","string");if(l.dtype!=="string")throw new Error("Data must be of datatype string");if(l.shape.length!==1)throw new Error(`Data must be a vector, saw: ${l.shape}`);let u=R(t,"dataSplits","stringNGrams");if(u.dtype!=="int32")throw new Error("Data splits must be of datatype int32");let p={separator:a,nGramWidths:n,leftPad:r,rightPad:s,padWidth:i,preserveShortSequences:o},c={data:l,dataSplits:u},d=L.runKernel(Ru,c,p);return{nGrams:d[0],nGramsSplits:d[1]}}var e$=z({stringNGrams_:Q_});function t$(e,t,a=!0){let n=R(e,"input","stringSplit","string"),r=R(t,"delimiter","stringSplit","string");if(n.rank!==1)throw new Error(`Input should be Tensor1D but received shape ${n.shape}`);if(r.rank!==0)throw new Error(`Delimiter should be a scalar but received shape ${r.shape}`);let s={skipEmpty:a},i={input:n,delimiter:r},o=L.runKernel(xp,i,s);return{indices:o[0],values:o[1],shape:o[2]}}var a$=z({stringSplit_:t$});function n$(e,t){let a=R(e,"input","stringToHashBucketFast","string"),n={numBuckets:t};if(t<=0)throw new Error("Number of buckets must be at least 1");let r={input:a};return L.runKernel(Ap,r,n)}var r$=z({stringToHashBucketFast_:n$});function s$(e,t,a,n=!0){let r=R(e,"input","staticRegexReplace","string"),s={pattern:t,rewrite:a,replaceGlobal:n};return L.runKernel(yp,{x:r},s)}var i$=z({staticRegexReplace_:s$}),j4={fft:zh,ifft:Vd,rfft:Lh,irfft:N3},q4={hammingWindow:$F,hannWindow:B4,frame:V4,stft:zF},fe={flipLeftRight:VF,grayscaleToRGB:GF,resizeNearestNeighbor:f_,resizeBilinear:c_,rotateWithOffset:jF,cropAndResize:WF,nonMaxSuppression:XF,nonMaxSuppressionAsync:a_,nonMaxSuppressionWithScore:r_,nonMaxSuppressionWithScoreAsync:i_,nonMaxSuppressionPadded:l_,nonMaxSuppressionPaddedAsync:d_,threshold:y_,transform:A_},X4={bandPart:v_,gramSchmidt:k_,qr:S_},K4={absoluteDifference:N_,computeWeightedLoss:Nr,cosineDistance:E_,hingeLoss:F_,huberLoss:$_,logLoss:P_,meanSquaredError:z_,sigmoidCrossEntropy:B_,softmaxCrossEntropy:G_},Y4={sparseFillEmptyRows:j_,sparseReshape:X_,sparseSegmentMean:Y_,sparseSegmentSum:J_},Z4={stringNGrams:e$,stringSplit:a$,stringToHashBucketFast:r$,staticRegexReplace:i$},J4={};Ze(J4,{Serializable:()=>Q4,SerializationMap:()=>Os,registerClass:()=>e7});var Q4=class{getClassName(){return this.constructor.className}static fromConfig(e,t){return new e(t)}},Os=class{constructor(){this.classNameMap={}}static getMap(){return Os.instance==null&&(Os.instance=new Os),Os.instance}static register(e){Os.getMap().classNameMap[e.className]=[e,e.fromConfig]}};function e7(e){D(e.className!=null,()=>"Class being registered does not have the static className property defined."),D(typeof e.className=="string",()=>"className is required to be a string, but got type "+typeof e.className),D(e.className.length>0,()=>"Class being registered has an empty-string as its className, which is disallowed."),Os.register(e)}var ls=class extends Q4{minimize(e,t=!1,a){let{value:n,grads:r}=this.computeGradients(e,a);if(a!=null){let s=a.map(i=>({name:i.name,tensor:r[i.name]}));this.applyGradients(s)}else this.applyGradients(r);return J(r),t?n:(n.dispose(),null)}get iterations(){return this.iterations_==null&&(this.iterations_=0),this.iterations_}incrementIterations(){this.iterations_=this.iterations+1}computeGradients(e,t){return Eb(e,t)}dispose(){this.iterations_!=null&&J(this.iterations_)}async saveIterations(){return this.iterations_==null&&(this.iterations_=0),{name:"iter",tensor:Ue(this.iterations_,"int32")}}async getWeights(){throw new Error("getWeights() is not implemented for this optimizer yet.")}async setWeights(e){throw new Error(`setWeights() is not implemented for this optimizer class ${this.getClassName()}`)}async extractIterations(e){return this.iterations_=(await e[0].tensor.data())[0],e.slice(1)}};Object.defineProperty(ls,Symbol.hasInstance,{value:e=>e.minimize!=null&&e.computeGradients!=null&&e.applyGradients!=null});var O3=class extends ls{static get className(){return"Adadelta"}constructor(e,t,a=null){super(),this.learningRate=e,this.rho=t,this.epsilon=a,this.accumulatedGrads=[],this.accumulatedUpdates=[],a==null&&(this.epsilon=L.backend.epsilon())}applyGradients(e){(Array.isArray(e)?e.map(t=>t.name):Object.keys(e)).forEach((t,a)=>{let n=L.registeredVariables[t],r=!1;this.accumulatedGrads[a]==null&&(this.accumulatedGrads[a]={originalName:`${t}/accum_grad`,variable:Pe(()=>Qa(n).variable(r))}),this.accumulatedUpdates[a]==null&&(this.accumulatedUpdates[a]={originalName:`${t}/accum_var`,variable:Pe(()=>Qa(n).variable(r))});let s=Array.isArray(e)?e[a].tensor:e[t];if(s==null)return;let i=this.accumulatedGrads[a].variable,o=this.accumulatedUpdates[a].variable;Pe(()=>{let l=ke(te(i,this.rho),te(Cn(s),1-this.rho)),u=te(ve(er(ke(o,this.epsilon)),er(ke(i,this.epsilon))),s),p=ke(te(o,this.rho),te(Cn(u),1-this.rho));i.assign(l),o.assign(p);let c=ke(te(u,-this.learningRate),n);n.assign(c)})}),this.incrementIterations()}dispose(){this.accumulatedUpdates!=null&&(J(this.accumulatedGrads.map(e=>e.variable)),J(this.accumulatedUpdates.map(e=>e.variable)))}async getWeights(){let e=[...this.accumulatedGrads,...this.accumulatedUpdates];return[await this.saveIterations()].concat(e.map(t=>({name:t.originalName,tensor:t.variable})))}async setWeights(e){e=await this.extractIterations(e);let t=e.length/2,a=!1;this.accumulatedGrads=e.slice(0,t).map(n=>({originalName:n.name,variable:n.tensor.variable(a)})),this.accumulatedUpdates=e.slice(t,t*2).map(n=>({originalName:n.name,variable:n.tensor.variable(a)}))}getConfig(){return{learningRate:this.learningRate,rho:this.rho,epsilon:this.epsilon}}static fromConfig(e,t){return new e(t.learningRate,t.rho,t.epsilon)}},z3=class extends ls{static get className(){return"Adagrad"}constructor(e,t=.1){super(),this.learningRate=e,this.initialAccumulatorValue=t,this.accumulatedGrads=[]}applyGradients(e){(Array.isArray(e)?e.map(t=>t.name):Object.keys(e)).forEach((t,a)=>{let n=L.registeredVariables[t];this.accumulatedGrads[a]==null&&(this.accumulatedGrads[a]={originalName:`${t}/accumulator`,variable:Pe(()=>sr(n.shape,this.initialAccumulatorValue).variable(!1))});let r=Array.isArray(e)?e[a].tensor:e[t];if(r==null)return;let s=this.accumulatedGrads[a].variable;Pe(()=>{let i=ke(s,Cn(r));s.assign(i);let o=ke(te(ve(r,er(ke(i,L.backend.epsilon()))),-this.learningRate),n);n.assign(o)})}),this.incrementIterations()}dispose(){this.accumulatedGrads!=null&&J(this.accumulatedGrads.map(e=>e.variable))}async getWeights(){return[await this.saveIterations()].concat(this.accumulatedGrads.map(e=>({name:e.originalName,tensor:e.variable})))}async setWeights(e){e=await this.extractIterations(e);let t=!1;this.accumulatedGrads=e.map(a=>({originalName:a.name,variable:a.tensor.variable(t)}))}getConfig(){return{learningRate:this.learningRate,initialAccumulatorValue:this.initialAccumulatorValue}}static fromConfig(e,t){return new e(t.learningRate,t.initialAccumulatorValue)}},L3=class extends ls{static get className(){return"Adam"}constructor(e,t,a,n=null){super(),this.learningRate=e,this.beta1=t,this.beta2=a,this.epsilon=n,this.accumulatedFirstMoment=[],this.accumulatedSecondMoment=[],Pe(()=>{this.accBeta1=Ue(t).variable(),this.accBeta2=Ue(a).variable()}),n==null&&(this.epsilon=L.backend.epsilon())}applyGradients(e){let t=Array.isArray(e)?e.map(a=>a.name):Object.keys(e);Pe(()=>{let a=ye(1,this.accBeta1),n=ye(1,this.accBeta2);t.forEach((r,s)=>{let i=L.registeredVariables[r],o=!1;this.accumulatedFirstMoment[s]==null&&(this.accumulatedFirstMoment[s]={originalName:`${r}/m`,variable:Pe(()=>Qa(i).variable(o))}),this.accumulatedSecondMoment[s]==null&&(this.accumulatedSecondMoment[s]={originalName:`${r}/v`,variable:Pe(()=>Qa(i).variable(o))});let l=Array.isArray(e)?e[s].tensor:e[r];if(l==null)return;let u=this.accumulatedFirstMoment[s].variable,p=this.accumulatedSecondMoment[s].variable,c=ke(te(u,this.beta1),te(l,1-this.beta1)),d=ke(te(p,this.beta2),te(Cn(l),1-this.beta2)),h=ve(c,a),f=ve(d,n);u.assign(c),p.assign(d);let m=ke(te(ve(h,ke(er(f),this.epsilon)),-this.learningRate),i);i.assign(m)}),this.accBeta1.assign(te(this.accBeta1,this.beta1)),this.accBeta2.assign(te(this.accBeta2,this.beta2))}),this.incrementIterations()}dispose(){this.accBeta1.dispose(),this.accBeta2.dispose(),this.accumulatedFirstMoment!=null&&J(this.accumulatedFirstMoment.map(e=>e.variable)),this.accumulatedSecondMoment!=null&&J(this.accumulatedSecondMoment.map(e=>e.variable))}async getWeights(){let e=[...this.accumulatedFirstMoment,...this.accumulatedSecondMoment];return[await this.saveIterations()].concat(e.map(t=>({name:t.originalName,tensor:t.variable})))}async setWeights(e){e=await this.extractIterations(e),Pe(()=>{this.accBeta1.assign(zl(this.beta1,this.iterations_+1)),this.accBeta2.assign(zl(this.beta2,this.iterations_+1))});let t=e.length/2,a=!1;this.accumulatedFirstMoment=e.slice(0,t).map(n=>({originalName:n.name,variable:n.tensor.variable(a)})),this.accumulatedSecondMoment=e.slice(t,t*2).map(n=>({originalName:n.name,variable:n.tensor.variable(a)}))}getConfig(){return{learningRate:this.learningRate,beta1:this.beta1,beta2:this.beta2,epsilon:this.epsilon}}static fromConfig(e,t){return new e(t.learningRate,t.beta1,t.beta2,t.epsilon)}},W3=class extends ls{static get className(){return"Adamax"}constructor(e,t,a,n=null,r=0){super(),this.learningRate=e,this.beta1=t,this.beta2=a,this.epsilon=n,this.decay=r,this.accumulatedFirstMoment=[],this.accumulatedWeightedInfNorm=[],Pe(()=>{this.iteration=Ue(0).variable(),this.accBeta1=Ue(t).variable()}),n==null&&(this.epsilon=L.backend.epsilon())}applyGradients(e){let t=Array.isArray(e)?e.map(a=>a.name):Object.keys(e);Pe(()=>{let a=ye(1,this.accBeta1),n=ve(-this.learningRate,ke(te(this.iteration,this.decay),1));t.forEach((r,s)=>{let i=L.registeredVariables[r],o=!1;this.accumulatedFirstMoment[s]==null&&(this.accumulatedFirstMoment[s]={originalName:`${r}/m`,variable:Qa(i).variable(o)}),this.accumulatedWeightedInfNorm[s]==null&&(this.accumulatedWeightedInfNorm[s]={originalName:`${r}/v`,variable:Qa(i).variable(o)});let l=Array.isArray(e)?e[s].tensor:e[r];if(l==null)return;let u=this.accumulatedFirstMoment[s].variable,p=this.accumulatedWeightedInfNorm[s].variable,c=ke(te(u,this.beta1),te(l,1-this.beta1)),d=te(p,this.beta2),h=Za(l),f=y3(d,h);u.assign(c),p.assign(f);let m=ke(te(ve(n,a),ve(c,ke(f,this.epsilon))),i);i.assign(m)}),this.iteration.assign(ke(this.iteration,1)),this.accBeta1.assign(te(this.accBeta1,this.beta1))}),this.incrementIterations()}dispose(){this.accBeta1.dispose(),this.iteration.dispose(),this.accumulatedFirstMoment!=null&&J(this.accumulatedFirstMoment.map(e=>e.variable)),this.accumulatedWeightedInfNorm!=null&&J(this.accumulatedWeightedInfNorm.map(e=>e.variable))}async getWeights(){throw new Error("getWeights() is not implemented for Adamax yet.")}async setWeights(e){throw new Error("setWeights() is not implemented for Adamax yet.")}getConfig(){return{learningRate:this.learningRate,beta1:this.beta1,beta2:this.beta2,epsilon:this.epsilon,decay:this.decay}}static fromConfig(e,t){return new e(t.learningRate,t.beta1,t.beta2,t.epsilon,t.decay)}},qh=class extends ls{static get className(){return"SGD"}constructor(e){super(),this.learningRate=e,this.setLearningRate(e)}applyGradients(e){(Array.isArray(e)?e.map(t=>t.name):Object.keys(e)).forEach((t,a)=>{let n=Array.isArray(e)?e[a].tensor:e[t];if(n==null)return;let r=L.registeredVariables[t];Pe(()=>{let s=ke(te(this.c,n),r);r.assign(s)})}),this.incrementIterations()}setLearningRate(e){this.learningRate=e,this.c!=null&&this.c.dispose(),this.c=zn(Ue(-e))}dispose(){this.c.dispose()}async getWeights(){return[await this.saveIterations()]}async setWeights(e){if(e=await this.extractIterations(e),e.length!==0)throw new Error("SGD optimizer does not have settable weights.")}getConfig(){return{learningRate:this.learningRate}}static fromConfig(e,t){return new e(t.learningRate)}},B3=class extends qh{static get className(){return"Momentum"}constructor(e,t,a=!1){super(e),this.learningRate=e,this.momentum=t,this.useNesterov=a,this.accumulations=[],this.m=Ue(this.momentum)}applyGradients(e){(Array.isArray(e)?e.map(t=>t.name):Object.keys(e)).forEach((t,a)=>{let n=L.registeredVariables[t];this.accumulations[a]==null&&(this.accumulations[a]={originalName:`${t}/momentum`,variable:Pe(()=>Qa(n).variable(!1))});let r=this.accumulations[a].variable,s=Array.isArray(e)?e[a].tensor:e[t];s!=null&&Pe(()=>{let i,o=ke(te(this.m,r),s);this.useNesterov?i=ke(te(this.c,ke(s,te(o,this.m))),n):i=ke(te(this.c,o),n),r.assign(o),n.assign(i)})}),this.incrementIterations()}dispose(){this.m.dispose(),this.accumulations!=null&&J(this.accumulations.map(e=>e.variable))}setMomentum(e){this.momentum=e}async getWeights(){return[await this.saveIterations()].concat(this.accumulations.map(e=>({name:e.originalName,tensor:e.variable})))}async setWeights(e){e=await this.extractIterations(e);let t=!1;this.accumulations=e.map(a=>({originalName:a.name,variable:a.tensor.variable(t)}))}getConfig(){return{learningRate:this.learningRate,momentum:this.momentum,useNesterov:this.useNesterov}}static fromConfig(e,t){return new e(t.learningRate,t.momentum,t.useNesterov)}},V3=class extends ls{static get className(){return"RMSProp"}constructor(e,t=.9,a=0,n=null,r=!1){if(super(),this.learningRate=e,this.decay=t,this.momentum=a,this.epsilon=n,this.accumulatedMeanSquares=[],this.accumulatedMoments=[],this.accumulatedMeanGrads=[],this.centered=r,n==null&&(this.epsilon=L.backend.epsilon()),e==null)throw new Error("learningRate for RMSPropOptimizer must be defined.")}applyGradients(e){(Array.isArray(e)?e.map(t=>t.name):Object.keys(e)).forEach((t,a)=>{let n=L.registeredVariables[t],r=!1;this.accumulatedMeanSquares[a]==null&&(this.accumulatedMeanSquares[a]={originalName:`${t}/rms`,variable:Pe(()=>Qa(n).variable(r))}),this.accumulatedMoments[a]==null&&(this.accumulatedMoments[a]={originalName:`${t}/momentum`,variable:Pe(()=>Qa(n).variable(r))}),this.accumulatedMeanGrads[a]==null&&this.centered&&(this.accumulatedMeanGrads[a]={originalName:`${t}/mg`,variable:Pe(()=>Qa(n).variable(r))});let s=Array.isArray(e)?e[a].tensor:e[t];if(s==null)return;let i=this.accumulatedMeanSquares[a].variable,o=this.accumulatedMoments[a].variable;Pe(()=>{let l=ke(te(i,this.decay),te(Cn(s),1-this.decay));if(this.centered){let u=this.accumulatedMeanGrads[a].variable,p=ke(te(u,this.decay),te(s,1-this.decay)),c=ve(te(s,this.learningRate),er(ye(l,ke(Cn(p),this.epsilon)))),d=ke(te(o,this.momentum),c);i.assign(l),u.assign(p),o.assign(d);let h=ye(n,d);n.assign(h)}else{let u=ke(te(i,this.decay),te(Cn(s),1-this.decay)),p=ke(te(o,this.momentum),ve(te(s,this.learningRate),er(ke(u,this.epsilon))));i.assign(u),o.assign(p);let c=ye(n,p);n.assign(c)}})}),this.incrementIterations()}dispose(){this.accumulatedMeanSquares!=null&&J(this.accumulatedMeanSquares.map(e=>e.variable)),this.accumulatedMeanGrads!=null&&this.centered&&J(this.accumulatedMeanGrads.map(e=>e.variable)),this.accumulatedMoments!=null&&J(this.accumulatedMoments.map(e=>e.variable))}async getWeights(){let e=[...this.accumulatedMeanSquares,...this.accumulatedMoments];return this.centered&&e.push(...this.accumulatedMeanGrads),[await this.saveIterations()].concat(e.map(t=>({name:t.originalName,tensor:t.variable})))}async setWeights(e){e=await this.extractIterations(e);let t=this.centered?e.length/3:e.length/2,a=!1;this.accumulatedMeanSquares=e.slice(0,t).map(n=>({originalName:n.name,variable:n.tensor.variable(a)})),this.accumulatedMoments=e.slice(t,t*2).map(n=>({originalName:n.name,variable:n.tensor.variable(a)})),this.centered&&(this.accumulatedMeanGrads=e.slice(t*2,t*3).map(n=>({originalName:n.name,variable:n.tensor.variable(a)})))}getConfig(){return{learningRate:this.learningRate,decay:this.decay,momentum:this.momentum,epsilon:this.epsilon,centered:this.centered}}static fromConfig(e,t){return new e(t.learningRate,t.decay,t.momentum,t.epsilon,t.centered)}},o$=[O3,z3,L3,W3,B3,V3,qh];function l$(){for(let e of o$)e7(e)}var Kn={};Ze(Kn,{browserFiles:()=>m$,browserHTTPRequest:()=>w$,concatenateArrayBuffers:()=>j1,copyModel:()=>nN,decodeWeights:()=>bA,encodeWeights:()=>FT,fromMemory:()=>I$,fromMemorySync:()=>s7,getLoadHandlers:()=>UT,getModelArtifactsForJSON:()=>q1,getModelArtifactsForJSONSync:()=>wA,getModelArtifactsInfoForJSON:()=>Cp,getSaveHandlers:()=>VT,getWeightSpecs:()=>kA,http:()=>G3,isHTTPScheme:()=>r1,listModels:()=>tN,loadWeights:()=>x$,moveModel:()=>rN,registerLoadRouter:()=>BT,registerSaveRouter:()=>WT,removeModel:()=>aN,weightsLoaderFactory:()=>a7,withSaveHandler:()=>S$,withSaveHandlerSync:()=>C$});var u$="model",d$=".json",p$=".weights.bin";function O5(e){return new Promise(t=>setTimeout(t)).then(e)}var Vl=class{constructor(e){if(!W().getBool("IS_BROWSER"))throw new Error("browserDownloads() cannot proceed because the current environment is not a browser.");e.startsWith(Vl.URL_SCHEME)&&(e=e.slice(Vl.URL_SCHEME.length)),(e==null||e.length===0)&&(e=u$),this.modelJsonFileName=e+d$,this.weightDataFileName=e+p$}async save(e){if(typeof document=="undefined")throw new Error("Browser downloads are not supported in this environment since `document` is not present");let t=window.URL.createObjectURL(new Blob([e.weightData],{type:"application/octet-stream"}));if(e.modelTopology instanceof ArrayBuffer)throw new Error("BrowserDownloads.save() does not support saving model topology in binary formats yet.");{let a=[{paths:["./"+this.weightDataFileName],weights:e.weightSpecs}],n=vA(e,a),r=window.URL.createObjectURL(new Blob([JSON.stringify(n)],{type:"application/json"})),s=this.modelJsonAnchor==null?document.createElement("a"):this.modelJsonAnchor;if(s.download=this.modelJsonFileName,s.href=r,await O5(()=>s.dispatchEvent(new MouseEvent("click"))),e.weightData!=null){let i=this.weightDataAnchor==null?document.createElement("a"):this.weightDataAnchor;i.download=this.weightDataFileName,i.href=t,await O5(()=>i.dispatchEvent(new MouseEvent("click")))}return{modelArtifactsInfo:Cp(e)}}}};Vl.URL_SCHEME="downloads://";var c$=class{constructor(e){if(e==null||e.length<1)throw new Error(`When calling browserFiles, at least 1 file is required, but received ${e}`);this.jsonFile=e[0],this.weightsFiles=e.slice(1)}async load(){return new Promise((e,t)=>{let a=new FileReader;a.onload=n=>{let r=JSON.parse(n.target.result),s=r.modelTopology;if(s==null){t(new Error(`modelTopology field is missing from file ${this.jsonFile.name}`));return}if(r.weightsManifest==null){t(new Error(`weightManifest field is missing from file ${this.jsonFile.name}`));return}if(this.weightsFiles.length===0){e({modelTopology:s});return}let i=q1(r,o=>this.loadWeights(o));e(i)},a.onerror=n=>t(`Failed to read model topology and weights manifest JSON from file '${this.jsonFile.name}'. BrowserFiles supports loading Keras-style tf.Model artifacts only.`),a.readAsText(this.jsonFile)})}loadWeights(e){let t=[],a=[];for(let s of e)t.push(...s.weights),a.push(...s.paths);let n=this.checkManifestAndWeightFiles(e),r=a.map(s=>this.loadWeightsFile(s,n[s]));return Promise.all(r).then(s=>[t,j1(s)])}loadWeightsFile(e,t){return new Promise((a,n)=>{let r=new FileReader;r.onload=s=>{let i=s.target.result;a(i)},r.onerror=s=>n(`Failed to weights data from file of path '${e}'.`),r.readAsArrayBuffer(t)})}checkManifestAndWeightFiles(e){let t=[],a=this.weightsFiles.map(r=>$5(r.name)),n={};for(let r of e)r.paths.forEach(s=>{let i=$5(s);if(t.indexOf(i)!==-1)throw new Error(`Duplicate file basename found in weights manifest: '${i}'`);if(t.push(i),a.indexOf(i)===-1)throw new Error(`Weight file with basename '${i}' is not provided.`);n[s]=this.weightsFiles[a.indexOf(i)]});if(t.length!==this.weightsFiles.length)throw new Error(`Mismatch in the number of files in weights manifest (${t.length}) and the number of weight files provided (${this.weightsFiles.length}).`);return n}},h$=e=>W().getBool("IS_BROWSER")&&!Array.isArray(e)&&e.startsWith(Vl.URL_SCHEME)?f$(e.slice(Vl.URL_SCHEME.length)):null;_t.registerSaveRouter(h$);function f$(e="model"){return new Vl(e)}function m$(e){return new c$(e)}var g$=class{constructor(e){if(this.shards=[],this.previousShardIndex=0,e instanceof Array||(e=[e]),e=e.map(a=>Zt(a)?a.buffer:a),e.length===0)return;this.bufferUniformSize=e[0].byteLength;let t=0;for(let a=0;a<e.length;a++){let n=e[a];a!==e.length-1&&n.byteLength!==this.bufferUniformSize&&(this.bufferUniformSize=void 0);let r=t+n.byteLength;this.shards.push({buffer:n,start:t,end:r}),t=r}this.shards.length===0&&(this.byteLength=0),this.byteLength=this.shards[this.shards.length-1].end}slice(e=0,t=this.byteLength){if(e=isNaN(Number(e))?0:e,t=isNaN(Number(t))?0:t,e=Math.max(0,e),t=Math.min(this.byteLength,t),t<=e)return new ArrayBuffer(0);let a=this.findShardForByte(e);if(a===-1)throw new Error(`Could not find start shard for byte ${e}`);let n=t-e,r=new ArrayBuffer(n),s=new Uint8Array(r),i=0;for(let o=a;o<this.shards.length;o++){let l=this.shards[o],u=e+i-l.start,p=i,c=Math.min(t,l.end)-l.start,d=new Uint8Array(l.buffer.slice(u,c));if(s.set(d,p),i+=d.length,t<l.end)break}return r}findShardForByte(e){if(this.shards.length===0||e<0||e>=this.byteLength)return-1;if(this.bufferUniformSize!=null)return this.previousShardIndex=Math.floor(e/this.bufferUniformSize),this.previousShardIndex;function t(n){return e<n.start?-1:e>=n.end?1:0}if(t(this.shards[this.previousShardIndex])===0)return this.previousShardIndex;let a=y$(this.shards,t);return a===-1?-1:(this.previousShardIndex=a,this.previousShardIndex)}};function y$(e,t){let a=0,n=e.length;for(;a<=n;){let r=Math.floor((n-a)/2)+a,s=t(e[r]);if(s===0)return r;s<0?n=r:a=r+1}return-1}function z5(e,t,a,n){i(e),a=a==null?0:a,n=n==null?1:n,o(a,n);let r=0,s=l=>(l.then(u=>{let p=a+ ++r/e.length*(n-a);return t(p),u}),l);function i(l){D(l!=null&&Array.isArray(l)&&l.length>0,()=>"promises must be a none empty array")}function o(l,u){D(l>=0&&l<=1,()=>`Progress fraction must be in range [0, 1], but got startFraction ${l}`),D(u>=0&&u<=1,()=>`Progress fraction must be in range [0, 1], but got endFraction ${u}`),D(u>=l,()=>`startFraction must be no more than endFraction, but got startFraction ${l} and endFraction ${u}`)}return Promise.all(e.map(s))}async function t7(e,t){t==null&&(t={});let a=t.fetchFunc==null?W().platform.fetch:t.fetchFunc,n=e.map(u=>a(u,t.requestInit,{isBinary:!0})),r=0,s=.5,i=(t.onProgress==null?await Promise.all(n):await z5(n,t.onProgress,r,s)).map(u=>u.arrayBuffer()),o=.5,l=1;return t.onProgress==null?await Promise.all(i):await z5(i,t.onProgress,o,l)}async function x$(e,t="",a,n){return a7(r=>t7(r,{requestInit:n}))(e,t,a)}function a7(e){return async(t,a="",n)=>{let r=t.map(()=>!1),s={},i=n!=null?n.map(()=>!1):[],o=[];if(t.forEach((h,f)=>{let m=0;h.weights.forEach(g=>{let y="quantization"in g?g.quantization.dtype:g.dtype,x=J2[y]*vt(g.shape),A=()=>{r[f]=!0,s[f]==null&&(s[f]=[]),s[f].push({manifestEntry:g,groupOffset:m,sizeBytes:x})};n!=null?n.forEach((b,w)=>{b===g.name&&(A(),i[w]=!0)}):A(),o.push(g.name),m+=x})}),!i.every(h=>h)){let h=n.filter((f,m)=>!i[m]);throw new Error(`Could not find weights in manifest with names: ${h.join(", ")}.
|
|
Manifest JSON has weights with names: ${o.join(", ")}.`)}let l=r.reduce((h,f,m)=>(f&&h.push(m),h),[]),u=[];l.forEach(h=>{t[h].paths.forEach(f=>{let m=a+(a.endsWith("/")?"":"/")+f;u.push(m)})});let p=await e(u),c={},d=0;return l.forEach(h=>{let f=t[h].paths.length,m=new g$(p.slice(d,d+f));s[h].forEach(g=>{let y=m.slice(g.groupOffset,g.groupOffset+g.sizeBytes),x=bA(y,[g.manifestEntry]);for(let A in x)c[A]=x[A]}),d+=f}),c}}var A$="application/octet-stream",b$="application/json",U3=class{constructor(e,t){if(this.DEFAULT_METHOD="POST",t==null&&(t={}),this.weightPathPrefix=t.weightPathPrefix,this.onProgress=t.onProgress,this.weightUrlConverter=t.weightUrlConverter,t.fetchFunc!=null?(D(typeof t.fetchFunc=="function",()=>"Must pass a function that matches the signature of `fetch` (see https://developer.mozilla.org/en-US/docs/Web/API/Fetch_API)"),this.fetch=t.fetchFunc):this.fetch=W().platform.fetch,D(e!=null&&e.length>0,()=>"URL path for http must not be null, undefined or empty."),Array.isArray(e)&&D(e.length===2,()=>`URL paths for http must have a length of 2, (actual length is ${e.length}).`),this.path=e,t.requestInit!=null&&t.requestInit.body!=null)throw new Error("requestInit is expected to have no pre-existing body, but has one.");this.requestInit=t.requestInit||{}}async save(e){if(e.modelTopology instanceof ArrayBuffer)throw new Error("BrowserHTTPRequest.save() does not support saving model topology in binary formats yet.");let t=Object.assign({method:this.DEFAULT_METHOD},this.requestInit);t.body=new FormData;let a=[{paths:["./model.weights.bin"],weights:e.weightSpecs}],n=vA(e,a);t.body.append("model.json",new Blob([JSON.stringify(n)],{type:b$}),"model.json"),e.weightData!=null&&t.body.append("model.weights.bin",new Blob([e.weightData],{type:A$}),"model.weights.bin");let r=await this.fetch(this.path,t);if(r.ok)return{modelArtifactsInfo:Cp(e),responses:[r]};throw new Error(`BrowserHTTPRequest.save() failed due to HTTP response status ${r.status}.`)}async load(){let e=await this.fetch(this.path,this.requestInit);if(!e.ok)throw new Error(`Request to ${this.path} failed with status code ${e.status}. Please verify this URL points to the model JSON of the model to load.`);let t;try{t=await e.json()}catch(r){let s=`Failed to parse model JSON of response from ${this.path}.`;throw this.path.endsWith(".pb")?s+=" Your path contains a .pb file extension. Support for .pb models have been removed in TensorFlow.js 1.0 in favor of .json models. You can re-convert your Python TensorFlow model using the TensorFlow.js 1.0 conversion scripts or you can convert your.pb models with the 'pb2json'NPM script in the tensorflow/tfjs-converter repository.":s+=" Please make sure the server is serving valid JSON for this request.",new Error(s)}let a=t.modelTopology,n=t.weightsManifest;if(a==null&&n==null)throw new Error(`The JSON from HTTP path ${this.path} contains neither model topology or manifest for weights.`);return q1(t,r=>this.loadWeights(r))}async loadWeights(e){let t=Array.isArray(this.path)?this.path[1]:this.path,[a,n]=v$(t),r=this.weightPathPrefix||a,s=kA(e),i=[],o=[];for(let u of e)for(let p of u.paths)this.weightUrlConverter!=null?o.push(this.weightUrlConverter(p)):i.push(r+p+n);this.weightUrlConverter&&i.push(...await Promise.all(o));let l=await t7(i,{requestInit:this.requestInit,fetchFunc:this.fetch,onProgress:this.onProgress});return[s,j1(l)]}};U3.URL_SCHEME_REGEX=/^https?:\/\//;function v$(e){let t=e.lastIndexOf("/"),a=e.lastIndexOf("?"),n=e.substring(0,t),r=a>t?e.substring(a):"";return[n+"/",r]}function r1(e){return e.match(U3.URL_SCHEME_REGEX)!=null}var n7=(e,t)=>{if(typeof fetch=="undefined"&&(t==null||t.fetchFunc==null))return null;{let a=!0;if(Array.isArray(e)?a=e.every(n=>r1(n)):a=r1(e),a)return G3(e,t)}return null};_t.registerSaveRouter(n7);_t.registerLoadRouter(n7);function G3(e,t){return new U3(e,t)}function w$(e,t){return G3(e,t)}var O2=class{constructor(e){this.modelArtifacts=e}load(){return this.modelArtifacts}},r7=class{constructor(e){this.saveHandler=e}save(e){return this.saveHandler(e)}},k$=class{constructor(e){e.load&&(this.load=()=>Promise.resolve(e.load())),e.save&&(this.save=t=>Promise.resolve(e.save(t)))}};function I$(e,t,a,n){let r=arguments;return new k$(s7(...r))}function s7(e,t,a,n){return arguments.length===1?e.modelTopology!=null||e.weightSpecs!=null?new O2(e):(console.warn("Please call tf.io.fromMemory() with only one argument. The argument should be of type ModelArtifacts. The multi-argument signature of tf.io.fromMemory() has been deprecated and will be removed in a future release."),new O2({modelTopology:e})):(console.warn("Please call tf.io.fromMemory() with only one argument. The argument should be of type ModelArtifacts. The multi-argument signature of tf.io.fromMemory() has been deprecated and will be removed in a future release."),new O2({modelTopology:e,weightSpecs:t,weightData:a,trainingConfig:n}))}function S$(e){return new r7(e)}function C$(e){return new r7(e)}var i7={};Ze(i7,{confusionMatrix:()=>N$});function T$(e,t,a){let n=R(e,"labels","confusionMatrix"),r=R(t,"predictions","confusionMatrix");D(a==null||a>0&&Number.isInteger(a),()=>`If provided, numClasses must be a positive integer, but got ${a}`),D(n.rank===1,()=>`Expected the rank of labels to be 1, but got ${n.rank}`),D(r.rank===1,()=>`Expected the rank of predictions to be 1, but got ${r.rank}`),D(n.shape[0]===r.shape[0],()=>`Mismatch in the number of examples: ${n.shape[0]} vs. ${r.shape[0]}. Labels and predictions should have the same number of elements.`),D(a>0&&Number.isInteger(a),()=>`numClasses is required to be a positive integer, but got ${a}`);let s=oh(Ke(n,"int32"),a),i=oh(Ke(r,"int32"),a),o=js(s),l=pt(o,i);return Ke(l,"int32")}var N$=z({confusionMatrix_:T$}),Rr={};Ze(Rr,{fromPixels:()=>D$,fromPixelsAsync:()=>_$,toPixels:()=>$$});var Fs;function o7(e,t=3){if(t>4)throw new Error("Cannot construct Tensor with more than 4 channels from pixels.");if(e==null)throw new Error("pixels passed to tf.browser.fromPixels() can not be null");let a=!1,n=!1,r=!1,s=!1,i=!1,o=!1;if(e.data instanceof Uint8Array)a=!0;else if(typeof ImageData!="undefined"&&e instanceof ImageData)n=!0;else if(typeof HTMLVideoElement!="undefined"&&e instanceof HTMLVideoElement)r=!0;else if(typeof HTMLImageElement!="undefined"&&e instanceof HTMLImageElement)s=!0;else if(e.getContext!=null)i=!0;else if(typeof ImageBitmap!="undefined"&&e instanceof ImageBitmap)o=!0;else throw new Error(`pixels passed to tf.browser.fromPixels() must be either an HTMLVideoElement, HTMLImageElement, HTMLCanvasElement, ImageData in browser, or OffscreenCanvas, ImageData in webworker or {data: Uint32Array, width: number, height: number}, but was ${e.constructor.name}`);if(ah(Ed,L.backendName)!=null){let d={pixels:e},h={numChannels:t};return L.runKernel(Ed,d,h)}let[l,u]=r?[e.videoWidth,e.videoHeight]:[e.width,e.height],p;if(i)p=e.getContext("2d").getImageData(0,0,l,u).data;else if(n||a)p=e.data;else if(s||r||o){if(Fs==null)if(typeof document=="undefined")if(typeof OffscreenCanvas!="undefined"&&typeof OffscreenCanvasRenderingContext2D!="undefined")Fs=new OffscreenCanvas(1,1).getContext("2d");else throw new Error("Cannot parse input in current context. Reason: OffscreenCanvas Context2D rendering is not supported.");else Fs=document.createElement("canvas").getContext("2d",{willReadFrequently:!0});Fs.canvas.width=l,Fs.canvas.height=u,Fs.drawImage(e,0,0,l,u),p=Fs.getImageData(0,0,l,u).data}let c;if(t===4)c=new Int32Array(p);else{let d=l*u;c=new Int32Array(d*t);for(let h=0;h<d;h++)for(let f=0;f<t;++f)c[h*t+f]=p[h*4+f]}return M3(c,[u,l,t],"int32")}function R$(e){return e!=null&&e.data instanceof Uint8Array}function E$(){return typeof window!="undefined"&&typeof ImageBitmap!="undefined"&&window.hasOwnProperty("createImageBitmap")}function M$(e){return e!=null&&e.width!==0&&e.height!==0}function F$(e){return E$()&&!(e instanceof ImageBitmap)&&M$(e)&&!R$(e)}async function _$(e,t=3){let a=null;if(W().getBool("WRAP_TO_IMAGEBITMAP")&&F$(e)){let n;try{n=await createImageBitmap(e,{premultiplyAlpha:"none"})}catch(r){n=null}n!=null&&n.width===e.width&&n.height===e.height?a=n:a=e}else a=e;return o7(a,t)}async function $$(e,t){let a=R(e,"img","toPixels");if(!(e instanceof gt)){let u=a;a=Ke(u,"int32"),u.dispose()}if(a.rank!==2&&a.rank!==3)throw new Error(`toPixels only supports rank 2 or 3 tensors, got rank ${a.rank}.`);let[n,r]=a.shape.slice(0,2),s=a.rank===2?1:a.shape[2];if(s>4||s===2)throw new Error(`toPixels only supports depth of size 1, 3 or 4 but got ${s}`);if(a.dtype!=="float32"&&a.dtype!=="int32")throw new Error(`Unsupported type for toPixels: ${a.dtype}. Please use float32 or int32 tensors.`);let i=await a.data(),o=a.dtype==="float32"?255:1,l=new Uint8ClampedArray(r*n*4);for(let u=0;u<n*r;++u){let p=[0,0,0,255];for(let d=0;d<s;d++){let h=i[u*s+d];if(a.dtype==="float32"){if(h<0||h>1)throw new Error(`Tensor values for a float32 Tensor must be in the range [0 - 1] but encountered ${h}.`)}else if(a.dtype==="int32"&&(h<0||h>255))throw new Error(`Tensor values for a int32 Tensor must be in the range [0 - 255] but encountered ${h}.`);s===1?(p[0]=h*o,p[1]=h*o,p[2]=h*o):p[d]=h*o}let c=u*4;l[c+0]=Math.round(p[0]),l[c+1]=Math.round(p[1]),l[c+2]=Math.round(p[2]),l[c+3]=Math.round(p[3])}if(t!=null){t.width=r,t.height=n;let u=t.getContext("2d"),p=new ImageData(l,r,n);u.putImageData(p,0,0)}return a!==e&&a.dispose(),l}var D$=z({fromPixels_:o7}),H3={};Ze(H3,{prepareAndValidate:()=>l7});function l7(e,t){let a=e.shape.length,n=t.shape.length;if(a<1)throw new Error(`tf.gatherND() expects the input to be rank 1 or higher, but the rank was ${a}.`);if(n<1)throw new Error(`tf.gatherND() expects the indices to be rank 1 or higher, but the rank was ${n}.`);if(t.dtype!=="int32")throw new Error(`tf.gatherND() expects the indices to be int32 type, but the dtype was ${t.dtype}.`);if(t.shape[n-1]>a)throw new Error(`index innermost dimension length must be <= tensor rank; saw: ${t.shape[n-1]} vs. ${a}`);if(vt(e.shape)===0)throw new Error(`Requested more than 0 entries, but input is empty. Input shape: ${e.shape}.`);let r=t.shape,s=r[r.length-1],i=1;for(let c=0;c<r.length-1;++c)i*=r[c];let o=e.shape,l=r.slice();l.pop();let u=1;for(let c=s;c<a;++c)u*=o[c],l.push(o[c]);let p=[...Yl(e.shape).map(c=>c/u),1].slice(0,s);return[l,i,u,p]}var Nt={};Ze(Nt,{assertParamsValid:()=>O$,computeFlatOffset:()=>V$,computeOutShape:()=>L$,getNormalizedAxes:()=>W$,isSliceContinous:()=>B$,maskToAxes:()=>z$,parseSliceParams:()=>U$,sliceInfo:()=>G$,startForAxis:()=>m7,startIndicesWithElidedDims:()=>c7,stopForAxis:()=>g7,stopIndicesWithElidedDims:()=>h7,stridesForAxis:()=>f7,stridesWithElidedDims:()=>u7});var s1=-2,P$=-1;function O$(e,t,a){let n=e.shape.length;D(n===t.length,()=>`Error in slice${n}D: Length of begin ${t} must match the rank of the array (${n}).`),D(n===a.length,()=>`Error in slice${n}D: Length of size ${a} must match the rank of the array (${n}).`);for(let r=0;r<n;++r)D(t[r]+a[r]<=e.shape[r],()=>`Error in slice${n}D: begin[${r}] + size[${r}] (${t[r]+a[r]}) would overflow input.shape[${r}] (${e.shape[r]})`)}function z$(e){let t=[],a=0;for(;e>0;)e&1&&t.push(a),e/=2,a++;return t}function L$(e,t,a){let n=[];for(let r=0;r<e.length;r++)n[r]=Math.ceil((t[r]-e[r])/a[r]);return n}function u7(e,t,a,n){let r=[...e];for(let s=r.length;s<n.length;s++)r.push(1);for(let s=0;s<a;s++)s===0?r[t]=1:(r.splice(t,0,1),r.pop());return r}function d7(e,t,a){return a<=e?a:a-(t-1)}function p7(e,t){let a=[];for(let n=0;n<e;n++)a.push(t+n);return a}function W$(e,t,a,n,r,s,i,o,l){let u=e.length,p=new Array(u),c=new Array(u),d=new Array(u);if(t.length&&a>0){let h=t[0],f=a+1;p=c7(i,h,f,n,e),c=h7(o,h,f,r,e),d=u7(s,h,f,e)}else for(let h=0;h<u;h++)p[h]=m7(i,n,s,e,h,l),c[h]=g7(o,r,s,e,h,l),d[h]=f7(s,h,l);return{begin:p,end:c,strides:d}}function c7(e,t,a,n,r){let s=[...r],i=p7(a,t);for(let o=0;o<s.length;o++)if(i.indexOf(o)>-1)s[o]=0;else{let l=d7(t,a,o),u=n[l];e&1<<l&&(u=0),s[o]=u}return s}function h7(e,t,a,n,r){let s=[...r],i=p7(a,t);for(let o=0;o<s.length;o++)if(i.indexOf(o)>-1)s[o]=Number.MAX_SAFE_INTEGER;else{let l=d7(t,a,o),u=n[l];e&1<<l&&(u=Number.MAX_SAFE_INTEGER),s[o]=u}for(let o=0;o<s.length;o++){let l=r[o];s[o]<0&&(s[o]+=l),s[o]=Rd(0,s[o],r[o])}return s}function f7(e,t,a){let n=e[t];return(a&1<<t||n==null)&&(n=1),n}function m7(e,t,a,n,r,s){let i=t[r],o=a[r]||1;(e&1<<r||s&1<<r||i==null)&&(o>0?i=Number.MIN_SAFE_INTEGER:i=Number.MAX_SAFE_INTEGER);let l=n[r];return i<0&&(i+=l),i=Rd(0,i,l-1),i}function g7(e,t,a,n,r,s){let i=t[r],o=a[r]||1;(e&1<<r||s&1<<r||i==null)&&(o>0?i=Number.MAX_SAFE_INTEGER:i=Number.MIN_SAFE_INTEGER);let l=n[r];return i<0&&(i+=l),o>0?i=Rd(0,i,l):i=Rd(-1,i,l-1),i}function B$(e,t,a){let n=a.length;for(let r=0;r<a.length;r++)if(a[r]>1){n=r;break}for(let r=n+1;r<a.length;r++)if(t[r]>0||a[r]!==e[r])return!1;return!0}function V$(e,t){let a=e.length>0?e[e.length-1]:1;for(let n=0;n<e.length-1;n++)a+=e[n]*t[n];return a}function U$(e,t,a){let n,r=e.shape.length;typeof t=="number"?n=[t,...new Array(r-1).fill(0)]:t.length<r?n=t.concat(new Array(r-t.length).fill(0)):n=t.slice(),n.forEach(i=>{D(i!==-1,()=>"slice() does not support negative begin indexing.")});let s;return a==null?s=new Array(r).fill(-1):typeof a=="number"?s=[a,...new Array(r-1).fill(-1)]:a.length<r?s=a.concat(new Array(r-a.length).fill(-1)):s=a,s=s.map((i,o)=>i>=0?i:(D(i===-1,()=>`Negative size values should be exactly -1 but got ${i} for the slice() size at index ${o}.`),e.shape[o]-n[o])),[n,s]}function G$(e,t,a,n,r,s,i,o,l){let u;if(n==null?(u=new Array(t.length),u.fill(1)):u=n,i!=null&&i&i-1)throw new Error("Multiple ellipses in slice is not allowed.");let p=!1,c={dims:u.length,numAddAxisAfterEllipsis:0,begin:t.slice(),end:a.slice(),strides:u.slice(),beginMask:r,endMask:s,ellipsisMask:i,newAxisMask:o,shrinkAxisMask:l};for(let x=0;x<c.dims;x++)p&&1<<x&o&&c.numAddAxisAfterEllipsis++,1<<x&i&&(p=!0);p||(c.ellipsisMask|=1<<c.dims,c.dims++);let d={dims:e.length,beginMask:0,endMask:0,beginValid:!1,endValid:!1};H$(c,d);let h=!0,f=!0,m=!0,g=[],y=[];for(let x=0;x<e.length;++x){if(d.strides[x]===0)throw Error(`strides[${x}] must be non-zero`);let A=!!(d.shrinkAxisMask&1<<x),b=e[x];if(b===-1){g.push(A?1:-1);continue}let w=[d.beginMask&1<<x,d.endMask&1<<x],C=[d.strides[x]>0?0:-1,d.strides[x]>0?b:b-1];if(A&&d.strides[x]<=0)throw Error("only stride 1 allowed on non-range indexing.");m=m&&d.strides[x]===1;let T=!!(d.beginMask&1<<x&&d.endMask&1<<x);if(d.beginValid&&d.endValid){if(A){let E=d.begin[x]<0?b+d.begin[x]:d.begin[x];if(d.begin[x]=E,d.end[x]=d.begin[x]+1,E<0||E>=b)throw Error(`slice index ${d.begin[x]} of dimension ${x} out of bounds.`)}else d.begin[x]=L5(d.begin[x],0,d.strides[x],b,w,C),d.end[x]=L5(d.end[x],1,d.strides[x],b,w,C);let F=d.strides[x]===1&&d.begin[x]===0&&d.end[x]===b;h=h&&F,f=f&&(x===0&&d.strides[x]===1||F)}else h=h&&d.strides[x]===1&&T,f=f&&(x===0&&d.strides[x]===1||T);let N,M=!1;if(d.beginValid&&d.endValid?(N=d.end[x]-d.begin[x],M=!0):A?(N=1,M=!0):T&&b>=0&&(d.strides[x]<0?N=-b:N=b,M=!0),M){let F;N===0||N<0!=d.strides[x]<0?F=0:F=Math.trunc(N/d.strides[x])+(N%d.strides[x]!==0?1:0),g.push(F)}else g.push(-1)}for(let x=0;x<d.finalShapeGatherIndices.length;++x){let A=d.finalShapeGatherIndices[x];A>=0?y.push(g[A]):A===s1&&y.push(1)}return{finalShapeSparse:y.filter((x,A)=>d.finalShapeGatherIndices[A]!==s1),finalShape:y,isIdentity:h,sliceDim0:f,isSimpleSlice:m,begin:d.begin,end:d.end,strides:d.strides}}function H$(e,t){t.beginMask=0,t.endMask=0,t.shrinkAxisMask=0;let a=0;t.beginValid=e.begin!=null,t.endValid=e.end!=null,t.begin=new Array(t.dims),t.end=new Array(t.dims),t.strides=new Array(t.dims),t.finalShapeGatherIndices=[],t.finalShapeGatherIndicesSparse=[],t.inputShapeGatherIndicesSparse=new Array(t.dims);for(let n=0;n<e.dims;n++)if(1<<n&e.ellipsisMask){let r=Math.min(t.dims-(e.dims-n)+1+e.numAddAxisAfterEllipsis,t.dims);for(;a<r;a++)t.begin[a]=0,t.end[a]=0,t.strides[a]=1,t.beginMask|=1<<a,t.endMask|=1<<a,t.finalShapeGatherIndices.push(a),t.finalShapeGatherIndicesSparse.push(-1),t.inputShapeGatherIndicesSparse[a]=n}else if(1<<n&e.newAxisMask)t.finalShapeGatherIndices.push(s1),t.finalShapeGatherIndicesSparse.push(-1);else{if(a===t.begin.length)throw Error(`Index out of range using input dim ${a}; input has only ${t.dims} dims, ${t.begin.length}.`);e.begin!=null&&(t.begin[a]=e.begin[n]),e.end!=null&&(t.end[a]=e.end[n]),t.strides[a]=e.strides[n],e.beginMask&1<<n&&(t.beginMask|=1<<a),e.endMask&1<<n&&(t.endMask|=1<<a),e.shrinkAxisMask&1<<n?(t.finalShapeGatherIndices.push(P$),t.finalShapeGatherIndicesSparse.push(-1),t.shrinkAxisMask|=1<<a):(t.finalShapeGatherIndices.push(a),t.finalShapeGatherIndicesSparse.push(n)),t.inputShapeGatherIndicesSparse[a]=n,a++}}function L5(e,t,a,n,r,s){if(r[t])return a>0?s[t]:s[t+1&1];{let i=e<0?n+e:e;return i<s[0]?s[0]:i>s[1]?s[1]:i}}var j3="4.5.0",y7=class{static sgd(e){return new qh(e)}static momentum(e,t,a=!1){return new B3(e,t,a)}static rmsprop(e,t=.9,a=0,n=null,r=!1){return new V3(e,t,a,n,r)}static adam(e=.001,t=.9,a=.999,n=null){return new L3(e,t,a,n)}static adadelta(e=.001,t=.95,a=null){return new O3(e,t,a)}static adamax(e=.002,t=.9,a=.999,n=null,r=0){return new W3(e,t,a,n,r)}static adagrad(e,t=.1){return new z3(e,t)}},j$=y7,q$=(()=>typeof requestAnimationFrame!="undefined"?requestAnimationFrame:typeof setImmediate!="undefined"?setImmediate:e=>e())();function x7(){return new Promise(e=>q$(()=>e()))}var S={};Ze(S,{ERF_A1:()=>dD,ERF_A2:()=>pD,ERF_A3:()=>cD,ERF_A4:()=>hD,ERF_A5:()=>fD,ERF_P:()=>uD,PARALLELIZE_THRESHOLD:()=>q3,RowPartitionType:()=>Xn,SELU_SCALE:()=>lD,SELU_SCALEALPHA:()=>oD,applyActivation:()=>Hh,assertAndGetBroadcastShape:()=>Bt,assertAxesAreInnerMostDims:()=>BR,assertParamsConsistent:()=>X$,assignToTypedArray:()=>bD,axesAreInnerMostDims:()=>s3,calculateShapes:()=>S4,checkEinsumDimSizes:()=>CD,checkPadOnDimRoundingMode:()=>Tn,combineLocations:()=>vb,combineRaggedTensorToTensorShapes:()=>Y$,complexWithEvenIndex:()=>yD,complexWithOddIndex:()=>xD,computeConv2DInfo:()=>Ep,computeConv3DInfo:()=>VA,computeDefaultPad:()=>J1,computeDilation2DInfo:()=>ON,computeOptimalWindowSize:()=>eD,computeOutAndReduceShapes:()=>WR,computeOutShape:()=>K$,computePool2DInfo:()=>BA,computePool3DInfo:()=>zN,convertConv2DDataFormat:()=>UA,decodeEinsumEquation:()=>ID,eitherStridesOrDilationsAreOne:()=>Tr,expandShapeToKeepDim:()=>_p,exponent:()=>wD,exponents:()=>vD,fromStringArrayToUint8:()=>qD,fromUint8ToStringArray:()=>jD,getAxesPermutation:()=>VR,getBroadcastDims:()=>mb,getComplexWithIndex:()=>AD,getEinsumComputePath:()=>TD,getEinsumPermutation:()=>SD,getFusedBiasGradient:()=>Gh,getFusedDyActivation:()=>Uh,getImageCenter:()=>tD,getInnerMostAxes:()=>GR,getPermuted:()=>nD,getRaggedRank:()=>J$,getReductionAxes:()=>a3,getReshaped:()=>aD,getReshapedPermuted:()=>rD,getRowPartitionTypesHelper:()=>Z$,getSliceBeginCoords:()=>sD,getSliceSize:()=>iD,getSparseFillEmptyRowsIndicesDenseShapeMismatch:()=>MD,getSparseFillEmptyRowsNegativeIndexErrorMessage:()=>FD,getSparseFillEmptyRowsOutOfRangeIndexErrorMessage:()=>_D,getSparseReshapeEmptyTensorZeroOutputDimErrorMessage:()=>PD,getSparseReshapeInputOutputMismatchErrorMessage:()=>zD,getSparseReshapeInputOutputMultipleErrorMessage:()=>OD,getSparseReshapeMultipleNegativeOneOutputDimErrorMessage:()=>$D,getSparseReshapeNegativeOutputDimErrorMessage:()=>DD,getSparseSegmentReductionIndicesOutOfRangeErrorMessage:()=>VD,getSparseSegmentReductionNegativeSegmentIdsErrorMessage:()=>LD,getSparseSegmentReductionNonIncreasingSegmentIdsErrorMessage:()=>WD,getSparseSegmentReductionSegmentIdOutOfRangeErrorMessage:()=>BD,getUndoAxesPermutation:()=>UR,isIdentityPermutation:()=>ND,log:()=>JC,mergeRealAndImagArrays:()=>mD,prepareAndValidate:()=>l7,prepareSplitSize:()=>ED,segment_util:()=>A7,shouldFuse:()=>jh,slice_util:()=>Nt,splitRealAndImagArrays:()=>gD,stridesOrDilationsArePositive:()=>Hs,tupleValuesAreOne:()=>zd,upcastType:()=>pa,validateDefaultValueShape:()=>Q$,validateInput:()=>Bh,validateUpdateShape:()=>F3,warn:()=>Wr});function X$(e,t){let a=e[0].length;e.forEach((r,s)=>{D(r.length===a,()=>`Error in concat${a}D: rank of tensors[${s}] must be the same as the rank of the rest (${a})`)}),D(t>=0&&t<a,()=>`Error in concat${a}D: axis must be between 0 and ${a-1}.`);let n=e[0];e.forEach((r,s)=>{for(let i=0;i<a;i++)D(i===t||r[i]===n[i],()=>`Error in concat${a}D: Shape of tensors[${s}] (${r}) does not match the shape of the rest (${n}) along the non-concatenated axis ${s}.`)})}function K$(e,t){let a=e[0].slice();for(let n=1;n<e.length;n++)a[t]+=e[n][t];return a}var Xn;(function(e){e[e.FIRST_DIM_SIZE=0]="FIRST_DIM_SIZE",e[e.VALUE_ROWIDS=1]="VALUE_ROWIDS",e[e.ROW_LENGTHS=2]="ROW_LENGTHS",e[e.ROW_SPLITS=3]="ROW_SPLITS",e[e.ROW_LIMITS=4]="ROW_LIMITS",e[e.ROW_STARTS=5]="ROW_STARTS"})(Xn||(Xn={}));function Y$(e,t,a){let n=new Array;if(a==null&&t==null)return n;if(t==null)for(;n.length<e+a.length;)n.push(-1);else n=t.slice();if(a==null)return n;if(e+a.length!==n.length)throw new Error(`rt input.shape and shape=${t} are incompatible: rt input.rank = ${e+a.length}, but shape.rank = ${n.length}`);for(let r=1;r<a.length;++r){let s=a[r],i=n[n.length-a.length+r],o=n[i];if(s>=0)if(o>=0){if(o!==s)throw new Error(`rt input.shape and shape=${t} are incompatible: rt input.shape[${r+e}] = ${s} but shape[${r+e}] = ${o}`)}else n[i]=s}return n}function Z$(e){let t={FIRST_DIM_SIZE:Xn.FIRST_DIM_SIZE,VALUE_ROWIDS:Xn.VALUE_ROWIDS,ROW_LENGTHS:Xn.ROW_LENGTHS,ROW_SPLITS:Xn.ROW_SPLITS,ROW_LIMITS:Xn.ROW_LIMITS,ROW_STARTS:Xn.ROW_STARTS},a=[];for(let n of e)if(n in t)a.push(t[n]);else break;return a}function J$(e){return e.length===0?0:e[0]===Xn.FIRST_DIM_SIZE?e.length-1:e.length}function Q$(e,t){if(e==null||t==null)return;let a=e.length,n=t.length;if(a>=n)throw new Error(`defaultValue.shape=${e} and ragged tensor flatValues.shape=${t}, are incompatible: defaultValue.rank = ${a} must be less than ragged tensor input flatValues.rank = ${n})`);for(let r=0;r<Math.min(a,n-1);++r){let s=e[r],i=t[r+1];if(s>=0&&i>=0&&s!==1&&s!==i)throw new Error(`defaultValue.shape=${e}, and ragged tensor input flatValues.shape=${t} are incompatible: defaultValue.shape[${r-e.length}] = ${s} but ragged tensor input.flatValues.shape[${r-e.length}] = ${i}`)}}var q3=30;function eD(e){return e<=q3?e:th(e,Math.floor(Math.sqrt(e)))}function tD(e,t,a){let n=a*(typeof e=="number"?e:e[0]),r=t*(typeof e=="number"?e:e[1]);return[n,r]}function aD(e,t,a,n=!0){let r=[];if(n)r=r.concat(t.slice(0)),r.push(e[0]/a),r=r.concat(e.slice(1));else{r=r.concat(e[0]);let s=t.length;for(let i=0;i<s;++i)r=r.concat([e[i+1]/t[i],t[i]]);r=r.concat(e.slice(s+1))}return r}function nD(e,t,a=!0){let n=[];if(a){n.push(t);for(let r=t+1;r<e;++r)r<=2*t?(n.push(r),n.push(r-(t+1))):n.push(r)}else{let r=[],s=[];for(let i=1;i<e;++i)i>=t*2+1||i%2===1?s.push(i):r.push(i);n.push(...r),n.push(0),n.push(...s)}return n}function rD(e,t,a,n=!0){let r=[];n?r.push(e[0]/a):r.push(e[0]*a);for(let s=1;s<e.length;++s)s<=t.length?n?r.push(t[s-1]*e[s]):r.push(e[s]/t[s-1]):r.push(e[s]);return r}function sD(e,t){let a=[0];for(let n=0;n<t;++n)a.push(e[n][0]);return a}function iD(e,t,a){let n=e.slice(0,1);for(let r=0;r<a;++r)n.push(e[r+1]-t[r][0]-t[r][1]);return n}var oD=1.7580993408473768,lD=1.0507009873554805,uD=.3275911,dD=.254829592,pD=-.284496736,cD=1.421413741,hD=-1.453152027,fD=1.061405429;function mD(e,t){if(e.length!==t.length)throw new Error(`Cannot merge real and imag arrays of different lengths. real:${e.length}, imag: ${t.length}.`);let a=new Float32Array(e.length*2);for(let n=0;n<a.length;n+=2)a[n]=e[n/2],a[n+1]=t[n/2];return a}function gD(e){let t=new Float32Array(e.length/2),a=new Float32Array(e.length/2);for(let n=0;n<e.length;n+=2)t[n/2]=e[n],a[n/2]=e[n+1];return{real:t,imag:a}}function yD(e){let t=Math.ceil(e.length/4),a=new Float32Array(t),n=new Float32Array(t);for(let r=0;r<e.length;r+=4)a[Math.floor(r/4)]=e[r],n[Math.floor(r/4)]=e[r+1];return{real:a,imag:n}}function xD(e){let t=Math.floor(e.length/4),a=new Float32Array(t),n=new Float32Array(t);for(let r=2;r<e.length;r+=4)a[Math.floor(r/4)]=e[r],n[Math.floor(r/4)]=e[r+1];return{real:a,imag:n}}function AD(e,t){let a=e[t*2],n=e[t*2+1];return{real:a,imag:n}}function bD(e,t,a,n){e[n*2]=t,e[n*2+1]=a}function vD(e,t){let a=new Float32Array(e/2),n=new Float32Array(e/2);for(let r=0;r<Math.ceil(e/2);r++){let s=(t?2:-2)*Math.PI*(r/e);a[r]=Math.cos(s),n[r]=Math.sin(s)}return{real:a,imag:n}}function wD(e,t,a){let n=(a?2:-2)*Math.PI*(e/t),r=Math.cos(n),s=Math.sin(n);return{real:r,imag:s}}var z2="->",kD=/->/g,W5=",",B5="...";function ID(e,t){e=e.replace(/\s/g,"");let a=(e.length-e.replace(kD,"").length)/z2.length;if(a<1)throw new Error("Equations without an arrow are not supported.");if(a>1)throw new Error(`Equation must contain exactly one arrow ("${z2}").`);let[n,r]=e.split(z2);D(n.indexOf(B5)===-1,()=>`The ellipsis notation ("${B5}") is not supported yet.`);let s=n.split(W5),i=s.length;if(t!==i)throw new Error(`Expected ${i} input tensors, received ${t}`);if(i>2)throw new Error("Support for more than 2 input tensors is not implemented yet.");let o=[];for(let d=0;d<r.length;++d){let h=r[d];if(!s.some(f=>f.indexOf(h)!==-1))throw new Error(`Output subscripts contain the label ${h} not present in the input subscripts.`);o.indexOf(h)===-1&&o.push(h)}for(let d=0;d<n.length;++d){let h=n[d];o.indexOf(h)===-1&&h!==W5&&o.push(h)}let l=new Array(s.length);for(let d=0;d<i;++d){if(new Set(s[d].split("")).size!==s[d].length)throw new Error(`Found duplicate axes in input component ${s[d]}. Support for duplicate axes in input is not implemented yet.`);l[d]=[];for(let h=0;h<s[d].length;++h)l[d].push(o.indexOf(s[d][h]))}let u=o.length,p=r.length,c=[];for(let d=p;d<u;++d)c.push(d);return{allDims:o,summedDims:c,idDims:l}}function SD(e,t){let a=new Array(e);a.fill(-1);for(let r=0;r<t.length;++r)a[t[r]]=r;let n=[];for(let r=0;r<e;++r)a[r]===-1&&n.push(r);return a=a.filter(r=>r!==-1),{permutationIndices:a,expandDims:n}}function CD(e,t,a){let n=new Array(e);for(let r=0;r<a.length;++r){let s=a[r].shape;for(let i=0;i<t[r].length;++i)n[t[r][i]]===void 0?n[t[r][i]]=s[i]:D(n[t[r][i]]===s[i],()=>`Expected dimension ${n[t[r][i]]} at axis ${i} of input shaped ${JSON.stringify(s)}, but got dimension ${s[i]}`)}}function TD(e,t){let a=e,n=[],r=0;e.length===0&&a.push(-1),r=e.length+1;for(let i=0;i<r;++i)n.push([]);let s=[];for(let i=0;i<a.length;++i){let o=a[i],l=RD(t,o);for(let u of l)s.indexOf(u)===-1&&(n[i].push(u),s.push(u))}return{path:a,steps:n}}function ND(e){return e.every((t,a)=>t===a)}function RD(e,t){let a=[];for(let n=0;n<e.length;++n)(e[n].length===0||e[n].indexOf(t)!==-1||t===-1)&&a.push(n);return a}function ED(e,t,a=0){let n=[];if(typeof t=="number")D(e.shape[a]%t===0,()=>"Number of splits must evenly divide the axis."),n=new Array(t).fill(e.shape[a]/t);else{let r=t.reduce((i,o)=>(o===-1&&(i+=1),i),0);D(r<=1,()=>"There should be only one negative value in split array.");let s=t.indexOf(-1);if(s!==-1){let i=t.reduce((o,l)=>l>0?o+l:o);t[s]=e.shape[a]-i}D(e.shape[a]===t.reduce((i,o)=>i+o),()=>"The sum of sizes must match the size of the axis dimension."),n=t}return n}function MD(e){return`Received SparseTensor with denseShape[0] = 0 but
|
|
indices.shape[0] = ${e}`}function FD(e,t){return`indices(${e}, 0) is invalid: ${t} < 0`}function _D(e,t,a){return`indices(${e}, 0) is invalid: ${t} >= ${a}`}function $D(e,t){return`only one output dimension may be -1, not both ${e} and ${t}`}function DD(e,t){return`size ${e} must be non-negative, not ${t}`}function PD(){return"reshape cannot infer the missing input size for an empty tensor unless all specified input sizes are non-zero"}function OD(e,t){let a=vt(e),n=vt(t);return`Input to reshape is a SparseTensor with ${a}
|
|
dense values, but the requested shape requires a multiple of ${n}. inputShape=${e} outputShape= ${t}`}function zD(e,t){let a=vt(e),n=vt(t);return`Input to reshape is a tensor with ${a} dense values, but the requested shape has ${n}. inputShape=${e} outputShape=${t}`}function LD(){return"segment ids must be >= 0"}function WD(){return"segment ids are not increasing"}function BD(e,t){return`Segment id ${e} out of range [0, ${t}), possibly because segmentIds input is not sorted.`}function VD(e,t,a){return`Bad: indices[${e}] == ${t} out of range [0, ${a})`}var A7={};Ze(A7,{collectGatherOpShapeInfo:()=>HD,computeOutShape:()=>GD,segOpComputeOptimalWindowSize:()=>UD});function UD(e,t){let a=!1,n;for(e<=q3?(n=e,a=!0):n=th(e,Math.floor(Math.sqrt(e)));!a;)n>t||n===e?a=!0:n=th(e,n+1);return n}function GD(e,t,a){let n=[],r=e.length;for(let s=0;s<r;s++)s!==t?n.push(e[s]):n.push(a);return n}function HD(e,t,a,n){let r=t.shape.length,s=e.shape.length;if(n!==0&&(n<-r||n>r))throw new Error(`Expect batchDims in the range of [-${r}, ${r}], but got ${n}`);if(n<0&&(n+=r),n>s)throw new Error(`batchDims (${n}) must be less than rank(x) (
|
|
${s}).`);if(a<n)throw new Error(`batchDims (${n}) must be less than or equal to axis (${a}).`);for(let c=0;c<n;++c)if(e.shape[c]!==t.shape[c])throw new Error(`x.shape[${c}]: ${e.shape[c]} should be equal to indices.shape[${c}]: ${t.shape[c]}.`);let i=e.shape[a],o=[],l=1,u=1,p=1;for(let c=0;c<n;++c)o.push(e.shape[c]),l*=e.shape[c];for(let c=n;c<a;c++)o.push(e.shape[c]),u*=e.shape[c];for(let c=n;c<r;c++)o.push(t.shape[c]);for(let c=a+1;c<s;c++)o.push(e.shape[c]),p*=e.shape[c];return{batchSize:l,sliceSize:p,outerSize:u,dimSize:i,outputShape:o}}function jD(e){try{return e.map(t=>nh(t))}catch(t){throw new Error(`Failed to decode encoded string bytes into utf-8, error: ${t}`)}}function qD(e){return e.map(t=>kp(t))}var Nn={};Ze(Nn,{nonMaxSuppressionV3Impl:()=>U4,nonMaxSuppressionV4Impl:()=>G4,nonMaxSuppressionV5Impl:()=>H4,whereImpl:()=>_4});l$();var XD=W();XD.registerFlag("KEEP_INTERMEDIATE_TENSORS",()=>!1,e=>{e&&console.warn("Keep intermediate tensors is ON. This will print the values of all intermediate tensors during model inference. Not all models support this mode. For details, check e2e/benchmarks/ model_config.js. This significantly impacts performance.")});var In;(function(e){e[e.DT_INVALID=0]="DT_INVALID",e[e.DT_FLOAT=1]="DT_FLOAT",e[e.DT_DOUBLE=2]="DT_DOUBLE",e[e.DT_INT32=3]="DT_INT32",e[e.DT_UINT8=4]="DT_UINT8",e[e.DT_INT16=5]="DT_INT16",e[e.DT_INT8=6]="DT_INT8",e[e.DT_STRING=7]="DT_STRING",e[e.DT_COMPLEX64=8]="DT_COMPLEX64",e[e.DT_INT64=9]="DT_INT64",e[e.DT_BOOL=10]="DT_BOOL",e[e.DT_QINT8=11]="DT_QINT8",e[e.DT_QUINT8=12]="DT_QUINT8",e[e.DT_QINT32=13]="DT_QINT32",e[e.DT_BFLOAT16=14]="DT_BFLOAT16",e[e.DT_QINT16=15]="DT_QINT16",e[e.DT_QUINT16=16]="DT_QUINT16",e[e.DT_UINT16=17]="DT_UINT16",e[e.DT_COMPLEX128=18]="DT_COMPLEX128",e[e.DT_HALF=19]="DT_HALF",e[e.DT_RESOURCE=20]="DT_RESOURCE",e[e.DT_VARIANT=21]="DT_VARIANT",e[e.DT_UINT32=22]="DT_UINT32",e[e.DT_UINT64=23]="DT_UINT64",e[e.DT_FLOAT_REF=101]="DT_FLOAT_REF",e[e.DT_DOUBLE_REF=102]="DT_DOUBLE_REF",e[e.DT_INT32_REF=103]="DT_INT32_REF",e[e.DT_UINT8_REF=104]="DT_UINT8_REF",e[e.DT_INT16_REF=105]="DT_INT16_REF",e[e.DT_INT8_REF=106]="DT_INT8_REF",e[e.DT_STRING_REF=107]="DT_STRING_REF",e[e.DT_COMPLEX64_REF=108]="DT_COMPLEX64_REF",e[e.DT_INT64_REF=109]="DT_INT64_REF",e[e.DT_BOOL_REF=110]="DT_BOOL_REF",e[e.DT_QINT8_REF=111]="DT_QINT8_REF",e[e.DT_QUINT8_REF=112]="DT_QUINT8_REF",e[e.DT_QINT32_REF=113]="DT_QINT32_REF",e[e.DT_BFLOAT16_REF=114]="DT_BFLOAT16_REF",e[e.DT_QINT16_REF=115]="DT_QINT16_REF",e[e.DT_QUINT16_REF=116]="DT_QUINT16_REF",e[e.DT_UINT16_REF=117]="DT_UINT16_REF",e[e.DT_COMPLEX128_REF=118]="DT_COMPLEX128_REF",e[e.DT_HALF_REF=119]="DT_HALF_REF",e[e.DT_RESOURCE_REF=120]="DT_RESOURCE_REF",e[e.DT_VARIANT_REF=121]="DT_VARIANT_REF",e[e.DT_UINT32_REF=122]="DT_UINT32_REF",e[e.DT_UINT64_REF=123]="DT_UINT64_REF"})(In||(In={}));var V5;(function(e){let t;(function(a){a[a.LEGACY=0]="LEGACY",a[a.V1=1]="V1",a[a.V2=2]="V2"})(t=e.CheckpointFormatVersion||(e.CheckpointFormatVersion={}))})(V5||(V5={}));var X3={};function KD(e,t){let a={tfOpName:e,category:"custom",inputs:[],attrs:[],customExecutor:t};X3[e]=a}function b7(e){return X3[e]}function YD(e){delete X3[e]}function k(e,t,a,n,r){let s=t.inputParams[e];if(s&&s.inputIndexStart!==void 0){let o=s.inputIndexStart,l=s.inputIndexEnd===0?void 0:s.inputIndexEnd===void 0?o+1:s.inputIndexEnd,u=o<0?t.inputNames.length+o:o;if(s.type==="tensor")return ua(t.inputNames[u],a,n,r);if(s.type==="tensors"){let d=t.inputs.slice(o,l);return t.inputNames.slice(o,l).filter((h,f)=>{var m;return((m=d[f])===null||m===void 0?void 0:m.op)!=="NoOp"}).map(h=>ua(h,a,n,r))}let p=ua(t.inputNames[u],a,n,r),c=p.dataSync();return s.type==="number"?c[0]:v.toNestedArray(p.shape,c)}let i=t.attrParams[e];return i&&i.value}function ua(e,t,a,n){let[r,s]=Ya(e,a);if(n!=null){let o=n.getHashTableHandleByName(r);if(o!=null)return o}let i=a.currentContextIds.find(o=>!!t[lh(r,o)]);return i!==void 0?t[lh(r,i)][s]:void 0}function U5(e,t,a){return t[lh(e,a.currentContextId)]}function Ar(e,t){let[a,n,r]=Ya(e,t);return[lh(a,t&&t.currentContextId),n,r]}function lh(e,t){return t?`${e}-${t}`:e}function Ya(e,t){if(e==="")return["",0,void 0];let a=t!=null&&t.parseNodeNameCache!=null;if(a){let s=t.parseNodeNameCache.get(e);if(s!=null)return s}let n=e.split(":"),r;if(n.length===1)r=[e,0,void 0];else{let s=n[0],i=n.length===3?n[1]:void 0,o=Number(n[n.length-1]);r=[s,o,i]}return a&&t.parseNodeNameCache.set(e,r),r}function qc(e,t,a){let n=k("pad",e,t,a);if(n==="explicit"){n=k("explicitPaddings",e,t,a);let r=[[0,0],[0,0],[0,0],[0,0]];for(let s=0;s<4;s++)r[s][0]=n[s*2],r[s][1]=n[s*2+1];return r}return n}function br(e){return e.kept?e:Sa(e)}var v7={};Ze(v7,{json:()=>ZD});var ZD=[{tfOpName:"Add",category:"arithmetic",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"AddV2",category:"arithmetic",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"AddN",category:"arithmetic",inputs:[{start:0,end:0,name:"tensors",type:"tensors"}]},{tfOpName:"BiasAdd",category:"arithmetic",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0},{tfName:"data_format",name:"dataFormat",type:"string",notSupported:!0}]},{tfOpName:"Sub",category:"arithmetic",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"RealDiv",category:"arithmetic",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Div",category:"arithmetic",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"DivNoNan",category:"arithmetic",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"FloorDiv",category:"arithmetic",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Mul",category:"arithmetic",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Maximum",category:"arithmetic",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Minimum",category:"arithmetic",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Pow",category:"arithmetic",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"SquaredDifference",category:"arithmetic",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Mod",category:"arithmetic",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"FloorMod",category:"arithmetic",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]}],w7={};Ze(w7,{json:()=>JD});var JD=[{tfOpName:"Abs",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Acos",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Asin",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Atan",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Atan2",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"y",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Ceil",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"ClipByValue",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"clipValueMin",type:"number"},{start:2,name:"clipValueMax",type:"number"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Complex",category:"basic_math",inputs:[{start:0,name:"real",type:"tensor"},{start:1,name:"imag",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"ComplexAbs",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Cos",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Cosh",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Elu",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Exp",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Floor",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Log",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Imag",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0},{tfName:"Tout",name:"outputType",type:"dtype",notSupported:!0}]},{tfOpName:"Neg",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Real",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0},{tfName:"Tout",name:"outputType",type:"dtype",notSupported:!0}]},{tfOpName:"Prelu",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"alpha",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Relu",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Relu6",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Selu",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Sigmoid",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Sin",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Sinh",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Sqrt",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Rsqrt",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Square",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Tan",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Tanh",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Sign",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Round",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Expm1",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Log1p",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Reciprocal",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Softplus",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Asinh",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Acosh",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Atanh",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Erf",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"LeakyRelu",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"alpha",name:"alpha",type:"number",defaultValue:.2},{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"IsNan",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"IsFinite",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"IsInf",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]}],k7={};Ze(k7,{json:()=>QD});var QD=[{tfOpName:"EmptyTensorList",category:"control",inputs:[{start:0,name:"elementShape",type:"shape"},{start:1,name:"maxNumElements",type:"number"}],attrs:[{tfName:"element_dtype",name:"elementDType",type:"dtype"}]},{tfOpName:"LoopCond",category:"control",inputs:[{start:0,name:"pred",type:"tensor"}]},{tfOpName:"Switch",category:"control",inputs:[{start:0,name:"data",type:"tensor"},{start:1,name:"pred",type:"tensor"}]},{tfOpName:"Merge",category:"control",inputs:[{start:0,end:0,name:"tensors",type:"tensors"}]},{tfOpName:"Enter",category:"control",inputs:[{start:0,name:"tensor",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0},{tfName:"frame_name",name:"frameName",type:"string"},{tfName:"is_constant",name:"isConstant",type:"bool"}]},{tfOpName:"Exit",category:"control",inputs:[{start:0,name:"tensor",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"NextIteration",category:"control",inputs:[{start:0,name:"tensor",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"TensorArrayV3",category:"control",inputs:[{start:0,name:"size",type:"number"}],attrs:[{tfName:"dtype",name:"dtype",type:"dtype"},{tfName:"element_shape",name:"elementShape",type:"shape"},{tfName:"dynamic_size",name:"dynamicSize",type:"bool"},{tfName:"clear_after_read",name:"clearAfterRead",type:"bool"},{tfName:"identical_element_shapes",name:"identicalElementShapes",type:"bool"},{tfName:"tensor_array_name",name:"name",type:"string"}]},{tfOpName:"TensorArrayWriteV3",category:"control",inputs:[{start:0,name:"tensorArrayId",type:"tensor"},{start:1,name:"index",type:"number"},{start:2,name:"tensor",type:"tensor"},{start:3,name:"flowIn",type:"number"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"TensorArrayReadV3",category:"control",inputs:[{start:0,name:"tensorArrayId",type:"tensor"},{start:1,name:"index",type:"number"},{start:2,name:"flowIn",type:"number"}],attrs:[{tfName:"dtype",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"TensorArrayGatherV3",category:"control",inputs:[{start:0,name:"tensorArrayId",type:"tensor"},{start:1,name:"indices",type:"number[]"},{start:2,name:"flowIn",type:"number"}],attrs:[{tfName:"dtype",name:"dtype",type:"dtype"},{tfName:"element_shape",name:"elementShape",type:"shape"}]},{tfOpName:"TensorArrayScatterV3",category:"control",inputs:[{start:0,name:"tensorArrayId",type:"tensor"},{start:1,name:"indices",type:"number[]"},{start:2,name:"tensor",type:"tensor"},{start:3,name:"flowIn",type:"number"}],attrs:[{tfName:"T",name:"dtype",type:"dtype"}]},{tfOpName:"TensorArrayConcatV3",category:"control",inputs:[{start:0,name:"tensorArrayId",type:"tensor"},{start:1,name:"flowIn",type:"number"}],attrs:[{tfName:"dtype",name:"dtype",type:"dtype"},{tfName:"element_shape_except0",name:"elementShapeExcept0",type:"shape",notSupported:!0}]},{tfOpName:"TensorArraySplitV3",category:"control",inputs:[{start:0,name:"tensorArrayId",type:"tensor"},{start:1,name:"tensor",type:"tensor"},{start:2,name:"lengths",type:"number[]"},{start:3,name:"flowIn",type:"number"}],attrs:[{tfName:"T",name:"dtype",type:"dtype"}]},{tfOpName:"TensorArraySizeV3",category:"control",inputs:[{start:0,name:"tensorArrayId",type:"tensor"},{start:1,name:"flowIn",type:"number"}]},{tfOpName:"TensorArrayCloseV3",category:"control",inputs:[{start:0,name:"tensorArrayId",type:"tensor"}]},{tfOpName:"StatelessIf",category:"control",inputs:[{start:0,name:"cond",type:"tensor"},{start:1,end:0,name:"args",type:"tensors"}],attrs:[{tfName:"then_branch",name:"thenBranch",type:"func"},{tfName:"else_branch",name:"elseBranch",type:"func"}]},{tfOpName:"If",category:"control",inputs:[{start:0,name:"cond",type:"tensor"},{start:1,end:0,name:"args",type:"tensors"}],attrs:[{tfName:"then_branch",name:"thenBranch",type:"func"},{tfName:"else_branch",name:"elseBranch",type:"func"}]},{tfOpName:"StatelessWhile",category:"control",inputs:[{start:0,end:0,name:"args",type:"tensors"}],attrs:[{tfName:"cond",name:"cond",type:"func"},{tfName:"body",name:"body",type:"func"}]},{tfOpName:"While",category:"control",inputs:[{start:0,end:0,name:"args",type:"tensors"}],attrs:[{tfName:"cond",name:"cond",type:"func"},{tfName:"body",name:"body",type:"func"}]},{tfOpName:"TensorListScatter",category:"control",inputs:[{start:0,name:"tensor",type:"tensor"},{start:1,name:"indices",type:"number[]"},{start:2,name:"elementShape",type:"shape"}],attrs:[{tfName:"element_dtype",name:"elementDType",type:"dtype"}]},{tfOpName:"TensorListScatterV2",category:"control",inputs:[{start:0,name:"tensor",type:"tensor"},{start:1,name:"indices",type:"number[]"},{start:2,name:"elementShape",type:"shape"},{start:3,name:"numElements",type:"number"}],attrs:[{tfName:"element_dtype",name:"elementDType",type:"dtype"}]},{tfOpName:"TensorListGather",category:"control",inputs:[{start:0,name:"tensorListId",type:"tensor"},{start:1,name:"indices",type:"number[]"},{start:2,name:"elementShape",type:"shape"}],attrs:[{tfName:"element_dtype",name:"elementDType",type:"dtype"}]},{tfOpName:"TensorListGetItem",category:"control",inputs:[{start:0,name:"tensorListId",type:"tensor"},{start:1,name:"index",type:"number"},{start:2,name:"elementShape",type:"shape"}],attrs:[{tfName:"element_dtype",name:"elementDType",type:"dtype"}]},{tfOpName:"TensorListSetItem",category:"control",inputs:[{start:0,name:"tensorListId",type:"tensor"},{start:1,name:"index",type:"number"},{start:2,name:"tensor",type:"tensor"}],attrs:[{tfName:"element_dtype",name:"elementDType",type:"dtype"}]},{tfOpName:"TensorListReserve",category:"control",inputs:[{start:0,name:"elementShape",type:"shape"},{start:1,name:"numElements",type:"number"}],attrs:[{tfName:"element_dtype",name:"elementDType",type:"dtype"}]},{tfOpName:"TensorListFromTensor",category:"control",inputs:[{start:0,name:"tensor",type:"tensor"},{start:1,name:"elementShape",type:"shape"}],attrs:[{tfName:"element_dtype",name:"elementDType",type:"dtype"}]},{tfOpName:"TensorListStack",category:"control",inputs:[{start:0,name:"tensorListId",type:"tensor"},{start:1,name:"elementShape",type:"shape"}],attrs:[{tfName:"element_dtype",name:"elementDType",type:"dtype"},{tfName:"num_elements",name:"numElements",type:"dtype"}]},{tfOpName:"TensorListSplit",category:"control",inputs:[{start:0,name:"tensor",type:"tensor"},{start:1,name:"elementShape",type:"shape"},{start:2,name:"lengths",type:"number[]"}],attrs:[{tfName:"element_dtype",name:"elementDType",type:"dtype"}]},{tfOpName:"TensorListConcat",category:"control",inputs:[{start:0,name:"tensorListId",type:"tensor"}],attrs:[{tfName:"element_shape",name:"elementShape",type:"shape"},{tfName:"element_dtype",name:"elementDType",type:"dtype"}]},{tfOpName:"TensorListConcatV2",category:"control",inputs:[{start:0,name:"tensorListId",type:"tensor"}],attrs:[{tfName:"element_shape",name:"elementShape",type:"shape"},{tfName:"element_dtype",name:"elementDType",type:"dtype"}]},{tfOpName:"TensorListPopBack",category:"control",inputs:[{start:0,name:"tensorListId",type:"tensor"},{start:1,name:"elementShape",type:"shape"}],attrs:[{tfName:"element_dtype",name:"elementDType",type:"dtype"}]},{tfOpName:"TensorListPushBack",category:"control",inputs:[{start:0,name:"tensorListId",type:"tensor"},{start:1,name:"tensor",type:"tensor"}],attrs:[{tfName:"element_dtype",name:"elementDType",type:"dtype"}]},{tfOpName:"TensorListLength",category:"control",inputs:[{start:0,name:"tensorListId",type:"tensor"}]},{tfOpName:"TensorListResize",category:"control",inputs:[{start:0,name:"tensorListId",type:"tensor"},{start:1,name:"size",type:"number"}]}],I7={};Ze(I7,{json:()=>eP});var eP=[{tfOpName:"AvgPool",category:"convolution",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"strides",name:"strides",type:"number[]"},{tfName:"padding",name:"pad",type:"string"},{tfName:"data_format",name:"dataFormat",type:"string",notSupported:!0},{tfName:"ksize",name:"kernelSize",type:"number[]"},{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"MaxPool",category:"convolution",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"strides",name:"strides",type:"number[]"},{tfName:"padding",name:"pad",type:"string"},{tfName:"data_format",name:"dataFormat",type:"string",notSupported:!0},{tfName:"ksize",name:"kernelSize",type:"number[]"},{tfName:"explicit_paddings",name:"explicitPaddings",type:"number[]",defaultValue:[],notSupported:!0},{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"MaxPoolWithArgmax",category:"convolution",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"strides",name:"strides",type:"number[]"},{tfName:"padding",name:"pad",type:"string"},{tfName:"ksize",name:"kernelSize",type:"number[]"},{tfName:"include_batch_in_index",name:"includeBatchInIndex",type:"bool"},{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"AvgPool3D",category:"convolution",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"strides",name:"strides",type:"number[]"},{tfName:"padding",name:"pad",type:"string"},{tfName:"data_format",name:"dataFormat",type:"string",notSupported:!0},{tfName:"ksize",name:"kernelSize",type:"number[]"},{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"MaxPool3D",category:"convolution",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"strides",name:"strides",type:"number[]"},{tfName:"padding",name:"pad",type:"string"},{tfName:"data_format",name:"dataFormat",type:"string",notSupported:!0},{tfName:"ksize",name:"kernelSize",type:"number[]"},{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Conv1D",category:"convolution",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"filter",type:"tensor"}],attrs:[{tfName:"stride",name:"stride",type:"number"},{tfName:"padding",name:"pad",type:"string"},{tfName:"data_format",name:"dataFormat",type:"string",defaultValue:"NWC"},{tfName:"T",name:"dtype",type:"dtype",notSupported:!0},{tfName:"dilation",name:"dilation",type:"number",defaultValue:1}]},{tfOpName:"Conv2D",category:"convolution",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"filter",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0},{tfName:"strides",name:"strides",type:"number[]"},{tfName:"padding",name:"pad",type:"string"},{tfName:"useCudnnOnGpu",name:"useCudnnOnGpu",type:"bool"},{tfName:"data_format",name:"dataFormat",type:"string",defaultValue:"NHWC"},{tfName:"explicit_paddings",name:"explicitPaddings",type:"number[]",defaultValue:[]},{tfName:"dilations",name:"dilations",type:"number[]"}]},{tfOpName:"_FusedConv2D",category:"convolution",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"filter",type:"tensor"},{start:2,end:0,name:"args",type:"tensors"}],attrs:[{tfName:"num_args",name:"numArgs",type:"number"},{tfName:"T",name:"dtype",type:"dtype",notSupported:!0},{tfName:"strides",name:"strides",type:"number[]"},{tfName:"padding",name:"pad",type:"string"},{tfName:"explicit_paddings",name:"explicitPaddings",type:"number[]",defaultValue:[]},{tfName:"use_cudnn_on_gpu",name:"useCudnnOnGpu",type:"bool",defaultValue:!0},{tfName:"data_format",name:"dataFormat",type:"string",defaultValue:"NHWC"},{tfName:"dilations",name:"dilations",type:"number[]",defaultValue:[1,1,1,1]},{tfName:"fused_ops",name:"fusedOps",type:"string[]",defaultValue:[]},{tfName:"epsilon",name:"epsilon",type:"number",defaultValue:1e-4},{tfName:"leakyrelu_alpha",name:"leakyreluAlpha",type:"number",defaultValue:.2}]},{tfOpName:"Conv2DBackpropInput",category:"convolution",inputs:[{start:2,name:"x",type:"tensor"},{start:1,name:"filter",type:"tensor"},{start:0,name:"outputShape",type:"number[]"}],attrs:[{tfName:"strides",name:"strides",type:"number[]"},{tfName:"padding",name:"pad",type:"string"},{tfName:"data_format",name:"dataFormat",type:"string",notSupported:!0},{tfName:"explicit_paddings",name:"explicitPaddings",type:"number[]",defaultValue:[]},{tfName:"dilations",name:"dilations",type:"number[]",notSupported:!0}]},{tfOpName:"DepthwiseConv2d",category:"convolution",inputs:[{start:0,name:"input",type:"tensor"},{start:1,name:"filter",type:"tensor"}],attrs:[{tfName:"strides",name:"strides",type:"number[]"},{tfName:"padding",name:"pad",type:"string"},{tfName:"data_format",name:"dataFormat",type:"string",defaultValue:"NHWC"},{tfName:"explicit_paddings",name:"explicitPaddings",type:"number[]",defaultValue:[]},{tfName:"dilations",name:"dilations",type:"number[]"}]},{tfOpName:"DepthwiseConv2dNative",category:"convolution",inputs:[{start:0,name:"input",type:"tensor"},{start:1,name:"filter",type:"tensor"}],attrs:[{tfName:"strides",name:"strides",type:"number[]"},{tfName:"padding",name:"pad",type:"string"},{tfName:"data_format",name:"dataFormat",type:"string",defaultValue:"NHWC"},{tfName:"explicit_paddings",name:"explicitPaddings",type:"number[]",defaultValue:[]},{tfName:"dilations",name:"dilations",type:"number[]"}]},{tfOpName:"FusedDepthwiseConv2dNative",category:"convolution",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"filter",type:"tensor"},{start:2,end:0,name:"args",type:"tensors"}],attrs:[{tfName:"num_args",name:"numArgs",type:"number"},{tfName:"T",name:"dtype",type:"dtype",notSupported:!0},{tfName:"strides",name:"strides",type:"number[]"},{tfName:"padding",name:"pad",type:"string"},{tfName:"data_format",name:"dataFormat",type:"string",defaultValue:"NHWC"},{tfName:"dilations",name:"dilations",type:"number[]",defaultValue:[1,1,1,1]},{tfName:"fused_ops",name:"fusedOps",type:"string[]",defaultValue:[]},{tfName:"explicit_paddings",name:"explicitPaddings",type:"number[]",defaultValue:[]}]},{tfOpName:"Conv3D",category:"convolution",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"filter",type:"tensor"}],attrs:[{tfName:"strides",name:"strides",type:"number[]"},{tfName:"padding",name:"pad",type:"string"},{tfName:"data_format",name:"dataFormat",type:"string",defaultValue:"NHWC"},{tfName:"dilations",name:"dilations",type:"number[]"}]},{tfOpName:"Dilation2D",category:"convolution",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"filter",type:"tensor"}],attrs:[{tfName:"strides",name:"strides",type:"number[]"},{tfName:"rates",name:"dilations",type:"number[]"},{tfName:"padding",name:"pad",type:"string"}]}],S7={};Ze(S7,{json:()=>tP});var tP=[{tfOpName:"Fill",category:"creation",inputs:[{start:0,name:"shape",type:"number[]"},{start:1,name:"value",type:"number"}],attrs:[{tfName:"T",name:"dtype",type:"dtype"}]},{tfOpName:"LinSpace",category:"creation",inputs:[{start:0,name:"start",type:"number"},{start:1,name:"stop",type:"number"},{start:2,name:"num",type:"number"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"OneHot",category:"creation",inputs:[{start:0,name:"indices",type:"tensor"},{start:1,name:"depth",type:"number"},{start:2,name:"onValue",type:"number",defaultValue:1},{start:3,name:"offValue",type:"number",defaultValue:0}],attrs:[{tfName:"axis",name:"axis",type:"number",notSupported:!0},{tfName:"T",name:"dtype",type:"dtype"}]},{tfOpName:"Ones",category:"creation",inputs:[{start:0,name:"shape",type:"number[]"}],attrs:[{tfName:"T",name:"dtype",type:"dtype"}]},{tfOpName:"OnesLike",category:"creation",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"dtype",name:"dtype",type:"dtype"}]},{tfOpName:"RandomStandardNormal",category:"creation",inputs:[{start:0,name:"shape",type:"number[]"}],attrs:[{tfName:"seed",name:"seed",type:"number",defaultValue:0},{tfName:"seed2",name:"seed2",type:"number",defaultValue:0,notSupported:!0},{tfName:"dtype",name:"dtype",type:"dtype"},{tfName:"T",name:"T",type:"number",notSupported:!0}]},{tfOpName:"RandomUniform",category:"creation",inputs:[{start:0,name:"shape",type:"number[]"}],attrs:[{tfName:"minval",name:"minval",type:"number",defaultValue:0},{tfName:"maxval",name:"maxval",type:"number",defaultValue:1},{tfName:"dtype",name:"dtype",type:"dtype"},{tfName:"seed",name:"seed",type:"number",defaultValue:0},{tfName:"seed2",name:"seed2",type:"number",defaultValue:0,notSupported:!0},{tfName:"T",name:"T",type:"number",notSupported:!0}]},{tfOpName:"RandomUniformInt",category:"creation",inputs:[{start:0,name:"shape",type:"number[]"}],attrs:[{tfName:"minval",name:"minval",type:"number"},{tfName:"maxval",name:"maxval",type:"number"},{tfName:"seed",name:"seed",type:"number",defaultValue:0},{tfName:"seed2",name:"seed2",type:"number",defaultValue:0,notSupported:!0}]},{tfOpName:"Range",category:"creation",inputs:[{start:0,name:"start",type:"number"},{start:1,name:"stop",type:"number"},{start:2,name:"step",type:"number",defaultValue:0}],attrs:[{tfName:"Tidx",name:"dtype",type:"dtype"}]},{tfOpName:"TruncatedNormal",category:"creation",inputs:[{start:0,name:"shape",type:"number[]"}],attrs:[{tfName:"means",name:"mean",type:"number",defaultValue:0},{tfName:"stddev",name:"stdDev",type:"number",defaultValue:1},{tfName:"seed",name:"seed",type:"number"},{tfName:"seed2",name:"seed2",type:"number",defaultValue:0,notSupported:!0},{tfName:"dtype",name:"dtype",type:"dtype"},{tfName:"T",name:"T",type:"number",notSupported:!0}]},{tfOpName:"Zeros",category:"creation",inputs:[{start:0,name:"shape",type:"number[]"}],attrs:[{tfName:"T",name:"dtype",type:"dtype"}]},{tfOpName:"ZerosLike",category:"creation",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype"}]},{tfOpName:"Multinomial",category:"creation",inputs:[{start:0,name:"logits",type:"tensor"},{start:1,name:"numSamples",type:"number"}],attrs:[{tfName:"seed",name:"seed",type:"number"},{tfName:"seed2",name:"seed2",type:"number"},{tfName:"T",name:"dtype",type:"dtype"},{tfName:"output_dtype",name:"output_dtype",type:"dtype"}]}],C7={};Ze(C7,{json:()=>aP});var aP=[{tfOpName:"NonMaxSuppressionV2",category:"dynamic",inputs:[{start:0,name:"boxes",type:"tensor"},{start:1,name:"scores",type:"tensor"},{start:2,name:"maxOutputSize",type:"number"},{start:3,name:"iouThreshold",type:"number"}]},{tfOpName:"NonMaxSuppressionV3",category:"dynamic",inputs:[{start:0,name:"boxes",type:"tensor"},{start:1,name:"scores",type:"tensor"},{start:2,name:"maxOutputSize",type:"number"},{start:3,name:"iouThreshold",type:"number"},{start:4,name:"scoreThreshold",type:"number"}]},{tfOpName:"NonMaxSuppressionV4",category:"dynamic",inputs:[{start:0,name:"boxes",type:"tensor"},{start:1,name:"scores",type:"tensor"},{start:2,name:"maxOutputSize",type:"number"},{start:3,name:"iouThreshold",type:"number"},{start:4,name:"scoreThreshold",type:"number"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0},{tfName:"T_threshold",name:"threshold",type:"dtype",notSupported:!0},{tfName:"pad_to_max_output_size",name:"padToMaxOutputSize",type:"bool"}]},{tfOpName:"NonMaxSuppressionV5",category:"dynamic",inputs:[{start:0,name:"boxes",type:"tensor"},{start:1,name:"scores",type:"tensor"},{start:2,name:"maxOutputSize",type:"number"},{start:3,name:"iouThreshold",type:"number"},{start:4,name:"scoreThreshold",type:"number"},{start:5,name:"softNmsSigma",type:"number"}]},{tfOpName:"Where",category:"dynamic",inputs:[{start:0,name:"condition",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"ListDiff",category:"dynamic",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"y",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]}],T7={};Ze(T7,{json:()=>nP});var nP=[{tfOpName:"LowerBound",category:"evaluation",inputs:[{start:0,name:"sortedSequence",type:"tensor"},{start:1,name:"values",type:"tensor"}]},{tfOpName:"TopKV2",category:"evaluation",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"k",type:"number"}],attrs:[{tfName:"sorted",name:"sorted",type:"bool"}]},{tfOpName:"UpperBound",category:"evaluation",inputs:[{start:0,name:"sortedSequence",type:"tensor"},{start:1,name:"values",type:"tensor"}]},{tfOpName:"Unique",category:"evaluation",inputs:[{start:0,name:"x",type:"tensor"}]},{tfOpName:"UniqueV2",category:"evaluation",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"axis",type:"number"}]}],N7={};Ze(N7,{json:()=>rP});var rP=[{tfOpName:"PlaceholderWithDefault",category:"graph",inputs:[{start:0,name:"default",type:"tensor"}],attrs:[{tfName:"shape",name:"shape",type:"shape"},{tfName:"dtype",name:"dtype",type:"dtype"}]},{tfOpName:"Placeholder",category:"graph",attrs:[{tfName:"shape",name:"shape",type:"shape"},{tfName:"dtype",name:"dtype",type:"dtype"}]},{tfOpName:"Const",category:"graph"},{tfOpName:"Identity",category:"graph",inputs:[{start:0,name:"x",type:"tensor"}]},{tfOpName:"IdentityN",category:"graph",inputs:[{start:0,end:0,name:"x",type:"tensors"}]},{tfOpName:"Snapshot",category:"graph",inputs:[{start:0,name:"x",type:"tensor"}]},{tfOpName:"Rank",category:"graph",inputs:[{start:0,name:"x",type:"tensor"}]},{tfOpName:"Size",category:"graph",inputs:[{start:0,name:"x",type:"tensor"}]},{tfOpName:"Shape",category:"graph",inputs:[{start:0,name:"x",type:"tensor"}]},{tfOpName:"ShapeN",category:"graph",inputs:[{start:0,end:0,name:"x",type:"tensors"}]},{tfOpName:"Print",category:"graph",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"data",type:"tensors"}],attrs:[{tfName:"message",name:"message",type:"string"},{tfName:"first_n",name:"firstN",type:"number",notSupported:!0},{tfName:"summarize",name:"summarize",type:"number",defaultValue:3}]},{tfOpName:"NoOp",category:"graph",inputs:[]},{tfOpName:"StopGradient",category:"graph",inputs:[{start:0,name:"x",type:"tensor"}]},{tfOpName:"FakeQuantWithMinMaxVars",category:"graph",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"min",name:"min",type:"number"},{tfName:"max",name:"max",type:"number"}]}],R7={};Ze(R7,{json:()=>sP});var sP=[{tfOpName:"HashTable",category:"hash_table",inputs:[],attrs:[{tfName:"shared_name",name:"sharedName",type:"string"},{tfName:"use_node_name_sharing",name:"useNodeNameSharing",type:"bool"},{tfName:"key_dtype",name:"keyDType",type:"dtype"},{tfName:"value_dtype",name:"valueDType",type:"dtype"}]},{tfOpName:"HashTableV2",category:"hash_table",inputs:[],attrs:[{tfName:"shared_name",name:"sharedName",type:"string"},{tfName:"use_node_name_sharing",name:"useNodeNameSharing",type:"bool"},{tfName:"key_dtype",name:"keyDType",type:"dtype"},{tfName:"value_dtype",name:"valueDType",type:"dtype"}]},{tfOpName:"LookupTableImport",category:"hash_table",inputs:[{start:0,name:"tableHandle",type:"tensor"},{start:1,name:"keys",type:"tensor"},{start:2,name:"values",type:"tensor"}],attrs:[{tfName:"Tin",name:"tIn",type:"dtype",notSupported:!0},{tfName:"Tout",name:"tOut",type:"dtype",notSupported:!0}]},{tfOpName:"LookupTableImportV2",category:"hash_table",inputs:[{start:0,name:"tableHandle",type:"tensor"},{start:1,name:"keys",type:"tensor"},{start:2,name:"values",type:"tensor"}],attrs:[{tfName:"Tin",name:"tIn",type:"dtype",notSupported:!0},{tfName:"Tout",name:"tOut",type:"dtype",notSupported:!0}]},{tfOpName:"LookupTableFind",category:"hash_table",inputs:[{start:0,name:"tableHandle",type:"tensor"},{start:1,name:"keys",type:"tensor"},{start:2,name:"defaultValue",type:"tensor"}],attrs:[{tfName:"Tin",name:"tIn",type:"dtype",notSupported:!0},{tfName:"Tout",name:"tOut",type:"dtype",notSupported:!0}]},{tfOpName:"LookupTableFindV2",category:"hash_table",inputs:[{start:0,name:"tableHandle",type:"tensor"},{start:1,name:"keys",type:"tensor"},{start:2,name:"defaultValue",type:"tensor"}],attrs:[{tfName:"Tin",name:"tIn",type:"dtype",notSupported:!0},{tfName:"Tout",name:"tOut",type:"dtype",notSupported:!0}]},{tfOpName:"LookupTableSize",category:"hash_table",inputs:[{start:0,name:"tableHandle",type:"tensor"}]},{tfOpName:"LookupTableSizeV2",category:"hash_table",inputs:[{start:0,name:"tableHandle",type:"tensor"}]},{tfOpName:"InitializeTable",category:"hash_table",inputs:[{start:0,name:"tableHandle",type:"tensor"},{start:1,name:"keys",type:"tensor"},{start:2,name:"values",type:"tensor"}]},{tfOpName:"InitializeTableV2",category:"hash_table",inputs:[{start:0,name:"tableHandle",type:"tensor"},{start:1,name:"keys",type:"tensor"},{start:2,name:"values",type:"tensor"}]}],E7={};Ze(E7,{json:()=>iP});var iP=[{tfOpName:"ResizeBilinear",category:"image",inputs:[{start:0,name:"images",type:"tensor"},{start:1,name:"size",type:"number[]"}],attrs:[{tfName:"align_corners",name:"alignCorners",type:"bool"},{tfName:"half_pixel_centers",name:"halfPixelCenters",type:"bool"},{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"ResizeNearestNeighbor",category:"image",inputs:[{start:0,name:"images",type:"tensor"},{start:1,name:"size",type:"number[]"}],attrs:[{tfName:"align_corners",name:"alignCorners",type:"bool"},{tfName:"half_pixel_centers",name:"halfPixelCenters",type:"bool"},{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"CropAndResize",category:"image",inputs:[{start:0,name:"image",type:"tensor"},{start:1,name:"boxes",type:"tensor"},{start:2,name:"boxInd",type:"tensor"},{start:3,name:"cropSize",type:"number[]"}],attrs:[{tfName:"method",name:"method",type:"string"},{tfName:"extrapolation_value",name:"extrapolationValue",type:"number"}]},{tfOpName:"ImageProjectiveTransformV3",category:"image",inputs:[{start:0,name:"images",type:"tensor"},{start:1,name:"transforms",type:"tensor"},{start:2,name:"outputShape",type:"number[]"},{start:3,name:"fillValue",type:"number"}],attrs:[{tfName:"interpolation",name:"interpolation",type:"string"},{tfName:"fill_mode",name:"fillMode",type:"string"}]}],M7={};Ze(M7,{json:()=>oP});var oP=[{tfOpName:"Equal",category:"logical",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"NotEqual",category:"logical",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Greater",category:"logical",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"GreaterEqual",category:"logical",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Less",category:"logical",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"LessEqual",category:"logical",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"LogicalAnd",category:"logical",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"LogicalNot",category:"logical",inputs:[{start:0,name:"a",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"LogicalOr",category:"logical",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Select",category:"logical",inputs:[{start:0,name:"condition",type:"tensor"},{start:1,name:"a",type:"tensor"},{start:2,name:"b",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"SelectV2",category:"logical",inputs:[{start:0,name:"condition",type:"tensor"},{start:1,name:"a",type:"tensor"},{start:2,name:"b",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"BitwiseAnd",category:"logical",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"y",type:"tensor"}]}],F7={};Ze(F7,{json:()=>lP});var lP=[{tfOpName:"_FusedMatMul",category:"matrices",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"},{start:2,end:0,name:"args",type:"tensors"}],attrs:[{tfName:"num_args",name:"numArgs",type:"number"},{tfName:"fused_ops",name:"fusedOps",type:"string[]",defaultValue:[]},{tfName:"epsilon",name:"epsilon",type:"number",defaultValue:1e-4},{tfName:"transpose_a",name:"transposeA",type:"bool",defaultValue:!1},{tfName:"transpose_b",name:"transposeB",type:"bool",defaultValue:!1},{tfName:"leakyrelu_alpha",name:"leakyreluAlpha",type:"number",defaultValue:.2},{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"MatMul",category:"matrices",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"transpose_a",name:"transposeA",type:"bool",defaultValue:!1},{tfName:"transpose_b",name:"transposeB",type:"bool",defaultValue:!1},{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"BatchMatMul",category:"matrices",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"adj_x",name:"transposeA",type:"bool",defaultValue:!1},{tfName:"adj_y",name:"transposeB",type:"bool",defaultValue:!1},{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"BatchMatMulV2",category:"matrices",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"adj_x",name:"transposeA",type:"bool",defaultValue:!1},{tfName:"adj_y",name:"transposeB",type:"bool",defaultValue:!1},{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Transpose",category:"matrices",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"perm",type:"number[]"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Einsum",category:"matrices",inputs:[{start:0,end:0,name:"tensors",type:"tensors"}],attrs:[{tfName:"equation",name:"equation",type:"string"},{tfName:"N",name:"n",type:"number",defaultValue:2},{tfName:"T",name:"dtype",type:"dtype"}]},{tfOpName:"MatrixBandPart",category:"matrices",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"numLower",type:"tensor"},{start:1,name:"numUpper",type:"tensor"}]}],_7={};Ze(_7,{json:()=>uP});var uP=[{tfOpName:"EuclideanNorm",category:"normalization",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"axis",type:"number[]"}],attrs:[{tfName:"keep_dims",name:"keepDims",type:"bool",defaultValue:!1}]},{tfOpName:"FusedBatchNorm",category:"normalization",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"scale",type:"tensor"},{start:2,name:"offset",type:"tensor"},{start:3,name:"mean",type:"tensor"},{start:4,name:"variance",type:"tensor"}],attrs:[{tfName:"epsilon",name:"epsilon",type:"number",defaultValue:.001},{tfName:"data_format",name:"dataFormat",type:"string",notSupported:!0}]},{tfOpName:"FusedBatchNormV2",category:"normalization",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"scale",type:"tensor"},{start:2,name:"offset",type:"tensor"},{start:3,name:"mean",type:"tensor"},{start:4,name:"variance",type:"tensor"}],attrs:[{tfName:"epsilon",name:"epsilon",type:"number",defaultValue:.001},{tfName:"data_format",name:"dataFormat",type:"string",notSupported:!0}]},{tfOpName:"FusedBatchNormV3",category:"normalization",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"scale",type:"tensor"},{start:2,name:"offset",type:"tensor"},{start:3,name:"mean",type:"tensor"},{start:4,name:"variance",type:"tensor"}],attrs:[{tfName:"epsilon",name:"epsilon",type:"number",defaultValue:.001},{tfName:"data_format",name:"dataFormat",type:"string",notSupported:!0}]},{tfOpName:"LRN",category:"normalization",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"depth_radius",name:"radius",type:"number",defaultValue:5},{tfName:"bias",name:"bias",type:"number",defaultValue:1},{tfName:"alpha",name:"alpha",type:"number",defaultValue:1},{tfName:"beta",name:"beta",type:"number",defaultValue:.5}]},{tfOpName:"Softmax",category:"normalization",inputs:[{start:0,name:"x",type:"tensor"}]},{tfOpName:"LogSoftmax",category:"normalization",inputs:[{start:0,name:"x",type:"tensor"}]}],$7={};Ze($7,{json:()=>dP});var dP=[{tfOpName:"Bincount",category:"reduction",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"size",type:"number"},{start:2,name:"weights",type:"tensor"}]},{tfOpName:"DenseBincount",category:"reduction",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"size",type:"number"},{start:2,name:"weights",type:"tensor"}],attrs:[{tfName:"binary_output",name:"binaryOutput",type:"bool"}]},{tfOpName:"Max",category:"reduction",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"axis",type:"number[]"}],attrs:[{tfName:"keep_dims",name:"keepDims",type:"bool"}]},{tfOpName:"Mean",category:"reduction",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"axis",type:"number[]"}],attrs:[{tfName:"keep_dims",name:"keepDims",type:"bool"}]},{tfOpName:"Min",category:"reduction",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"axis",type:"number[]"}],attrs:[{tfName:"keep_dims",name:"keepDims",type:"bool"}]},{tfOpName:"Sum",category:"reduction",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"axis",type:"number[]"}],attrs:[{tfName:"keep_dims",name:"keepDims",type:"bool"}]},{tfOpName:"All",category:"reduction",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"axis",type:"number[]"}],attrs:[{tfName:"keep_dims",name:"keepDims",type:"bool"}]},{tfOpName:"Any",category:"reduction",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"axis",type:"number[]"}],attrs:[{tfName:"keep_dims",name:"keepDims",type:"bool"}]},{tfOpName:"ArgMax",category:"reduction",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"axis",type:"number"}]},{tfOpName:"ArgMin",category:"reduction",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"axis",type:"number"}]},{tfOpName:"Prod",category:"reduction",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"axis",type:"number[]"}],attrs:[{tfName:"keep_dims",name:"keepDims",type:"bool"},{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Cumprod",category:"reduction",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"axis",type:"number"}],attrs:[{tfName:"exclusive",name:"exclusive",type:"bool"},{tfName:"reverse",name:"reverse",type:"bool"}]},{tfOpName:"Cumsum",category:"reduction",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"axis",type:"number"}],attrs:[{tfName:"exclusive",name:"exclusive",type:"bool"},{tfName:"reverse",name:"reverse",type:"bool"}]}],D7={};Ze(D7,{json:()=>pP});var pP=[{tfOpName:"ConcatV2",category:"slice_join",inputs:[{start:0,end:-1,name:"tensors",type:"tensors"},{start:-1,name:"axis",type:"number"}],attrs:[{tfName:"N",name:"n",type:"number",defaultValue:2}]},{tfOpName:"Concat",category:"slice_join",inputs:[{start:1,end:0,name:"tensors",type:"tensors"},{start:0,name:"axis",type:"number"}],attrs:[{tfName:"N",name:"n",type:"number",defaultValue:2}]},{tfOpName:"GatherV2",category:"slice_join",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"indices",type:"tensor"},{start:2,name:"axis",type:"number",defaultValue:0}],attrs:[{tfName:"batch_dims",name:"batchDims",type:"number",defaultValue:0}]},{tfOpName:"Gather",category:"slice_join",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"indices",type:"tensor"}],attrs:[{tfName:"validate_indices",name:"validateIndices",type:"bool",notSupported:!0}]},{tfOpName:"Reverse",category:"slice_join",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"dims",type:"bool[]"}]},{tfOpName:"ReverseV2",category:"slice_join",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"axis",type:"number[]"}]},{tfOpName:"Slice",category:"slice_join",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"begin",type:"number[]"},{start:2,name:"size",type:"number[]"}]},{tfOpName:"StridedSlice",category:"slice_join",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"begin",type:"number[]"},{start:2,name:"end",type:"number[]"},{start:3,name:"strides",type:"number[]"}],attrs:[{tfName:"begin_mask",name:"beginMask",type:"number",defaultValue:0},{tfName:"end_mask",name:"endMask",type:"number",defaultValue:0},{tfName:"new_axis_mask",name:"newAxisMask",type:"number",defaultValue:0},{tfName:"ellipsis_mask",name:"ellipsisMask",type:"number",defaultValue:0},{tfName:"shrink_axis_mask",name:"shrinkAxisMask",type:"number",defaultValue:0}]},{tfOpName:"Pack",category:"slice_join",inputs:[{start:0,end:0,name:"tensors",type:"tensors"}],attrs:[{tfName:"axis",name:"axis",type:"number",defaultValue:0}]},{tfOpName:"Unpack",category:"slice_join",inputs:[{start:0,name:"tensor",type:"tensor"}],attrs:[{tfName:"axis",name:"axis",type:"number",defaultValue:0},{tfName:"num",name:"num",type:"number",defaultValue:0,notSupported:!0}]},{tfOpName:"Tile",category:"slice_join",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"reps",type:"number[]"}]},{tfOpName:"Split",category:"slice_join",inputs:[{start:0,name:"axis",type:"number",defaultValue:0},{start:1,name:"x",type:"tensor"}],attrs:[{tfName:"num_split",name:"numOrSizeSplits",type:"number",defaultValue:1}]},{tfOpName:"SplitV",category:"slice_join",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"numOrSizeSplits",type:"number[]"},{start:2,name:"axis",type:"number",defaultValue:0}]},{tfOpName:"ScatterNd",category:"slice_join",inputs:[{start:0,name:"indices",type:"tensor"},{start:1,name:"values",type:"tensor"},{start:2,name:"shape",type:"number[]"}]},{tfOpName:"GatherNd",category:"slice_join",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"indices",type:"tensor"}]},{tfOpName:"SparseToDense",category:"slice_join",inputs:[{start:0,name:"sparseIndices",type:"tensor"},{start:1,name:"outputShape",type:"number[]"},{start:2,name:"sparseValues",type:"tensor"},{start:3,name:"defaultValue",type:"tensor"}],attrs:[{tfName:"validate_indices",name:"validateIndices",type:"bool",defaultValue:!1,notSupported:!0}]},{tfOpName:"TensorScatterUpdate",category:"slice_join",inputs:[{start:0,name:"tensor",type:"tensor"},{start:1,name:"indices",type:"tensor"},{start:2,name:"values",type:"tensor"}]}],P7={};Ze(P7,{json:()=>cP});var cP=[{tfOpName:"SparseFillEmptyRows",category:"sparse",inputs:[{start:0,name:"indices",type:"tensor"},{start:1,name:"values",type:"tensor"},{start:2,name:"denseShape",type:"tensor"},{start:3,name:"defaultValue",type:"tensor"}]},{tfOpName:"SparseReshape",category:"sparse",inputs:[{start:0,name:"inputIndices",type:"tensor"},{start:1,name:"inputShape",type:"tensor"},{start:2,name:"newShape",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"SparseSegmentMean",category:"sparse",inputs:[{start:0,name:"data",type:"tensor"},{start:1,name:"indices",type:"tensor"},{start:2,name:"segmentIds",type:"tensor"}]},{tfOpName:"SparseSegmentSum",category:"sparse",inputs:[{start:0,name:"data",type:"tensor"},{start:1,name:"indices",type:"tensor"},{start:2,name:"segmentIds",type:"tensor"}]}],O7={};Ze(O7,{json:()=>hP});var hP=[{tfOpName:"FFT",category:"spectral",inputs:[{start:0,name:"x",type:"tensor"}]},{tfOpName:"IFFT",category:"spectral",inputs:[{start:0,name:"x",type:"tensor"}]},{tfOpName:"RFFT",category:"spectral",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"fft_length",type:"number",notSupported:!0}]},{tfOpName:"IRFFT",category:"spectral",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"fft_length",type:"number",notSupported:!0}]}],z7={};Ze(z7,{json:()=>fP});var fP=[{tfOpName:"StaticRegexReplace",category:"string",inputs:[{start:0,name:"input",type:"tensor"}],attrs:[{tfName:"pattern",name:"pattern",type:"string"},{tfName:"rewrite",name:"rewrite",type:"string"},{tfName:"replace_global",name:"replaceGlobal",type:"bool"}]},{tfOpName:"StringNGrams",category:"string",inputs:[{start:0,name:"data",type:"tensor"},{start:1,name:"dataSplits",type:"tensor"}],attrs:[{tfName:"separator",name:"separator",type:"string"},{tfName:"ngram_widths",name:"nGramWidths",type:"number[]"},{tfName:"left_pad",name:"leftPad",type:"string"},{tfName:"right_pad",name:"rightPad",type:"string"},{tfName:"pad_width",name:"padWidth",type:"number"},{tfName:"preserve_short_sequences",name:"preserveShortSequences",type:"bool"}],outputs:["ngrams","ngrams_splits"]},{tfOpName:"StringSplit",category:"string",inputs:[{start:0,name:"input",type:"tensor"},{start:1,name:"delimiter",type:"tensor"}],attrs:[{tfName:"skip_empty",name:"skipEmpty",type:"bool"}],outputs:["indices","values","shape"]},{tfOpName:"StringToHashBucketFast",category:"string",inputs:[{start:0,name:"input",type:"tensor"}],attrs:[{tfName:"num_buckets",name:"numBuckets",type:"number"}]}],L7={};Ze(L7,{json:()=>mP});var mP=[{tfOpName:"Cast",category:"transformation",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"SrcT",name:"sdtype",type:"dtype",notSupported:!0},{tfName:"DstT",name:"dtype",type:"dtype"}]},{tfOpName:"ExpandDims",category:"transformation",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"axis",type:"number"}]},{tfOpName:"MirrorPad",category:"transformation",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"padding",type:"number[]"}],attrs:[{tfName:"mode",name:"mode",type:"string"}]},{tfOpName:"Pad",category:"transformation",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"padding",type:"number[]"}],attrs:[{tfName:"constant_value",name:"constantValue",type:"number",defaultValue:0}]},{tfOpName:"PadV2",category:"transformation",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"padding",type:"number[]"},{start:2,name:"constantValue",type:"number",defaultValue:0}]},{tfOpName:"Reshape",category:"transformation",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"shape",type:"number[]"}]},{tfOpName:"EnsureShape",category:"transformation",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"shape",type:"number[]"}]},{tfOpName:"Squeeze",category:"transformation",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"axis",tfDeprecatedName:"squeeze_dims",name:"axis",type:"number[]"}]},{tfOpName:"SpaceToBatchND",category:"transformation",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"blockShape",type:"number[]"},{start:2,name:"paddings",type:"number[]"}]},{tfOpName:"BatchToSpaceND",category:"transformation",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"blockShape",type:"number[]"},{start:2,name:"crops",type:"number[]"}]},{tfOpName:"DepthToSpace",category:"transformation",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"block_size",name:"blockSize",type:"number"},{tfName:"data_format",name:"dataFormat",type:"string"}]},{tfOpName:"BroadcastTo",category:"transformation",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"shape",type:"number[]"}],attrs:[]},{tfOpName:"BroadcastArgs",category:"transformation",inputs:[{start:0,name:"s0",type:"tensor"},{start:1,name:"s1",type:"tensor"}],attrs:[]}],G5=class{static get Instance(){return this._instance||(this._instance=new this)}constructor(){let e=[v7,w7,k7,I7,S7,C7,T7,N7,R7,E7,M7,F7,_7,$7,D7,P7,O7,z7,L7],t=[].concat(...e.map(a=>a.json));this.opMappers=t.reduce((a,n)=>(a[n.tfOpName]=n,a),{})}transformGraph(e,t={}){let a=e.node,n=[],r=[],s=[],i=a.reduce((f,m)=>(f[m.name]=this.mapNode(m),m.op.startsWith("Placeholder")?n.push(f[m.name]):m.op==="Const"?r.push(f[m.name]):(m.input==null||m.input.length===0)&&s.push(f[m.name]),f),{}),o=[],l=[],u={},p={};t!=null&&(u=this.mapSignatureEntries(t.inputs),p=this.mapSignatureEntries(t.outputs));let c=Object.keys(i);c.forEach(f=>{let m=i[f];m.inputNames.forEach((g,y)=>{let[x,,A]=Ar(g),b=i[x];if(b.outputs!=null){let w=b.outputs.indexOf(A);if(w!==-1){let C=`${x}:${w}`;m.inputNames[y]=C}}m.inputs.push(b),b.children.push(m)})}),Object.keys(p).length===0?c.forEach(f=>{let m=i[f];m.children.length===0&&l.push(m)}):Object.keys(p).forEach(f=>{let[m]=Ar(f),g=i[m];g!=null&&(g.signatureKey=p[f],l.push(g))}),Object.keys(u).length>0?Object.keys(u).forEach(f=>{let[m]=Ar(f),g=i[m];g&&(g.signatureKey=u[f],o.push(g))}):o=n;let d={};e.library!=null&&e.library.function!=null&&(d=e.library.function.reduce((f,m)=>(f[m.signature.name]=this.mapFunction(m),f),{}));let h={nodes:i,inputs:o,outputs:l,weights:r,placeholders:n,signature:t,functions:d};return s.length>0&&(h.initNodes=s),h}mapSignatureEntries(e){return Object.keys(e||{}).reduce((t,a)=>(t[e[a].name]=a,t),{})}mapNode(e){let t=b7(e.op)||this.opMappers[e.op]||{};e.attr==null&&(e.attr={});let a={name:e.name,op:e.op,category:t.category,inputNames:(e.input||[]).map(n=>n.startsWith("^")?n.slice(1):n),inputs:[],children:[],inputParams:{},attrParams:{},rawAttrs:e.attr,outputs:t.outputs};return t.inputs!=null&&(a.inputParams=t.inputs.reduce((n,r)=>(n[r.name]={type:r.type,inputIndexStart:r.start,inputIndexEnd:r.end},n),{})),t.attrs!=null&&(a.attrParams=t.attrs.reduce((n,r)=>{let s=r.type,i;switch(r.type){case"string":i=i1(e.attr,r.tfName,r.defaultValue),i===void 0&&r.tfDeprecatedName&&(i=i1(e.attr,r.tfDeprecatedName,r.defaultValue));break;case"string[]":i=h1(e.attr,r.tfName,r.defaultValue),i===void 0&&r.tfDeprecatedName&&(i=h1(e.attr,r.tfDeprecatedName,r.defaultValue));break;case"number":i=l1(e.attr,r.tfName,r.defaultValue||0),i===void 0&&r.tfDeprecatedName&&(i=l1(e.attr,r.tfDeprecatedName,r.defaultValue));break;case"number[]":i=c1(e.attr,r.tfName,r.defaultValue),i===void 0&&r.tfDeprecatedName&&(i=c1(e.attr,r.tfDeprecatedName,r.defaultValue));break;case"bool":i=o1(e.attr,r.tfName,r.defaultValue),i===void 0&&r.tfDeprecatedName&&(i=o1(e.attr,r.tfDeprecatedName,r.defaultValue));break;case"bool[]":i=m1(e.attr,r.tfName,r.defaultValue),i===void 0&&r.tfDeprecatedName&&(i=m1(e.attr,r.tfDeprecatedName,r.defaultValue));break;case"shape":i=p1(e.attr,r.tfName,r.defaultValue),i===void 0&&r.tfDeprecatedName&&(i=p1(e.attr,r.tfDeprecatedName,r.defaultValue));break;case"shape[]":i=f1(e.attr,r.tfName,r.defaultValue),i===void 0&&r.tfDeprecatedName&&(i=f1(e.attr,r.tfDeprecatedName,r.defaultValue));break;case"dtype":i=u1(e.attr,r.tfName,r.defaultValue),i===void 0&&r.tfDeprecatedName&&(i=u1(e.attr,r.tfDeprecatedName,r.defaultValue));break;case"dtype[]":i=d1(e.attr,r.tfName,r.defaultValue),i===void 0&&r.tfDeprecatedName&&(i=d1(e.attr,r.tfDeprecatedName,r.defaultValue));break;case"func":i=H5(e.attr,r.tfName,r.defaultValue),i===void 0&&r.tfDeprecatedName&&(i=H5(e.attr,r.tfDeprecatedName,r.defaultValue));break;case"tensor":case"tensors":break;default:throw new Error(`Unsupported param type: ${r.type} for op: ${e.op}`)}return n[r.name]={value:i,type:s},n},{})),a}mapFunction(e){let t=e.nodeDef,a=[],n=[],r={};t!=null&&(r=t.reduce((u,p)=>(u[p.name]=this.mapNode(p),p.op==="Const"&&n.push(u[p.name]),u),{}));let s=[],i=[];e.signature.inputArg.forEach(u=>{let[p]=Ar(u.name),c={name:p,op:"Placeholder",inputs:[],inputNames:[],category:"graph",inputParams:{},attrParams:{dtype:{value:K3(u.type),type:"dtype"}},children:[]};c.signatureKey=u.name,s.push(c),r[p]=c}),Object.keys(r).forEach(u=>{let p=r[u];p.inputNames.forEach((c,d)=>{let[h,,f]=Ar(c),m=r[h];if(m.outputs!=null){let g=m.outputs.indexOf(f);if(g!==-1){let y=`${h}:${g}`;p.inputNames[d]=y}}p.inputs.push(m),m.children.push(p)})});let o=e.ret;e.signature.outputArg.forEach(u=>{let[p,c]=Ar(o[u.name]),d=r[p];d!=null&&(d.defaultOutput=c,i.push(d))});let l=this.mapArgsToSignature(e);return{nodes:r,inputs:s,outputs:i,weights:n,placeholders:a,signature:l}}mapArgsToSignature(e){return{methodName:e.signature.name,inputs:e.signature.inputArg.reduce((t,a)=>(t[a.name]=this.mapArgToTensorInfo(a),t),{}),outputs:e.signature.outputArg.reduce((t,a)=>(t[a.name]=this.mapArgToTensorInfo(a,e.ret),t),{})}}mapArgToTensorInfo(e,t){let a=e.name;return t!=null&&(a=t[a]),{name:a,dtype:e.type}}};function gP(e){let t=W().global;if(typeof t.atob!="undefined")return t.atob(e);if(typeof Buffer!="undefined")return new Buffer(e,"base64").toString();throw new Error("Unable to decode base64 in this environment. Missing built-in atob() or Buffer()")}function W7(e,t){let a=Array.isArray(e)?String.fromCharCode.apply(null,e):gP(e);return t?a:a.toLowerCase()}function i1(e,t,a,n=!1){let r=e[t];return r!=null?W7(r.s,n):a}function o1(e,t,a){let n=e[t];return n?n.b:a}function l1(e,t,a){let n=e[t]||{},r=n.i!=null?n.i:n.f!=null?n.f:a;return typeof r=="number"?r:parseInt(r,10)}function K3(e){switch(typeof e=="string"&&(e=In[e]),e){case In.DT_FLOAT:case In.DT_HALF:return"float32";case In.DT_INT32:case In.DT_INT64:case In.DT_INT8:case In.DT_UINT8:return"int32";case In.DT_BOOL:return"bool";case In.DT_DOUBLE:return"float32";case In.DT_STRING:return"string";default:return null}}function H5(e,t,a){let n=e[t];return n&&n.func?n.func.name:a}function u1(e,t,a){let n=e[t];return n&&n.type?K3(n.type):a}function d1(e,t,a){let n=e[t];return n&&n.list&&n.list.type?n.list.type.map(r=>K3(r)):a}function B7(e){if(!e.unknownRank)return e.dim!=null?e.dim.map(t=>typeof t.size=="number"?t.size:parseInt(t.size,10)):[]}function p1(e,t,a){let n=e[t];return n&&n.shape?B7(n.shape):a}function c1(e,t,a){let n=e[t];return n?((n.list.f&&n.list.f.length?n.list.f:n.list.i)||[]).map(r=>typeof r=="number"?r:parseInt(r,10)):a}function h1(e,t,a,n=!1){let r=e[t];return r&&r.list&&r.list.s?r.list.s.map(s=>W7(s,n)):a}function f1(e,t,a){let n=e[t];return n&&n.list&&n.list.shape?n.list.shape.map(r=>B7(r)):a}function m1(e,t,a){let n=e[t];return n&&n.list&&n.list.b?n.list.b:a}var yP=class{constructor(e,t,a){this.node=e,this.tensorMap=t,this.context=a,this.inputs=[],this.attrs={},this.inputs=e.inputNames.map(n=>this.getInput(n)),e.rawAttrs!=null&&(this.attrs=Object.keys(e.rawAttrs).reduce((n,r)=>(n[r]=this.getAttr(r),n),{}))}getInput(e){return ua(e,this.tensorMap,this.context)}getAttr(e,t){let a=this.node.rawAttrs[e];if(a.tensor!=null)return ua(e,this.tensorMap,this.context);if(a.i!=null||a.f!=null)return l1(this.node.rawAttrs,e,t);if(a.s!=null)return i1(this.node.rawAttrs,e,t);if(a.b!=null)return o1(this.node.rawAttrs,e,t);if(a.shape!=null)return p1(this.node.rawAttrs,e,t);if(a.type!=null)return u1(this.node.rawAttrs,e,t);if(a.list!=null){if(a.list.i!=null||a.list.f!=null)return c1(this.node.rawAttrs,e,t);if(a.list.s!=null)return h1(this.node.rawAttrs,e,t);if(a.list.shape!=null)return f1(this.node.rawAttrs,e,t);if(a.list.b!=null)return m1(this.node.rawAttrs,e,t);if(a.list.type!=null)return d1(this.node.rawAttrs,e,t)}return t}},Qt={};Ze(Qt,{OP_SCOPE_SUFFIX:()=>G1,abs:()=>Za,acos:()=>MA,acosh:()=>FA,add:()=>ke,addN:()=>Eh,all:()=>_A,any:()=>$A,argMax:()=>rr,argMin:()=>DA,asin:()=>PA,asinh:()=>OA,atan:()=>zA,atan2:()=>LA,atanh:()=>WA,avgPool:()=>Q1,avgPool3d:()=>GA,basicLSTMCell:()=>HA,batchNorm:()=>Mp,batchNorm2d:()=>jA,batchNorm3d:()=>qA,batchNorm4d:()=>XA,batchToSpaceND:()=>e3,bincount:()=>t3,bitwiseAnd:()=>KA,booleanMaskAsync:()=>$4,broadcastArgs:()=>YA,broadcastTo:()=>Fl,buffer:()=>$e,cast:()=>Ke,ceil:()=>ZA,clipByValue:()=>JA,clone:()=>Sa,complex:()=>Sr,concat:()=>lt,concat1d:()=>QA,concat2d:()=>Fu,concat3d:()=>eb,concat4d:()=>tb,conv1d:()=>ab,conv2d:()=>Fp,conv2dTranspose:()=>rb,conv3d:()=>sb,conv3dTranspose:()=>ib,cos:()=>ob,cosh:()=>lb,cosineWindow:()=>Vh,cumprod:()=>ub,cumsum:()=>db,denseBincount:()=>pb,depthToSpace:()=>cb,depthwiseConv2d:()=>Mh,diag:()=>hb,dilation2d:()=>fb,div:()=>ve,divNoNan:()=>gb,dot:()=>yb,dropout:()=>L4,einsum:()=>xb,elu:()=>r3,enclosingPowerOfTwo:()=>$3,ensureShape:()=>Ab,equal:()=>n3,erf:()=>bb,euclideanNorm:()=>kb,exp:()=>Qr,expandDims:()=>Ht,expm1:()=>Ib,eye:()=>i3,fft:()=>zh,fill:()=>sr,floor:()=>o3,floorDiv:()=>Rp,fused:()=>D3,gather:()=>l3,gatherND:()=>z4,greater:()=>Dp,greaterEqual:()=>u3,ifft:()=>Vd,imag:()=>Pp,image:()=>fe,inTopKAsync:()=>W4,irfft:()=>N3,isFinite:()=>Sb,isInf:()=>Cb,isNaN:()=>Tb,leakyRelu:()=>d3,less:()=>ih,lessEqual:()=>Fh,linalg:()=>X4,linspace:()=>Nb,localResponseNormalization:()=>Rb,log:()=>Ll,log1p:()=>p3,logSigmoid:()=>Mb,logSoftmax:()=>Fb,logSumExp:()=>h3,logicalAnd:()=>Ld,logicalNot:()=>f3,logicalOr:()=>m3,logicalXor:()=>_b,losses:()=>K4,lowerBound:()=>$b,matMul:()=>pt,max:()=>ga,maxPool:()=>g3,maxPool3d:()=>Db,maxPoolWithArgmax:()=>Pb,maximum:()=>y3,mean:()=>Wd,meshgrid:()=>Ob,min:()=>Jr,minimum:()=>Bd,mirrorPad:()=>zb,mod:()=>_u,moments:()=>Lb,movingAverage:()=>D4,mul:()=>te,multiRNNCell:()=>Wb,multinomial:()=>Bb,neg:()=>Ln,norm:()=>$p,notEqual:()=>x3,oneHot:()=>oh,ones:()=>Ur,onesLike:()=>Vb,op:()=>z,outerProduct:()=>Ub,pad:()=>ir,pad1d:()=>Gb,pad2d:()=>Hb,pad3d:()=>jb,pad4d:()=>qb,pool:()=>Xb,pow:()=>zl,prelu:()=>b3,print:()=>X1,prod:()=>Kb,raggedGather:()=>Yb,raggedRange:()=>Zb,raggedTensorToTensor:()=>Jb,rand:()=>Qb,randomGamma:()=>n4,randomNormal:()=>S3,randomStandardNormal:()=>r4,randomUniform:()=>Dh,randomUniformInt:()=>s4,range:()=>Wl,real:()=>Bl,reciprocal:()=>i4,relu:()=>Op,relu6:()=>C3,reshape:()=>Q,reverse:()=>es,reverse1d:()=>o4,reverse2d:()=>l4,reverse3d:()=>u4,reverse4d:()=>d4,rfft:()=>Lh,round:()=>T3,rsqrt:()=>p4,scalar:()=>Ue,scatterND:()=>P4,searchSorted:()=>$h,selu:()=>c4,separableConv2d:()=>h4,setdiff1dAsync:()=>f4,sigmoid:()=>Wa,sign:()=>m4,signal:()=>q4,sin:()=>g4,sinh:()=>y4,slice:()=>De,slice1d:()=>x4,slice2d:()=>A4,slice3d:()=>zp,slice4d:()=>Ph,softmax:()=>Oh,softplus:()=>c3,spaceToBatchND:()=>A3,sparse:()=>Y4,sparseToDense:()=>O4,spectral:()=>j4,split:()=>Ca,sqrt:()=>er,square:()=>Cn,squaredDifference:()=>R3,squeeze:()=>Oe,stack:()=>ca,step:()=>E3,stridedSlice:()=>b4,string:()=>Z4,sub:()=>ye,sum:()=>ot,tan:()=>v4,tanh:()=>sh,tensor:()=>Ve,tensor1d:()=>jt,tensor2d:()=>Zn,tensor3d:()=>M3,tensor4d:()=>w4,tensor5d:()=>k4,tensor6d:()=>I4,tensorScatterUpdate:()=>C4,tile:()=>jr,topk:()=>T4,transpose:()=>js,truncatedNormal:()=>N4,unique:()=>R4,unsortedSegmentSum:()=>E4,unstack:()=>Ra,upperBound:()=>M4,variable:()=>F4,where:()=>kr,whereAsync:()=>_3,zeros:()=>gn,zerosLike:()=>Qa});var xP=(e,t,a,n=Qt)=>{switch(e.op){case"BiasAdd":case"AddV2":case"Add":return[n.add(k("a",e,t,a),k("b",e,t,a))];case"AddN":return[n.addN(k("tensors",e,t,a))];case"FloorMod":case"Mod":return[n.mod(k("a",e,t,a),k("b",e,t,a))];case"Mul":return[n.mul(k("a",e,t,a),k("b",e,t,a))];case"RealDiv":case"Div":return[n.div(k("a",e,t,a),k("b",e,t,a))];case"DivNoNan":return[n.divNoNan(k("a",e,t,a),k("b",e,t,a))];case"FloorDiv":return[n.floorDiv(k("a",e,t,a),k("b",e,t,a))];case"Sub":return[n.sub(k("a",e,t,a),k("b",e,t,a))];case"Minimum":return[n.minimum(k("a",e,t,a),k("b",e,t,a))];case"Maximum":return[n.maximum(k("a",e,t,a),k("b",e,t,a))];case"Pow":return[n.pow(k("a",e,t,a),k("b",e,t,a))];case"SquaredDifference":return[n.squaredDifference(k("a",e,t,a),k("b",e,t,a))];default:throw TypeError(`Node type ${e.op} is not implemented`)}},AP=(e,t,a,n=Qt)=>{switch(e.op){case"Abs":case"ComplexAbs":return[n.abs(k("x",e,t,a))];case"Acos":return[n.acos(k("x",e,t,a))];case"Acosh":return[n.acosh(k("x",e,t,a))];case"Asin":return[n.asin(k("x",e,t,a))];case"Asinh":return[n.asinh(k("x",e,t,a))];case"Atan":return[n.atan(k("x",e,t,a))];case"Atan2":return[n.atan2(k("x",e,t,a),k("y",e,t,a))];case"Atanh":return[n.atanh(k("x",e,t,a))];case"Ceil":return[n.ceil(k("x",e,t,a))];case"Complex":return[n.complex(k("real",e,t,a),k("imag",e,t,a))];case"Cos":return[n.cos(k("x",e,t,a))];case"Cosh":return[n.cosh(k("x",e,t,a))];case"Elu":return[n.elu(k("x",e,t,a))];case"Erf":return[n.erf(k("x",e,t,a))];case"Exp":return[n.exp(k("x",e,t,a))];case"Expm1":return[n.expm1(k("x",e,t,a))];case"Floor":return[n.floor(k("x",e,t,a))];case"Log":return[n.log(k("x",e,t,a))];case"Log1p":return[n.log1p(k("x",e,t,a))];case"Imag":return[n.imag(k("x",e,t,a))];case"Neg":return[n.neg(k("x",e,t,a))];case"Reciprocal":return[n.reciprocal(k("x",e,t,a))];case"Real":return[n.real(k("x",e,t,a))];case"Relu":return[n.relu(k("x",e,t,a))];case"Round":return[n.round(k("x",e,t,a))];case"Selu":return[n.selu(k("x",e,t,a))];case"Sigmoid":return[n.sigmoid(k("x",e,t,a))];case"Sin":return[n.sin(k("x",e,t,a))];case"Sign":return[n.sign(k("x",e,t,a))];case"Sinh":return[n.sinh(k("x",e,t,a))];case"Softplus":return[n.softplus(k("x",e,t,a))];case"Sqrt":return[n.sqrt(k("x",e,t,a))];case"Square":return[n.square(k("x",e,t,a))];case"Tanh":return[n.tanh(k("x",e,t,a))];case"Tan":return[n.tan(k("x",e,t,a))];case"ClipByValue":return[n.clipByValue(k("x",e,t,a),k("clipValueMin",e,t,a),k("clipValueMax",e,t,a))];case"Relu6":return[n.relu6(k("x",e,t,a))];case"Rsqrt":return[n.rsqrt(ua(e.inputNames[0],t,a))];case"LeakyRelu":return[n.leakyRelu(k("x",e,t,a),k("alpha",e,t,a))];case"Prelu":return[n.prelu(k("x",e,t,a),k("alpha",e,t,a))];case"IsNan":return[n.isNaN(ua(e.inputNames[0],t,a))];case"IsInf":return[n.isInf(ua(e.inputNames[0],t,a))];case"IsFinite":return[n.isFinite(ua(e.inputNames[0],t,a))];default:throw TypeError(`Node type ${e.op} is not implemented`)}};function Sn(e,t,a=""){if(!(typeof e=="number"||typeof t=="number")){v.assert(e.length===t.length,()=>a+` Shapes ${e} and ${t} must match`);for(let n=0;n<e.length;n++){let r=e[n],s=t[n];v.assert(r<0||s<0||r===s,()=>a+` Shapes ${e} and ${t} must match`)}}}function j5(e){return!(typeof e=="number"||e.some(t=>t<0))}function fd(e,t,a){let n=g1(e,a),r=!j5(n);if(r&&t.length===0)throw new Error(`Tried to calculate elements of an empty list with non-fully-defined elementShape: ${n}`);if(r&&t.forEach(s=>{n=g1(s.shape,n)}),!j5(n))throw new Error(`Non-fully-defined elementShape: ${n}`);return n}function g1(e,t){if(typeof e=="number")return t;if(typeof t=="number")return e;if(e.length!==t.length)throw new Error(`Incompatible ranks during merge: ${e} vs. ${t}`);let a=[];for(let n=0;n<e.length;++n){let r=e[n],s=t[n];if(r>=0&&s>=0&&r!==s)throw new Error(`Incompatible shape during merge: ${e} vs. ${t}`);a[n]=r>=0?r:s}return a}var bP=class{constructor(e,t,a,n,r,s,i){this.name=e,this.dtype=t,this.maxSize=a,this.elementShape=n,this.identicalElementShapes=r,this.dynamicSize=s,this.clearAfterRead=i,this.tensors=[],this.closed_=!1,this.idTensor=Ue(0),zn(this.idTensor)}get id(){return this.idTensor.id}get closed(){return this.closed_}clearAndClose(e){this.tensors.forEach(t=>{(e==null||!e.has(t.tensor.id))&&t.tensor.dispose()}),this.tensors=[],this.closed_=!0,this.idTensor.dispose()}size(){return this.tensors.length}read(e){if(this.closed_)throw new Error(`TensorArray ${this.name} has already been closed.`);if(e<0||e>=this.size())throw new Error(`Tried to read from index ${e}, but array size is: ${this.size()}`);let t=this.tensors[e];if(t.cleared)throw new Error(`TensorArray ${this.name}: Could not read index ${e} twice because it was cleared after a previous read (perhaps try setting clear_after_read = false?).`);return this.clearAfterRead&&(t.cleared=!0),t.read=!0,t.tensor}readMany(e){return e.map(t=>this.read(t))}write(e,t){if(this.closed_)throw new Error(`TensorArray ${this.name} has already been closed.`);if(e<0||!this.dynamicSize&&e>=this.maxSize)throw new Error(`Tried to write to index ${e}, but array is not resizeable and size is: ${this.maxSize}`);let a=this.tensors[e]||{};if(t.dtype!==this.dtype)throw new Error(`TensorArray ${this.name}: Could not write to TensorArray index ${e},
|
|
because the value dtype is ${t.dtype}, but TensorArray dtype is ${this.dtype}.`);if(this.size()===0&&(this.elementShape==null||this.elementShape.length===0)&&(this.elementShape=t.shape),Sn(this.elementShape,t.shape,`TensorArray ${this.name}: Could not write to TensorArray index ${e}.`),a.read)throw new Error(`TensorArray ${this.name}: Could not write to TensorArray index ${e}, because it has already been read.`);if(a.written)throw new Error(`TensorArray ${this.name}: Could not write to TensorArray index ${e}, because it has already been written.`);a.tensor=t,zn(t),a.written=!0,this.tensors[e]=a}writeMany(e,t){if(e.length!==t.length)throw new Error(`TensorArray ${this.name}: could not write multiple tensors,because the index size: ${e.length} is not the same as tensors size: ${t.length}.`);e.forEach((a,n)=>this.write(a,t[n]))}gather(e,t){if(t&&t!==this.dtype)throw new Error(`TensorArray dtype is ${this.dtype} but gather requested dtype ${t}`);if(e)e=e.slice(0,this.size());else{e=[];for(let n=0;n<this.size();n++)e.push(n)}if(e.length===0)return Ve([],[0].concat(this.elementShape));let a=this.readMany(e);return Sn(this.elementShape,a[0].shape,"TensorArray shape mismatch: "),ca(a,0)}concat(e){if(e&&e!==this.dtype)throw new Error(`TensorArray dtype is ${this.dtype} but concat requested dtype ${e}`);if(this.size()===0)return Ve([],[0].concat(this.elementShape));let t=[];for(let n=0;n<this.size();n++)t.push(n);let a=this.readMany(t);return Sn(this.elementShape,a[0].shape,`TensorArray shape mismatch: tensor array shape (${this.elementShape}) vs first tensor shape (${a[0].shape})`),lt(a,0)}scatter(e,t){if(t.dtype!==this.dtype)throw new Error(`TensorArray dtype is ${this.dtype} but tensor has dtype ${t.dtype}`);if(e.length!==t.shape[0])throw new Error(`Expected len(indices) == tensor.shape[0], but saw: ${e.length} vs. ${t.shape[0]}`);let a=Math.max(...e);if(!this.dynamicSize&&a>=this.maxSize)throw new Error(`Max index must be < array size (${a} vs. ${this.maxSize})`);this.writeMany(e,Ra(t,0))}split(e,t){if(t.dtype!==this.dtype)throw new Error(`TensorArray dtype is ${this.dtype} but tensor has dtype ${t.dtype}`);let a=0,n=e.map(o=>(a+=o,a));if(a!==t.shape[0])throw new Error(`Expected sum of lengths to be equal to
|
|
tensor.shape[0], but sum of lengths is
|
|
${a}, and tensor's shape is: ${t.shape}`);if(!this.dynamicSize&&e.length!==this.maxSize)throw new Error(`TensorArray's size is not equal to the size of lengths (${this.maxSize} vs. ${e.length}), and the TensorArray is not marked as dynamically resizeable`);let r=a===0?0:t.size/a,s=[];Pe(()=>{t=Q(t,[1,a,r]);for(let o=0;o<e.length;++o){let l=[0,o===0?0:n[o-1],0],u=[1,e[o],r];s[o]=Q(De(t,l,u),this.elementShape)}return s});let i=[];for(let o=0;o<e.length;o++)i[o]=o;this.writeMany(i,s)}},Ul=class{get id(){return this.idTensor.id}constructor(e,t,a,n=-1){this.tensors=e,this.elementShape=t,this.elementDtype=a,e!=null&&e.forEach(r=>{if(a!==r.dtype)throw new Error(`Invalid data types; op elements ${a}, but list elements ${r.dtype}`);Sn(t,r.shape,"TensorList shape mismatch: "),zn(r)}),this.idTensor=Ue(0),this.maxNumElements=n,zn(this.idTensor)}copy(){return new Ul([...this.tensors],this.elementShape,this.elementDtype)}clearAndClose(e){this.tensors.forEach(t=>{(e==null||!e.has(t.id))&&t.dispose()}),this.tensors.length=0,this.idTensor.dispose()}size(){return this.tensors.length}stack(e,t,a=-1){if(t!==this.elementDtype)throw new Error(`Invalid data types; op elements ${t}, but list elements ${this.elementDtype}`);if(a!==-1&&this.tensors.length!==a)throw new Error(`Operation expected a list with ${a} elements but got a list with ${this.tensors.length} elements.`);Sn(e,this.elementShape,"TensorList shape mismatch: ");let n=fd(this.elementShape,this.tensors,e);return Pe(()=>{let r=this.tensors.map(s=>Q(s,n));return ca(r,0)})}popBack(e,t){if(t!==this.elementDtype)throw new Error(`Invalid data types; op elements ${t}, but list elements ${this.elementDtype}`);if(this.size()===0)throw new Error("Trying to pop from an empty list.");let a=fd(this.elementShape,this.tensors,e),n=this.tensors.pop();return n.kept=!1,Sn(n.shape,e,"TensorList shape mismatch: "),Q(n,a)}pushBack(e){if(e.dtype!==this.elementDtype)throw new Error(`Invalid data types; op elements ${e.dtype}, but list elements ${this.elementDtype}`);if(Sn(e.shape,this.elementShape,"TensorList shape mismatch: "),this.maxNumElements===this.size())throw new Error("Trying to push element into a full list.");zn(e),this.tensors.push(e)}resize(e){if(e<0)throw new Error(`TensorListResize expects size to be non-negative. Got: ${e}`);if(this.maxNumElements!==-1&&e>this.maxNumElements)throw new Error(`TensorListResize input size ${e} is greater maxNumElement ${this.maxNumElements}.`);let t=new Ul([],this.elementShape,this.elementDtype,this.maxNumElements);t.tensors.length=e;for(let a=0;a<Math.min(this.tensors.length,e);++a)t.tensors[a]=this.tensors[a];return t}getItem(e,t,a){if(a!==this.elementDtype)throw new Error(`Invalid data types; op elements ${a}, but list elements ${this.elementDtype}`);if(e<0||e>this.tensors.length)throw new Error(`Trying to access element ${e} in a list with ${this.tensors.length} elements.`);if(this.tensors[e]==null)throw new Error(`element at index ${e} is null.`);Sn(this.tensors[e].shape,t,"TensorList shape mismatch: ");let n=fd(this.elementShape,this.tensors,t);return Q(this.tensors[e],n)}setItem(e,t){if(t.dtype!==this.elementDtype)throw new Error(`Invalid data types; op elements ${t.dtype}, but list elements ${this.elementDtype}`);if(e<0||this.maxNumElements!==-1&&e>=this.maxNumElements)throw new Error(`Trying to set element ${e} in a list with max ${this.maxNumElements} elements.`);Sn(this.elementShape,t.shape,"TensorList shape mismatch: "),zn(t),this.tensors[e]!=null&&(this.tensors[e].kept=!1),this.tensors[e]=t}gather(e,t,a){if(t!==this.elementDtype)throw new Error(`Invalid data types; op elements ${t}, but list elements ${this.elementDtype}`);Sn(this.elementShape,a,"TensorList shape mismatch: "),e=e.slice(0,this.size());let n=fd(this.elementShape,this.tensors,a);return e.length===0?Ve([],[0].concat(n)):Pe(()=>{let r=e.map(s=>Q(this.tensors[s],n));return ca(r,0)})}concat(e,t){if(e&&e!==this.elementDtype)throw new Error(`TensorList dtype is ${this.elementDtype} but concat requested dtype ${e}`);Sn(this.elementShape,t,"TensorList shape mismatch: ");let a=fd(this.elementShape,this.tensors,t);return this.size()===0?Ve([],[0].concat(a)):Pe(()=>{let n=this.tensors.map(r=>Q(r,a));return lt(n,0)})}};function vP(e,t,a){let n=e.dtype;if(e.shape.length<1)throw new Error(`Tensor must be at least a vector, but saw shape: ${e.shape}`);if(e.dtype!==a)throw new Error(`Invalid data types; op elements ${e.dtype}, but list elements ${a}`);let r=e.shape.slice(1);Sn(r,t,"TensorList shape mismatch: ");let s=Ra(e);return new Ul(s,t,n)}function wP(e,t,a,n){return new Ul([],e,t,n)}function kP(e,t,a,n){if(t.length!==e.shape[0])throw new Error(`Expected len(indices) == tensor.shape[0], but saw: ${t.length} vs. ${e.shape[0]}`);let r=Math.max(...t);if(n!=null&&n!==-1&&r>=n)throw new Error(`Max index must be < array size (${r} vs. ${n})`);let s=new Ul([],a,e.dtype,n),i=Ra(e,0);return t.forEach((o,l)=>{s.setItem(o,i[l])}),s}function IP(e,t,a){let n=0,r=t.map(p=>(n+=p,n));if(n!==e.shape[0])throw new Error(`Expected sum of lengths to be equal to
|
|
tensor.shape[0], but sum of lengths is
|
|
${n}, and tensor's shape is: ${e.shape}`);let s=e.shape.slice(1),i=g1(s,a),o=n===0?0:e.size/n,l=Pe(()=>{let p=[];e=Q(e,[1,n,o]);for(let c=0;c<t.length;++c){let d=[0,c===0?0:r[c-1],0],h=[1,t[c],o];p[c]=Q(De(e,d,h),i)}return e.dispose(),p}),u=new Ul([],a,e.dtype,t.length);for(let p=0;p<l.length;p++)u.setItem(p,l[p]);return u}var SP=async(e,t,a)=>{switch(e.op){case"If":case"StatelessIf":{let n=k("thenBranch",e,t,a),r=k("elseBranch",e,t,a),s=k("cond",e,t,a),i=k("args",e,t,a);return(await s.data())[0]?a.functionMap[n].executeFunctionAsync(i,a.tensorArrayMap,a.tensorListMap):a.functionMap[r].executeFunctionAsync(i,a.tensorArrayMap,a.tensorListMap)}case"While":case"StatelessWhile":{let n=k("body",e,t,a),r=k("cond",e,t,a),s=k("args",e,t,a),i=await a.functionMap[r].executeFunctionAsync(s,a.tensorArrayMap,a.tensorListMap),o=s.map(p=>p.id),l=await i[0].data();i.forEach(p=>{!p.kept&&o.indexOf(p.id)===-1&&p.dispose()});let u=s;for(;l[0];){let p=u;u=await a.functionMap[n].executeFunctionAsync(u,a.tensorArrayMap,a.tensorListMap);let c=u.map(h=>h.id);p.forEach(h=>{!h.kept&&o.indexOf(h.id)===-1&&c.indexOf(h.id)===-1&&h.dispose()});let d=await a.functionMap[r].executeFunctionAsync(u,a.tensorArrayMap,a.tensorListMap);l=await d[0].data(),d.forEach(h=>{!h.kept&&o.indexOf(h.id)===-1&&c.indexOf(h.id)===-1&&h.dispose()})}return u}case"LoopCond":{let n=k("pred",e,t,a);return[br(n)]}case"Switch":{let n=k("pred",e,t,a),r=k("data",e,t,a);return r.kept||(r=br(r)),(await n.data())[0]?[void 0,r]:[r,void 0]}case"Merge":{let n=e.inputNames.find(r=>ua(r,t,a)!==void 0);if(n){let r=ua(n,t,a);return[br(r)]}return}case"Enter":{let n=k("frameName",e,t,a),r=k("tensor",e,t,a);return a.enterFrame(n),[br(r)]}case"Exit":{let n=k("tensor",e,t,a);return a.exitFrame(),[br(n)]}case"NextIteration":{let n=k("tensor",e,t,a);return a.nextIteration(),[br(n)]}case"TensorArrayV3":{let n=k("size",e,t,a),r=k("dtype",e,t,a),s=k("elementShape",e,t,a),i=k("dynamicSize",e,t,a),o=k("clearAfterRead",e,t,a),l=k("identicalElementShapes",e,t,a),u=k("name",e,t,a),p=new bP(u,r,n,s,l,i,o);return a.addTensorArray(p),[p.idTensor,Ue(1)]}case"TensorArrayWriteV3":{let n=k("tensorArrayId",e,t,a),r=k("index",e,t,a),s=k("tensor",e,t,a),i=a.getTensorArray(n.id);return i.write(r,s),[i.idTensor]}case"TensorArrayReadV3":{let n=k("tensorArrayId",e,t,a),r=k("index",e,t,a);return[a.getTensorArray(n.id).read(r)]}case"TensorArrayGatherV3":{let n=k("tensorArrayId",e,t,a),r=k("indices",e,t,a),s=k("dtype",e,t,a);return[a.getTensorArray(n.id).gather(r,s)]}case"TensorArrayScatterV3":{let n=k("tensorArrayId",e,t,a),r=k("indices",e,t,a),s=k("tensor",e,t,a),i=a.getTensorArray(n.id);return i.scatter(r,s),[i.idTensor]}case"TensorArrayConcatV3":{let n=k("tensorArrayId",e,t,a),r=a.getTensorArray(n.id),s=k("dtype",e,t,a);return[r.concat(s)]}case"TensorArraySplitV3":{let n=k("tensorArrayId",e,t,a),r=k("tensor",e,t,a),s=k("lengths",e,t,a),i=a.getTensorArray(n.id);return i.split(s,r),[i.idTensor]}case"TensorArraySizeV3":{let n=k("tensorArrayId",e,t,a),r=a.getTensorArray(n.id);return[Ue(r.size(),"int32")]}case"TensorArrayCloseV3":{let n=k("tensorArrayId",e,t,a),r=a.getTensorArray(n.id);return r.clearAndClose(),[r.idTensor]}case"TensorListSetItem":{let n=k("tensorListId",e,t,a),r=k("index",e,t,a),s=k("tensor",e,t,a),i=a.getTensorList(n.id);return i.setItem(r,s),[i.idTensor]}case"TensorListGetItem":{let n=k("tensorListId",e,t,a),r=k("index",e,t,a),s=k("elementShape",e,t,a),i=k("elementDType",e,t,a);return[a.getTensorList(n.id).getItem(r,s,i)]}case"TensorListScatterV2":case"TensorListScatter":{let n=k("indices",e,t,a),r=k("tensor",e,t,a),s=k("elementShape",e,t,a),i=k("numElements",e,t,a),o=kP(r,n,s,i);return a.addTensorList(o),[o.idTensor]}case"TensorListReserve":case"EmptyTensorList":{let n=k("elementShape",e,t,a),r=k("elementDType",e,t,a),s;e.op==="TensorListReserve"?s="numElements":s="maxNumElements";let i=k(s,e,t,a),o=e.op==="TensorListReserve"?-1:i,l=wP(n,r,i,o);return a.addTensorList(l),[l.idTensor]}case"TensorListGather":{let n=k("tensorListId",e,t,a),r=k("indices",e,t,a),s=k("elementShape",e,t,a),i=k("elementDType",e,t,a);return[a.getTensorList(n.id).gather(r,i,s)]}case"TensorListStack":{let n=k("tensorListId",e,t,a),r=k("elementShape",e,t,a),s=k("elementDType",e,t,a),i=k("numElements",e,t,a);return[a.getTensorList(n.id).stack(r,s,i)]}case"TensorListFromTensor":{let n=k("tensor",e,t,a),r=k("elementShape",e,t,a),s=k("elementDType",e,t,a),i=vP(n,r,s);return a.addTensorList(i),[i.idTensor]}case"TensorListConcat":case"TensorListConcatV2":{let n=k("tensorListId",e,t,a),r=a.getTensorList(n.id),s=k("dtype",e,t,a),i=k("elementShape",e,t,a);return[r.concat(s,i)]}case"TensorListPushBack":{let n=k("tensorListId",e,t,a),r=k("tensor",e,t,a),s=a.getTensorList(n.id);return s.pushBack(r),[s.idTensor]}case"TensorListPopBack":{let n=k("tensorListId",e,t,a),r=k("elementShape",e,t,a),s=k("elementDType",e,t,a);return[a.getTensorList(n.id).popBack(r,s)]}case"TensorListSplit":{let n=k("tensor",e,t,a),r=k("elementShape",e,t,a),s=k("lengths",e,t,a),i=IP(n,s,r);return a.addTensorList(i),[i.idTensor]}case"TensorListLength":{let n=k("tensorListId",e,t,a),r=a.getTensorList(n.id);return[Ue(r.size(),"int32")]}case"TensorListResize":{let n=k("tensorListId",e,t,a),r=k("size",e,t,a),s=a.getTensorList(n.id).resize(r);return a.addTensorList(s),[s.idTensor]}default:throw TypeError(`Node type ${e.op} is not implemented`)}};function q5(e,t,a){let[n,r]=k("fusedOps",e,t,a),s=n==="biasadd",i=!s,o=r==="prelu",l=n==="fusedbatchnorm",u=k("numArgs",e,t,a);if(s){if(o&&u!==2)throw new Error("FusedConv2d and DepthwiseConv2d with BiasAdd and Prelu must have two extra arguments: bias and alpha.");if(!o&&s&&u!==1)throw new Error("FusedConv2d and DepthwiseConv2d with BiasAdd must have one extra argument: bias.")}if(l)throw new Error("FusedConv2d and DepthwiseConv2d with FusedBatchNorm is not supported");let p=k("strides",e,t,a),c=qc(e,t,a),d=k("dataFormat",e,t,a).toUpperCase(),h=k("dilations",e,t,a),[f,m]=k("args",e,t,a);i&&(m=f,f=void 0);let g=k("leakyreluAlpha",e,t,a);return{stride:p,pad:c,dataFormat:d,dilations:h,biasArg:f,preluArg:m,activationFunc:r,leakyreluAlpha:g}}var CP=(e,t,a,n=Qt)=>{switch(e.op){case"Conv1D":{let r=k("stride",e,t,a),s=k("pad",e,t,a),i=k("dataFormat",e,t,a).toUpperCase(),o=k("dilation",e,t,a);return[n.conv1d(k("x",e,t,a),k("filter",e,t,a),r,s,i,o)]}case"Conv2D":{let r=k("strides",e,t,a),s=qc(e,t,a),i=k("dataFormat",e,t,a).toUpperCase(),o=k("dilations",e,t,a);return[n.conv2d(k("x",e,t,a),k("filter",e,t,a),[r[1],r[2]],s,i,[o[1],o[2]])]}case"_FusedConv2D":{let{stride:r,pad:s,dataFormat:i,dilations:o,biasArg:l,preluArg:u,activationFunc:p,leakyreluAlpha:c}=q5(e,t,a);return[n.fused.conv2d({x:k("x",e,t,a),filter:k("filter",e,t,a),strides:[r[1],r[2]],pad:s,dataFormat:i,dilations:[o[1],o[2]],bias:l,activation:p,preluActivationWeights:u,leakyreluAlpha:c})]}case"FusedDepthwiseConv2dNative":{let{stride:r,pad:s,dataFormat:i,dilations:o,biasArg:l,preluArg:u,activationFunc:p,leakyreluAlpha:c}=q5(e,t,a);return[n.fused.depthwiseConv2d({x:k("x",e,t,a),filter:k("filter",e,t,a),strides:[r[1],r[2]],pad:s,dataFormat:i,dilations:[o[1],o[2]],bias:l,activation:p,preluActivationWeights:u,leakyreluAlpha:c})]}case"Conv2DBackpropInput":case"Conv2dTranspose":{let r=k("outputShape",e,t,a),s=k("strides",e,t,a),i=qc(e,t,a);return[n.conv2dTranspose(k("x",e,t,a),k("filter",e,t,a),r,[s[1],s[2]],i)]}case"DepthwiseConv2dNative":case"DepthwiseConv2d":{let r=k("strides",e,t,a),s=qc(e,t,a),i=k("dilations",e,t,a),o=k("dataFormat",e,t,a).toUpperCase();return[n.depthwiseConv2d(k("input",e,t,a),k("filter",e,t,a),[r[1],r[2]],s,o,[i[1],i[2]])]}case"Conv3D":{let r=k("strides",e,t,a),s=k("pad",e,t,a),i=k("dataFormat",e,t,a).toUpperCase(),o=k("dilations",e,t,a);return[n.conv3d(k("x",e,t,a),k("filter",e,t,a),[r[1],r[2],r[3]],s,i,[o[1],o[2],o[3]])]}case"AvgPool":{let r=k("strides",e,t,a),s=k("pad",e,t,a),i=k("kernelSize",e,t,a);return[n.avgPool(k("x",e,t,a),[i[1],i[2]],[r[1],r[2]],s)]}case"MaxPool":{let r=k("strides",e,t,a),s=k("pad",e,t,a),i=k("kernelSize",e,t,a);return[n.maxPool(k("x",e,t,a),[i[1],i[2]],[r[1],r[2]],s)]}case"MaxPoolWithArgmax":{let r=k("strides",e,t,a),s=k("pad",e,t,a),i=k("kernelSize",e,t,a),o=k("includeBatchInIndex",e,t,a),{result:l,indexes:u}=n.maxPoolWithArgmax(k("x",e,t,a),[i[1],i[2]],[r[1],r[2]],s,o);return[l,u]}case"AvgPool3D":{let r=k("strides",e,t,a),s=k("pad",e,t,a),i=k("kernelSize",e,t,a);return[n.avgPool3d(k("x",e,t,a),[i[1],i[2],i[3]],[r[1],r[2],r[3]],s)]}case"MaxPool3D":{let r=k("strides",e,t,a),s=k("pad",e,t,a),i=k("kernelSize",e,t,a);return[n.maxPool3d(k("x",e,t,a),[i[1],i[2],i[3]],[r[1],r[2],r[3]],s)]}case"Dilation2D":{let r=k("strides",e,t,a),s=k("pad",e,t,a),i=k("dilations",e,t,a),o=r[1],l=r[2],u=i[1],p=i[2];return[n.dilation2d(k("x",e,t,a),k("filter",e,t,a),[o,l],s,[u,p],"NHWC")]}default:throw TypeError(`Node type ${e.op} is not implemented`)}},TP=(e,t,a,n=Qt)=>{switch(e.op){case"Fill":{let r=k("shape",e,t,a),s=k("dtype",e,t,a),i=k("value",e,t,a);return[n.fill(r,i,s)]}case"LinSpace":{let r=k("start",e,t,a),s=k("stop",e,t,a),i=k("num",e,t,a);return[n.linspace(r,s,i)]}case"Multinomial":{let r=k("logits",e,t,a),s=k("numSamples",e,t,a),i=k("seed",e,t,a);return[n.multinomial(r,s,i)]}case"OneHot":{let r=k("indices",e,t,a),s=k("depth",e,t,a),i=k("onValue",e,t,a),o=k("offValue",e,t,a),l=k("dtype",e,t,a);return[n.oneHot(r,s,i,o,l)]}case"Ones":return[n.ones(k("shape",e,t,a),k("dtype",e,t,a))];case"OnesLike":return[n.onesLike(k("x",e,t,a))];case"RandomStandardNormal":return[n.randomStandardNormal(k("shape",e,t,a),k("dtype",e,t,a),k("seed",e,t,a))];case"RandomUniform":return[n.randomUniform(k("shape",e,t,a),k("minval",e,t,a),k("maxval",e,t,a),k("dtype",e,t,a))];case"RandomUniformInt":return[n.randomUniformInt(k("shape",e,t,a),k("minval",e,t,a),k("maxval",e,t,a),k("seed",e,t,a))];case"Range":{let r=k("start",e,t,a),s=k("stop",e,t,a),i=k("step",e,t,a);return[n.range(r,s,i,k("dtype",e,t,a))]}case"TruncatedNormal":{let r=k("shape",e,t,a),s=k("mean",e,t,a),i=k("stdDev",e,t,a),o=k("seed",e,t,a);return[n.truncatedNormal(r,s,i,k("dtype",e,t,a),o)]}case"Zeros":return[n.zeros(k("shape",e,t,a),k("dtype",e,t,a))];case"ZerosLike":return[n.zerosLike(k("x",e,t,a))];default:throw TypeError(`Node type ${e.op} is not implemented`)}};function L2(e,t,a){let n=k("boxes",e,t,a),r=k("scores",e,t,a),s=k("maxOutputSize",e,t,a),i=k("iouThreshold",e,t,a),o=k("scoreThreshold",e,t,a),l=k("softNmsSigma",e,t,a);return{boxes:n,scores:r,maxOutputSize:s,iouThreshold:i,scoreThreshold:o,softNmsSigma:l}}var NP=async(e,t,a,n,r=Qt)=>{switch(e.op){case"NonMaxSuppressionV5":{let{boxes:s,scores:i,maxOutputSize:o,iouThreshold:l,scoreThreshold:u,softNmsSigma:p}=L2(e,t,a),c=await r.image.nonMaxSuppressionWithScoreAsync(s,i,o,l,u,p);return[c.selectedIndices,c.selectedScores]}case"NonMaxSuppressionV4":{let{boxes:s,scores:i,maxOutputSize:o,iouThreshold:l,scoreThreshold:u}=L2(e,t,a),p=k("padToMaxOutputSize",e,t,a),c=await r.image.nonMaxSuppressionPaddedAsync(s,i,o,l,u,p);return[c.selectedIndices,c.validOutputs]}case"NonMaxSuppressionV3":case"NonMaxSuppressionV2":{let{boxes:s,scores:i,maxOutputSize:o,iouThreshold:l,scoreThreshold:u}=L2(e,t,a);return[await r.image.nonMaxSuppressionAsync(s,i,o,l,u)]}case"Where":{let s=r.cast(k("condition",e,t,a),"bool"),i=[await r.whereAsync(s)];return s.dispose(),i}case"ListDiff":return r.setdiff1dAsync(k("x",e,t,a),k("y",e,t,a));default:throw TypeError(`Node type ${e.op} is not implemented`)}},RP=(e,t,a,n=Qt)=>{switch(e.op){case"LowerBound":{let r=k("sortedSequence",e,t,a),s=k("values",e,t,a);return[n.lowerBound(r,s)]}case"TopKV2":{let r=k("x",e,t,a),s=k("k",e,t,a),i=k("sorted",e,t,a),o=n.topk(r,s,i);return[o.values,o.indices]}case"UpperBound":{let r=k("sortedSequence",e,t,a),s=k("values",e,t,a);return[n.upperBound(r,s)]}case"Unique":{let r=k("x",e,t,a),s=n.unique(r);return[s.values,s.indices]}case"UniqueV2":{let r=k("x",e,t,a),s=k("axis",e,t,a),i=n.unique(r,s);return[i.values,i.indices]}default:throw TypeError(`Node type ${e.op} is not implemented`)}},EP=(e,t,a,n=Qt)=>{switch(e.op){case"Const":return t[e.name];case"PlaceholderWithDefault":let r=k("default",e,t,a);return[ua(e.name,t,a)||r];case"Placeholder":return[ua(e.name,t,a)];case"Identity":case"StopGradient":case"FakeQuantWithMinMaxVars":{let p=k("x",e,t,a);return[br(p)]}case"IdentityN":return k("x",e,t,a).map(p=>br(p));case"Snapshot":let s=k("x",e,t,a);return[br(s)];case"Shape":return[n.tensor1d(k("x",e,t,a).shape,"int32")];case"ShapeN":return k("x",e,t,a).map(p=>n.tensor1d(p.shape));case"Size":return[n.scalar(k("x",e,t,a).size,"int32")];case"Rank":return[n.scalar(k("x",e,t,a).rank,"int32")];case"NoOp":return[n.scalar(1)];case"Print":let i=k("x",e,t,a),o=k("data",e,t,a),l=k("message",e,t,a),u=k("summarize",e,t,a);console.warn("The graph has a tf.print() operation,usually used for debugging, which slows down performance."),console.log(l);for(let p=0;p<o.length;p++)console.log(Array.prototype.slice.call(o[p].dataSync()).slice(0,u));return[i];default:throw TypeError(`Node type ${e.op} is not implemented`)}},MP=class{get id(){return this.handle.id}constructor(e,t){this.keyDType=e,this.valueDType=t,this.handle=Ue(0),this.tensorMap=new Map,zn(this.handle)}clearAndClose(){this.tensorMap.forEach(e=>e.dispose()),this.tensorMap.clear(),this.handle.dispose()}size(){return this.tensorMap.size}tensorSize(){return Ue(this.size(),"int32")}async import(e,t){this.checkKeyAndValueTensor(e,t);let a=await e.data();return this.tensorMap.forEach(n=>n.dispose()),this.tensorMap.clear(),Pe(()=>{let n=Ra(t),r=a.length,s=n.length;v.assert(r===s,()=>`The number of elements doesn't match, keys has ${r} elements, the values has ${s} elements.`);for(let i=0;i<r;i++){let o=a[i],l=n[i];zn(l),this.tensorMap.set(o,l)}return this.handle})}async find(e,t){this.checkKeyAndValueTensor(e,t);let a=await e.data();return Pe(()=>{let n=[];for(let r=0;r<a.length;r++){let s=a[r],i=this.findWithDefault(s,t);n.push(i)}return ca(n)})}findWithDefault(e,t){let a=this.tensorMap.get(e);return a!=null?a:t}checkKeyAndValueTensor(e,t){if(e.dtype!==this.keyDType)throw new Error(`Expect key dtype ${this.keyDType}, but got ${e.dtype}`);if(t.dtype!==this.valueDType)throw new Error(`Expect value dtype ${this.valueDType}, but got ${t.dtype}`)}},FP=async(e,t,a,n)=>{switch(e.op){case"HashTable":case"HashTableV2":{let r=n.getHashTableHandleByName(e.name);if(r!=null)return[r];{let s=k("keyDType",e,t,a),i=k("valueDType",e,t,a),o=new MP(s,i);return n.addHashTable(e.name,o),[o.handle]}}case"InitializeTable":case"InitializeTableV2":case"LookupTableImport":case"LookupTableImportV2":{let r=k("tableHandle",e,t,a,n),s=k("keys",e,t,a),i=k("values",e,t,a);return[await n.getHashTableById(r.id).import(s,i)]}case"LookupTableFind":case"LookupTableFindV2":{let r=k("tableHandle",e,t,a,n),s=k("keys",e,t,a),i=k("defaultValue",e,t,a);return[await n.getHashTableById(r.id).find(s,i)]}case"LookupTableSize":case"LookupTableSizeV2":{let r=k("tableHandle",e,t,a,n);return[n.getHashTableById(r.id).tensorSize()]}default:throw TypeError(`Node type ${e.op} is not implemented`)}},_P=(e,t,a,n=Qt)=>{switch(e.op){case"ResizeBilinear":{let r=k("images",e,t,a),s=k("size",e,t,a),i=k("alignCorners",e,t,a),o=k("halfPixelCenters",e,t,a);return[n.image.resizeBilinear(r,[s[0],s[1]],i,o)]}case"ResizeNearestNeighbor":{let r=k("images",e,t,a),s=k("size",e,t,a),i=k("alignCorners",e,t,a),o=k("halfPixelCenters",e,t,a);return[n.image.resizeNearestNeighbor(r,[s[0],s[1]],i,o)]}case"CropAndResize":{let r=k("image",e,t,a),s=k("boxes",e,t,a),i=k("boxInd",e,t,a),o=k("cropSize",e,t,a),l=k("method",e,t,a),u=k("extrapolationValue",e,t,a);return[n.image.cropAndResize(r,s,i,o,l,u)]}case"ImageProjectiveTransformV3":{let r=k("images",e,t,a),s=k("transforms",e,t,a),i=k("outputShape",e,t,a),o=k("fillValue",e,t,a),l=k("interpolation",e,t,a),u=k("fillMode",e,t,a);return[n.image.transform(r,s,l.toLowerCase(),u.toLowerCase(),o,i)]}default:throw TypeError(`Node type ${e.op} is not implemented`)}},$P=(e,t,a,n=Qt)=>{switch(e.op){case"Equal":return[n.equal(k("a",e,t,a),k("b",e,t,a))];case"NotEqual":return[n.notEqual(k("a",e,t,a),k("b",e,t,a))];case"Greater":return[n.greater(k("a",e,t,a),k("b",e,t,a))];case"GreaterEqual":return[n.greaterEqual(k("a",e,t,a),k("b",e,t,a))];case"Less":return[n.less(k("a",e,t,a),k("b",e,t,a))];case"LessEqual":return[n.lessEqual(k("a",e,t,a),k("b",e,t,a))];case"LogicalAnd":return[n.logicalAnd(k("a",e,t,a),k("b",e,t,a))];case"LogicalNot":return[n.logicalNot(k("a",e,t,a))];case"LogicalOr":return[n.logicalOr(k("a",e,t,a),k("b",e,t,a))];case"Select":case"SelectV2":return[n.where(k("condition",e,t,a),k("a",e,t,a),k("b",e,t,a))];case"BitwiseAnd":return[n.bitwiseAnd(k("a",e,t,a),k("b",e,t,a))];default:throw TypeError(`Node type ${e.op} is not implemented`)}},DP=(e,t,a,n=Qt)=>{switch(e.op){case"BatchMatMul":case"BatchMatMulV2":case"MatMul":return[n.matMul(k("a",e,t,a),k("b",e,t,a),k("transposeA",e,t,a),k("transposeB",e,t,a))];case"Einsum":return[n.einsum(k("equation",e,t,a),...k("tensors",e,t,a))];case"Transpose":return[n.transpose(k("x",e,t,a),k("perm",e,t,a))];case"_FusedMatMul":let[r,s]=k("fusedOps",e,t,a),i=r==="biasadd",o=s==="prelu",l=k("numArgs",e,t,a),u=k("leakyreluAlpha",e,t,a);if(i){if(o&&l!==2)throw new Error("Fused MatMul with BiasAdd and Prelu must have two extra arguments: bias and alpha.");if(!o&&l!==1)throw new Error("Fused MatMul with BiasAdd must have one extra argument: bias.")}let[p,c]=k("args",e,t,a);return[n.fused.matMul({a:k("a",e,t,a),b:k("b",e,t,a),transposeA:k("transposeA",e,t,a),transposeB:k("transposeB",e,t,a),bias:p,activation:s,preluActivationWeights:c,leakyreluAlpha:u})];case"MatrixBandPart":return[n.linalg.bandPart(k("a",e,t,a),k("numLower",e,t,a),k("numUpper",e,t,a))];default:throw TypeError(`Node type ${e.op} is not implemented`)}},PP=(e,t,a,n=Qt)=>{switch(e.op){case"EuclideanNorm":return[n.euclideanNorm(k("x",e,t,a),k("axis",e,t,a),k("keepDims",e,t,a))];case"FusedBatchNorm":case"FusedBatchNormV2":return[n.batchNorm(k("x",e,t,a),k("mean",e,t,a),k("variance",e,t,a),k("offset",e,t,a),k("scale",e,t,a),k("epsilon",e,t,a))];case"FusedBatchNormV3":return[n.batchNorm(k("x",e,t,a),k("mean",e,t,a),k("variance",e,t,a),k("offset",e,t,a),k("scale",e,t,a),k("epsilon",e,t,a))];case"LRN":return[n.localResponseNormalization(k("x",e,t,a),k("radius",e,t,a),k("bias",e,t,a),k("alpha",e,t,a),k("beta",e,t,a))];case"Softmax":return[n.softmax(k("x",e,t,a))];case"LogSoftmax":return[n.logSoftmax(k("x",e,t,a))];default:throw TypeError(`Node type ${e.op} is not implemented`)}},OP=(e,t,a,n=Qt)=>{switch(e.op){case"RaggedGather":{let{outputNestedSplits:r,outputDenseValues:s}=n.raggedGather(k("paramsNestedSplits",e,t,a),k("paramsDenseValues",e,t,a),k("indices",e,t,a),k("outputRaggedRank",e,t,a));return r.concat(s)}case"RaggedRange":{let{rtNestedSplits:r,rtDenseValues:s}=n.raggedRange(k("starts",e,t,a),k("limits",e,t,a),k("splits",e,t,a));return[r,s]}case"RaggedTensorToTensor":return[n.raggedTensorToTensor(k("shape",e,t,a),k("values",e,t,a),k("defaultValue",e,t,a),k("rowPartitionTensors",e,t,a),k("rowPartitionTypes",e,t,a))];default:throw TypeError(`Node type ${e.op} is not implemented`)}},zP=(e,t,a,n=Qt)=>{switch(e.op){case"Max":{let o=k("axis",e,t,a),l=k("keepDims",e,t,a);return[n.max(k("x",e,t,a),o,l)]}case"Mean":{let o=k("axis",e,t,a),l=k("keepDims",e,t,a);return[n.mean(k("x",e,t,a),o,l)]}case"Min":{let o=k("axis",e,t,a),l=k("keepDims",e,t,a);return[n.min(k("x",e,t,a),o,l)]}case"Sum":{let o=k("axis",e,t,a),l=k("keepDims",e,t,a);return[n.sum(k("x",e,t,a),o,l)]}case"All":{let o=k("axis",e,t,a),l=k("keepDims",e,t,a);return[n.all(k("x",e,t,a),o,l)]}case"Any":{let o=k("axis",e,t,a),l=k("keepDims",e,t,a);return[n.any(k("x",e,t,a),o,l)]}case"ArgMax":{let o=k("axis",e,t,a);return[n.argMax(k("x",e,t,a),o)]}case"ArgMin":{let o=k("axis",e,t,a);return[n.argMin(k("x",e,t,a),o)]}case"Prod":{let o=k("axis",e,t,a),l=k("keepDims",e,t,a);return[n.prod(k("x",e,t,a),o,l)]}case"Cumprod":{let o=k("axis",e,t,a),l=k("exclusive",e,t,a),u=k("reverse",e,t,a);return[n.cumprod(k("x",e,t,a),o,l,u)]}case"Cumsum":{let o=k("axis",e,t,a),l=k("exclusive",e,t,a),u=k("reverse",e,t,a);return[n.cumsum(k("x",e,t,a),o,l,u)]}case"Bincount":let r=k("x",e,t,a),s=k("weights",e,t,a),i=k("size",e,t,a);return[n.bincount(r,s,i)];case"DenseBincount":{let o=k("x",e,t,a),l=k("weights",e,t,a),u=k("size",e,t,a),p=k("binaryOutput",e,t,a);return[n.denseBincount(o,l,u,p)]}default:throw TypeError(`Node type ${e.op} is not implemented`)}},LP=(e,t,a,n=Qt)=>{switch(e.op){case"ConcatV2":case"Concat":{let r=k("n",e,t,a),s=k("axis",e,t,a),i=k("tensors",e,t,a);return i=i.slice(0,r),[n.concat(i,s)]}case"Gather":{let r=k("x",e,t,a),s=k("indices",e,t,a);return[n.gather(r,n.cast(s,"int32"),0)]}case"GatherV2":{let r=k("axis",e,t,a),s=k("batchDims",e,t,a),i=k("x",e,t,a),o=k("indices",e,t,a);return[n.gather(i,n.cast(o,"int32"),r,s)]}case"Reverse":{let r=k("dims",e,t,a),s=[];for(let o=0;o<r.length;o++)r[o]&&s.push(o);let i=k("x",e,t,a);return[n.reverse(i,s)]}case"ReverseV2":{let r=k("axis",e,t,a),s=k("x",e,t,a);return[n.reverse(s,r)]}case"Slice":{let r=k("begin",e,t,a),s=k("size",e,t,a);return[n.slice(k("x",e,t,a),r,s)]}case"StridedSlice":{let r=k("begin",e,t,a),s=k("end",e,t,a),i=k("strides",e,t,a),o=k("beginMask",e,t,a),l=k("endMask",e,t,a),u=k("ellipsisMask",e,t,a),p=k("newAxisMask",e,t,a),c=k("shrinkAxisMask",e,t,a),d=k("x",e,t,a);return[n.stridedSlice(d,r,s,i,o,l,u,p,c)]}case"Pack":return Pe(()=>{let r=k("axis",e,t,a),s=k("tensors",e,t,a),i=s[0].shape,o=n.squeeze(s[0]).shape,l=s.map(u=>{let p=v.arraysEqual(u.shape,i);if(!p&&!v.arraysEqual(n.squeeze(u).shape,o))throw new Error("the input tensors shape does not match");return p?u:n.reshape(u,i)});return[n.stack(l,r)]});case"Unpack":{let r=k("axis",e,t,a),s=k("tensor",e,t,a);return n.unstack(s,r)}case"Tile":{let r=k("reps",e,t,a);return[n.tile(k("x",e,t,a),r)]}case"Split":case"SplitV":{let r=k("axis",e,t,a),s=k("numOrSizeSplits",e,t,a),i=k("x",e,t,a);return n.split(i,s,r)}case"ScatterNd":{let r=k("indices",e,t,a),s=k("values",e,t,a),i=k("shape",e,t,a);return[n.scatterND(r,s,i)]}case"GatherNd":{let r=k("x",e,t,a),s=k("indices",e,t,a);return[n.gatherND(r,s)]}case"SparseToDense":{let r=k("sparseIndices",e,t,a),s=k("outputShape",e,t,a),i=k("sparseValues",e,t,a),o=k("defaultValue",e,t,a);return[n.sparseToDense(r,i,s,i.dtype===o.dtype?o:n.cast(o,i.dtype))]}case"TensorScatterUpdate":{let r=k("indices",e,t,a),s=k("values",e,t,a),i=k("tensor",e,t,a);return[n.tensorScatterUpdate(i,r,s)]}default:throw TypeError(`Node type ${e.op} is not implemented`)}},WP=(e,t,a,n=Qt)=>{switch(e.op){case"SparseFillEmptyRows":{let{outputIndices:r,outputValues:s,emptyRowIndicator:i,reverseIndexMap:o}=n.sparse.sparseFillEmptyRows(k("indices",e,t,a),k("values",e,t,a),k("denseShape",e,t,a),k("defaultValue",e,t,a));return[r,s,i,o]}case"SparseReshape":{let{outputIndices:r,outputShape:s}=n.sparse.sparseReshape(k("inputIndices",e,t,a),k("inputShape",e,t,a),k("newShape",e,t,a));return[r,s]}case"SparseSegmentMean":return[n.sparse.sparseSegmentMean(k("data",e,t,a),k("indices",e,t,a),k("segmentIds",e,t,a))];case"SparseSegmentSum":return[n.sparse.sparseSegmentSum(k("data",e,t,a),k("indices",e,t,a),k("segmentIds",e,t,a))];default:throw TypeError(`Node type ${e.op} is not implemented`)}},BP=(e,t,a,n=Qt)=>{switch(e.op){case"FFT":return[n.fft(k("x",e,t,a))];case"IFFT":return[n.ifft(k("x",e,t,a))];case"RFFT":return[n.rfft(k("x",e,t,a))];case"IRFFT":return[n.irfft(k("x",e,t,a))];default:throw TypeError(`Node type ${e.op} is not implemented`)}},VP=(e,t,a,n=Qt)=>{switch(e.op){case"StaticRegexReplace":return[n.string.staticRegexReplace(k("input",e,t,a),k("pattern",e,t,a),k("rewrite",e,t,a),k("replaceGlobal",e,t,a))];case"StringNGrams":{let{nGrams:r,nGramsSplits:s}=n.string.stringNGrams(k("data",e,t,a),k("dataSplits",e,t,a),k("separator",e,t,a),k("nGramWidths",e,t,a),k("leftPad",e,t,a),k("rightPad",e,t,a),k("padWidth",e,t,a),k("preserveShortSequences",e,t,a));return[r,s]}case"StringSplit":{let{indices:r,values:s,shape:i}=n.string.stringSplit(k("input",e,t,a),k("delimiter",e,t,a),k("skipEmpty",e,t,a));return[r,s,i]}case"StringToHashBucketFast":return[n.string.stringToHashBucketFast(k("input",e,t,a),k("numBuckets",e,t,a))];default:throw TypeError(`Node type ${e.op} is not implemented`)}},UP=(e,t,a,n=Qt)=>{switch(e.op){case"Cast":return[n.cast(k("x",e,t,a),k("dtype",e,t,a))];case"ExpandDims":{let r=k("axis",e,t,a);return[n.expandDims(k("x",e,t,a),r)]}case"Squeeze":{let r=k("axis",e,t,a);return[n.squeeze(k("x",e,t,a),r)]}case"Reshape":return[n.reshape(k("x",e,t,a),k("shape",e,t,a))];case"EnsureShape":return[n.ensureShape(k("x",e,t,a),k("shape",e,t,a))];case"MirrorPad":return[n.mirrorPad(k("x",e,t,a),k("padding",e,t,a),k("mode",e,t,a))];case"PadV2":case"Pad":return[n.pad(k("x",e,t,a),k("padding",e,t,a),k("constantValue",e,t,a))];case"SpaceToBatchND":{let r=k("blockShape",e,t,a),s=k("paddings",e,t,a);return[n.spaceToBatchND(k("x",e,t,a),r,s)]}case"BatchToSpaceND":{let r=k("blockShape",e,t,a),s=k("crops",e,t,a);return[n.batchToSpaceND(k("x",e,t,a),r,s)]}case"DepthToSpace":{let r=k("blockSize",e,t,a),s=k("dataFormat",e,t,a).toUpperCase();return[n.depthToSpace(k("x",e,t,a),r,s)]}case"BroadcastTo":return[n.broadcastTo(k("x",e,t,a),k("shape",e,t,a))];case"BroadcastArgs":return[n.broadcastArgs(k("s0",e,t,a),k("s1",e,t,a))];default:throw TypeError(`Node type ${e.op} is not implemented`)}};function X5(e,t,a,n,r=Pe){let s=((i,o,l)=>{switch(i.category){case"arithmetic":return r(()=>xP(i,o,l));case"basic_math":return r(()=>AP(i,o,l));case"control":return SP(i,o,l);case"convolution":return r(()=>CP(i,o,l));case"creation":return r(()=>TP(i,o,l));case"dynamic":return NP(i,o,l);case"evaluation":return r(()=>RP(i,o,l));case"image":return r(()=>_P(i,o,l));case"graph":return r(()=>EP(i,o,l));case"logical":return r(()=>$P(i,o,l));case"matrices":return r(()=>DP(i,o,l));case"normalization":return r(()=>PP(i,o,l));case"ragged":return r(()=>OP(i,o,l));case"reduction":return r(()=>zP(i,o,l));case"slice_join":return r(()=>LP(i,o,l));case"sparse":return r(()=>WP(i,o,l));case"spectral":return r(()=>BP(i,o,l));case"string":return r(()=>VP(i,o,l));case"transformation":return r(()=>UP(i,o,l));case"hash_table":return FP(i,o,l,n);case"custom":let u=b7(i.op);if(u&&u.customExecutor)return u.customExecutor(new yP(i,o,l));throw TypeError(`Custom op ${i.op} is not registered.`);default:throw TypeError(`Unknown op '${i.op}'. File an issue at https://github.com/tensorflow/tfjs/issues so we can add it, or register a custom execution with tf.registerOp()`)}})(e,t,a);return v.isPromise(s)?s.then(i=>[].concat(i)):[].concat(s)}var K5=class{constructor(e={},t={},a={},n={},r){this.weightMap=e,this.tensorArrayMap=t,this.tensorListMap=a,this.functionMap=n,this.parseNodeNameCache=r,this.rootContext={id:0,frameName:"",iterationId:0},this.contexts=[this.rootContext],this.lastId=0,this.generateCurrentContextIds()}newFrame(e,t){return{id:e,frameName:t,iterationId:0}}set currentContext(e){this.contexts!==e&&(this.contexts=e,this.generateCurrentContextIds())}get currentContext(){return this.contexts}get currentContextId(){return this._currentContextIds[0]}get currentContextIds(){return this._currentContextIds}generateCurrentContextIds(){let e=[];for(let t=0;t<this.contexts.length-1;t++){let a=this.contexts.slice(0,this.contexts.length-t);e.push(this.contextIdforContexts(a))}e.push(""),this._currentContextIds=e}contextIdforContexts(e){return e?e.map(t=>t.id===0&&t.iterationId===0?"":`${t.frameName}-${t.iterationId}`).join("/"):""}enterFrame(e){this.contexts&&(this.lastId++,this.contexts=this.contexts.slice(),this.contexts.push(this.newFrame(this.lastId,e)),this._currentContextIds.unshift(this.contextIdforContexts(this.contexts)))}exitFrame(){if(this.contexts&&this.contexts.length>1)this.contexts=this.contexts.slice(),this.contexts.splice(-1),this.currentContextIds.shift();else throw new Error("Cannot exit frame, the context is empty")}nextIteration(){if(this.contexts&&this.contexts.length>0){this.contexts=this.contexts.slice(),this.lastId++;let e=Object.assign({},this.contexts[this.contexts.length-1]);e.iterationId+=1,e.id=this.lastId,this.contexts.splice(-1,1,e),this._currentContextIds.splice(0,1,this.contextIdforContexts(this.contexts))}else throw new Error("Cannot increase frame iteration, the context is empty")}getWeight(e){return this.weightMap[e]}addTensorArray(e){this.tensorArrayMap[e.id]=e}getTensorArray(e){return this.tensorArrayMap[e]}addTensorList(e){this.tensorListMap[e.id]=e}getTensorList(e){return this.tensorListMap[e]}dispose(e){for(let t in this.tensorArrayMap)this.tensorArrayMap[t].clearAndClose(e);for(let t in this.tensorListMap)this.tensorListMap[t].clearAndClose(e)}};function Y5(e,t,a,n){let r=new Set,s=[],i=null,o=null,l=new Set,u=new Set(Object.keys(e).map(d=>Ya(d)[0]));n=n||[];let p=new Set(n.map(d=>Ya(d.name)[0])),c=[...t];for(;c.length>0;){let d=c.pop();if((zs(d)||ZP(d)||JP(d))&&i==null&&(i=d,o=i.children.map(h=>h.name).filter(h=>r.has(h))),r.add(d.name),a[d.name]==null&&!u.has(d.name)&&!p.has(d.name)){if(d.inputs.length===0){s.push(d.name);continue}d.inputs.forEach(h=>{l.has(h.name)||(l.add(h.name),c.push(h))})}}return{inputs:e,outputs:t,usedNodes:r,missingInputs:s,dynamicNode:i,syncInputs:o}}function GP(e,t){let{usedNodes:a,inputs:n}=t,r=Object.keys(n).map(g=>Ya(g)[0]).map(g=>e.nodes[g]),s=e.initNodes||[],i=g=>a.has(typeof g=="string"?g:g.name);function o(g){return[...new Map(g.map(y=>[y.name,y])).values()]}let l=o([...r,...e.weights,...s]).filter(i),u=o([...l,...Object.values(e.nodes)]).filter(i),p=new Map(u.map(g=>[g.name,g])),c={};for(let g of u){c[g.name]=c[g.name]||0;for(let y of g.children)i(y)||(c[y.name]=Number.POSITIVE_INFINITY),c[y.name]=(c[y.name]||0)+1}let d=Object.entries(c).filter(([,g])=>g===0).map(([g])=>g),h=[...d];for(;d.length>0;){let g=d.pop(),y=p.get(g);for(let x of y.children.filter(i))--c[x.name]===0&&(h.push(x.name),d.push(x.name))}let f=h.map(g=>p.get(g)),m=HP(f,l);return jP(m,l),m}function HP(e,t){let a=new Map(e.map(s=>[s.name,s])),n=t.map(s=>s.name),r=new Set(n);for(;n.length>0;){let s=n.pop(),i=a.get(s);for(let o of i.children)!a.has(o.name)||r.has(o.name)||(r.add(o.name),n.push(o.name))}return e.filter(s=>r.has(s.name))}var Wc=class extends Error{constructor(e){super(`NodesExecutionOrderError: ${e}`)}};function jP(e,t){let a=new Map(e.map((o,l)=>[o.name,l])),n=new Set(t.map(o=>o.name)),r=o=>n.has(typeof o=="string"?o:o.name),s=new Set(e.map(o=>o.name)),i=o=>s.has(typeof o=="string"?o:o.name);for(let o of e){for(let l of o.children.filter(i)){if(!a.has(l.name))throw new Wc(`Child ${l.name} of node ${o.name} is unreachable.`);if(a.get(o.name)>a.get(l.name))throw new Wc(`Node ${o.name} is scheduled to run after its child ${l.name}.`)}if(!r(o))for(let l of o.inputs){if(!a.has(l.name))throw new Wc(`Input ${l.name} of node ${o.name} is unreachable.`);if(a.get(l.name)>a.get(o.name))throw new Wc(`Node ${o.name} is scheduled to run before its input ${l.name}.`)}}}function qP(e){let t=new Map(e.map((o,l)=>[o.name,l])),a=Number.MAX_SAFE_INTEGER,n=e.map((o,l)=>zs(o)?a:l),r=o=>{let l=n[t.get(o.name)];return l==null?-1:l},s=e.map((o,l)=>o.children.map(r).reduce((u,p)=>Math.max(u,p),n[l])),i=new Map;for(let o=0;o<e.length;++o){let l=s[o];if(l===a)continue;let u=e[o],p=e[l];i.has(p.name)||i.set(p.name,[]),i.get(p.name).push(u)}return i}var XP=new Set(["Switch","Merge","Enter","Exit","NextIteration","StatelessIf","StatelessWhile","if","While"]),KP=new Set(["NonMaxSuppressionV2","NonMaxSuppressionV3","NonMaxSuppressionV5","Where"]),YP=new Set(["HashTable","HashTableV2","LookupTableImport","LookupTableImportV2","LookupTableFind","LookupTableFindV2","LookupTableSize","LookupTableSizeV2"]);function zs(e){return XP.has(e.op)}function ZP(e){return KP.has(e.op)}function JP(e){return YP.has(e.op)}var y1=class{get weightIds(){return this.parent?this.parent.weightIds:this._weightIds}get functionExecutorMap(){return this.parent?this.parent.functionExecutorMap:this._functionExecutorMap}get weightMap(){return this.parent?this.parent.weightMap:this._weightMap}set weightMap(e){let t=Object.keys(e).map(a=>e[a].map(n=>n.id));this._weightIds=[].concat(...t),this._weightMap=e}set resourceManager(e){this._resourceManager=e}get inputs(){return this._inputs.map(e=>({name:e.name,shape:e.attrParams.shape?e.attrParams.shape.value:void 0,dtype:e.attrParams.dtype?e.attrParams.dtype.value:void 0}))}get outputs(){return this._outputs.map(e=>({name:e.name,shape:e.attrParams.shape?e.attrParams.shape.value:void 0,dtype:e.attrParams.dtype?e.attrParams.dtype.value:void 0}))}get inputNodes(){return this._inputs.map(e=>e.signatureKey||e.name)}get outputNodes(){return this._outputs.map(e=>{let t=e.signatureKey||e.name;return e.defaultOutput?`${t}:${e.defaultOutput}`:t})}get functions(){return Object.keys(this._functions).reduce((e,t)=>(e[t]=this._functions[t].signature,e),{})}constructor(e,t){this.graph=e,this.parent=t,this.compiledMap=new Map,this.parseNodeNameCache=new Map,this._weightMap={},this.SEPARATOR=",",this._functions={},this._functionExecutorMap={},this.keepIntermediateTensors=!1,this._outputs=e.outputs,this._inputs=e.inputs,this._initNodes=e.initNodes,this._signature=e.signature,this._functions=e.functions,e.functions!=null&&Object.keys(e.functions).forEach(a=>{this._functionExecutorMap[a]=new y1(e.functions[a],this)})}getCompilationKey(e,t){let a=e.map(r=>r.name).sort(),n=t.map(r=>r.name).sort();return a.join(this.SEPARATOR)+"--"+n.join(this.SEPARATOR)}compile(e,t){let a=Y5(e,t,this.weightMap,this._initNodes),{missingInputs:n,dynamicNode:r,syncInputs:s}=a;if(r!=null)throw new Error(`This execution contains the node '${r.name}', which has the dynamic op '${r.op}'. Please use model.executeAsync() instead. Alternatively, to avoid the dynamic ops, specify the inputs [${s}]`);if(n.length>0){let l=t.map(p=>p.name),u=Object.keys(e);throw new Error(`Cannot compute the outputs [${l}] from the provided inputs [${u}]. Missing the following inputs: [${n}]`)}let i=GP(this.graph,a),o=qP(i);return{orderedNodes:i,nodeLiveUntilMap:o}}cloneAndKeepTensor(e){if(e==null)return null;let t=e.clone();return zn(t),t}cloneTensorList(e){return e?e.map(t=>this.cloneAndKeepTensor(t)):null}cloneTensorMap(e){return Object.fromEntries(Object.entries(e).map(([t,a])=>[t,this.cloneTensorList(a)]))}execute(e,t){this.disposeIntermediateTensors(),e=this.mapInputs(e);let a=Object.keys(e).sort();this.checkInputs(e),this.checkInputShapeAndType(e),t=this.mapOutputs(t),this.checkOutputs(t);let n=a.map(c=>this.graph.nodes[Ya(c)[0]]),r=t.map(c=>Ya(c)[0]),s=new Set(r),i=r.map(c=>this.graph.nodes[c]);i.length===0&&(i=this._outputs);let o=this.getCompilationKey(n,i),l=this.compiledMap.get(o);l==null&&(l=this.compile(e,i),this.compiledMap.set(o,l));try{this.keepIntermediateTensors=W().getBool("KEEP_INTERMEDIATE_TENSORS")}catch(c){this.keepIntermediateTensors=!1,console.warn(c.message)}let u={},p={};return Pe(()=>{let c=new K5(this.weightMap,u,p,this.functionExecutorMap,this.parseNodeNameCache),d=Object.assign({},this.weightMap);this.keepIntermediateTensors&&(this.clonedTensorsMap=this.cloneTensorMap(this.weightMap)),Object.keys(e).forEach(g=>{let[y,x]=Ya(g,c),A=[];A[x]=e[g],d[y]=A,this.keepIntermediateTensors&&(this.clonedTensorsMap[y]=this.cloneTensorList(A))});let h=this.getFrozenTensorIds(d),{orderedNodes:f,nodeLiveUntilMap:m}=l;for(let g of f){if(d[g.name])continue;let y=X5(g,d,c,this._resourceManager);if(v.isPromise(y))throw new Error(`The execution of the op '${g.op}' returned a promise. Please use model.executeAsync() instead.`);d[g.name]=y,this.keepIntermediateTensors&&(this.clonedTensorsMap[g.name]=this.cloneTensorList(y)),this.checkTensorForDisposalWithNodeLiveUntilInfo(g,d,c,h,s,m.get(g.name))}return this.parent==null&&c.dispose(h),t.map(g=>ua(g,d,c))})}getFrozenTensorIds(e){let t=[].concat.apply([],Object.keys(e).map(a=>e[a]).map(a=>a.map(n=>n.id)));return new Set(t)}checkTensorForDisposal(e,t,a,n,r,s,i){if(!(zs(t)||s.has(e))){for(let o of a[e])o!=null&&(i[o.id]=(i[o.id]||0)+t.children.length);for(let o of t.inputs){if(zs(o))continue;let l=U5(o.name,a,n);if(l!=null)for(let u of l){if(!u||u.kept||r.has(u.id))continue;let p=i[u.id];p===1?(u.dispose(),delete i[u.id]):p!=null&&i[u.id]--}}}}checkTensorForDisposalWithNodeLiveUntilInfo(e,t,a,n,r,s){function i(o){return zs(o)||r.has(o.name)}if(!(zs(e)||s==null))for(let o of s){if(i(o))continue;let l=U5(o.name,t,a);for(let u of l)!u||u.kept||n.has(u.id)||u.dispose()}}async executeAsync(e,t){return this._executeAsync(e,t)}disposeIntermediateTensors(){this.clonedTensorsMap&&(Object.values(this.clonedTensorsMap).forEach(e=>{for(let t of e)t&&!t.isDisposed&&t.dispose()}),this.clonedTensorsMap=null)}getIntermediateTensors(){return this.clonedTensorsMap}async _executeAsync(e,t,a=!1,n={},r={}){this.disposeIntermediateTensors(),a||(e=this.mapInputs(e),this.checkInputs(e),this.checkInputShapeAndType(e),t=this.mapOutputs(t),this.checkOutputs(t));try{this.keepIntermediateTensors=W().getBool("KEEP_INTERMEDIATE_TENSORS")}catch(c){this.keepIntermediateTensors=!1,console.warn(c.message)}let s=new K5(this.weightMap,n,r,this.functionExecutorMap,this.parseNodeNameCache);this.keepIntermediateTensors&&(this.clonedTensorsMap=this.cloneTensorMap(this.weightMap));let i=await this.executeWithControlFlow(e,s,t,a),o=t.map(c=>ua(c,i,s)),l=o.map(c=>c.id),u=Object.keys(e).map(c=>e[c].id),p=new Set([...l,...u,...this.weightIds]);return Object.values(i).forEach(c=>{c.forEach(d=>{d&&!d.isDisposed&&!p.has(d.id)&&d.dispose()})}),this.parent==null&&s.dispose(p),o}async executeFunctionAsync(e,t,a){let n=e.reduce((r,s,i)=>(r[this.inputs[i].name]=s,r),{});return this._executeAsync(n,this.outputNodes,!0,t,a)}async executeWithControlFlow(e,t,a,n){let r=Object.keys(e),s=r.map(A=>this.graph.nodes[Ya(A)[0]]),i=a.map(A=>Ya(A)[0]),o=new Set(i),l=i.map(A=>this.graph.nodes[A]);l.length===0&&(l=this._outputs);let{usedNodes:u,missingInputs:p,dynamicNode:c,syncInputs:d}=Y5(e,l,this.weightMap,this._initNodes),h=[...s,...this.graph.weights,...this._initNodes||[]].map(A=>({node:A,contexts:t.currentContext})),f=Object.assign({},this.weightMap);Object.keys(e).forEach(A=>{let[b,w]=Ya(A),C=[];C[w]=e[A],f[b]=C});let m={},g=this.getFrozenTensorIds(f),y={};for(;h.length>0;){let A=this.processStack(s,h,t,f,y,g,o,m,u);await Promise.all(A)}c==null&&!n&&console.warn("This model execution did not contain any nodes with control flow or dynamic output shapes. You can use model.execute() instead.");let x=l.filter(A=>!zs(A)&&!ua(A.name,f,t)).map(A=>A.name);if(x.length>0){let A="";throw c!=null&&(A=`Alternatively, to avoid the dynamic ops, use model.execute() and specify the inputs [${d}]`),new Error(`Cannot compute the outputs [${x}] from the provided inputs [${r}]. Consider providing the following inputs: [${p}]. ${A}`)}return f}processStack(e,t,a,n,r,s,i,o,l){let u=[];for(;t.length>0;){let p=t.pop();a.currentContext=p.contexts;let c="";if(p.node.op==="Enter"&&k("isConstant",p.node,n,a)&&([c]=Ar(p.node.name,a)),n[p.node.name]==null){let d=X5(p.node,n,a,this._resourceManager);c||([c]=Ar(p.node.name,a));let h=a.currentContext;v.isPromise(d)?u.push(d.then(f=>(n[c]=f,this.keepIntermediateTensors&&(this.clonedTensorsMap[c]=this.cloneTensorList(f)),a.currentContext=h,this.checkTensorForDisposal(c,p.node,n,a,s,i,o),this.processChildNodes(p.node,t,a,n,r,l),f))):(n[c]=d,this.keepIntermediateTensors&&(this.clonedTensorsMap[c]=this.cloneTensorList(d)),this.checkTensorForDisposal(c,p.node,n,a,s,i,o),this.processChildNodes(p.node,t,a,n,r,l))}else this.processChildNodes(p.node,t,a,n,r,l)}return u}processChildNodes(e,t,a,n,r,s){e.children.forEach(i=>{let[o]=Ar(i.name,a);r[o]||!s.has(i.name)||(i.op==="Merge"?i.inputNames.some(l=>!!ua(l,n,a))&&(r[o]=!0,t.push({contexts:a.currentContext,node:i})):i.inputNames.every(l=>!!ua(l,n,a))&&(r[o]=!0,t.push({contexts:a.currentContext,node:i})))})}dispose(){Object.keys(this.weightMap).forEach(e=>this.weightMap[e].forEach(t=>t.dispose()))}checkInputShapeAndType(e){Object.keys(e).forEach(t=>{let a=e[t],[n]=Ya(t),r=this.graph.nodes[n];if(r.attrParams.shape&&r.attrParams.shape.value){let s=r.attrParams.shape.value,i=s.length===a.shape.length&&a.shape.every((o,l)=>s[l]===-1||s[l]===o);v.assert(i,()=>`The shape of dict['${r.name}'] provided in model.execute(dict) must be [${s}], but was [${a.shape}]`)}r.attrParams.dtype&&r.attrParams.dtype.value&&v.assert(a.dtype===r.attrParams.dtype.value,()=>`The dtype of dict['${r.name}'] provided in model.execute(dict) must be ${r.attrParams.dtype.value}, but was ${a.dtype}`)})}mapInputs(e){var t,a;let n={};for(let r in e){let s=(a=(t=this._signature)===null||t===void 0?void 0:t.inputs)===null||a===void 0?void 0:a[r];s!=null?n[s.name]=e[r]:n[r]=e[r]}return n}checkInputs(e){let t=Object.keys(e).filter(a=>{let[n]=Ya(a);return this.graph.nodes[n]==null});if(t.length>0)throw new Error(`The dict provided in model.execute(dict) has keys: [${t}] that are not part of graph`)}mapOutputs(e){return e.map(t=>{var a,n;let r=(n=(a=this._signature)===null||a===void 0?void 0:a.outputs)===null||n===void 0?void 0:n[t];return r!=null?r.name:t},{})}checkOutputs(e){e.forEach(t=>{let[a]=Ya(t);if(!this.graph.nodes[a])throw new Error(`The output '${t}' is not found in the graph`)})}},QP=class{constructor(e={},t={}){this.hashTableNameToHandle=e,this.hashTableMap=t}addHashTable(e,t){this.hashTableNameToHandle[e]=t.handle,this.hashTableMap[t.id]=t}getHashTableHandleByName(e){return this.hashTableNameToHandle[e]}getHashTableById(e){return this.hashTableMap[e]}dispose(){for(let e in this.hashTableMap)this.hashTableMap[e].clearAndClose(),delete this.hashTableMap[e];for(let e in this.hashTableNameToHandle)this.hashTableNameToHandle[e].dispose(),delete this.hashTableNameToHandle[e]}},eO="?tfjs-format=file",tO="model.json",Lp=class{get modelVersion(){return this.version}get inputNodes(){return this.executor.inputNodes}get outputNodes(){return this.executor.outputNodes}get inputs(){return this.executor.inputs}get outputs(){return this.executor.outputs}get weights(){return this.executor.weightMap}get metadata(){return this.artifacts.userDefinedMetadata}get modelSignature(){return this.signature}get modelStructuredOutputKeys(){return this.structuredOutputKeys}constructor(e,t={},a=Kn){this.modelUrl=e,this.loadOptions=t,this.version="n/a",this.io=a,t==null&&(this.loadOptions={}),this.resourceManager=new QP}findIOHandler(){let e=this.modelUrl;if(e.load!=null)this.handler=e;else if(this.loadOptions.requestInit!=null)this.handler=this.io.browserHTTPRequest(e,this.loadOptions);else{let t=this.io.getLoadHandlers(e,this.loadOptions);if(t.length===0)t.push(this.io.browserHTTPRequest(e,this.loadOptions));else if(t.length>1)throw new Error(`Found more than one (${t.length}) load handlers for URL '${[e]}'`);this.handler=t[0]}}load(){if(this.findIOHandler(),this.handler.load==null)throw new Error("Cannot proceed with model loading because the IOHandler provided does not have the `load` method implemented.");let e=this.handler.load();return v.isPromise(e)?e.then(t=>this.loadSync(t)):this.loadSync(e)}loadSync(e){this.artifacts=e;let t=this.artifacts.modelTopology,a=this.artifacts.signature;if(this.artifacts.userDefinedMetadata!=null){let r=this.artifacts.userDefinedMetadata;r.signature!=null&&(a=r.signature),r.structuredOutputKeys!=null&&(this.structuredOutputKeys=r.structuredOutputKeys)}this.signature=a,this.version=`${t.versions.producer}.${t.versions.minConsumer}`;let n=this.io.decodeWeights(this.artifacts.weightData,this.artifacts.weightSpecs);if(this.executor=new y1(G5.Instance.transformGraph(t,this.signature)),this.executor.weightMap=this.convertTensorMapToTensorsMap(n),this.executor.resourceManager=this.resourceManager,e.modelInitializer!=null&&e.modelInitializer.node!=null){let r=G5.Instance.transformGraph(e.modelInitializer);this.initializer=new y1(r),this.initializer.weightMap=this.executor.weightMap,this.initializer.resourceManager=this.resourceManager,this.initializerSignature=e.initializerSignature}return!0}async save(e,t){if(typeof e=="string"){let a=this.io.getSaveHandlers(e);if(a.length===0)throw new Error(`Cannot find any save handlers for URL '${e}'`);if(a.length>1)throw new Error(`Found more than one (${a.length}) save handlers for URL '${e}'`);e=a[0]}if(e.save==null)throw new Error("GraphModel.save() cannot proceed because the IOHandler provided does not have the `save` attribute defined.");return e.save(this.artifacts)}addStructuredOutputNames(e){if(this.structuredOutputKeys){let t=e instanceof gt?[e]:e,a={};return t.forEach((n,r)=>a[this.structuredOutputKeys[r]]=n),a}return e}predict(e,t){let a=this.execute(e,this.outputNodes);return this.addStructuredOutputNames(a)}async predictAsync(e,t){let a=await this.executeAsync(e,this.outputNodes);return this.addStructuredOutputNames(a)}normalizeInputs(e){var t;if(!(e instanceof gt)&&!Array.isArray(e)){let r=(t=this.signature)===null||t===void 0?void 0:t.inputs;if(r!=null)for(let s in r){let i=r[s];i.resourceId!=null&&(e[s]=this.resourceIdToCapturedInput[i.resourceId])}return e}e=Array.isArray(e)?e:[e];let a=Object.keys(this.resourceIdToCapturedInput).length;if(e.length+a!==this.inputNodes.length)throw new Error(`Input tensor count mismatch, the graph model has ${this.inputNodes.length-a} non-resource placeholders, while there are ${e.length} input tensors provided.`);let n=0;return this.inputNodes.reduce((r,s)=>{var i,o,l;let u=(l=(o=(i=this.signature)===null||i===void 0?void 0:i.inputs)===null||o===void 0?void 0:o[s])===null||l===void 0?void 0:l.resourceId;return u!=null?r[s]=this.resourceIdToCapturedInput[u]:r[s]=e[n++],r},{})}normalizeOutputs(e){return e=e||this.outputNodes,Array.isArray(e)?e:[e]}executeInitializerGraph(){return this.initializer==null?[]:this.initializerSignature==null?this.initializer.execute({},[]):this.initializer.execute({},Object.keys(this.initializerSignature.outputs))}async executeInitializerGraphAsync(){return this.initializer==null?[]:this.initializerSignature==null?this.initializer.executeAsync({},[]):this.initializer.executeAsync({},Object.keys(this.initializerSignature.outputs))}setResourceIdToCapturedInput(e){if(this.resourceIdToCapturedInput={},this.initializerSignature){let t=this.initializerSignature.outputs,a=Object.keys(t);for(let n=0;n<a.length;n++){let r=a[n],s=t[r];this.resourceIdToCapturedInput[s.resourceId]=e[n]}}}execute(e,t){this.resourceIdToCapturedInput==null&&this.setResourceIdToCapturedInput(this.executeInitializerGraph()),e=this.normalizeInputs(e),t=this.normalizeOutputs(t);let a=this.executor.execute(e,t);return a.length>1?a:a[0]}async executeAsync(e,t){this.resourceIdToCapturedInput==null&&this.setResourceIdToCapturedInput(await this.executeInitializerGraphAsync()),e=this.normalizeInputs(e),t=this.normalizeOutputs(t);let a=await this.executor.executeAsync(e,t);return a.length>1?a:a[0]}getIntermediateTensors(){return this.executor.getIntermediateTensors()}disposeIntermediateTensors(){this.executor.disposeIntermediateTensors()}convertTensorMapToTensorsMap(e){return Object.keys(e).reduce((t,a)=>(t[a]=[e[a]],t),{})}dispose(){this.executor.dispose(),this.initializer&&(this.initializer.dispose(),this.resourceIdToCapturedInput&&J(this.resourceIdToCapturedInput)),this.resourceManager.dispose()}};async function Y3(e,t={},a=Kn){if(e==null)throw new Error("modelUrl in loadGraphModel() cannot be null. Please provide a url or an IOHandler that loads the model");t==null&&(t={}),t.fromTFHub&&typeof e=="string"&&(e=nO(e));let n=new Lp(e,t,a);return await n.load(),n}function aO(e){if(e==null)throw new Error("modelUrl in loadGraphModelSync() cannot be null. Please provide model artifacts or an IOHandler that loads the model");let t;if(e instanceof Array){let[n,r]=e;if(!n)throw new Error("modelJSON must be the first element of the array");if(!r||!(r instanceof ArrayBuffer))throw new Error("An ArrayBuffer of weights must be the second element of the array");if(!("modelTopology"in n))throw new Error("Model JSON is missing 'modelTopology'");if(!("weightsManifest"in n))throw new Error("Model JSON is missing 'weightsManifest'");let s=Kn.getWeightSpecs(n.weightsManifest),i=Kn.getModelArtifactsForJSONSync(n,s,r);t=Kn.fromMemorySync(i)}else if("load"in e)t=e;else if("modelTopology"in e&&"weightSpecs"in e&&"weightData"in e)t=Kn.fromMemorySync(e);else throw new Error("Unknown model format");let a=new Lp(t);return a.load(),a}function nO(e){return e.endsWith("/")||(e=e+"/"),`${e}${tO}${eO}`}var rO="4.5.0";function Ie(e,t){Array.isArray(e)||(e=[e]),e.forEach(a=>{a!=null&&v.assert(a.dtype!=="complex64",()=>`${t} does not support complex64 tensors in the CPU backend.`)})}var sO=Nn.whereImpl,Xh=class extends Kl{nextDataId(){return Xh.nextDataId++}constructor(){super(),this.blockSize=48,this.firstUse=!0,this.data=new Zd(this,It())}write(e,t,a){this.firstUse&&(this.firstUse=!1,W().get("IS_NODE")&&S.warn(`
|
|
============================
|
|
Hi, looks like you are running TensorFlow.js in Node.js. To speed things up dramatically, install our node backend, visit https://github.com/tensorflow/tfjs-node for more details.
|
|
============================`));let n={id:this.nextDataId()};return this.data.set(n,{values:e,dtype:a,refCount:1}),n}makeTensorInfo(e,t,a){let n;if(t==="string"&&a!=null&&a.length>0&&v.isString(a[0])){let r=a.map(s=>v.encodeString(s));n=this.write(r,e,t)}else n=this.write(a,e,t);return{dataId:n,shape:e,dtype:t}}refCount(e){return this.data.has(e)?this.data.get(e).refCount:0}incRef(e){let t=this.data.get(e);t.refCount++}decRef(e){if(this.data.has(e)){let t=this.data.get(e);t.refCount--}}move(e,t,a,n,r){this.data.set(e,{values:t,dtype:n,refCount:r})}numDataIds(){return this.data.numDataIds()}async read(e){return this.readSync(e)}readSync(e){let{dtype:t,complexTensorInfos:a}=this.data.get(e);if(t==="complex64"){let n=this.readSync(a.real.dataId),r=this.readSync(a.imag.dataId);return S.mergeRealAndImagArrays(n,r)}return v.convertBackendValuesAndArrayBuffer(this.data.get(e).values,t)}bufferSync(e){let t=this.readSync(e.dataId);if(e.dtype==="string")try{let a=t.map(n=>v.decodeString(n));return $e(e.shape,e.dtype,a)}catch(a){throw new Error("Failed to decode encoded string bytes into utf-8")}return $e(e.shape,e.dtype,t)}makeOutput(e,t,a){return It().makeTensorFromTensorInfo(this.makeTensorInfo(t,a,e),this)}disposeData(e,t=!1){if(this.data.has(e)){if(this.data.get(e).refCount--,!t&&this.data.get(e).refCount>0)return!1;let{complexTensorInfos:a}=this.data.get(e);a!=null&&(this.disposeData(a.real.dataId,!0),this.disposeData(a.imag.dataId,!0)),this.data.delete(e)}return!0}disposeIntermediateTensorInfo(e){this.disposeData(e.dataId)}async time(e){let t=v.now();return e(),{kernelMs:v.now()-t}}memory(){return{unreliable:!0,reasons:["The reported memory is an upper bound. Due to automatic garbage collection, the true allocated memory may be less."]}}where(e){Ie([e],"where");let t=this.readSync(e.dataId);return sO(e.shape,t)}dispose(){}floatPrecision(){return 32}epsilon(){return super.epsilon()}};Xh.nextDataId=0;var Kh={};Ze(Kh,{addImpl:()=>G7,bincountImpl:()=>J3,bincountReduceImpl:()=>H7,bitwiseAndImpl:()=>j7,castImpl:()=>U7,ceilImpl:()=>q7,concatImpl:()=>Q3,equalImpl:()=>X7,expImpl:()=>Y7,expm1Impl:()=>J7,floorDivImpl:()=>e6,floorImpl:()=>Q7,gatherNdImpl:()=>t6,gatherV2Impl:()=>a6,greaterEqualImpl:()=>r6,greaterImpl:()=>n6,lessEqualImpl:()=>i6,lessImpl:()=>s6,linSpaceImpl:()=>o6,logImpl:()=>l6,maxImpl:()=>u6,maximumImpl:()=>d6,minimumImpl:()=>p6,multiplyImpl:()=>eg,negImpl:()=>c6,notEqualImpl:()=>h6,prodImpl:()=>f6,raggedGatherImpl:()=>m6,raggedRangeImpl:()=>g6,raggedTensorToTensorImpl:()=>y6,rangeImpl:()=>ag,rsqrtImpl:()=>x6,scatterImpl:()=>Bs,sigmoidImpl:()=>nz,simpleAbsImpl:()=>V7,sliceImpl:()=>dh,sparseFillEmptyRowsImpl:()=>b6,sparseReshapeImpl:()=>v6,sparseSegmentReductionImpl:()=>ng,sqrtImpl:()=>iz,squaredDifferenceImpl:()=>w6,staticRegexReplaceImpl:()=>k6,stridedSliceImpl:()=>I6,stringNGramsImpl:()=>rg,stringSplitImpl:()=>sg,stringToHashBucketFastImpl:()=>ig,subImpl:()=>S6,tileImpl:()=>C6,topKImpl:()=>N6,transposeImpl:()=>tg,uniqueImpl:()=>lg});function V7(e){let t=new Float32Array(e.length);for(let a=0;a<e.length;++a)t[a]=Math.abs(e[a]);return t}var iO=e=>{let{x:t}=e.inputs,a=e.backend;Ie(t,"abs");let n=new Float32Array(v.sizeFromShape(t.shape)),r=a.data.get(t.dataId).values;return n=V7(r),a.makeOutput(n,t.shape,t.dtype)},oO={kernelName:Zl,backendName:"cpu",kernelFunc:iO};function $t(e){return(t,a,n,r,s)=>{let i=S.assertAndGetBroadcastShape(t,a),o=i.length,l=v.computeStrides(i),u=v.sizeFromShape(i),p=v.getTypedArrayFromDType(s,u),c=t.length,d=a.length,h=v.computeStrides(t),f=v.computeStrides(a),m=S.getBroadcastDims(t,i),g=S.getBroadcastDims(a,i);if(m.length+g.length===0)for(let y=0;y<p.length;++y)p[y]=e(n[y%n.length],r[y%r.length]);else for(let y=0;y<p.length;++y){let x=v.indexToLoc(y,o,l),A=x.slice(-c);m.forEach(T=>A[T]=0);let b=v.locToIndex(A,c,h),w=x.slice(-d);g.forEach(T=>w[T]=0);let C=v.locToIndex(w,d,f);p[y]=e(n[b],r[C])}return[p,i]}}function Ja(e){let{inputs:t,backend:a}=e,{real:n,imag:r}=t,s=a.data.get(n.dataId).values,i=a.data.get(r.dataId).values,o=a.makeTensorInfo(n.shape,"complex64"),l=a.data.get(o.dataId);return l.complexTensorInfos={real:a.makeTensorInfo(n.shape,"float32",s),imag:a.makeTensorInfo(r.shape,"float32",i)},o}var lO={kernelName:tp,backendName:"cpu",kernelFunc:Ja};function uh(e,t,a="float32"){if(a==="complex64"){let r=uh(e,t,"float32"),s=uh(e,t,"float32");return Ja({inputs:{real:r,imag:s},backend:e})}let n=v.makeZerosTypedArray(v.sizeFromShape(t),a);return e.makeTensorInfo(t,a,n)}function ar(e){let{inputs:t,backend:a}=e,{x:n}=t;return a.incRef(n.dataId),{dataId:n.dataId,shape:n.shape,dtype:n.dtype}}var uO={kernelName:zi,backendName:"cpu",kernelFunc:ar};function qs(e){let{inputs:t,backend:a}=e,{input:n}=t,r=a.data.get(n.dataId).complexTensorInfos.real,s=a.data.get(r.dataId).values;return a.makeTensorInfo(r.shape,r.dtype,s)}var dO={kernelName:cp,backendName:"cpu",kernelFunc:qs};function U7(e,t,a,n){if(n==="int32"){let r=Int32Array.from(e);return[t,"int32",r]}if(n==="bool"){let r=v.toTypedArray([0],a),[s,i]=$t((o,l)=>o!==l?1:0)(t,[],e,r,"bool");return[i,"bool",s]}throw new Error(`Error in Cast: failed to cast ${a} to ${n}`)}function ts(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{dtype:s}=n;if(s==="complex64"){if(r.dtype==="complex64")return ar({inputs:{x:r},backend:a});let p=uh(a,r.shape,r.dtype),c=ts({inputs:{x:r},backend:a,attrs:{dtype:"float32"}}),d=Ja({inputs:{real:c,imag:p},backend:a});return a.disposeIntermediateTensorInfo(p),a.disposeIntermediateTensorInfo(c),d}if(r.dtype==="complex64"){let p=qs({inputs:{input:r},backend:a}),c=ts({inputs:{x:p},backend:a,attrs:{dtype:s}});return a.disposeIntermediateTensorInfo(p),c}if(!v.hasEncodingLoss(r.dtype,s)){let p=ar({inputs:{x:r},backend:a});return{dataId:p.dataId,shape:p.shape,dtype:s}}let i=a.data.get(r.dataId).values,[o,l,u]=U7(i,r.shape,r.dtype,s);return a.makeTensorInfo(o,l,u)}var pO={kernelName:ci,backendName:"cpu",kernelFunc:ts};function Xt(e,t,a,n){return a==null?({inputs:r,backend:s})=>{let{a:i,b:o}=r,l=s;Ie([i,o],e);let u=l.data.get(i.dataId).values,p=l.data.get(o.dataId).values,c=i.dtype==="string"?S.fromUint8ToStringArray(u):u,d=i.dtype==="string"?S.fromUint8ToStringArray(p):p,h=n||i.dtype,[f,m]=t(i.shape,o.shape,c,d,h);return l.makeTensorInfo(m,h,f)}:({inputs:r,backend:s})=>{let{a:i,b:o}=r,l=s;if(i.dtype==="complex64"||o.dtype==="complex64"){let u=ts({inputs:{x:i},backend:l,attrs:{dtype:"complex64"}}),p=l.data.get(u.dataId),c=p.complexTensorInfos.real,d=p.complexTensorInfos.imag,h=l.data.get(c.dataId).values,f=l.data.get(d.dataId).values,m=ts({inputs:{x:o},backend:l,attrs:{dtype:"complex64"}}),g=l.data.get(m.dataId),y=g.complexTensorInfos.real,x=g.complexTensorInfos.imag,A=l.data.get(y.dataId).values,b=l.data.get(x.dataId).values,[w,C,T]=a(i.shape,o.shape,h,f,A,b),N=l.makeTensorInfo(T,"float32",w),M=l.makeTensorInfo(T,"float32",C),F=Ja({inputs:{real:N,imag:M},backend:l});return l.disposeIntermediateTensorInfo(u),l.disposeIntermediateTensorInfo(m),l.disposeIntermediateTensorInfo(N),l.disposeIntermediateTensorInfo(M),F}else{let u=l.data.get(i.dataId).values,p=l.data.get(o.dataId).values,c=n||i.dtype,[d,h]=t(i.shape,o.shape,u,p,c);return l.makeTensorInfo(h,c,d)}}}function Z3(e){return(t,a,n,r,s,i)=>{let o=S.assertAndGetBroadcastShape(t,a),l=v.sizeFromShape(o),u=o.length,p=v.computeStrides(o),c=v.getTypedArrayFromDType("float32",l),d=v.getTypedArrayFromDType("float32",l),h=S.getBroadcastDims(t,o),f=S.getBroadcastDims(a,o),m=S.mergeRealAndImagArrays(n,r),g=S.mergeRealAndImagArrays(s,i),y=t.length,x=v.computeStrides(t),A=a.length,b=v.computeStrides(a);if(h.length+f.length===0)for(let w=0;w<c.length;w++){let C=w%m.length,T=w%g.length,N=e(m[C*2],m[C*2+1],g[T*2],g[T*2+1]);c[w]=N.real,d[w]=N.imag}else for(let w=0;w<c.length;w++){let C=v.indexToLoc(w,u,p),T=C.slice(-y);h.forEach(I=>T[I]=0);let N=v.locToIndex(T,y,x),M=C.slice(-A);f.forEach(I=>M[I]=0);let F=v.locToIndex(M,A,b),E=e(m[N*2],m[N*2+1],g[F*2],g[F*2+1]);c[w]=E.real,d[w]=E.imag}return[c,d,o]}}var G7=$t((e,t)=>e+t),cO=Z3((e,t,a,n)=>({real:e+a,imag:t+n})),Gl=Xt(ns,G7,cO),hO={kernelName:ns,backendName:"cpu",kernelFunc:Gl};function J3(e,t,a,n,r){let s=v.sizeFromShape(n),i=v.makeZerosTypedArray(r,a);for(let o=0;o<e.length;o++){let l=e[o];if(l<0)throw new Error("Input x must be non-negative!");l>=r||(s>0?i[l]+=t[o]:i[l]+=1)}return i}function H7(e,t,a,n=!1){let r=e.shape[0],s=e.shape[1],i=$e([r,a],t.dtype);for(let o=0;o<r;o++)for(let l=0;l<s;l++){let u=e.get(o,l);if(u<0)throw new Error("Input x must be non-negative!");u>=a||(n?i.set(1,o,u):t.size>0?i.set(i.get(o,u)+t.get(o,l),o,u):i.set(i.get(o,u)+1,o,u))}return i}var j7=$t((e,t)=>e&t),fO=Xt(Ih,j7),mO={kernelName:Ih,backendName:"cpu",kernelFunc:fO};function or(e){return(t,a,n)=>{let r=v.getArrayFromDType(a,t.length);for(let s=0;s<t.length;++s)r[s]=e(t[s],n);return r}}function ct(e,t,a){let n=or(t);return us(e,n,a)}function us(e,t,a){return({inputs:n,attrs:r,backend:s})=>{let{x:i}=n;Ie(i,e);let o=s,l=o.data.get(i.dataId).values,u;if(i.dtype==="string"){if(!Array.isArray(l))throw new Error("String tensor's value was not an instance of Array");u=S.fromUint8ToStringArray(l)}else u=l;let p=a||i.dtype,c=t(u,p,r);return o.makeTensorInfo(i.shape,p,c)}}var q7=or(e=>Math.ceil(e)),gO=us(hi,q7),yO={kernelName:hi,backendName:"cpu",kernelFunc:gO};function Q3(e,t,a,n){let r=v.getArrayFromDType(a,v.sizeFromShape(t));if(n&&a!=="string"){let s=0;e.forEach(i=>{let o=v.sizeFromShape(i.shape);r.set(i.vals,s),s+=o})}else{let s=0;e.forEach(i=>{let o=a==="string"?S.fromUint8ToStringArray(i.vals):i.vals,l=0;for(let u=0;u<i.shape[0];++u){let p=u*t[1]+s;for(let c=0;c<i.shape[1];++c)r[p+c]=o[l++]}s+=i.shape[1]})}return r}var X7=$t((e,t)=>e===t?1:0),K7=Xt(Ni,X7,null,"bool"),xO={kernelName:Ni,backendName:"cpu",kernelFunc:K7},Y7=or(e=>Math.exp(e)),Z7=us(Ri,Y7,"float32"),AO={kernelName:Ri,backendName:"cpu",kernelFunc:Z7},J7=or(e=>Math.expm1(e)),bO=us(Ei,J7),vO={kernelName:Ei,backendName:"cpu",kernelFunc:bO},Q7=or(e=>Math.floor(e)),wO=us(Fi,Q7),kO={kernelName:Fi,backendName:"cpu",kernelFunc:wO},e6=$t((e,t)=>Math.floor(e/t)),IO=Xt(_i,e6,null,"int32"),SO={kernelName:_i,backendName:"cpu",kernelFunc:IO};function t6(e,t,a,n,r,s,i,o,l){let u=$e([n,s],a);for(let p=0;p<n;p++){let c=[],d=0;for(let h=0;h<r;h++){let f=e[p*r+h];d+=f*i[h],c.push(f)}if(d<0||d>=l/s)throw new Error(`Invalid indices: ${c} does not index into ${o}`);for(let h=0;h<s;h++)u.values[p*s+h]=t.get(...t.indexToLoc(d*s+h))}return u}function a6(e,t,a){let n=$e(a,e.dtype);for(let r=0;r<n.size;++r){let s=n.indexToLoc(r).slice(),i=s[0],o=s[2],l=t.locToIndex([i,o]);s[2]=t.values[l];let u=e.locToIndex(s);0<=u&&u<e.values.length&&(n.values[r]=e.values[u])}return n}var n6=$t((e,t)=>e>t?1:0),CO=Xt(Pi,n6,null,"bool"),TO={kernelName:Pi,backendName:"cpu",kernelFunc:CO},r6=$t((e,t)=>e>=t?1:0),NO=Xt(Oi,r6,null,"bool"),RO={kernelName:Oi,backendName:"cpu",kernelFunc:NO},s6=$t((e,t)=>e<t?1:0),EO=Xt(Ui,s6,null,"bool"),MO={kernelName:Ui,backendName:"cpu",kernelFunc:EO},i6=$t((e,t)=>e<=t?1:0),FO=Xt(Gi,i6,null,"bool"),_O={kernelName:Gi,backendName:"cpu",kernelFunc:FO};function o6(e,t,a){let n=(t-e)/(a-1),r=v.makeZerosTypedArray(a,"float32");r[0]=e;for(let s=1;s<r.length;s++)r[s]=r[s-1]+n;return r}var l6=or(e=>Math.log(e)),$O=us(ji,l6),DO={kernelName:ji,backendName:"cpu",kernelFunc:$O};function u6(e,t,a,n){let r=v.getTypedArrayFromDType(n,v.sizeFromShape(a));for(let s=0;s<r.length;++s){let i=s*t,o=e[i];for(let l=0;l<t;++l){let u=e[i+l];(Number.isNaN(u)||u>o)&&(o=u)}r[s]=o}return r}var d6=$t((e,t)=>Math.max(e,t)),PO=Xt(Qi,d6),OO={kernelName:Qi,backendName:"cpu",kernelFunc:PO},p6=$t((e,t)=>Math.min(e,t)),zO=Xt(no,p6),LO={kernelName:no,backendName:"cpu",kernelFunc:zO},eg=$t((e,t)=>e*t),WO=Z3((e,t,a,n)=>({real:e*a-t*n,imag:e*n+t*a})),Yh=Xt(io,eg,WO),BO={kernelName:io,backendName:"cpu",kernelFunc:Yh};function c6(e,t,a){let n=v.createScalarValue(-1,a);return eg([],t,n,e,a)}function VO(e){let{inputs:t,backend:a}=e,{x:n}=t;Ie(n,"neg");let r=a.data.get(n.dataId).values,[s,i]=c6(r,n.shape,n.dtype);return a.makeTensorInfo(i,n.dtype,s)}var UO={kernelName:mu,backendName:"cpu",kernelFunc:VO},h6=$t((e,t)=>e!==t?1:0),GO=Xt(oo,h6,null,"bool"),HO={kernelName:oo,backendName:"cpu",kernelFunc:GO};function tg(e,t,a,n,r){let s=t.length,i=v.sizeFromShape(t),o=v.computeStrides(t),l=v.computeStrides(r),u=v.getTypedArrayFromDType(a,v.sizeFromShape(r));for(let p=0;p<i;++p){let c=v.indexToLoc(p,s,o),d=new Array(c.length);for(let f=0;f<d.length;f++)d[f]=c[n[f]];let h=v.locToIndex(d,s,l);u[h]=e[p]}return u}function Ua(e){let{inputs:t,attrs:a,backend:n}=e,{x:r}=t,{perm:s}=a;Ie(r,"transpose");let i=r.shape.length,o=new Array(i);for(let p=0;p<o.length;p++)o[p]=r.shape[s[p]];let l=n.data.get(r.dataId).values,u=tg(l,r.shape,r.dtype,s,o);return{dataId:n.write(u,o,r.dtype),shape:o,dtype:r.dtype}}var jO={kernelName:wr,backendName:"cpu",kernelFunc:Ua};function f6(e,t,a,n){let[r,s]=S.computeOutAndReduceShapes(e,n),i=pa(t,"int32"),o=v.makeZerosTypedArray(v.sizeFromShape(r),i),l=v.sizeFromShape(s);for(let u=0;u<o.length;++u){let p=u*l,c=1;for(let d=0;d<l;++d)c*=a[p+d];o[u]=c}return{outVals:o,outShape:r,outDtype:i}}function qO(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{axis:s,keepDims:i}=n;Ie(r,"prod");let o=r.shape.length,l=v.parseAxisParam(s,r.shape),u=S.getAxesPermutation(l,o),p=l,c=r,d=[];u!=null&&(c=Ua({inputs:{x:r},backend:a,attrs:{perm:u}}),d.push(c),p=S.getInnerMostAxes(p.length,o));let h=a.data.get(c.dataId).values,{outVals:f,outShape:m,outDtype:g}=f6(c.shape,c.dtype,h,p),y=m;return i&&(y=S.expandShapeToKeepDim(m,l)),d.forEach(x=>a.disposeIntermediateTensorInfo(x)),a.makeTensorInfo(y,g,f)}var XO={kernelName:mo,backendName:"cpu",kernelFunc:qO};function KO(e,t,a){e.forEach((n,r)=>{if(n<0||n>=a){let s=v.indexToLoc(r,t.length,v.computeStrides(t)).join(",");throw new Error(`indices[${s}] = ${n} is not in [0, ${a})`)}})}function YO(e,t){for(let a=0;a<e.length;++a){let n=e[a],r=a===e.length-1?t:e[a+1].length;if(n.length===0)throw new Error("Ragged splits may not be empty");if(n[0]<0)throw new Error("Ragged splits must be non-negative");if(n[n.length-1]>r)throw new Error("Ragged splits must not point past values");for(let s=1;s<n.length;++s)if(n[s-1]>n[s])throw new Error("Ragged splits must be sorted in ascending order")}}function ZO(e,t,a,n){let r=[],s=0,i=t.length-1+a.length,o=new Array(i).fill(null).map(()=>[0]);YO(a,n);let l=1;for(let u=0;u<t.length-1;++u){l*=t[u];let p=t[u+1];for(let c=1;c<l+1;++c)o[u].push(c*p)}for(let u=0;u<e.length;++u){let p=e[u],c=e[u]+1;for(let d=0;d<a.length;++d){let h=a[d],f=d+t.length-1;if(f>=0){let m=o[f],g=m[m.length-1]-h[p];for(let y=p;y<c;++y)o[f].push(h[y+1]+g)}p=h[p],c=h[c]}c!==p&&(r.push([p,c]),s+=c-p)}return{outSplits:o,valueSlices:r,numValues:s}}function JO(e){let t=[];for(let a=0;a<e.length;++a){let n=e[a].length,r=v.getArrayFromDType("int32",n);t.push(r),e[a].forEach((s,i)=>r[i]=s)}return t}function Z5(e,t){let a=e.slice(0,t);for(;a.length<t;)a.push(1);for(let n=t;n<e.length;n++)a[t-1]*=e[n];return a}function QO(e,t,a,n,r,s){let i=Z5(t,2)[1],o=Z5(s,2)[1],l=0;for(let u of a)for(let p=u[0];p<u[1];++p){for(let c=0;c<n;++c)r[l*o+c]=e[p*i+c];++l}}function ez(e,t,a,n,r){let s=t.slice();s[0]=r;let i=v.getArrayFromDType(a,v.sizeFromShape(s)),o=e.length,l=o===0?0:o/t[0];return QO(e,t,n,l,i,s),[i,s]}function m6(e,t,a,n,r,s,i,o){if(e.length===0)throw new Error("paramsNestedSplits must be non empty");if(t[0].length===0)throw new Error("Split tensors must not be scalars");let l=t[0][0]-1;if(KO(s,i,l),n.length===0)throw new Error("params.rank must be nonzero");let u=n[0],{outSplits:p,valueSlices:c,numValues:d}=ZO(s,i,e,u),h=JO(p),f=ez(a,n,r,c,d);return[h,f[0],f[1]]}var J5=2147483647;function g6(e,t,a,n,r,s,i){if(t.length>1)throw new Error("starts must be a scalar or vector");if(r.length>1)throw new Error("limits must be a scalar or vector");if(i.length>1)throw new Error("deltas must be a scalar or vector");let o=t.length===0,l=r.length===0,u=i.length===0,p=[];o||p.push(t[0]),l||p.push(r[0]),u||p.push(i[0]);for(let g=1;g<p.length;++g)if(p[g]!==p[g-1])throw new Error("starts, limits, and deltas must have the same shape");let c=p.length===0?1:p[0],d=v.getArrayFromDType("int32",c+1);d[0]=0;for(let g=0;g<c;++g){let y=o?e[0]:e[g],x=l?n[0]:n[g],A=u?s[0]:s[g];if(A===0)throw new Error("Requires delta != 0");let b;if(A>0&&x<y||A<0&&x>y)b=0;else if(b=Math.ceil(Math.abs((x-y)/A)),b>J5)throw new Error(`Requires ((limit - start) / delta) <= ${J5}`);d[g+1]=d[g]+b}let h=d[c],f=v.getArrayFromDType(a,h),m=0;for(let g=0;g<c;++g){let y=d[g+1]-d[g],x=o?e[0]:e[g],A=u?s[0]:s[g];for(let b=0;b<y;++b)f[m++]=x,x+=A}return[d,f]}var kn=S.RowPartitionType,x1=class{constructor(e,t,a,n,r,s,i,o,l,u){this.shape=e,this.shapeShape=t,this.values=a,this.valuesShape=n,this.valuesDType=r,this.defaultValue=s,this.defaultValueShape=i,this.rowPartitionValues=o,this.rowPartitionValuesShapes=l,this.rowPartitionTypes=S.getRowPartitionTypesHelper(u),this.raggedRank=S.getRaggedRank(this.rowPartitionTypes)}getRowPartitionTypeByDimension(e){return this.rowPartitionTypes[0]===kn.FIRST_DIM_SIZE?this.rowPartitionTypes[e+1]:this.rowPartitionTypes[e]}getRowPartitionTensor(e){return this.rowPartitionTypes[0]===kn.FIRST_DIM_SIZE?this.rowPartitionValues[e+1]:this.rowPartitionValues[e]}getMaxWidth(e){let t=this.getRowPartitionTensor(e-1);switch(this.getRowPartitionTypeByDimension(e-1)){case kn.VALUE_ROWIDS:return x1.getMaxWidthValueRowID(t);case kn.ROW_SPLITS:return x1.getMaxWidthRowSplit(t);default:throw new Error(`Cannot handle partition type ${kn[this.getRowPartitionTypeByDimension(e-1)]}`)}}static getMaxWidthRowSplit(e){let t=e.length;if(t===0||t===1)return 0;let a=0;for(let n=0;n<t-1;++n){let r=e[n+1]-e[n];r>a&&(a=r)}return a}static getMaxWidthValueRowID(e){let t=e.length;if(t===0)return 0;let a=0,n=e[0],r=0;for(let s=1;s<t;++s){let i=e[s];i!==n&&(n=i,r=Math.max(s-a,r),a=s)}return Math.max(t-a,r)}tensorShapeFromTensor(e,t,a=!0){if(t.length===0){if(e[0]===-1)return[];throw new Error("The only valid scalar shape tensor is the fully unknown shape specified as -1.")}return ex(e,a)}calculateOutputSize(e){let t=this.valuesShape,a=this.defaultValueShape;S.validateDefaultValueShape(a,t);let n=this.tensorShapeFromTensor(this.shape,this.shapeShape),r=S.combineRaggedTensorToTensorShapes(this.raggedRank,n,t);r[0]<0&&(r[0]=e);for(let s=1;s<=this.raggedRank;++s)r[s]<0&&(r[s]=this.getMaxWidth(s));return r}calculateFirstParentOutputIndex(e,t,a){let n=Math.min(e,a),r=[],s=0;for(let i=0;i<n;++i,s+=t)r.push(s);for(let i=n;i<e;++i)r.push(-1);return v.assert(r.length===e,()=>"Final length of result must be equal to firstDimension."),r}calculateOutputIndexRowSplit(e,t,a,n){let r=e.length,s=[];for(let i=0;i<r-1;++i){let o=e[i+1]-e[i],l=Math.min(n,o),u=t[i];u===-1&&(l=0);for(let p=0;p<l;++p)s.push(u),u+=a;for(let p=0;p<o-l;++p)s.push(-1)}if(r>0&&s.length!==e[r-1])throw new Error("Invalid row split size.");return s}calculateOutputIndexValueRowID(e,t,a,n){let r=e.length,s=[];if(r===0)return[];let i=0,o=e[0];if(o>=t.length)throw new Error(`Got currentValueRowId=${o}, which is not less than ${t.length}`);let l=t[o];s.push(l);for(let u=1;u<r;++u){let p=e[u];if(p===o)l>=0&&(++i,i<n?l+=a:l=-1);else{if(i=0,o=p,p>=t.length)throw new Error(`Got nextValueRowId=${p} which is not less than ${t.length}`);l=t[p]}s.push(l)}if(s.length!==e.length)throw new Error("Invalid row ids.");return s}calculateOutputIndex(e,t,a,n){let r=this.getRowPartitionTensor(e),s=this.getRowPartitionTypeByDimension(e);switch(s){case kn.VALUE_ROWIDS:return this.calculateOutputIndexValueRowID(r,t,a,n);case kn.ROW_SPLITS:if(r.length-1>t.length)throw new Error(`Row partition size is greater than output size: ${r.length-1} > ${t.length}`);return this.calculateOutputIndexRowSplit(r,t,a,n);default:throw new Error(`Unsupported partition type: ${kn[s]}`)}}getFirstDimensionSize(){let e=this.rowPartitionValues[0];if(this.rowPartitionTypes.length===0)throw new Error("No row_partition_types given.");let t=this.rowPartitionTypes[0];switch(t){case kn.FIRST_DIM_SIZE:return e[0];case kn.VALUE_ROWIDS:throw new Error("Cannot handle VALUE_ROWIDS in first dimension.");case kn.ROW_SPLITS:return this.rowPartitionValuesShapes[0][0]-1;default:throw new Error(`Cannot handle type ${kn[t]}`)}}compute(){if(this.rowPartitionValues[0].length<=0)throw new Error("Invalid first partition input. Tensor requires at least one element.");let e=this.getFirstDimensionSize(),t=this.calculateOutputSize(e),a=new Array(this.raggedRank+1);a[a.length-1]=1;for(let s=a.length-2;s>=0;--s)a[s]=a[s+1]*t[s+1];let n=ex(t,!1),r=v.getArrayFromDType(this.valuesDType,v.sizeFromShape(n));if(a[0]*t[0]>0){let s=this.calculateFirstParentOutputIndex(e,a[0],t[0]);for(let i=1;i<=this.raggedRank;++i)s=this.calculateOutputIndex(i-1,s,a[i],t[i]);this.setOutput(this.raggedRank,s,r,n)}return[n,r]}setOutput(e,t,a,n){if(a.length===0)return;let r=this.values,s=a,i=n.slice();i=i.slice(e+1);let o=v.sizeFromShape(i),l=t.length,u=this.defaultValue;if(u.length!==o&&u.length!==1){let h=this.defaultValueShape;Pe(()=>{let f=Q(u,h);u=Fl(f,i).dataSync()})}let p=0,c=0,d=0;for(let h=0;h<=l;++h){let f=h<l?t[h]:-1;if(f===d){++d;continue}if(c<d){let m=r.subarray(p*o),g=s.subarray(c*o),y=(d-c)*o;Q5(g,m,y)}if(h>=l){let m=a.length;f=Math.floor(m/o)}if(f>d)if(this.defaultValue.length===1)s.subarray(d*o,f*o).fill(this.defaultValue[0]),d=f;else for(;f>d;){let m=s.slice(d*o);Q5(m,u,o),++d}f<0?(p=h+1,c=d):(p=h,c=d,d=c+1)}}};function Q5(e,t,a){for(let n=0;n<a;n++)e[n]=t[n]}function ex(e,t){let a=[];for(let n of e){if(n<0){if(!t)throw new Error(`Dimension ${n} must be >= 0`);if(n<-1)throw new Error(`Dimension ${n} must be >= -1`);n=-1}a.push(n)}return a}function y6(e,t,a,n,r,s,i,o,l,u){return new x1(e,t,a,n,r,s,i,o,l,u).compute()}function ag(e,t,a,n){let r=e===t,s=e<t&&a<0,i=t<e&&a>1;if(r||s||i)return v.makeZerosTypedArray(0,n);let o=Math.abs(Math.ceil((t-e)/a)),l=v.makeZerosTypedArray(o,n);t<e&&a===1&&(a=-1),l[0]=e;for(let u=1;u<l.length;u++)l[u]=l[u-1]+a;return l}var x6=or(e=>1/Math.sqrt(e)),tz=us(ko,x6),az={kernelName:ko,backendName:"cpu",kernelFunc:tz};function Bs(e,t,a,n,r,s,i,o,l,u){let p=[n/r,r],c=e.values,d=t.values;if(n===0)return $e(a,t.dtype);let h=l instanceof Wt?l:$e(p,t.dtype);typeof l=="string"||typeof l=="number"?h.values.fill(l):typeof l=="boolean"&&h.values.fill(+l);for(let f=0;f<s;f++){let m=[],g=0;for(let y=0;y<i;y++){let x=c[f*i+y];m.push(x),g+=x*o[y]}if(g<0||g>=n/r)throw new Error(`Invalid indices: ${m} does not index into ${a}`);for(let y=0;y<r;y++)u?h.values[g*r+y]+=d[f*r+y]:h.values[g*r+y]=t.rank===0?d[0]:d[f*r+y]}return h}var nz=or(e=>1/(1+Math.exp(-e))),A6=ct(Eo,e=>1/(1+Math.exp(-e))),rz={kernelName:Eo,backendName:"cpu",kernelFunc:A6};function dh(e,t,a,n,r){let s=Nt.isSliceContinous(n,t,a),i=v.sizeFromShape(a),o=v.computeStrides(n);if(s){let c=Nt.computeFlatOffset(t,o);return r==="string"?e.slice(c,c+i):e.subarray(c,c+i)}let l=r==="string"?S.fromUint8ToStringArray(e):e,u=$e(n,r,l),p=$e(a,r);for(let c=0;c<p.size;++c){let d=p.indexToLoc(c),h=d.map((f,m)=>f+t[m]);p.set(u.get(...h),...d)}return r==="string"?S.fromStringArrayToUint8(p.values):p.values}function Xs(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{begin:s,size:i}=n;Ie(r,"slice");let[o,l]=Nt.parseSliceParams(r,s,i);Nt.assertParamsValid(r,o,l);let u=a.data.get(r.dataId).values,p=dh(u,o,l,r.shape,r.dtype);return a.makeTensorInfo(l,r.dtype,p)}var sz={kernelName:Iu,backendName:"cpu",kernelFunc:Xs};function b6(e,t,a,n,r,s,i){let o=t[0],l=s[0],u=new Array(l),p=new Array(o),c=t[1];if(l===0){if(o!==0)throw new Error(S.getSparseFillEmptyRowsIndicesDenseShapeMismatch(o));let g=v.getArrayFromDType(a,0),y=v.getArrayFromDType(r,0);return[g,[0,c],y,u,p]}let d=!0,h=0,f=new Array(l).fill(0);for(let g=0;g<o;++g){let y=e[g*c];if(y<0)throw new Error(S.getSparseFillEmptyRowsNegativeIndexErrorMessage(g,y));if(y>=l)throw new Error(S.getSparseFillEmptyRowsOutOfRangeIndexErrorMessage(g,y,l));++f[y],d=d&&y>=h,h=y}let m=!0;for(let g=0;g<l;++g){let y=f[g]===0;u[g]=y,m=m&&!y,f[g]=Math.max(f[g],1),g>0&&(f[g]+=f[g-1])}if(m&&d){let g=e,y=n;for(let x=0;x<o;++x)p[x]=x;return[g,[o,c],y,u,p]}else{let g=f[l-1],y=v.getArrayFromDType(a,g*c),x=v.getArrayFromDType(r,g),A=new Array(l).fill(0);for(let b=0;b<o;++b){let w=e[b*c],C=A[w],T=(w===0?0:f[w-1])+C;A[w]++;for(let N=0;N<c;++N)y[T*c+N]=e[b*c+N];x[T]=n[b],p[b]=T}for(let b=0;b<l;++b)if(A[b]===0){let w=b===0?0:f[b-1];y[w*c+0]=b;for(let C=1;C<c;++C)y[w*c+C]=0;x[w]=i}return[y,[g,c],x,u,p]}}function v6(e,t,a,n,r){let s=v.sizeFromShape(n),i=t[0],o=r.length,l=[],u=1,p=-1;for(let m=0;m<o;++m){let g=r[m];if(g===-1){if(p!==-1)throw new Error(S.getSparseReshapeMultipleNegativeOneOutputDimErrorMessage(p,m));p=m,l.push(1)}else{if(g<0)throw new Error(S.getSparseReshapeNegativeOutputDimErrorMessage(m,g));u*=g,l.push(g)}}if(p!==-1){if(u<=0)throw new Error(S.getSparseReshapeEmptyTensorZeroOutputDimErrorMessage());let m=Math.trunc(s/u);if(u*m!==s)throw new Error(S.getSparseReshapeInputOutputMultipleErrorMessage(n,l));l[p]=m}if(v.sizeFromShape(l)!==s)throw new Error(S.getSparseReshapeInputOutputMismatchErrorMessage(n,l));let c=n.length,d=[];if(c>0){d[c-1]=1;for(let m=c-2;m>=0;--m)d[m]=d[m+1]*n[m+1]}let h=[];if(o>0){h[o-1]=1;for(let m=o-2;m>=0;--m)h[m]=h[m+1]*l[m+1]}let f=v.getArrayFromDType(a,i*o);for(let m=0;m<i;++m){let g=0;for(let y=0;y<c;++y)g+=e[m*c+y]*d[y];for(let y=0;y<o;++y)f[m*o+y]=Math.trunc(g/h[y]),g%=h[y]}return[f,[i,o],l]}function ng(e,t,a,n,r,s=!1,i=0){let o=n.length,l=[t[0],e.length/t[0]],u=l[1],p=o>0?r[o-1]+1:0;if(p<0)throw new Error(S.getSparseSegmentReductionNegativeSegmentIdsErrorMessage());let c=t.slice();c[0]=p;let d=c.reduce((x,A)=>x*A,1),h=v.getArrayFromDType(a,d);if(o===0)return p>0&&h.fill(i),[h,c];if(p<=0)throw new Error(S.getSparseSegmentReductionNegativeSegmentIdsErrorMessage());let f=0,m=1,g=0,y=r[f];for(;;){let x=0;if(m<o){if(x=r[m],y===x){++m;continue}if(y>=x)throw new Error(S.getSparseSegmentReductionNonIncreasingSegmentIdsErrorMessage())}if(y<0||y>=p)throw new Error(S.getSparseSegmentReductionSegmentIdOutOfRangeErrorMessage(y,p));y>g&&h.fill(i,g*u,y*u);for(let A=f;A<m;++A){let b=n[A];if(b<0||b>=l[0])throw new Error(S.getSparseSegmentReductionIndicesOutOfRangeErrorMessage(A,n[A],l[0]));for(let w=0;w<u;w++)h[y*u+w]+=e[b*u+w]}if(s)for(let A=0;A<u;A++)h[y*u+A]/=m-f;if(f=m,++m,g=y+1,y=x,m>o)break}return g<p&&h.fill(i,g*u,p*u),[h,c]}var iz=or(e=>Math.sqrt(e)),oz=ct(Fo,e=>Math.sqrt(e)),lz={kernelName:Fo,backendName:"cpu",kernelFunc:oz},w6=$t((e,t)=>{let a=e-t;return a*a}),uz=Xt(Po,w6),dz={kernelName:Po,backendName:"cpu",kernelFunc:uz},k6=or((e,t)=>{let{pattern:a,replaceGlobal:n,rewrite:r}=t;return e.replace(new RegExp(a,n?"g":""),r)}),pz=us(yp,k6),cz={kernelName:yp,backendName:"cpu",kernelFunc:pz};function I6(e,t,a,n){let r=$e(e,t.dtype);for(let s=0;s<r.size;s++){let i=r.indexToLoc(s),o=new Array(i.length);for(let l=0;l<o.length;l++)o[l]=i[l]*a[l]+n[l];r.set(t.get(...o),...i)}return r}var hz=class{constructor(e,t,a,n,r,s){this.separator=v.encodeString(e),this.nGramWidths=t,this.leftPad=v.encodeString(a),this.rightPad=v.encodeString(n),this.padWidth=r,this.preserveShort=s}getPadWidth(e){return Math.min(this.padWidth<0?e-1:this.padWidth,e-1)}getNumNGrams(e,t){let a=this.getPadWidth(t);return Math.max(0,e+2*a-t+1)}createNGrams(e,t,a,n,r,s){for(let i=0;i<r;++i){let o=this.getPadWidth(s),l=Math.max(0,o-i),u=Math.max(0,o-(r-(i+1))),p=s-(l+u),c=t+(l>0?0:i-o),d=0;d+=l*this.leftPad.length;for(let y=0;y<p;++y)d+=e[c+y].length;d+=u*this.rightPad.length;let h=l+u+p-1;d+=h*this.separator.length,a[n+i]=new Uint8Array(d);let f=a[n+i],m=0,g=y=>y.forEach(x=>f[m++]=x);for(let y=0;y<l;++y)g(this.leftPad),g(this.separator);for(let y=0;y<p-1;++y)g(e[c+y]),g(this.separator);if(p>0){g(e[c+p-1]);for(let y=0;y<u;++y)g(this.separator),g(this.rightPad)}else{for(let y=0;y<u-1;++y)g(this.rightPad),g(this.separator);g(this.rightPad)}}}compute(e,t){let a=e.length,n=t.length;if(n>0){let o=t[0];if(o!==0)throw new Error(`First split value must be 0, got ${o}`);for(let l=1;l<n;++l){let u=t[l]>=o;if(u=u&&t[l]<=a,!u)throw new Error(`Invalid split value ${t[l]}, must be in [${o}, ${a}]`);o=t[l]}if(o!==a)throw new Error(`Last split value must be data size. Expected ${a}, got ${o}`)}let r=n-1,s=v.getArrayFromDType("int32",n);if(a===0||n===0){let o=new Array(a);for(let l=0;l<=r;++l)s[l]=0;return[o,s]}s[0]=0;for(let o=1;o<=r;++o){let l=t[o]-t[o-1],u=0;this.nGramWidths.forEach(p=>{u+=this.getNumNGrams(l,p)}),this.preserveShort&&l>0&&u===0&&(u=1),s[o]=s[o-1]+u}let i=new Array(s[r]);for(let o=0;o<r;++o){let l=t[o],u=s[o];if(this.nGramWidths.forEach(p=>{let c=t[o+1]-t[o],d=this.getNumNGrams(c,p);this.createNGrams(e,l,i,u,d,p),u+=d}),this.preserveShort&&u===s[o]){let p=t[o+1]-t[o];if(p===0)continue;let c=p+2*this.padWidth,d=1;this.createNGrams(e,l,i,u,d,c)}}return[i,s]}};function rg(e,t,a,n,r,s,i,o){return new hz(a,n,r,s,i,o).compute(e,t)}function fz(e,t,a,n){if(!e.length)return;if(t.length===0){for(let s=0;s<e.length;++s)n.push(e.subarray(s,s+1));return}if(t.length===1){let s=t[0],i=e.indexOf(s);for(;i!==-1;){let o=e.subarray(0,i);(!a||o.length!==0)&&n.push(o),e=e.subarray(i+1),i=e.indexOf(s)}(!a||e.length!==0)&&n.push(e);return}let r=0;for(let s=0;s<e.length+1;s++)if(s===e.length||t.indexOf(e[s])!==-1){let i=e.subarray(r,s);(!a||i.length!==0)&&n.push(i),r=s+1}}function sg(e,t,a){let n=e.length,r=[],s=0,i=0,o=new Array(n);for(let d=0;d<n;++d){let h=r.length;fz(e[d],t,a,r);let f=r.length-h;o[d]=f,s+=f,i=Math.max(i,f)}let l=v.getArrayFromDType("int32",s*2),u=new Array(s),p=[n,i],c=0;for(let d=0;d<n;++d)for(let h=0;h<o[d];++h)l[c*2]=d,l[c*2+1]=h,u[c]=r[c],++c;return[l,u,p]}function ig(e,t){let a=v.getArrayFromDType("int32",e.length);for(let n=0;n<e.length;++n)a[n]=v.fingerPrint64(e[n]).modulo(t).getLowBitsUnsigned();return a}var S6=$t((e,t)=>e-t),mz=Z3((e,t,a,n)=>({real:e-a,imag:t-n})),og=Xt(zo,S6,mz),gz={kernelName:zo,backendName:"cpu",kernelFunc:og};function C6(e,t){let a=new Array(e.rank);for(let r=0;r<a.length;r++)a[r]=e.shape[r]*t[r];let n=$e(a,e.dtype);for(let r=0;r<n.values.length;++r){let s=n.indexToLoc(r),i=new Array(e.rank);for(let l=0;l<i.length;l++)i[l]=s[l]%e.shape[l];let o=e.locToIndex(i);n.values[r]=e.values[o]}return n}var yd=(e,t)=>{let a=t.value-e.value;return a===0?e.index-t.index:a};function T6(e,t,a=0,n=e.length-1){for(;n>a;){if(n-a>600){let o=n-a+1,l=t-a+1,u=Math.log(o),p=.5*Math.exp(2*u/3),c=.5*Math.sqrt(u*p*(o-p)/o)*Math.sign(l-o/2),d=Math.max(a,Math.floor(t-l*p/o+c)),h=Math.min(n,Math.floor(t+(o-l)*p/o+c));T6(e,t,d,h)}let r=e[t],s=a,i=n;for(v.swap(e,a,t),yd(e[n],r)>0&&v.swap(e,a,n);s<i;){for(v.swap(e,s,i),s++,i--;yd(e[s],r)<0;)s=s+1;for(;yd(e[i],r)>0;)i=i-1}yd(e[a],r)===0?v.swap(e,a,i):(i=i+1,v.swap(e,i,n)),i<=t&&(a=i+1),t<=i&&(n=i-1)}}function N6(e,t,a,n,r){let s=t[t.length-1],[i,o]=[e.length/s,s],l=v.getTypedArrayFromDType(a,i*n),u=v.getTypedArrayFromDType("int32",i*n);for(let c=0;c<i;c++){let d=c*o,h=e.subarray(d,d+o),f=new Array(h.length);h.forEach((x,A)=>f[A]={value:x,index:A}),n<f.length&&(T6(f,n),f=f.slice(0,n)),r&&f.sort(yd);let m=c*n,g=l.subarray(m,m+n),y=u.subarray(m,m+n);for(let x=0;x<n;x++)g[x]=f[x].value,y[x]=f[x].index}let p=t.slice();return p[p.length-1]=n,[$e(p,a,l),$e(p,"int32",u)]}function lg(e,t,a,n){let r=v.parseAxisParam(t,a)[0],s=[1,a[0],1];for(let f=0;f<r;f++)s[0]*=a[f];s[1]=a[r];for(let f=r+1;f<a.length;f++)s[2]*=a[f];let i=new Map,o=new Int32Array(a[r]),l=new Wt(s,n,e),u=[],p=s[0]===1&&s[2]===1;for(let f=0;f<a[r];f++){let m;if(p)m=e[f].toString();else{let y=[];for(let x=0;x<s[0];x++)for(let A=0;A<s[2];A++)y.push(l.get(x,f,A));m=y.join(",")}let g=i.get(m);if(g!=null)o[f]=g;else{let y=i.size;i.set(m,y),o[f]=y,u.push(f)}}let c=s.slice();c[1]=i.size;let d=new Wt(c,n);u.forEach((f,m)=>{for(let g=0;g<s[0];g++)for(let y=0;y<s[2];y++)d.set(l.get(g,f,y),g,m,y)});let h=a.slice();return h[r]=c[1],{outputValues:d.values,outputShape:h,indices:o}}var yz="4.5.0";Go("cpu",()=>new Xh,1);var R6=ct(Ti,e=>e>=0?e:Math.exp(e)-1),xz={kernelName:Ti,backendName:"cpu",kernelFunc:R6};function E6(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{alpha:s}=n;Ie([r],"leakyRelu");let i=v.sizeFromShape(r.shape),o=a.data.get(r.dataId).values,l=v.getTypedArrayFromDType("float32",i);for(let u=0;u<o.length;u++)l[u]=o[u]<0?s*o[u]:o[u];return a.makeTensorInfo(r.shape,"float32",l)}var Az={kernelName:Vi,backendName:"cpu",kernelFunc:E6},bz=$t((e,t)=>e<0?t*e:e);function M6(e){let{inputs:t,backend:a}=e,{x:n,alpha:r}=t;Ie([n,r],"prelu");let s=a.data.get(n.dataId).values,i=a.data.get(r.dataId).values,[o,l]=bz(n.shape,r.shape,s,i,"float32");return a.makeTensorInfo(l,"float32",o)}var vz={kernelName:fo,backendName:"cpu",kernelFunc:M6},F6=ct(yo,e=>Math.max(0,e)),wz={kernelName:yo,backendName:"cpu",kernelFunc:F6},_6=ct(bo,e=>Math.min(Math.max(0,e),6)),kz={kernelName:bo,backendName:"cpu",kernelFunc:_6};function ph(e,t,a,n,r){if(a==="linear")return ar({inputs:{x:t},backend:e});if(a==="relu")return F6({inputs:{x:t},backend:e});if(a==="elu")return R6({inputs:{x:t},backend:e});if(a==="relu6")return _6({inputs:{x:t},backend:e});if(a==="prelu")return M6({inputs:{x:t,alpha:n},backend:e});if(a==="leakyrelu")return E6({inputs:{x:t},backend:e,attrs:{alpha:r}});if(a==="sigmoid")return A6({inputs:{x:t},backend:e});throw new Error(`Activation ${a} has not been implemented for the CPU backend.`)}function At(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{shape:s}=n,i=v.sizeFromShape(r.shape),o=v.inferFromImplicitShape(s,i),l=v.sizeFromShape(o);v.assert(i===l,()=>`The new shape (${o}) has ${l} elements and the old shape (${r.shape}) has ${i} elements. The new shape and old shape must have the same number of elements.`),a.incRef(r.dataId);let u=a.data.get(r.dataId);if(u.complexTensorInfos!=null){let p=u.complexTensorInfos.real,c=u.complexTensorInfos.imag;p.shape=o,c.shape=o}return{dataId:r.dataId,shape:o,dtype:r.dtype}}var Iz={kernelName:bu,backendName:"cpu",kernelFunc:At};function $6(e){let{inputs:t,backend:a,attrs:n}=e,{a:r,b:s}=t,{transposeA:i,transposeB:o}=n;Ie([r,s],"matMul");let l=r.shape.length,u=s.shape.length,p=i?r.shape[l-2]:r.shape[l-1],c=o?s.shape[u-1]:s.shape[u-2],d=i?r.shape[l-1]:r.shape[l-2],h=o?s.shape[u-2]:s.shape[u-1],f=r.shape.slice(0,-2),m=s.shape.slice(0,-2),g=v.sizeFromShape(f),y=v.sizeFromShape(m),x=Ho.assertAndGetBroadcastShape(r.shape.slice(0,-2),s.shape.slice(0,-2)).concat([d,h]);v.assert(p===c,()=>`Error in matMul: inner shapes (${p}) and (${c}) of Tensors with shapes ${r.shape} and ${s.shape} and transposeA=${i} and transposeB=${o} must match.`);let A=i?[g,p,d]:[g,d,p],b=o?[y,h,c]:[y,c,h],w=At({inputs:{x:r},backend:a,attrs:{shape:A}}),C=At({inputs:{x:s},backend:a,attrs:{shape:b}}),T=i?w.shape[1]:w.shape[2],N=i?w.shape[2]:w.shape[1],M=o?C.shape[1]:C.shape[2],F=Math.max(g,y),E=a.data.get(w.dataId).values,I=a.data.get(C.dataId).values,_=v.computeStrides(w.shape),O=v.computeStrides(C.shape),[B,$,U]=i?[_[0],1,_[1]]:[_[0],_[1],1],[G,q,H]=o?[1,O[1],O[0]]:[O[1],1,O[0]],V=N*M,Z=$e([F,N,M],w.dtype),X=Z.values,re=a.blockSize;for(let ee=0;ee<F;ee++){let me=ee%g,ie=ee%y;for(let Ae=0;Ae<N;Ae+=re){let Ce=Math.min(Ae+re,N);for(let Re=0;Re<M;Re+=re){let Le=Math.min(Re+re,M);for(let qe=0;qe<T;qe+=re){let mt=Math.min(qe+re,T);for(let dt=Ae;dt<Ce;dt++)for(let st=Re;st<Le;st++){let it=0;for(let Ge=qe;Ge<mt;Ge++){let yt=E[me*B+dt*$+Ge*U],ja=I[Ge*G+st*q+ie*H];it+=yt*ja}X[ee*V+(dt*M+st)]+=it}}}}}return a.disposeIntermediateTensorInfo(w),a.disposeIntermediateTensorInfo(C),a.makeTensorInfo(x,Z.dtype,Z.values)}var Sz={kernelName:di,backendName:"cpu",kernelFunc:$6};function Cz(e){let{inputs:t,backend:a,attrs:n}=e,{a:r,b:s,bias:i,preluActivationWeights:o}=t,{transposeA:l,transposeB:u,activation:p,leakyreluAlpha:c}=n,d,h,f,m=[];d=$6({inputs:{a:r,b:s},attrs:{transposeA:l,transposeB:u},backend:a}),i&&(h=Gl({inputs:{a:d,b:i},backend:a}),m.push(d),d=h),p&&(f=ph(a,d,p,o,c),m.push(d),d=f);for(let g of m)a.disposeIntermediateTensorInfo(g);return d}var Tz={kernelName:Xr,backendName:"cpu",kernelFunc:Cz},Nz=ct(Qs,e=>Math.acos(e)),Rz={kernelName:Qs,backendName:"cpu",kernelFunc:Nz},Ez=ct(ei,e=>Math.acosh(e)),Mz={kernelName:ei,backendName:"cpu",kernelFunc:Ez};function Fz(e){let{inputs:t,backend:a}=e,n=t;Ie(t,"addN");let r=n.map(o=>a.data.get(o.dataId).values),s=$e(n[0].shape,n[0].dtype),i=s.values;for(let o=0;o<n.length;o++){let l=r[o];for(let u=0;u<i.length;u++)i[u]+=l[u]}return a.makeTensorInfo(s.shape,s.dtype,s.values)}var _z={kernelName:ti,backendName:"cpu",kernelFunc:Fz};function $z(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{axis:s,keepDims:i}=n;Ie(r,"all");let o=v.parseAxisParam(s,r.shape),l=o,u=S.getAxesPermutation(l,r.shape.length),p=r;u!=null&&(p=Ua({inputs:{x:r},backend:a,attrs:{perm:u}}),l=S.getInnerMostAxes(l.length,r.shape.length)),S.assertAxesAreInnerMostDims("all",l,p.shape.length);let[c,d]=S.computeOutAndReduceShapes(p.shape,l),h=v.sizeFromShape(d),f=v.makeZerosTypedArray(v.sizeFromShape(c),p.dtype),m=a.data.get(p.dataId).values;for(let y=0;y<f.length;++y){let x=y*h,A=m[x];for(let b=0;b<h;++b){let w=m[x+b];A=A&&w}f[y]=A}u!=null&&a.disposeIntermediateTensorInfo(p);let g=a.makeTensorInfo(c,p.dtype,f);if(i){let y=S.expandShapeToKeepDim(c,o),x=At({inputs:{x:g},backend:a,attrs:{shape:y}});return a.disposeIntermediateTensorInfo(g),x}return g}var Dz={kernelName:ai,backendName:"cpu",kernelFunc:$z};function Pz(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{axis:s,keepDims:i}=n;Ie(r,"any");let o=v.parseAxisParam(s,r.shape),l=o,u=S.getAxesPermutation(l,r.shape.length),p=r;u!=null&&(p=Ua({inputs:{x:r},backend:a,attrs:{perm:u}}),l=S.getInnerMostAxes(l.length,r.shape.length)),S.assertAxesAreInnerMostDims("any",l,p.shape.length);let[c,d]=S.computeOutAndReduceShapes(p.shape,l),h=v.sizeFromShape(d),f=v.makeZerosTypedArray(v.sizeFromShape(c),p.dtype),m=a.data.get(p.dataId).values;for(let y=0;y<f.length;++y){let x=y*h,A=m[x];for(let b=0;b<h;++b){let w=m[x+b];A=A||w}f[y]=A}u!=null&&a.disposeIntermediateTensorInfo(p);let g=a.makeTensorInfo(c,p.dtype,f);if(i){let y=S.expandShapeToKeepDim(c,o),x=At({inputs:{x:g},backend:a,attrs:{shape:y}});return a.disposeIntermediateTensorInfo(g),x}return g}var Oz={kernelName:ni,backendName:"cpu",kernelFunc:Pz};function zz(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{axis:s}=n;Ie(r,"argMax");let i=v.parseAxisParam(s,r.shape),o=S.getAxesPermutation(i,r.shape.length),l=r,u=[];o!=null&&(l=Ua({inputs:{x:r},backend:a,attrs:{perm:o}}),u.push(l),i=S.getInnerMostAxes(i.length,l.shape.length)),i=[i[0]],S.assertAxesAreInnerMostDims("argMax",i,l.shape.length);let[p,c]=S.computeOutAndReduceShapes(l.shape,i),d=v.sizeFromShape(p),h=v.makeZerosTypedArray(d,"int32"),f=v.sizeFromShape(c),m=a.data.get(l.dataId).values;for(let g=0;g<h.length;++g){let y=g*f,x=m[y],A=0;for(let b=0;b<f;++b){let w=m[y+b];w>x&&(x=w,A=b)}h[g]=A}return u.forEach(g=>a.disposeIntermediateTensorInfo(g)),a.makeTensorInfo(p,"int32",h)}var Lz={kernelName:Jl,backendName:"cpu",kernelFunc:zz};function Wz(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{axis:s}=n;Ie(r,"argMin");let i=v.parseAxisParam(s,r.shape),o=S.getAxesPermutation(i,r.shape.length),l=r,u=[];o!=null&&(l=Ua({inputs:{x:r},backend:a,attrs:{perm:o}}),u.push(l),i=S.getInnerMostAxes(i.length,l.shape.length)),i=[i[0]],S.assertAxesAreInnerMostDims("argMin",i,l.shape.length);let[p,c]=S.computeOutAndReduceShapes(l.shape,i),d=v.sizeFromShape(p),h=v.makeZerosTypedArray(d,"int32"),f=v.sizeFromShape(c),m=a.data.get(l.dataId).values;for(let g=0;g<h.length;++g){let y=g*f,x=m[y],A=0;for(let b=0;b<f;++b){let w=m[y+b];w<x&&(x=w,A=b)}h[g]=A}return u.forEach(g=>a.disposeIntermediateTensorInfo(g)),a.makeTensorInfo(p,"int32",h)}var Bz={kernelName:Ql,backendName:"cpu",kernelFunc:Wz},Vz=ct(ri,e=>Math.asin(e)),Uz={kernelName:ri,backendName:"cpu",kernelFunc:Vz},Gz=ct(si,e=>Math.asinh(e)),Hz={kernelName:si,backendName:"cpu",kernelFunc:Gz},jz=ct(ii,e=>Math.atan(e)),qz={kernelName:ii,backendName:"cpu",kernelFunc:jz},Xz=$t((e,t)=>Math.atan2(e,t)),Kz=Xt(li,Xz),Yz={kernelName:li,backendName:"cpu",kernelFunc:Kz},Zz=ct(oi,e=>Math.atanh(e)),Jz={kernelName:oi,backendName:"cpu",kernelFunc:Zz};function ug(e,t,a,n,r,s){let i=r.strideHeight,o=r.strideWidth,l=r.dilationHeight,u=r.dilationWidth,p=r.effectiveFilterHeight,c=r.effectiveFilterWidth,d=r.padInfo.top,h=r.padInfo.left,f=s==="max"?Number.NEGATIVE_INFINITY:Number.POSITIVE_INFINITY,m=$e(r.outShape,a),g=m.values,y=r.outShape[1]*r.outShape[2]*r.outShape[3],x=r.outShape[2]*r.outShape[3],A=r.outShape[3];for(let b=0;b<r.batchSize;++b){let w=b*y,C=b*n[0];for(let T=0;T<r.inChannels;++T)for(let N=0;N<r.outHeight;++N){let M=N*i-d,F=Math.max(0,M),E=Math.min(r.inHeight,p+M),I=w+N*x;for(let _=0;_<r.outWidth;++_){let O=_*o-h,B=Math.max(0,O),$=Math.min(r.inWidth,c+O),U=f,G=0,q=0;for(let V=F;V<E;V+=l){let Z=C+V*n[1];for(let X=B;X<$;X+=u){let re=Z+X*n[2],ee=e[re+T];s==="max"&&ee>U?U=ee:s==="avg"&&(G+=ee,q++)}if(isNaN(U))break}let H=I+_*A+T;g[H]=s==="avg"?G/q:U}}}return m}function D6(e,t,a,n,r=!1,s=!1){let i=$e(n.outShape,"int32"),o=n.strideHeight,l=n.strideWidth,u=n.dilationHeight,p=n.dilationWidth,c=n.effectiveFilterHeight,d=n.effectiveFilterWidth,h=n.padInfo.top,f=n.padInfo.left,m=$e(t,a,e);for(let g=0;g<n.batchSize;++g)for(let y=0;y<n.inChannels;++y)for(let x=0;x<n.outHeight;++x){let A=x*o-h,b=A;for(;b<0;)b+=u;let w=Math.min(n.inHeight,c+A);for(let C=0;C<n.outWidth;++C){let T=C*l-f,N=T;for(;N<0;)N+=p;let M=Math.min(n.inWidth,d+T),F=Number.NEGATIVE_INFINITY,E=-1;for(let I=b;I<w;I+=u){let _=I-A;for(let O=N;O<M;O+=p){let B=O-T,$=m.get(g,I,O,y);$>F&&(F=$,r?E=s?((g*n.inHeight+I)*n.inWidth+O)*n.inChannels+y:(I*n.inWidth+O)*n.inChannels+y:E=_*d+B)}}i.set(E,g,x,C,y)}}return i}function P6(e,t,a,n,r,s){let i=r.strideDepth,o=r.strideHeight,l=r.strideWidth,u=r.dilationDepth,p=r.dilationHeight,c=r.dilationWidth,d=r.effectiveFilterDepth,h=r.effectiveFilterHeight,f=r.effectiveFilterWidth,m=r.padInfo.front,g=r.padInfo.top,y=r.padInfo.left,x=s==="max"?Number.NEGATIVE_INFINITY:Number.POSITIVE_INFINITY,A=$e(r.outShape,a),b=A.values,w=r.outShape[1]*r.outShape[2]*r.outShape[3]*r.outShape[4],C=r.outShape[2]*r.outShape[3]*r.outShape[4],T=r.outShape[3]*r.outShape[4],N=r.outShape[4];for(let M=0;M<r.batchSize;++M){let F=M*w,E=M*n[0];for(let I=0;I<r.inChannels;++I)for(let _=0;_<r.outDepth;++_){let O=_*i-m,B=O;for(;B<0;)B+=u;let $=Math.min(r.inDepth,d+O),U=F+_*C;for(let G=0;G<r.outHeight;++G){let q=G*o-g,H=q;for(;H<0;)H+=p;let V=Math.min(r.inHeight,h+q),Z=U+G*T;for(let X=0;X<r.outWidth;++X){let re=X*l-y,ee=re;for(;ee<0;)ee+=c;let me=Math.min(r.inWidth,f+re),ie=Z+X*N,Ae=x,Ce=0,Re=0;for(let qe=B;qe<$;qe+=u){let mt=E+qe*n[1];for(let dt=H;dt<V;dt+=p){let st=mt+dt*n[2];for(let it=ee;it<me;it+=c){let Ge=st+it*n[3],yt=e[Ge+I];if(s==="max"&&yt>Ae?Ae=yt:s==="avg"&&(Ce+=yt,Re++),isNaN(Ae))break}if(isNaN(Ae))break}if(isNaN(Ae))break}let Le=ie+I;b[Le]=s==="avg"?Ce/Math.max(Re,1):Ae}}}}return A}function Qz(e,t){let a=$e(t.outShape,"int32"),n=t.strideDepth,r=t.strideHeight,s=t.strideWidth,i=t.dilationDepth,o=t.dilationHeight,l=t.dilationWidth,u=t.effectiveFilterDepth,p=t.effectiveFilterHeight,c=t.effectiveFilterWidth,d=t.padInfo.front,h=t.padInfo.top,f=t.padInfo.left;for(let m=0;m<t.batchSize;++m)for(let g=0;g<t.inChannels;++g)for(let y=0;y<t.outDepth;++y){let x=y*n-d,A=x;for(;A<0;)A+=i;let b=Math.min(t.inDepth,u+x);for(let w=0;w<t.outHeight;++w){let C=w*r-h,T=C;for(;T<0;)T+=o;let N=Math.min(t.inHeight,p+C);for(let M=0;M<t.outWidth;++M){let F=M*s-f,E=F;for(;E<0;)E+=l;let I=Math.min(t.inWidth,c+F),_=Number.NEGATIVE_INFINITY,O=-1;for(let B=A;B<b;B+=i){let $=B-x;for(let U=T;U<N;U+=o){let G=U-C;for(let q=E;q<I;q+=l){let H=q-F,V=e.get(m,B,U,q,g);V>=_&&(_=V,O=$*p*c+G*p+H)}}}a.set(O,m,y,w,M,g)}}}return a}function eL(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t;Ie(r,"avgPool");let{filterSize:s,strides:i,pad:o,dimRoundingMode:l}=n,u=1;v.assert(S.eitherStridesOrDilationsAreOne(i,u),()=>`Error in avgPool: Either strides or dilations must be 1. Got strides ${i} and dilations '${u}'`);let p=S.computePool2DInfo(r.shape,s,i,u,o,l),c;if(p.filterWidth===1&&p.filterHeight===1&&v.arraysEqual(p.inShape,p.outShape))c=ar({inputs:{x:r},backend:a});else{let d=a.data.get(r.dataId).values,h=v.computeStrides(r.shape),f=ug(d,r.shape,r.dtype,h,p,"avg");c=a.makeTensorInfo(p.outShape,r.dtype,f.values)}return c}var tL={kernelName:ui,backendName:"cpu",kernelFunc:eL};function aL(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{filterSize:s,strides:i,pad:o,dimRoundingMode:l,dataFormat:u}=n;Ie(r,"avgPool3d");let p=S.computePool3DInfo(r.shape,s,i,1,o,l,u),c=a.data.get(r.dataId).values,d=P6(c,r.shape,r.dtype,v.computeStrides(r.shape),p,"avg");return a.makeTensorInfo(d.shape,"float32",d.values)}var nL={kernelName:eu,backendName:"cpu",kernelFunc:aL};function rL(e){let{inputs:t,backend:a,attrs:n}=e,{dy:r,input:s}=t,{filterSize:i,strides:o,pad:l,dimRoundingMode:u}=n;Ie([r,s],"avgPool3DGrad");let p=S.computePool3DInfo(s.shape,i,o,1,l,u),c=p.strideDepth,d=p.strideHeight,h=p.strideWidth,f=p.filterDepth,m=p.filterHeight,g=p.filterWidth,y=p.dilationDepth,x=p.dilationHeight,A=p.dilationWidth,b=p.effectiveFilterDepth,w=p.effectiveFilterHeight,C=p.effectiveFilterWidth,T=b-1-p.padInfo.front,N=C-1-p.padInfo.left,M=w-1-p.padInfo.top,F=$e(s.shape,"float32"),E=1/(f*m*g),I=a.bufferSync(r);for(let _=0;_<p.batchSize;++_)for(let O=0;O<p.inChannels;++O)for(let B=0;B<p.inDepth;++B)for(let $=0;$<p.inHeight;++$)for(let U=0;U<p.inWidth;++U){let G=B-T,q=$-M,H=U-N,V=0;for(let Z=0;Z<b;Z+=y){let X=(G+Z)/c;if(!(X<0||X>=p.outDepth||Math.floor(X)!==X))for(let re=0;re<w;re+=x){let ee=(q+re)/d;if(!(ee<0||ee>=p.outHeight||Math.floor(ee)!==ee))for(let me=0;me<C;me+=A){let ie=(H+me)/h;if(ie<0||ie>=p.outWidth||Math.floor(ie)!==ie)continue;let Ae=I.get(_,X,ee,ie,O);V+=Ae}}}F.set(V*E,_,B,$,U,O)}return a.makeTensorInfo(F.shape,F.dtype,F.values)}var sL={kernelName:ep,backendName:"cpu",kernelFunc:rL};function iL(e){let{inputs:t,backend:a,attrs:n}=e,{dy:r,input:s}=t,i=s;Ie([r,s],"avgPoolGrad");let{filterSize:o,strides:l,pad:u}=n,p=S.computePool2DInfo(i.shape,o,l,1,u),c=p.strideHeight,d=p.strideWidth,h=p.filterHeight,f=p.filterWidth,m=p.dilationHeight,g=p.dilationWidth,y=p.effectiveFilterHeight,x=p.effectiveFilterWidth,A=x-1-p.padInfo.left,b=y-1-p.padInfo.top,w=$e(i.shape,"float32"),C=1/(h*f),T=a.data.get(r.dataId).values,N=$e(r.shape,"float32",T);for(let M=0;M<p.batchSize;++M)for(let F=0;F<p.inChannels;++F)for(let E=0;E<p.inHeight;++E)for(let I=0;I<p.inWidth;++I){let _=E-b,O=I-A,B=0;for(let $=0;$<y;$+=m){let U=(_+$)/c;if(!(U<0||U>=p.outHeight||Math.floor(U)!==U))for(let G=0;G<x;G+=g){let q=(O+G)/d;if(q<0||q>=p.outWidth||Math.floor(q)!==q)continue;let H=N.get(M,U,q,F);B+=H}}w.set(B*C,M,E,I,F)}return a.makeTensorInfo(w.shape,w.dtype,w.values)}var oL={kernelName:kh,backendName:"cpu",kernelFunc:iL};function lL(e){let{inputs:t,backend:a,attrs:n}=e,{x:r,scale:s,offset:i,mean:o,variance:l}=t;v.assert(o.shape.length===l.shape.length,()=>"Batch normalization gradient requires mean and variance to have equal ranks."),v.assert(i==null||o.shape.length===i.shape.length,()=>"Batch normalization gradient requires mean and offset to have equal ranks."),v.assert(s==null||o.shape.length===s.shape.length,()=>"Batch normalization gradient requires mean and scale to have equal ranks."),Ie([r,o,l,s,i],"batchNorm");let{varianceEpsilon:u}=n;u==null&&(u=.001);let p=a.data.get(r.dataId).values,c=a.data.get(o.dataId).values,d=a.data.get(l.dataId).values,h=s?a.data.get(s.dataId).values:new Float32Array([1]),f=i?a.data.get(i.dataId).values:new Float32Array([0]),m=new Float32Array(p.length),g=f.length,y=h.length,x=d.length,A=c.length,b=0,w=0,C=0,T=0;for(let N=0;N<p.length;++N)m[N]=f[b++]+(p[N]-c[w++])*h[C++]/Math.sqrt(d[T++]+u),b>=g&&(b=0),w>=A&&(w=0),C>=y&&(C=0),T>=x&&(T=0);return a.makeTensorInfo(r.shape,r.dtype,m)}var uL={kernelName:$i,backendName:"cpu",kernelFunc:lL};function dL(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{blockShape:s,crops:i}=n;Ie([r],"batchToSpaceND");let o=s.reduce((y,x)=>y*x),l=S.getReshaped(r.shape,s,o),u=S.getPermuted(l.length,s.length),p=S.getReshapedPermuted(r.shape,s,o),c=S.getSliceBeginCoords(i,s.length),d=S.getSliceSize(p,i,s.length),h=At({inputs:{x:r},backend:a,attrs:{shape:l}}),f=Ua({inputs:{x:h},backend:a,attrs:{perm:u}}),m=At({inputs:{x:f},backend:a,attrs:{shape:p}}),g=Xs({inputs:{x:m},backend:a,attrs:{begin:c,size:d}});return a.disposeIntermediateTensorInfo(h),a.disposeIntermediateTensorInfo(f),a.disposeIntermediateTensorInfo(m),g}var pL={kernelName:tu,backendName:"cpu",kernelFunc:dL};function cL(e){let{inputs:t,backend:a,attrs:n}=e,{x:r,weights:s}=t,{size:i}=n,o=a.data.get(r.dataId).values,l=a.data.get(s.dataId).values,u=J3(o,l,s.dtype,s.shape,i);return a.makeTensorInfo([i],s.dtype,u)}var hL={kernelName:pi,backendName:"cpu",kernelFunc:cL};function fL(e){let{inputs:t,backend:a}=e,{s0:n,s1:r}=t,s=a.data.get(n.dataId).values,i=a.data.get(r.dataId).values,o=S.assertAndGetBroadcastShape(Array.from(s),Array.from(i));return a.makeTensorInfo([o.length],"int32",Int32Array.from(o))}var mL={kernelName:au,backendName:"cpu",kernelFunc:fL},gL=ct(rs,(e,t)=>{let a=t;return e>a.clipValueMax?a.clipValueMax:e<a.clipValueMin?a.clipValueMin:e}),yL={kernelName:rs,backendName:"cpu",kernelFunc:gL},xL=e=>{let{x:t}=e.inputs,a=e.backend,n=new Float32Array(v.sizeFromShape(t.shape)),r=a.data.get(t.dataId),s=r.complexTensorInfos.real,i=r.complexTensorInfos.imag,o=a.data.get(s.dataId).values,l=a.data.get(i.dataId).values;for(let u=0;u<o.length;u++){let p=o[u],c=l[u];n[u]=Math.hypot(p,c)}return a.makeOutput(n,t.shape,"float32")},AL={kernelName:ap,backendName:"cpu",kernelFunc:xL};function Hl(e){let{inputs:t,backend:a}=e,{input:n}=t,r=a.data.get(n.dataId).complexTensorInfos.imag,s=a.data.get(r.dataId).values;return a.makeTensorInfo(r.shape,r.dtype,s)}var bL={kernelName:up,backendName:"cpu",kernelFunc:Hl};function jl(e){let{inputs:t,backend:a,attrs:n}=e,{axis:r}=n,s=v.parseAxisParam(r,t[0].shape)[0],i=t.map(m=>m.shape);S.assertParamsConsistent(i,s);let o=S.computeOutShape(t.map(m=>m.shape),s);if(v.sizeFromShape(o)===0)return a.makeTensorInfo(o,t[0].dtype,[]);let l=t.filter(m=>v.sizeFromShape(m.shape)>0);if(l.length===1)return ar({inputs:{x:l[0]},backend:a});if(l[0].dtype==="complex64"){let m=l.map(b=>qs({inputs:{input:b},backend:a})),g=l.map(b=>Hl({inputs:{input:b},backend:a})),y=jl({inputs:m,backend:a,attrs:{axis:s}}),x=jl({inputs:g,backend:a,attrs:{axis:s}}),A=Ja({inputs:{real:y,imag:x},backend:a});return m.forEach(b=>a.disposeIntermediateTensorInfo(b)),g.forEach(b=>a.disposeIntermediateTensorInfo(b)),a.disposeIntermediateTensorInfo(y),a.disposeIntermediateTensorInfo(x),A}let u=l.map(m=>{let g=[-1,v.sizeFromShape(m.shape.slice(s))];return At({inputs:{x:m},backend:a,attrs:{shape:g}})}),p=u.map(m=>({vals:a.data.get(m.dataId).values,shape:m.shape}));o=S.computeOutShape(u.map(m=>m.shape),1);let c=u[0].shape[0]===1,d=Q3(p,o,t[0].dtype,c),h=S.computeOutShape(l.map(m=>m.shape),s),f=a.makeTensorInfo(h,t[0].dtype,d);return u.forEach(m=>a.disposeIntermediateTensorInfo(m)),f}var vL={kernelName:nu,backendName:"cpu",kernelFunc:jl};function O6(e){let{inputs:t,backend:a,attrs:n}=e,{x:r,filter:s}=t,{strides:i,pad:o,dataFormat:l,dilations:u,dimRoundingMode:p}=n;Ie([r,s],"conv2d");let c=S.convertConv2DDataFormat(l),d=S.computeConv2DInfo(r.shape,s.shape,i,u,o,p,!1,c),h=d.filterHeight,f=d.filterWidth,m=d.dilationHeight,g=d.dilationWidth,y=d.padInfo.left,x=d.padInfo.top,A=d.dataFormat==="channelsLast",b=new Wt(d.outShape,r.dtype),w=v.computeStrides(r.shape),C=v.computeStrides(s.shape),T=w[0],N=A?w[1]:w[2],M=A?w[2]:1,F=A?1:w[1],E=b.strides[0],I=A?b.strides[1]:b.strides[2],_=A?b.strides[2]:1,O=A?1:b.strides[1],B=a.data.get(r.dataId).values,$=a.data.get(s.dataId).values,U=b.values;for(let G=0;G<d.batchSize;++G){let q=G*T,H=G*E;for(let V=0;V<d.outHeight;++V){let Z=H+V*I,X=V*d.strideHeight-x;for(let re=0;re<h;++re){let ee=X+re*m;if(ee<0||ee>=d.inHeight)continue;let me=re*C[0],ie=q+ee*N;for(let Ae=0;Ae<d.outWidth;++Ae){let Ce=Z+Ae*_,Re=Ae*d.strideWidth-y;for(let Le=0;Le<f;++Le){let qe=Re+Le*g;if(qe<0||qe>=d.inWidth)continue;let mt=me+Le*C[1],dt=ie+qe*M,st=mt;for(let it=0;it<d.inChannels;++it){let Ge=B[dt+it*F];for(let yt=0;yt<d.outChannels;++yt)U[Ce+yt*O]+=Ge*$[st+yt];st+=d.outChannels}}}}}}return a.makeTensorInfo(b.shape,b.dtype,U)}var wL={kernelName:fi,backendName:"cpu",kernelFunc:O6};function kL(e){let{inputs:t,backend:a,attrs:n}=e,{x:r,dy:s}=t,{strides:i,pad:o,dataFormat:l,dimRoundingMode:u,filterShape:p}=n;Ie([r,s],"conv2dBackpropFilter");let c=S.convertConv2DDataFormat(l),d=S.computeConv2DInfo(r.shape,p,i,1,o,u,!1,c),{strideHeight:h,strideWidth:f,filterHeight:m,filterWidth:g}=d,y=d.dataFormat==="channelsLast",x=new Wt(d.filterShape,"float32"),A=d.padInfo.left,b=d.padInfo.top,w=a.data.get(r.dataId).values,C=a.data.get(s.dataId).values,T=new Wt(r.shape,r.dtype,w),N=new Wt(s.shape,s.dtype,C);for(let M=0;M<m;++M){let F=Math.max(0,Math.ceil((b-M)/h)),E=Math.min(d.outHeight,(d.inHeight+b-M)/h);for(let I=0;I<g;++I){let _=Math.max(0,Math.ceil((A-I)/f)),O=Math.min(d.outWidth,(d.inWidth+A-I)/f);for(let B=0;B<d.inChannels;++B)for(let $=0;$<d.outChannels;++$){let U=0;for(let G=0;G<d.batchSize;++G)for(let q=F;q<E;++q){let H=M+q*h-b;for(let V=_;V<O;++V){let Z=I+V*f-A;y?U+=T.get(G,H,Z,B)*N.get(G,q,V,$):U+=T.get(G,B,H,Z)*N.get(G,$,q,V)}}x.set(U,M,I,B,$)}}}return a.makeTensorInfo(x.shape,x.dtype,x.values)}var IL={kernelName:np,backendName:"cpu",kernelFunc:kL};function SL(e){let{inputs:t,backend:a,attrs:n}=e,{dy:r,filter:s}=t,{inputShape:i,strides:o,pad:l,dataFormat:u,dimRoundingMode:p}=n;Ie([r,s],"conv2dBackpropInput");let c=v.computeStrides(s.shape),d=v.computeStrides(r.shape),h=S.convertConv2DDataFormat(u),f=S.computeConv2DInfo(i,s.shape,o,1,l,p,!1,h),m=new Wt(f.inShape,"float32"),g=m.values,y=a.data.get(r.dataId).values,x=a.data.get(s.dataId).values,[A,b,w]=c,{batchSize:C,filterHeight:T,filterWidth:N,inChannels:M,inHeight:F,inWidth:E,outChannels:I,outHeight:_,outWidth:O,strideHeight:B,strideWidth:$}=f;h=f.dataFormat;let U=T-1-f.padInfo.top,G=N-1-f.padInfo.left,q=h==="channelsLast",H=m.strides[0],V=q?m.strides[1]:m.strides[2],Z=q?m.strides[2]:1,X=q?1:m.strides[1],re=d[0],ee=q?d[1]:d[2],me=q?d[2]:1,ie=q?1:d[1];for(let Ae=0;Ae<C;++Ae)for(let Ce=0;Ce<M;++Ce)for(let Re=0;Re<F;++Re){let Le=Re-U,qe=Math.max(0,Math.ceil(Le/B)),mt=Math.min(_,(T+Le)/B);for(let dt=0;dt<E;++dt){let st=dt-G,it=Math.max(0,Math.ceil(st/$)),Ge=Math.min(O,(N+st)/$),yt=0;for(let zt=qe;zt<mt;++zt){let un=zt*B-Le;for(let la=it;la<Ge;++la){let Da=la*$-st,dn=re*Ae+ee*zt+me*la,Pa=A*(T-1-un)+b*(N-1-Da)+w*Ce;for(let ht=0;ht<I;++ht){let Oa=y[dn+ie*ht],qa=x[Pa+ht];yt+=Oa*qa}}}let ja=H*Ae+V*Re+Z*dt+X*Ce;g[ja]=yt}}return a.makeTensorInfo(m.shape,m.dtype,m.values)}var CL={kernelName:mi,backendName:"cpu",kernelFunc:SL};function TL(e){let{inputs:t,backend:a,attrs:n}=e,{x:r,filter:s}=t,{strides:i,pad:o,dilations:l}=n;Ie([r,s],"conv3d");let u=S.computeConv3DInfo(r.shape,s.shape,i,l,o),{filterDepth:p,filterHeight:c,filterWidth:d,dilationDepth:h,dilationHeight:f,dilationWidth:m,padInfo:g}=u,y=g.front,x=g.left,A=g.top,b=new Wt(u.outShape,r.dtype),w=a.data.get(r.dataId).values,C=a.data.get(s.dataId).values,T=b.values,N=v.computeStrides(r.shape),M=v.computeStrides(s.shape);for(let F=0;F<u.batchSize;++F){let E=F*N[0],I=F*b.strides[0];for(let _=0;_<u.outDepth;++_){let O=I+_*b.strides[1],B=_*u.strideDepth-y;for(let $=0;$<p;++$){let U=B+$*h;if(U<0||U>=u.inDepth)continue;let G=$*M[0],q=E+U*N[1];for(let H=0;H<u.outHeight;++H){let V=O+H*b.strides[2],Z=H*u.strideHeight-A;for(let X=0;X<c;++X){let re=Z+X*f;if(re<0||re>=u.inHeight)continue;let ee=G+X*M[1],me=q+re*N[2];for(let ie=0;ie<u.outWidth;++ie){let Ae=V+ie*u.outChannels,Ce=ie*u.strideWidth-x;for(let Re=0;Re<d;++Re){let Le=Ce+Re*m;if(Le<0||Le>=u.inWidth)continue;let qe=ee+Re*M[2],mt=me+Le*u.inChannels,dt=qe;for(let st=0;st<u.inChannels;++st){let it=w[mt+st];for(let Ge=0;Ge<u.outChannels;++Ge)T[Ae+Ge]+=it*C[dt+Ge];dt+=u.outChannels}}}}}}}}return a.makeTensorInfo(b.shape,b.dtype,b.values)}var NL={kernelName:gi,backendName:"cpu",kernelFunc:TL};function RL(e){let{inputs:t,backend:a,attrs:n}=e,{x:r,dy:s}=t,{strides:i,pad:o,filterShape:l}=n;Ie([r,s],"conv3dBackpropFilterV2");let u=v.computeStrides(r.shape),p=v.computeStrides(s.shape),c=S.computeConv3DInfo(r.shape,l,i,1,o),d=c.strideDepth,h=c.strideHeight,f=c.strideWidth,m=c.filterDepth,g=c.filterHeight,y=c.filterWidth,x=new Wt(c.filterShape,"float32"),A=x.values,[b,w,C,T]=x.strides,N=a.data.get(s.dataId).values,[M,F,E,I]=p,_=a.data.get(r.dataId).values,[O,B,$,U]=u,G=c.padInfo.front,q=c.padInfo.left,H=c.padInfo.top;for(let V=0;V<m;++V){let Z=Math.max(0,Math.ceil((G-V)/d)),X=Math.min(c.outDepth,(c.inDepth+G-V)/d),re=V*b;for(let ee=0;ee<g;++ee){let me=Math.max(0,Math.ceil((H-ee)/h)),ie=Math.min(c.outHeight,(c.inHeight+H-ee)/h),Ae=ee*w+re;for(let Ce=0;Ce<y;++Ce){let Re=Math.max(0,Math.ceil((q-Ce)/f)),Le=Math.min(c.outWidth,(c.inWidth+q-Ce)/f),qe=Ce*C+Ae;for(let mt=0;mt<c.inChannels;++mt){let dt=mt*T+qe;for(let st=0;st<c.outChannels;++st){let it=0;for(let Ge=0;Ge<c.batchSize;++Ge){let yt=Ge*O,ja=Ge*M;for(let zt=Z;zt<X;++zt){let un=(V+zt*d-G)*B+yt,la=zt*F+ja;for(let Da=me;Da<ie;++Da){let dn=(ee+Da*h-H)*$+un,Pa=Da*E+la;for(let ht=Re;ht<Le;++ht){let Oa=(Ce+ht*f-q)*U+dn,qa=ht*I+Pa;it+=_[Oa+mt]*N[qa+st]}}}}A[dt+st]=it}}}}}return a.makeTensorInfo(x.shape,x.dtype,x.values)}var EL={kernelName:ru,backendName:"cpu",kernelFunc:RL};function ML(e){let{inputs:t,backend:a,attrs:n}=e,{dy:r,filter:s}=t,{pad:i,strides:o,inputShape:l}=n;Ie([r],"conv3dBackpropInputV2");let u=v.computeStrides(r.shape),p=v.computeStrides(s.shape),c=S.computeConv3DInfo(l,s.shape,o,1,i),d=new Wt(c.inShape,"float32"),h=d.values,[f,m,g,y]=d.strides,x=a.data.get(r.dataId).values,[A,b,w,C]=u,T=a.data.get(s.dataId).values,[N,M,F,E]=p,{batchSize:I,filterDepth:_,filterHeight:O,filterWidth:B,inChannels:$,inDepth:U,inHeight:G,inWidth:q,outChannels:H,outDepth:V,outHeight:Z,outWidth:X,strideDepth:re,strideHeight:ee,strideWidth:me}=c,ie=_-1-c.padInfo.front,Ae=O-1-c.padInfo.top,Ce=B-1-c.padInfo.left;for(let Re=0;Re<I;++Re)for(let Le=0;Le<$;++Le)for(let qe=0;qe<U;++qe){let mt=qe-ie,dt=Math.max(0,Math.ceil(mt/re)),st=Math.min(V,(_+mt)/re);for(let it=0;it<G;++it){let Ge=it-Ae,yt=Math.max(0,Math.ceil(Ge/ee)),ja=Math.min(Z,(O+Ge)/ee);for(let zt=0;zt<q;++zt){let un=zt-Ce,la=Math.max(0,Math.ceil(un/me)),Da=Math.min(X,(B+un)/me),dn=0;for(let Pa=dt;Pa<st;++Pa){let ht=Pa*re-mt;for(let Oa=yt;Oa<ja;++Oa){let qa=Oa*ee-Ge;for(let hr=la;hr<Da;++hr){let ml=hr*me-un,jn=A*Re+b*Pa+w*Oa+C*hr,sd=N*(_-1-ht)+M*(O-1-qa)+F*(B-1-ml)+E*Le;for(let wn=0;wn<H;++wn){let Dr=x[jn+wn],Kt=T[sd+wn];dn+=Dr*Kt}}}}h[f*Re+m*qe+g*it+y*zt+Le]=dn}}}return a.makeTensorInfo(d.shape,d.dtype,d.values)}var FL={kernelName:yi,backendName:"cpu",kernelFunc:ML},_L=ct(xi,e=>Math.cos(e)),$L={kernelName:xi,backendName:"cpu",kernelFunc:_L},DL=ct(Ai,e=>Math.cosh(e)),PL={kernelName:Ai,backendName:"cpu",kernelFunc:DL};function OL(e){let{inputs:t,backend:a,attrs:n}=e,{image:r,boxes:s,boxInd:i}=t,{cropSize:o,method:l,extrapolationValue:u}=n,[p,c,d,h]=r.shape,f=s.shape[0],[m,g]=o,y=$e([f,m,g,h],"float32"),x=a.data.get(s.dataId).values,A=a.data.get(i.dataId).values,b=a.data.get(r.dataId).values,w=v.computeStrides(r.shape),C=v.computeStrides(y.shape);for(let T=0;T<f;T++){let N=T*4,M=x[N],F=x[N+1],E=x[N+2],I=x[N+3],_=A[T];if(_>=p)continue;let O=m>1?(E-M)*(c-1)/(m-1):0,B=g>1?(I-F)*(d-1)/(g-1):0;for(let $=0;$<m;$++){let U=m>1?M*(c-1)+$*O:.5*(M+E)*(c-1);if(U<0||U>c-1){for(let G=0;G<g;G++)for(let q=0;q<h;q++){let H=q+G*C[2]+$*C[1]+T*C[0];y.values[H]=u}continue}if(l==="bilinear"){let G=Math.floor(U),q=Math.ceil(U),H=U-G;for(let V=0;V<g;V++){let Z=g>1?F*(d-1)+V*B:.5*(F+I)*(d-1);if(Z<0||Z>d-1){for(let me=0;me<h;me++){let ie=me+V*C[2]+$*C[1]+T*C[0];y.values[ie]=u}continue}let X=Math.floor(Z),re=Math.ceil(Z),ee=Z-X;for(let me=0;me<h;me++){let ie=me+X*w[2]+G*w[1]+_*w[0],Ae=b[ie];ie=me+re*w[2]+G*w[1]+_*w[0];let Ce=b[ie];ie=me+X*w[2]+q*w[1]+_*w[0];let Re=b[ie];ie=me+re*w[2]+q*w[1]+_*w[0];let Le=b[ie],qe=Ae+(Ce-Ae)*ee,mt=Re+(Le-Re)*ee;ie=me+V*C[2]+$*C[1]+T*C[0],y.values[ie]=qe+(mt-qe)*H}}}else for(let G=0;G<g;++G){let q=g>1?F*(d-1)+G*B:.5*(F+I)*(d-1);if(q<0||q>d-1){for(let Z=0;Z<h;Z++){let X=Z+G*C[2]+$*C[1]+T*C[0];y.values[X]=u}continue}let H=Math.round(q),V=Math.round(U);for(let Z=0;Z<h;Z++){let X=Z+H*w[2]+V*w[1]+_*w[0],re=Z+G*C[2]+$*C[1]+T*C[0];y.values[re]=b[X]}}}}return a.makeTensorInfo(y.shape,y.dtype,y.values)}var zL={kernelName:wi,backendName:"cpu",kernelFunc:OL};function LL(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{axis:s,exclusive:i,reverse:o}=n;Ie(r,"cumprod");let l=S.getAxesPermutation([s],r.shape.length),u=r;l!=null&&(u=Ua({inputs:{x:r},backend:a,attrs:{perm:l}}));let p=S.getInnerMostAxes(1,r.shape.length)[0];if(p!==u.shape.length-1)throw new Error(`backend.cumprod in CPU expects an inner-most axis=${u.shape.length-1} but got axis=${p}`);let c=pa(u.dtype,"int32"),d=v.makeOnesTypedArray(v.sizeFromShape(u.shape),c),h=a.data.get(u.dataId).values,f=u.shape[u.shape.length-1],m=o?(y,x)=>y+f-x-1:(y,x)=>y+x;for(let y=0;y<h.length;y+=f)for(let x=0;x<f;x++){let A=m(y,x);if(x===0)d[A]=i?1:h[A];else{let b=m(y,x-1);d[A]=i?h[b]*d[b]:h[A]*d[b]}}let g=a.makeTensorInfo(u.shape,c,d);if(l!=null){let y=S.getUndoAxesPermutation(l),x=Ua({inputs:{x:g},backend:a,attrs:{perm:y}});return a.disposeIntermediateTensorInfo(g),a.disposeIntermediateTensorInfo(u),x}return g}var WL={kernelName:bi,backendName:"cpu",kernelFunc:LL};function BL(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{axis:s,exclusive:i,reverse:o}=n;Ie(r,"cumsum");let l=S.getAxesPermutation([s],r.shape.length),u=r;l!=null&&(u=Ua({inputs:{x:r},backend:a,attrs:{perm:l}}));let p=S.getInnerMostAxes(1,r.shape.length)[0];if(p!==u.shape.length-1)throw new Error(`backend.cumsum in CPU expects an inner-most axis=${u.shape.length-1} but got axis=${p}`);let c=pa(u.dtype,"int32"),d=v.makeZerosTypedArray(v.sizeFromShape(u.shape),c),h=a.data.get(u.dataId).values,f=u.shape[u.shape.length-1],m=o?(y,x)=>y+f-x-1:(y,x)=>y+x;for(let y=0;y<h.length;y+=f)for(let x=0;x<f;x++){let A=m(y,x);if(x===0)d[A]=i?0:h[A];else{let b=m(y,x-1);d[A]=i?h[b]+d[b]:h[A]+d[b]}}let g=a.makeTensorInfo(u.shape,c,d);if(l!=null){let y=S.getUndoAxesPermutation(l),x=Ua({inputs:{x:g},backend:a,attrs:{perm:y}});return a.disposeIntermediateTensorInfo(g),a.disposeIntermediateTensorInfo(u),x}return g}var VL={kernelName:vi,backendName:"cpu",kernelFunc:BL};function UL(e){let{inputs:t,backend:a,attrs:n}=e,{x:r,weights:s}=t,{size:i,binaryOutput:o}=n;if(r.shape.length===1){let l=a.data.get(r.dataId).values,u=a.data.get(s.dataId).values,p=J3(l,u,s.dtype,s.shape,i);return a.makeTensorInfo([i],s.dtype,p)}else if(r.shape.length===2){let l=a.bufferSync(r),u=a.bufferSync(s),p=H7(l,u,i,o);return a.makeTensorInfo(p.shape,s.dtype,p.values)}throw new Error(`Error in denseBincount: input must be at most rank 2, but got rank${r.shape.length}.`)}var GL={kernelName:su,backendName:"cpu",kernelFunc:UL};function HL(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{blockSize:s,dataFormat:i}=n;v.assert(i==="NHWC",()=>`Only NHWC dataFormat supported on CPU for depthToSpace. Got ${i}`);let o=r.shape[0],l=r.shape[1],u=r.shape[2],p=r.shape[3],c=l*s,d=u*s,h=p/(s*s),f=a.data.get(r.dataId).values,m=new Float32Array(o*c*d*h),g=0;for(let y=0;y<o;++y)for(let x=0;x<c;++x){let A=Math.floor(x/s),b=x%s;for(let w=0;w<d;++w){let C=Math.floor(w/s),T=w%s,N=(b*s+T)*h;for(let M=0;M<h;++M){let F=M+N+p*(C+u*(A+l*y));m[g++]=f[F]}}}return a.makeTensorInfo([o,c,d,h],r.dtype,m)}var jL={kernelName:ki,backendName:"cpu",kernelFunc:HL};function z6(e){let{inputs:t,backend:a,attrs:n}=e,{x:r,filter:s}=t,{strides:i,pad:o,dilations:l,dimRoundingMode:u}=n;Ie([r,s],"depthwiseConv2DNative");let p=v.computeStrides(r.shape),c=v.computeStrides(s.shape),d=l;d==null&&(d=[1,1]),v.assert(S.eitherStridesOrDilationsAreOne(i,d),()=>`Error in depthwiseConv2d: Either strides or dilations must be 1. Got strides ${i} and dilations '${d}'`);let h=S.computeConv2DInfo(r.shape,s.shape,i,d,o,u,!0),{filterHeight:f,filterWidth:m,dilationHeight:g,dilationWidth:y,padInfo:x}=h,A=x.left,b=x.top,w=h.outChannels/h.inChannels,C=new Wt(h.outShape,r.dtype),T=a.data.get(r.dataId).values,N=a.data.get(s.dataId).values,M=C.values;for(let F=0;F<h.batchSize;++F){let E=F*p[0],I=F*C.strides[0];for(let _=0;_<h.outHeight;++_){let O=I+_*C.strides[1],B=_*h.strideHeight-b;for(let $=0;$<f;++$){let U=B+$*g;if(U<0||U>=h.inHeight)continue;let G=$*c[0],q=E+U*p[1];for(let H=0;H<h.outWidth;++H){let V=O+H*C.strides[2],Z=H*h.strideWidth-A;for(let X=0;X<m;++X){let re=Z+X*y;if(re<0||re>=h.inWidth)continue;let ee=G+X*c[1],me=q+re*h.inChannels,ie=V,Ae=ee;for(let Ce=0;Ce<h.inChannels;++Ce){let Re=T[me+Ce];for(let Le=0;Le<w;++Le)M[ie+Le]+=Re*N[Ae+Le];ie+=w,Ae+=w}}}}}}return a.makeTensorInfo(C.shape,C.dtype,C.values)}var qL={kernelName:Ii,backendName:"cpu",kernelFunc:z6};function XL(e){let{inputs:t,backend:a,attrs:n}=e,{x:r,dy:s}=t,{strides:i,dilations:o,pad:l,dimRoundingMode:u,filterShape:p}=n;Ie([r,s],"depthwiseConv2dNativeBackpropFilter");let c=S.computeConv2DInfo(r.shape,p,i,o,l,u,!0),{strideHeight:d,strideWidth:h,filterHeight:f,filterWidth:m}=c,g=new Wt(c.filterShape,"float32"),y=c.padInfo.left,x=c.padInfo.top,A=c.outChannels/c.inChannels,b=a.data.get(r.dataId).values,w=new Wt(r.shape,r.dtype,b),C=a.data.get(s.dataId).values,T=new Wt(s.shape,s.dtype,C);for(let N=0;N<f;++N){let M=Math.max(0,Math.ceil((x-N)/d)),F=Math.min(c.outHeight,(c.inHeight+x-N)/d);for(let E=0;E<m;++E){let I=Math.max(0,Math.ceil((y-E)/h)),_=Math.min(c.outWidth,(c.inWidth+y-E)/h);for(let O=0;O<c.outChannels;++O){let B=Math.trunc(O/A),$=O%A,U=0;for(let G=0;G<c.batchSize;++G)for(let q=M;q<F;++q){let H=N+q*d-x;for(let V=I;V<_;++V){let Z=E+V*h-y;U+=w.get(G,H,Z,B)*T.get(G,q,V,O)}}g.set(U,N,E,B,$)}}}return a.makeTensorInfo(g.shape,g.dtype,g.values)}var KL={kernelName:rp,backendName:"cpu",kernelFunc:XL};function YL(e){let{inputs:t,backend:a,attrs:n}=e,{dy:r,filter:s}=t,{strides:i,dilations:o,pad:l,dimRoundingMode:u,inputShape:p}=n;Ie([r,s],"depthwiseConv2DNativeBackpropInput");let c=v.computeStrides(r.shape),d=v.computeStrides(s.shape),h=S.computeConv2DInfo(p,s.shape,i,o,l,u,!0),f=new Wt(h.inShape,"float32"),m=f.values,[g,y,x]=f.strides,A=a.data.get(r.dataId).values,[b,w,C]=c,T=a.data.get(s.dataId).values,[N,M,F]=d,{batchSize:E,filterHeight:I,filterWidth:_,inChannels:O,inHeight:B,inWidth:$,outChannels:U,outHeight:G,outWidth:q,strideHeight:H,strideWidth:V}=h,Z=I-1-h.padInfo.top,X=_-1-h.padInfo.left,re=U/O;for(let ee=0;ee<E;++ee)for(let me=0;me<O;++me)for(let ie=0;ie<B;++ie){let Ae=ie-Z,Ce=Math.max(0,Math.ceil(Ae/H)),Re=Math.min(G,(I+Ae)/H);for(let Le=0;Le<$;++Le){let qe=Le-X,mt=Math.max(0,Math.ceil(qe/V)),dt=Math.min(q,(_+qe)/V),st=0;for(let it=Ce;it<Re;++it){let Ge=it*H-Ae;for(let yt=mt;yt<dt;++yt){let ja=yt*V-qe,zt=b*ee+w*it+C*yt,un=N*(I-1-Ge)+M*(_-1-ja)+F*me;for(let la=0;la<re;++la){let Da=me*re+la,dn=A[zt+Da],Pa=T[un+la];st+=dn*Pa}}}m[g*ee+y*ie+x*Le+me]=st}}return a.makeTensorInfo(f.shape,f.dtype,f.values)}var ZL={kernelName:sp,backendName:"cpu",kernelFunc:YL};function JL(e){let{inputs:t,backend:a}=e,{x:n}=t,r=v.sizeFromShape(n.shape),s=a.data.get(n.dataId).values,i=$e([r,r],n.dtype),o=i.values;for(let u=0;u<s.length;u++)o[u*r+u]=s[u];let l=[...n.shape,...n.shape];return a.makeTensorInfo(l,i.dtype,i.values)}var QL={kernelName:iu,backendName:"cpu",kernelFunc:JL},eW={kernelName:Si,backendName:"cpu",kernelFunc:({inputs:e,backend:t,attrs:a})=>{let{x:n,filter:r}=e,{strides:s,pad:i,dilations:o}=a,l=t,u=l.data.get(n.dataId).values,p=n.shape.length,c=l.data.get(r.dataId).values,d=r.shape.length,{batchSize:h,inHeight:f,inWidth:m,inChannels:g,outHeight:y,outWidth:x,padInfo:A,strideHeight:b,strideWidth:w,filterHeight:C,filterWidth:T,dilationHeight:N,dilationWidth:M,outShape:F}=S.computeDilation2DInfo(n.shape,r.shape,s,i,"NHWC",o),E=v.sizeFromShape(F),I=F.length,_=v.getArrayFromDType(n.dtype,E);for(let O=0;O<h;++O)for(let B=0;B<y;++B){let $=B*b-A.top;for(let U=0;U<x;++U){let G=U*w-A.left;for(let q=0;q<g;++q){let H=Number.MIN_SAFE_INTEGER;for(let Z=0;Z<C;++Z){let X=$+Z*N;if(X>=0&&X<f)for(let re=0;re<T;++re){let ee=G+re*M;if(ee>=0&&ee<m){let me=v.locToIndex([O,X,ee,q],p,v.computeStrides(n.shape)),ie=v.locToIndex([Z,re,q],d,v.computeStrides(r.shape)),Ae=u[me]+c[ie];Ae>H&&(H=Ae)}}}let V=v.locToIndex([O,B,U,q],I,v.computeStrides(F));_[V]=H}}}return{dataId:l.write(v.toTypedArray(_,n.dtype),F,n.dtype),shape:F,dtype:n.dtype}}},tW={kernelName:Pl,backendName:"cpu",kernelFunc:({inputs:e,backend:t,attrs:a})=>{let{x:n,filter:r,dy:s}=e,{strides:i,pad:o,dilations:l}=a,u=t,p=v.toNestedArray(n.shape,u.data.get(n.dataId).values),c=v.toNestedArray(r.shape,u.data.get(r.dataId).values),{batchSize:d,inHeight:h,inWidth:f,inChannels:m,outHeight:g,outWidth:y,padInfo:x,strideHeight:A,strideWidth:b,filterHeight:w,filterWidth:C,dilationHeight:T,dilationWidth:N,outShape:M}=S.computeDilation2DInfo(n.shape,r.shape,i,o,"NHWC",l);v.assert(s.rank===M.length,()=>`Error in ${Pl}, dy must have the same rank as output ${M.length}, but got ${s.rank}`);let F=v.toNestedArray(M,u.data.get(s.dataId).values),E=v.makeZerosNestedTypedArray(r.shape,r.dtype);for(let I=0;I<d;++I)for(let _=0;_<g;++_){let O=_*A-x.top;for(let B=0;B<y;++B){let $=B*b-x.left;for(let U=0;U<m;++U){let G=Number.MIN_SAFE_INTEGER,q=0,H=0;for(let V=0;V<w;++V){let Z=O+V*T;if(Z>=0&&Z<h)for(let X=0;X<C;++X){let re=$+X*N;if(re>=0&&re<f){let ee=p[I][Z][re][U]+c[V][X][U];ee>G&&(G=ee,q=V,H=X)}}}E[q][H][U]+=F[I][_][B][U]}}}return{dataId:u.write(v.toTypedArray(E,n.dtype),r.shape,r.dtype),shape:r.shape,dtype:r.dtype}}},aW={kernelName:Dl,backendName:"cpu",kernelFunc:({inputs:e,backend:t,attrs:a})=>{let{x:n,filter:r,dy:s}=e,{strides:i,pad:o,dilations:l}=a,u=t,p=v.toNestedArray(n.shape,u.data.get(n.dataId).values),c=v.toNestedArray(r.shape,u.data.get(r.dataId).values),{batchSize:d,inHeight:h,inWidth:f,inChannels:m,outHeight:g,outWidth:y,padInfo:x,strideHeight:A,strideWidth:b,filterHeight:w,filterWidth:C,dilationHeight:T,dilationWidth:N,outShape:M}=S.computeDilation2DInfo(n.shape,r.shape,i,o,"NHWC",l);v.assert(s.rank===M.length,()=>`Error in ${Dl}, dy must have the same rank as output ${M.length}, but got ${s.rank}`);let F=v.toNestedArray(M,u.data.get(s.dataId).values),E=v.makeZerosNestedTypedArray(n.shape,n.dtype);for(let I=0;I<d;++I)for(let _=0;_<g;++_){let O=_*A-x.top;for(let B=0;B<y;++B){let $=B*b-x.left;for(let U=0;U<m;++U){let G=Number.MIN_SAFE_INTEGER,q=O<0?0:O,H=$<0?0:$;for(let V=0;V<w;++V){let Z=O+V*T;if(Z>=0&&Z<h)for(let X=0;X<C;++X){let re=$+X*N;if(re>=0&&re<f){let ee=p[I][Z][re][U]+c[V][X][U];ee>G&&(G=ee,q=Z,H=re)}}}E[I][q][H][U]+=F[I][_][B][U]}}}return{dataId:u.write(v.toTypedArray(E,n.dtype),n.shape,n.dtype),shape:n.shape,dtype:n.dtype}}};function Wp(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{axis:s,keepDims:i}=n;Ie(r,"sum");let o;r.dtype==="bool"?o=ts({inputs:{x:r},backend:a,attrs:{dtype:"int32"}}):o=ar({inputs:{x:r},backend:a});let l=o.shape.length,u=v.parseAxisParam(s,o.shape),p=S.getAxesPermutation(u,l),c=u,d=o;p!=null&&(d=Ua({inputs:{x:o},backend:a,attrs:{perm:p}}),c=S.getInnerMostAxes(c.length,l)),S.assertAxesAreInnerMostDims("sum",c,d.shape.length);let[h,f]=S.computeOutAndReduceShapes(d.shape,c),m=S.upcastType(d.dtype,"int32"),g=uh(a,h,m),y=v.sizeFromShape(f),x=a.data.get(g.dataId).values,A=a.data.get(d.dataId).values;for(let b=0;b<x.length;++b){let w=b*y,C=0;for(let T=0;T<y;++T)C+=A[w+T];x[b]=C}if(i){let b=S.expandShapeToKeepDim(g.shape,u),w=g;g=At({inputs:{x:g},backend:a,attrs:{shape:b}}),a.disposeIntermediateTensorInfo(w)}return a.disposeIntermediateTensorInfo(o),p!=null&&a.disposeIntermediateTensorInfo(d),g}var nW={kernelName:_o,backendName:"cpu",kernelFunc:Wp};function rW(e){let{inputs:t,backend:a,attrs:n}=e,{equation:r}=n,s=t,{allDims:i,summedDims:o,idDims:l}=S.decodeEinsumEquation(r,s.length);S.checkEinsumDimSizes(i.length,l,s);let{path:u,steps:p}=S.getEinsumComputePath(o,l),c=p.length,d=null,h=i.length,f=[];for(let m=0;m<c;++m){for(let g of p[m]){let{permutationIndices:y,expandDims:x}=S.getEinsumPermutation(h,l[g]),A;S.isIdentityPermutation(y)?A=s[g]:(A=Ua({inputs:{x:s[g]},backend:a,attrs:{perm:y}}),f.push(A));let b=A.shape.slice();for(let w=0;w<x.length;++w)b.splice(x[w],0,1);v.arraysEqual(A.shape,b)||(A=At({inputs:{x:A},backend:a,attrs:{shape:b}}),f.push(A)),d===null?d=A:(d=Yh({inputs:{a:A,b:d},backend:a}),f.push(d))}m<c-1&&(u[m]>=0&&(d=Wp({inputs:{x:d},backend:a,attrs:{axis:u[m]-(i.length-h),keepDims:!1}}),f.push(d)),h--)}for(let m of f)m!==d&&a.disposeIntermediateTensorInfo(m);return d}var sW={kernelName:ip,backendName:"cpu",kernelFunc:rW};function iW(e){let{inputs:t,backend:a}=e,{dy:n,y:r}=t;Ie([n,r],"eluGrad");let s=new Float32Array(v.sizeFromShape(r.shape)),i=a.data.get(r.dataId).values,o=a.data.get(n.dataId).values;for(let l=0;l<i.length;++l){let u=i[l];u>=0?s[l]=o[l]:s[l]=o[l]*(u+1)}return a.makeTensorInfo(r.shape,"float32",s)}var oW={kernelName:ou,backendName:"cpu",kernelFunc:iW},lW=S.ERF_P,uW=S.ERF_A1,dW=S.ERF_A2,pW=S.ERF_A3,cW=S.ERF_A4,hW=S.ERF_A5,fW=ct(lu,e=>{let t=Math.sign(e),a=Math.abs(e),n=1/(1+lW*a);return t*(1-((((hW*n+cW)*n+pW)*n+dW)*n+uW)*n*Math.exp(-a*a))}),mW={kernelName:lu,backendName:"cpu",kernelFunc:fW};function ch(e){let{inputs:t,backend:a,attrs:n}=e,{input:r}=t,{dim:s}=n,i=r.shape.length,o=r.shape.slice(),l=s;return s<0&&(v.assert(-(i+1)<=s,()=>`Axis must be in the interval [${-(i+1)}, ${i}]`),l=i+s+1),o.splice(l,0,1),At({inputs:{x:r},backend:a,attrs:{shape:o}})}var gW={kernelName:uu,backendName:"cpu",kernelFunc:ch},yW=$t((e,t)=>e/t),dg=Xt(Ci,yW),A1={kernelName:Ci,backendName:"cpu",kernelFunc:dg};function L6(e,t,a){let n=e.shape,r=n[0],s=n[1],i=a.data.get(e.dataId),o=i.complexTensorInfos.real,l=i.complexTensorInfos.imag,u=[r,s],p=v.sizeFromShape(u),c=v.getTypedArrayFromDType("float32",p),d=v.getTypedArrayFromDType("float32",p);for(let g=0;g<r;g++){let y=Xs({inputs:{x:o},backend:a,attrs:{begin:[g,0],size:[1,s]}}),x=Xs({inputs:{x:l},backend:a,attrs:{begin:[g,0],size:[1,s]}}),A=Ja({inputs:{real:y,imag:x},backend:a}),{real:b,imag:w}=xW(A,t,a),C=S.mergeRealAndImagArrays(b,w);for(let T=0;T<s;T++){let N=S.getComplexWithIndex(C,T);c[g*s+T]=N.real,d[g*s+T]=N.imag}a.disposeIntermediateTensorInfo(y),a.disposeIntermediateTensorInfo(x),a.disposeIntermediateTensorInfo(A)}let h=a.makeTensorInfo(u,"float32",c),f=a.makeTensorInfo(u,"float32",d),m=Ja({inputs:{real:h,imag:f},backend:a});return a.disposeIntermediateTensorInfo(h),a.disposeIntermediateTensorInfo(f),m}function xW(e,t,a){let n=v.sizeFromShape(e.shape),r=a.data.get(e.dataId),s=a.data.get(r.complexTensorInfos.real.dataId).values,i=a.data.get(r.complexTensorInfos.imag.dataId).values;if(AW(n)){let o=b1(s,i,n,t,a),l=[e.shape[0],e.shape[1]];if(t){let u=a.makeTensorInfo(l,"float32",o.real),p=a.makeTensorInfo(l,"float32",o.imag),c=a.makeTensorInfo([],"float32",v.createScalarValue(n,"float32")),d=ar({inputs:{x:c},backend:a}),h=A1.kernelFunc({inputs:{a:u,b:c},backend:a}),f=A1.kernelFunc({inputs:{a:p,b:d},backend:a}),m=a.data.get(h.dataId).values,g=a.data.get(f.dataId).values;return a.disposeIntermediateTensorInfo(u),a.disposeIntermediateTensorInfo(p),a.disposeIntermediateTensorInfo(c),a.disposeIntermediateTensorInfo(d),a.disposeIntermediateTensorInfo(h),a.disposeIntermediateTensorInfo(f),{real:m,imag:g}}return o}else{let o=S.mergeRealAndImagArrays(s,i),l=bW(o,n,t);return S.splitRealAndImagArrays(l)}}function AW(e){return(e&e-1)===0}function b1(e,t,a,n,r){if(a===1)return{real:e,imag:t};let s=S.mergeRealAndImagArrays(e,t),i=a/2,o=S.complexWithEvenIndex(s),l=o.real,u=o.imag,p=[l.length],c=r.makeTensorInfo(p,"float32",l),d=r.makeTensorInfo(p,"float32",u),h=Ja({inputs:{real:c,imag:d},backend:r}),f=S.complexWithOddIndex(s),m=f.real,g=f.imag,y=[m.length],x=r.makeTensorInfo(y,"float32",m),A=r.makeTensorInfo(y,"float32",g),b=Ja({inputs:{real:x,imag:A},backend:r}),w=b1(l,u,i,n,r),C=w.real,T=w.imag,N=[C.length],M=r.makeTensorInfo(N,"float32",C),F=r.makeTensorInfo(N,"float32",T),E=Ja({inputs:{real:M,imag:F},backend:r}),I=b1(m,g,i,n,r),_=I.real,O=I.imag,B=[_.length],$=r.makeTensorInfo(B,"float32",_),U=r.makeTensorInfo(B,"float32",O),G=Ja({inputs:{real:$,imag:U},backend:r}),q=S.exponents(a,n),H=[q.real.length],V=r.makeTensorInfo(H,"float32",q.real),Z=r.makeTensorInfo(H,"float32",q.imag),X=Ja({inputs:{real:V,imag:Z},backend:r}),re=Yh({inputs:{a:X,b:G},backend:r}),ee=Gl({inputs:{a:E,b:re},backend:r}),me=og({inputs:{a:E,b:re},backend:r}),ie=qs({inputs:{input:ee},backend:r}),Ae=qs({inputs:{input:me},backend:r}),Ce=Hl({inputs:{input:ee},backend:r}),Re=Hl({inputs:{input:me},backend:r}),Le=jl({inputs:[ie,Ae],backend:r,attrs:{axis:0}}),qe=jl({inputs:[Ce,Re],backend:r,attrs:{axis:0}}),mt=r.data.get(Le.dataId).values,dt=r.data.get(qe.dataId).values;return r.disposeIntermediateTensorInfo(c),r.disposeIntermediateTensorInfo(d),r.disposeIntermediateTensorInfo(h),r.disposeIntermediateTensorInfo(x),r.disposeIntermediateTensorInfo(A),r.disposeIntermediateTensorInfo(b),r.disposeIntermediateTensorInfo(M),r.disposeIntermediateTensorInfo(F),r.disposeIntermediateTensorInfo(E),r.disposeIntermediateTensorInfo($),r.disposeIntermediateTensorInfo(U),r.disposeIntermediateTensorInfo(G),r.disposeIntermediateTensorInfo(V),r.disposeIntermediateTensorInfo(Z),r.disposeIntermediateTensorInfo(X),r.disposeIntermediateTensorInfo(re),r.disposeIntermediateTensorInfo(ee),r.disposeIntermediateTensorInfo(me),r.disposeIntermediateTensorInfo(ie),r.disposeIntermediateTensorInfo(Ce),r.disposeIntermediateTensorInfo(Ae),r.disposeIntermediateTensorInfo(Re),r.disposeIntermediateTensorInfo(Le),r.disposeIntermediateTensorInfo(qe),{real:mt,imag:dt}}function bW(e,t,a){let n=new Float32Array(t*2);for(let r=0;r<t;r++){let s=0,i=0;for(let o=0;o<t;o++){let l=S.exponent(r*o,t,a),u=S.getComplexWithIndex(e,o);s+=u.real*l.real-u.imag*l.imag,i+=u.real*l.imag+u.imag*l.real}a&&(s/=t,i/=t),S.assignToTypedArray(n,s,i,r)}return n}function vW(e){let{inputs:t,backend:a}=e,{input:n}=t,r=v.sizeFromShape(n.shape),s=n.shape[n.shape.length-1],i=r/s,o=At({inputs:{x:n},backend:a,attrs:{shape:[i,s]}}),l=L6(o,!1,a),u=At({inputs:{x:l},backend:a,attrs:{shape:n.shape}});return a.disposeIntermediateTensorInfo(o),a.disposeIntermediateTensorInfo(l),u}var wW={kernelName:op,backendName:"cpu",kernelFunc:vW};function pg(e){let{backend:t,attrs:a}=e,{shape:n,value:r,dtype:s}=a,i=s||v.inferDtype(r),o=v.getArrayFromDType(i,v.sizeFromShape(n));return IW(o,r,i),t.makeTensorInfo(n,i,o)}var kW={kernelName:du,backendName:"cpu",kernelFunc:pg};function IW(e,t,a){e.fill(t)}var SW={kernelName:Mi,backendName:"cpu",kernelFunc:({inputs:e,attrs:t,backend:a})=>{let{image:n}=e,r=a,s=v.getTypedArrayFromDType(n.dtype,v.sizeFromShape(n.shape)),[i,o,l,u]=n.shape,p=r.data.get(n.dataId).values;for(let c=0;c<i;c++){let d=c*l*o*u;for(let h=0;h<o;h++){let f=h*(l*u);for(let m=0;m<l;m++){let g=m*u;for(let y=0;y<u;y++){let x=Math.round(l-m-1),A=d+f+g+y,b=p[A];if(x>=0&&x<l){let w=x*u,C=d+f+w+y;b=p[C]}s[A]=b}}}}return{dataId:r.write(s,n.shape,n.dtype),shape:n.shape,dtype:n.dtype}}};function CW(e){let{inputs:t,backend:a,attrs:n}=e,{x:r,filter:s,bias:i,preluActivationWeights:o}=t,{strides:l,pad:u,dataFormat:p,dilations:c,dimRoundingMode:d,activation:h,leakyreluAlpha:f}=n,m=O6({inputs:{x:r,filter:s},backend:a,attrs:{strides:l,pad:u,dataFormat:p,dilations:c,dimRoundingMode:d}});if(i){let g=m;if(p==="NCHW"&&i.shape.length===1&&i.shape[0]!==1){let y=At({inputs:{x:i},backend:a,attrs:{shape:[i.shape[0],1,1]}});m=Gl({inputs:{a:m,b:y},backend:a}),a.disposeIntermediateTensorInfo(y)}else m=Gl({inputs:{a:m,b:i},backend:a});a.disposeIntermediateTensorInfo(g)}if(h){let g=m;if(p==="NCHW"&&h==="prelu"&&o.shape.length===1&&o.shape[0]!==1){let y=At({inputs:{x:o},backend:a,attrs:{shape:[o.shape[0],1,1]}});m=ph(a,m,h,y,f),a.disposeIntermediateTensorInfo(y)}else m=ph(a,m,h,o,f);a.disposeIntermediateTensorInfo(g)}return m}var TW={kernelName:Kr,backendName:"cpu",kernelFunc:CW};function NW(e){let{inputs:t,backend:a,attrs:n}=e,{x:r,filter:s,bias:i,preluActivationWeights:o}=t,{strides:l,pad:u,dataFormat:p,dilations:c,dimRoundingMode:d,activation:h,leakyreluAlpha:f}=n,m=z6({inputs:{x:r,filter:s},backend:a,attrs:{strides:l,pad:u,dataFormat:p,dilations:c,dimRoundingMode:d}});if(i){let g=m;m=Gl({inputs:{a:m,b:i},backend:a}),a.disposeIntermediateTensorInfo(g)}if(h){let g=m;m=ph(a,m,h,o,f),a.disposeIntermediateTensorInfo(g)}return m}var RW={kernelName:Yr,backendName:"cpu",kernelFunc:NW};function EW(e){let{inputs:t,backend:a}=e,{params:n,indices:r}=t,s=v.sizeFromShape(n.shape),i=r.shape,o=i[i.length-1],[l,u,p,c]=S.prepareAndValidate(n,r);if(u===0)return a.makeTensorInfo(l,n.dtype,[]);let d=a.data.get(r.dataId).values,h=a.bufferSync(n),f=t6(d,h,n.dtype,u,o,p,c,n.shape,s);return a.makeTensorInfo(l,n.dtype,f.values)}var MW={kernelName:Di,backendName:"cpu",kernelFunc:EW};function FW(e){let{inputs:t,backend:a,attrs:n}=e,{x:r,indices:s}=t,{axis:i,batchDims:o}=n;Ie([r,s],"gatherV2");let l=v.parseAxisParam(i,r.shape)[0],u=a.data.get(s.dataId).values,p=r.shape[l];for(let b=0;b<u.length;++b){let w=u[b];v.assert(w<=p-1&&w>=0,()=>`GatherV2: the index value ${w} is not in [0, ${p-1}]`)}let c=o;o==null&&(c=0);let d=v.sizeFromShape(s.shape),h=S.segment_util.collectGatherOpShapeInfo(r,s,l,c),f=At({inputs:{x:r},backend:a,attrs:{shape:[h.batchSize,h.outerSize,h.dimSize,h.sliceSize]}}),m=At({inputs:{x:s},backend:a,attrs:{shape:[h.batchSize,d/h.batchSize]}}),g=[h.batchSize,h.outerSize,d/h.batchSize,h.sliceSize],y=a.bufferSync(m),x=a.bufferSync(f),A=a6(x,y,g);return a.disposeIntermediateTensorInfo(f),a.disposeIntermediateTensorInfo(m),a.makeTensorInfo(h.outputShape,A.dtype,A.values)}var _W={kernelName:pu,backendName:"cpu",kernelFunc:FW};function $W(e){let{inputs:t,backend:a}=e,{input:n}=t,r=v.sizeFromShape(n.shape),s=n.shape[n.shape.length-1],i=r/s,o=At({inputs:{x:n},backend:a,attrs:{shape:[i,s]}}),l=L6(o,!0,a),u=At({inputs:{x:l},backend:a,attrs:{shape:n.shape}});return a.disposeIntermediateTensorInfo(o),a.disposeIntermediateTensorInfo(l),u}var DW={kernelName:lp,backendName:"cpu",kernelFunc:$W},PW=ct(Li,e=>Number.isFinite(e)?1:0,"bool"),OW={kernelName:Li,backendName:"cpu",kernelFunc:PW},zW=ct(Wi,e=>Math.abs(e)===1/0?1:0,"bool"),LW={kernelName:Wi,backendName:"cpu",kernelFunc:zW},WW=ct(Bi,e=>Number.isNaN(e)?1:0,"bool"),BW={kernelName:Bi,backendName:"cpu",kernelFunc:WW};function VW(e){let{backend:t,attrs:a}=e,{start:n,stop:r,num:s}=a,i=o6(n,r,s);return t.makeTensorInfo([i.length],"float32",i)}var UW={kernelName:Hi,backendName:"cpu",kernelFunc:VW},GW=ct(qi,e=>Math.log1p(e)),HW={kernelName:qi,backendName:"cpu",kernelFunc:GW},jW=$t((e,t)=>e&&t),qW=Xt(Xi,jW,null,"bool"),XW={kernelName:Xi,backendName:"cpu",kernelFunc:qW},KW=ct(Ki,e=>e?0:1,"bool"),YW={kernelName:Ki,backendName:"cpu",kernelFunc:KW},ZW=$t((e,t)=>e||t),JW=Xt(Yi,ZW,null,"bool"),QW={kernelName:Yi,backendName:"cpu",kernelFunc:JW};function eB(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{depthRadius:s,bias:i,alpha:o,beta:l}=n;Ie(r,"LRN");let u=r.shape[3],p=u-1,c=a.data.get(r.dataId).values,d=v.sizeFromShape(r.shape),h=new Float32Array(d);function f(m){let g=m%u,y=m-g+Math.max(0,g-s),x=m-g+Math.min(g+s,p),A=0;for(;y<=x;y++){let b=c[y];A+=b*b}return A}for(let m=0;m<d;m++){let g=f(m),y=c[m]*Math.pow(i+o*g,-l);h[m]=y}return a.makeTensorInfo(r.shape,r.dtype,h)}var tB={kernelName:Zi,backendName:"cpu",kernelFunc:eB};function aB(e){let{inputs:t,backend:a,attrs:n}=e,{x:r,y:s,dy:i}=t,{depthRadius:o,bias:l,alpha:u,beta:p}=n;Ie(i,"LRNGrad");let c=v.sizeFromShape(i.shape),d=i.shape[3],h=a.data.get(i.dataId).values,f=a.data.get(r.dataId).values,m=a.data.get(s.dataId).values,g=new Float32Array(c),y=c;for(let x=0;x<y;x++){let A=x%d,b=x-A+Math.max(0,A-o),w=x-A+Math.min(d,A+o+1),C=0;for(let T=b;T<w;T++)C+=Math.pow(f[T],2);C=u*C+l;for(let T=b;T<w;T++){let N=-2*u*p*f[T]*m[x]/C;x===T&&(N+=Math.pow(C,-p)),N*=h[x],g[T]+=N}}return a.makeTensorInfo(i.shape,r.dtype,g)}var nB={kernelName:cu,backendName:"cpu",kernelFunc:aB};function W6(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{reductionIndices:s,keepDims:i}=n,o=a,l=r.shape,u=l.length,p=v.parseAxisParam(s,l),c=p,d=S.getAxesPermutation(c,u),h=o.data.get(r.dataId).values;if(d!=null){let b=new Array(u);for(let w=0;w<b.length;w++)b[w]=l[d[w]];h=tg(h,l,r.dtype,d,b),c=S.getInnerMostAxes(c.length,u),l=b}Ie(r,"max"),S.assertAxesAreInnerMostDims("max",c,u);let[f,m]=S.computeOutAndReduceShapes(l,c),g=v.sizeFromShape(m),y=u6(h,g,f,r.dtype),x=o.write(y,f,r.dtype),A=f;return i&&(A=S.expandShapeToKeepDim(f,p)),{dataId:x,shape:A,dtype:r.dtype}}var rB={kernelName:Ji,backendName:"cpu",kernelFunc:W6};function sB(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t;Ie(r,"maxPool");let{filterSize:s,strides:i,pad:o,dimRoundingMode:l}=n,u=1;v.assert(S.eitherStridesOrDilationsAreOne(i,u),()=>`Error in maxPool: Either strides or dilations must be 1. Got strides ${i} and dilations '${u}'`);let p=S.computePool2DInfo(r.shape,s,i,u,o,l),c;if(p.filterWidth===1&&p.filterHeight===1&&v.arraysEqual(p.inShape,p.outShape))c=ar({inputs:{x:r},backend:a});else{let d=a.data.get(r.dataId).values,h=v.computeStrides(r.shape),f=ug(d,r.shape,r.dtype,h,p,"max");c=a.makeTensorInfo(p.outShape,r.dtype,f.values)}return c}var iB={kernelName:eo,backendName:"cpu",kernelFunc:sB};function oB(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{filterSize:s,strides:i,pad:o,dimRoundingMode:l,dataFormat:u}=n;Ie(r,"maxPool3d");let p=S.computePool3DInfo(r.shape,s,i,1,o,l,u),c=a.data.get(r.dataId).values,d=P6(c,r.shape,r.dtype,v.computeStrides(r.shape),p,"max");return a.makeTensorInfo(d.shape,"float32",d.values)}var lB={kernelName:hu,backendName:"cpu",kernelFunc:oB};function uB(e){let{inputs:t,backend:a,attrs:n}=e,{dy:r,input:s}=t,{filterSize:i,strides:o,pad:l,dimRoundingMode:u}=n;Ie([r,s],"maxPool3DGrad");let p=S.computePool3DInfo(s.shape,i,o,1,l,u),c=a.bufferSync(s),d=Qz(c,p),h=p.strideDepth,f=p.strideHeight,m=p.strideWidth,g=p.dilationDepth,y=p.dilationHeight,x=p.dilationWidth,A=p.effectiveFilterDepth,b=p.effectiveFilterHeight,w=p.effectiveFilterWidth,C=A-1-p.padInfo.front,T=w-1-p.padInfo.left,N=b-1-p.padInfo.top,M=$e(s.shape,"float32"),F=a.bufferSync(r);for(let E=0;E<p.batchSize;++E)for(let I=0;I<p.inChannels;++I)for(let _=0;_<p.inDepth;++_)for(let O=0;O<p.inHeight;++O)for(let B=0;B<p.inWidth;++B){let $=_-C,U=O-N,G=B-T,q=0;for(let H=0;H<A;H+=g){let V=($+H)/h;if(!(V<0||V>=p.outDepth||Math.floor(V)!==V))for(let Z=0;Z<b;Z+=y){let X=(U+Z)/f;if(!(X<0||X>=p.outHeight||Math.floor(X)!==X))for(let re=0;re<w;re+=x){let ee=(G+re)/m;if(ee<0||ee>=p.outWidth||Math.floor(ee)!==ee)continue;let me=A*b*w-1-d.get(E,V,X,ee,I),ie=H*b*w+Z*w+re,Ae=me===ie?1:0;if(Ae===0)continue;let Ce=F.get(E,V,X,ee,I);q+=Ce*Ae}}}M.set(q,E,_,O,B,I)}return a.makeTensorInfo(M.shape,M.dtype,M.values)}var dB={kernelName:dp,backendName:"cpu",kernelFunc:uB};function pB(e){let{inputs:t,backend:a,attrs:n}=e,{dy:r,input:s,output:i}=t,o=s;Ie([s,i],"maxPoolGrad");let{filterSize:l,strides:u,pad:p,dimRoundingMode:c}=n,d=S.computePool2DInfo(o.shape,l,u,1,p,c),h=a.data.get(o.dataId).values,f=$e(d.outShape,o.dtype,D6(h,o.shape,o.dtype,d).values),m=d.strideHeight,g=d.strideWidth,y=d.dilationHeight,x=d.dilationWidth,A=d.effectiveFilterHeight,b=d.effectiveFilterWidth,w=b-1-d.padInfo.left,C=A-1-d.padInfo.top,T=$e(o.shape,"float32"),N=a.data.get(r.dataId).values,M=$e(r.shape,"float32",N);for(let F=0;F<d.batchSize;++F)for(let E=0;E<d.inChannels;++E)for(let I=0;I<d.inHeight;++I)for(let _=0;_<d.inWidth;++_){let O=I-C,B=_-w,$=0;for(let U=0;U<A;U+=y){let G=(O+U)/m;if(!(G<0||G>=d.outHeight||Math.floor(G)!==G))for(let q=0;q<b;q+=x){let H=(B+q)/g;if(H<0||H>=d.outWidth||Math.floor(H)!==H)continue;let V=A*b-1-f.get(F,G,H,E),Z=U*b+q,X=V===Z?1:0;if(X===0)continue;let re=M.get(F,G,H,E);$+=re*X}}T.set($,F,I,_,E)}return a.makeTensorInfo(T.shape,T.dtype,T.values)}var cB={kernelName:Sh,backendName:"cpu",kernelFunc:pB};function hB(e,t,a,n,r){let s=v.computeStrides(t),i=ug(e,t,a,s,r,"max"),o=D6(e,t,a,r,!0,n);return[i.values,o.values]}var fB={kernelName:pp,backendName:"cpu",kernelFunc:({inputs:e,attrs:t,backend:a})=>{let{x:n}=e,{filterSize:r,strides:s,pad:i,includeBatchInIndex:o}=t,l=a;Ie(n,"MaxPoolWithArgmax");let u=l.data.get(n.dataId).values,p=S.computePool2DInfo(n.shape,r,s,[1,1],i),[c,d]=hB(u,n.shape,n.dtype,o,p),h=l.write(c,p.outShape,n.dtype),f=l.write(d,p.outShape,n.dtype);return[{dataId:h,shape:p.outShape,dtype:n.dtype},{dataId:f,shape:p.outShape,dtype:"int32"}]}};function mB(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{axis:s,keepDims:i}=n,o=v.parseAxisParam(s,r.shape),l=S.computeOutAndReduceShapes(r.shape,o)[1],u=v.sizeFromShape(l),p=[],c=a.makeTensorInfo([],"float32",new Float32Array([u]));p.push(c);let d=ts({inputs:{x:r},backend:a,attrs:{dtype:"float32"}});p.push(d);let h=dg({inputs:{a:d,b:c},backend:a});p.push(h);let f=Wp({inputs:{x:h},backend:a,attrs:{axis:s,keepDims:i}});return p.forEach(m=>a.disposeIntermediateTensorInfo(m)),f}var gB={kernelName:to,backendName:"cpu",kernelFunc:mB};function yB(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{axis:s,keepDims:i}=n;Ie(r,"min");let o=v.parseAxisParam(s,r.shape),l=o,u=S.getAxesPermutation(l,r.shape.length),p=r;u!=null&&(p=Ua({inputs:{x:r},backend:a,attrs:{perm:u}}),l=S.getInnerMostAxes(l.length,r.shape.length)),S.assertAxesAreInnerMostDims("min",l,p.shape.length);let[c,d]=S.computeOutAndReduceShapes(p.shape,l),h=v.sizeFromShape(d),f=v.makeZerosTypedArray(v.sizeFromShape(c),p.dtype),m=a.data.get(p.dataId).values;for(let y=0;y<f.length;++y){let x=y*h,A=m[x];for(let b=0;b<h;++b){let w=m[x+b];(Number.isNaN(w)||w<A)&&(A=w)}f[y]=A}u!=null&&a.disposeIntermediateTensorInfo(p);let g=a.makeTensorInfo(c,p.dtype,f);if(i){let y=S.expandShapeToKeepDim(c,o),x=At({inputs:{x:g},backend:a,attrs:{shape:y}});return a.disposeIntermediateTensorInfo(g),x}return g}var xB={kernelName:ao,backendName:"cpu",kernelFunc:yB};function AB(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{paddings:s,mode:i}=n;Ie(r,"mirrorPad");let o=s.map((x,A)=>x[0]+r.shape[A]+x[1]),l=s.map(x=>x[0]),u=s.map((x,A)=>x[0]+r.shape[A]),p=i==="reflect"?0:1,c=a.data.get(r.dataId).values,d=r.shape.length,h=v.computeStrides(r.shape),f=v.sizeFromShape(o),m=o.length,g=v.computeStrides(o),y=v.getTypedArrayFromDType(r.dtype,f);for(let x=0;x<f;x++){let A=v.indexToLoc(x,m,g);for(let w=0;w<m;w++)A[w]<l[w]?A[w]=l[w]*2-A[w]-p:A[w]>=u[w]&&(A[w]=(u[w]-1)*2-A[w]+p);A=A.map((w,C)=>w-l[C]);let b=v.locToIndex(A,d,h);y[x]=c[b]}return{dataId:a.write(y,o,r.dtype),shape:o,dtype:r.dtype}}var bB={kernelName:ro,backendName:"cpu",kernelFunc:AB},vB=$t((e,t)=>{let a=e%t;return e<0&&t<0||e>=0&&t>=0?a:(a+t)%t}),wB=Xt(fu,vB),kB={kernelName:fu,backendName:"cpu",kernelFunc:wB},IB=Xl(jx());function B6(e){let{inputs:t,backend:a,attrs:n}=e,{logits:r}=t,{dim:s}=n,i=r.shape.length,o=s;if(o===-1&&(o=i-1),o!==i-1)throw Error(`Softmax along a non-last dimension is not yet supported. Logits was rank ${i} and dim was ${o}`);let l=v.parseAxisParam([o],r.shape),u=W6({inputs:{x:r},backend:a,attrs:{reductionIndices:l,keepDims:!1}}),p=S.expandShapeToKeepDim(u.shape,l),c=At({inputs:{x:u},backend:a,attrs:{shape:p}}),d=og({inputs:{a:r,b:c},backend:a}),h=Z7({inputs:{x:d},backend:a}),f=Wp({inputs:{x:h},backend:a,attrs:{axis:l,keepDims:!1}}),m=At({inputs:{x:f},backend:a,attrs:{shape:p}}),g=dg({inputs:{a:h,b:m},backend:a});return a.disposeIntermediateTensorInfo(u),a.disposeIntermediateTensorInfo(c),a.disposeIntermediateTensorInfo(d),a.disposeIntermediateTensorInfo(h),a.disposeIntermediateTensorInfo(f),a.disposeIntermediateTensorInfo(m),g}var SB={kernelName:$o,backendName:"cpu",kernelFunc:B6};function CB(e){let{inputs:t,backend:a,attrs:n}=e,{logits:r}=t,{numSamples:s,seed:i,normalized:o}=n;Ie(r,"multinomial");let l=o?r:B6({inputs:{logits:r},backend:a,attrs:{dim:-1}}),u=l.shape[0],p=l.shape[1],c=a.data.get(l.dataId).values,d=[u,s],h=v.makeZerosTypedArray(v.sizeFromShape(d),"int32");for(let f=0;f<u;++f){let m=f*p,g=new Float32Array(p-1);g[0]=c[m];for(let A=1;A<g.length;++A)g[A]=g[A-1]+c[m+A];let y=IB.alea(i.toString()),x=f*s;for(let A=0;A<s;++A){let b=y();h[x+A]=g.length;for(let w=0;w<g.length;w++)if(b<g[w]){h[x+A]=w;break}}}return o||a.disposeIntermediateTensorInfo(l),a.makeTensorInfo(d,"int32",h)}var TB={kernelName:so,backendName:"cpu",kernelFunc:CB},NB=Nn.nonMaxSuppressionV3Impl;function RB(e){let{inputs:t,backend:a,attrs:n}=e,{boxes:r,scores:s}=t,{maxOutputSize:i,iouThreshold:o,scoreThreshold:l}=n;Ie(r,"NonMaxSuppression");let u=a.data.get(r.dataId).values,p=a.data.get(s.dataId).values,{selectedIndices:c}=NB(u,p,i,o,l);return a.makeTensorInfo([c.length],"int32",new Int32Array(c))}var EB={kernelName:lo,backendName:"cpu",kernelFunc:RB},MB=Nn.nonMaxSuppressionV4Impl;function FB(e){let{inputs:t,backend:a,attrs:n}=e,{boxes:r,scores:s}=t,{maxOutputSize:i,iouThreshold:o,scoreThreshold:l,padToMaxOutputSize:u}=n;Ie(r,"NonMaxSuppressionPadded");let p=a.data.get(r.dataId).values,c=a.data.get(s.dataId).values,{selectedIndices:d,validOutputs:h}=MB(p,c,i,o,l,u);return[a.makeTensorInfo([d.length],"int32",new Int32Array(d)),a.makeTensorInfo([],"int32",new Int32Array([h]))]}var _B={kernelName:gu,backendName:"cpu",kernelFunc:FB},$B=Nn.nonMaxSuppressionV5Impl;function DB(e){let{inputs:t,backend:a,attrs:n}=e,{boxes:r,scores:s}=t,{maxOutputSize:i,iouThreshold:o,scoreThreshold:l,softNmsSigma:u}=n;Ie(r,"NonMaxSuppressionWithScore");let p=a.data.get(r.dataId).values,c=a.data.get(s.dataId).values,d=i,h=o,f=l,m=u,{selectedIndices:g,selectedScores:y}=$B(p,c,d,h,f,m);return[a.makeTensorInfo([g.length],"int32",new Int32Array(g)),a.makeTensorInfo([y.length],"float32",new Float32Array(y))]}var PB={kernelName:uo,backendName:"cpu",kernelFunc:DB};function OB(e){let{inputs:t,backend:a,attrs:n}=e,{indices:r}=t,{dtype:s,depth:i,onValue:o,offValue:l}=n;Ie(r,"oneHot");let u=v.sizeFromShape(r.shape),p=new Float32Array(u*i);p.fill(l);let c=a.data.get(r.dataId).values;for(let d=0;d<u;++d)c[d]>=0&&c[d]<i&&(p[d*i+c[d]]=o);return a.makeTensorInfo([...r.shape,i],s,p)}var zB={kernelName:po,backendName:"cpu",kernelFunc:OB};function hh(e){let{inputs:t,backend:a}=e,{x:n}=t;if(n.dtype==="string")throw new Error("zerosLike is not supported for string tensors");if(n.dtype==="complex64"){let r=qs({inputs:{input:n},backend:a}),s=hh({inputs:{x:r},backend:a}),i=Hl({inputs:{input:n},backend:a}),o=hh({inputs:{x:i},backend:a}),l=Ja({inputs:{real:s,imag:o},backend:a});return a.disposeIntermediateTensorInfo(r),a.disposeIntermediateTensorInfo(s),a.disposeIntermediateTensorInfo(i),a.disposeIntermediateTensorInfo(o),l}else return pg({backend:a,attrs:{shape:n.shape,value:0,dtype:n.dtype}})}var LB={kernelName:Mu,backendName:"cpu",kernelFunc:hh};function V6(e){let{inputs:t,backend:a}=e,{x:n}=t;if(n.dtype==="string")throw new Error("onesLike is not supported for string tensors");if(n.dtype==="complex64"){let r=qs({inputs:{input:n},backend:a}),s=V6({inputs:{x:r},backend:a}),i=Hl({inputs:{input:n},backend:a}),o=hh({inputs:{x:i},backend:a}),l=Ja({inputs:{real:s,imag:o},backend:a});return a.disposeIntermediateTensorInfo(r),a.disposeIntermediateTensorInfo(s),a.disposeIntermediateTensorInfo(i),a.disposeIntermediateTensorInfo(o),l}else return pg({backend:a,attrs:{shape:n.shape,value:1,dtype:n.dtype}})}var WB={kernelName:yu,backendName:"cpu",kernelFunc:V6};function U6(e){let{inputs:t,backend:a,attrs:n}=e,{axis:r}=n;if(t.length===1)return ch({inputs:{input:t[0]},backend:a,attrs:{dim:r}});let s=t[0].shape,i=t[0].dtype;t.forEach(p=>{v.assertShapesMatch(s,p.shape,"All tensors passed to stack must have matching shapes"),v.assert(i===p.dtype,()=>"All tensors passed to stack must have matching dtypes")});let o=[],l=t.map(p=>{let c=ch({inputs:{input:p},backend:a,attrs:{dim:r}});return o.push(c),c}),u=jl({inputs:l,backend:a,attrs:{axis:r}});return o.forEach(p=>a.disposeIntermediateTensorInfo(p)),u}var BB={kernelName:xu,backendName:"cpu",kernelFunc:U6};function VB(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{paddings:s,constantValue:i}=n;Ie(r,"pad");let o=s.map((y,x)=>y[0]+r.shape[x]+y[1]),l=s.map(y=>y[0]),u=a.data.get(r.dataId).values,p=v.sizeFromShape(r.shape),c=r.shape.length,d=v.computeStrides(r.shape),h=v.sizeFromShape(o),f=o.length,m=v.computeStrides(o),g=v.getTypedArrayFromDType(r.dtype,h);i!==0&&g.fill(i);for(let y=0;y<p;y++){let x=v.indexToLoc(y,c,d).map((b,w)=>b+l[w]),A=v.locToIndex(x,f,m);g[A]=u[y]}return{dataId:a.write(g,o,r.dtype),shape:o,dtype:r.dtype}}var G6={kernelName:co,backendName:"cpu",kernelFunc:VB},UB=$t((e,t)=>Math.pow(e,t)),GB=Xt(ho,UB),HB={kernelName:ho,backendName:"cpu",kernelFunc:GB};function jB(e){let{inputs:t,backend:a,attrs:n}=e,{paramsNestedSplits:r,paramsDenseValues:s,indices:i}=t,{outputRaggedRank:o}=n,l=r.map(y=>a.data.get(y.dataId).values),u=r.map(y=>y.shape),p=a.data.get(s.dataId).values,c=a.data.get(i.dataId).values,[d,h,f]=m6(l,u,p,s.shape,s.dtype,c,i.shape,o),m=d.map(y=>a.makeTensorInfo([y.length],"int32",y)),g=a.makeTensorInfo(f,s.dtype,h);return m.concat([g])}var qB={kernelName:Ch,backendName:"cpu",kernelFunc:jB};function XB(e){let{inputs:t,backend:a}=e,{starts:n,limits:r,deltas:s}=t,i=a.data.get(n.dataId).values,o=a.data.get(r.dataId).values,l=a.data.get(s.dataId).values,[u,p]=g6(i,n.shape,n.dtype,o,r.shape,l,s.shape),c=a.makeTensorInfo([u.length],"int32",u),d=a.makeTensorInfo([p.length],n.dtype,p);return[c,d]}var KB={kernelName:Th,backendName:"cpu",kernelFunc:XB};function YB(e){let{inputs:t,backend:a,attrs:n}=e,{shape:r,values:s,defaultValue:i,rowPartitionTensors:o}=t,{rowPartitionTypes:l}=n,u=a.data.get(r.dataId).values,p=a.data.get(s.dataId).values,c=a.data.get(i.dataId).values,d=o.map(g=>a.data.get(g.dataId).values),h=o.map(g=>g.shape),[f,m]=y6(u,r.shape,p,s.shape,s.dtype,c,i.shape,d,h,l);return a.makeTensorInfo(f,s.dtype,m)}var ZB={kernelName:Nh,backendName:"cpu",kernelFunc:YB};function JB(e){let{backend:t,attrs:a}=e,{start:n,stop:r,dtype:s,step:i}=a,o=ag(n,r,i,s);return t.makeTensorInfo([o.length],s,o)}var QB={kernelName:Au,backendName:"cpu",kernelFunc:JB},eV=ct(go,e=>1/e),tV={kernelName:go,backendName:"cpu",kernelFunc:eV};function aV(e){let{inputs:t,backend:a,attrs:n}=e,{images:r}=t,{alignCorners:s,halfPixelCenters:i,size:o}=n;Ie(r,"resizeBilinear");let l=v.computeStrides(r.shape),[u,p]=o,[c,d,h,f]=r.shape,m=a.data.get(r.dataId).values,g=new Float32Array(v.sizeFromShape([c,u,p,f])),y=[s&&u>1?d-1:d,s&&p>1?h-1:h],x=[s&&u>1?u-1:u,s&&p>1?p-1:p],A=0,b=y[0]/x[0],w=y[1]/x[1];for(let C=0;C<c;C++)for(let T=0;T<u;T++){let N;i?N=b*(T+.5)-.5:N=b*T;let M=Math.max(0,Math.floor(N)),F=N-M,E=Math.min(d-1,Math.ceil(N)),I=C*l[0]+M*l[1],_=C*l[0]+E*l[1];for(let O=0;O<p;O++){let B;i?B=w*(O+.5)-.5:B=w*O;let $=Math.max(0,Math.floor(B)),U=B-$,G=Math.min(h-1,Math.ceil(B)),q=I+$*l[2],H=_+$*l[2],V=I+G*l[2],Z=_+G*l[2];for(let X=0;X<f;X++){let re=m[q+X],ee=m[H+X],me=m[V+X],ie=m[Z+X],Ae=re+(me-re)*U,Ce=ee+(ie-ee)*U,Re=Ae+(Ce-Ae)*F;g[A++]=Re}}}return a.makeTensorInfo([c,u,p,f],"float32",g)}var nV={kernelName:Ao,backendName:"cpu",kernelFunc:aV};function rV(e){let{inputs:t,backend:a,attrs:n}=e,{images:r,dy:s}=t,{alignCorners:i}=n;Ie([s,r],"resizeBilinearGrad");let o=v.computeStrides(r.shape),[l,u,p,c]=r.shape,[,d,h]=s.shape,f=new Float32Array(l*u*p*c),m=[i&&d>1?u-1:u,i&&h>1?p-1:p],g=[i&&d>1?d-1:d,i&&h>1?h-1:h],y=m[0]/g[0],x=m[1]/g[1],A=a.data.get(s.dataId).values,b=0;for(let w=0;w<l;w++){let C=w*o[0];for(let T=0;T<d;T++){let N=T*y,M=Math.floor(N),F=Math.min(Math.ceil(N),u-1),E=C+M*o[1],I=C+F*o[1],_=N-M,O=1-_;for(let B=0;B<h;B++){let $=B*x,U=Math.floor($),G=Math.min(Math.ceil($),p-1),q=$-U,H=1-q,V=E+U*o[2],Z=E+G*o[2],X=I+U*o[2],re=I+G*o[2],ee=O*H,me=O*q,ie=_*H,Ae=_*q;for(let Ce=0;Ce<c;Ce++){let Re=A[b++];f[V+Ce]+=Re*ee,f[Z+Ce]+=Re*me,f[X+Ce]+=Re*ie,f[re+Ce]+=Re*Ae}}}}return a.makeTensorInfo([l,p,u,c],"float32",f)}var sV={kernelName:wu,backendName:"cpu",kernelFunc:rV};function iV(e){let{inputs:t,backend:a,attrs:n}=e,{images:r}=t,{alignCorners:s,halfPixelCenters:i,size:o}=n;Ie(r,"resizeNearestNeighbor");let l=v.computeStrides(r.shape),[u,p]=o,[c,d,h,f]=r.shape,m=a.data.get(r.dataId).values,g=new Float32Array(c*u*p*f),y=[s&&u>1?d-1:d,s&&p>1?h-1:h],x=[s&&u>1?u-1:u,s&&p>1?p-1:p],A=y[0]/x[0],b=y[1]/x[1],w=0;for(let C=0;C<c;C++){let T=C*l[0];for(let N=0;N<u;N++){let M=i?A*(N+.5):A*N,F=Math.min(d-1,s?Math.round(M):Math.floor(M));i&&(F=Math.max(0,F));let E=T+F*l[1];for(let I=0;I<p;I++){let _=i?b*(I+.5):b*I,O=Math.min(h-1,s?Math.round(_):Math.floor(_));i&&(O=Math.max(0,O));let B=E+O*l[2];for(let $=0;$<f;$++){let U=m[B+$];g[w++]=U}}}}return a.makeTensorInfo([c,u,p,f],r.dtype,g)}var oV={kernelName:xo,backendName:"cpu",kernelFunc:iV};function lV(e){let{inputs:t,backend:a,attrs:n}=e,{images:r,dy:s}=t,{alignCorners:i}=n;Ie([s,r],"resizeNearestNeighborGrad");let o=v.computeStrides(r.shape),l=v.computeStrides(s.shape),[u,p,c,d]=r.shape,[,h,f]=s.shape,m=new Float32Array(u*p*c*d),g=a.data.get(s.dataId).values,y=[i&&h>1?p-1:p,i&&f>1?c-1:c],x=[i&&h>1?h-1:h,i&&f>1?f-1:f],A=y[0]/x[0],b=y[1]/x[1],w=1/A,C=1/b,T=Math.ceil(w)*2+2,N=Math.ceil(C)*2+2;for(let M=0;M<u;M++){let F=M*o[0];for(let E=0;E<p;E++){let I=F+E*o[1],_=Math.floor(E*w),O=Math.floor(_-T/2);for(let B=0;B<c;B++){let $=I+B*o[2],U=Math.floor(B*C),G=Math.floor(U-N/2);for(let q=0;q<d;q++){let H=0;for(let V=0;V<T;V++){let Z=V+O;if(Z<0||Z>=h)continue;let X=F+Z*l[1],re=Z*A,ee=Math.min(p-1,i?Math.round(re):Math.floor(re));if(E===ee)for(let me=0;me<N;me++){let ie=me+G;if(ie<0||ie>=f)continue;let Ae=X+ie*l[2],Ce=ie*b,Re=Math.min(c-1,i?Math.round(Ce):Math.floor(Ce));B===Re&&(H+=g[Ae+q])}}m[$+q]=H}}}}return a.makeTensorInfo(r.shape,r.dtype,m)}var uV={kernelName:vu,backendName:"cpu",kernelFunc:lV};function dV(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{dims:s}=n;Ie(r,"reverse");let i=r.shape.length,o=v.parseAxisParam(s,r.shape);if(i===0)return ar({inputs:{x:r},backend:a});let l=new Wt(r.shape,r.dtype),u=a.bufferSync(r);for(let p=0;p<l.size;p++){let c=l.indexToLoc(p),d=c.slice();o.forEach(h=>d[h]=r.shape[h]-1-d[h]),l.set(u.get(...d),...c)}return a.makeTensorInfo(l.shape,l.dtype,l.values)}var pV={kernelName:vo,backendName:"cpu",kernelFunc:dV},cV={kernelName:Uo,backendName:"cpu",kernelFunc:({inputs:e,attrs:t,backend:a})=>{let{image:n}=e,{radians:r,fillValue:s,center:i}=t,o=a,l=v.getTypedArrayFromDType(n.dtype,v.sizeFromShape(n.shape)),[u,p,c,d]=n.shape,[h,f]=S.getImageCenter(i,p,c),m=255,g=Math.sin(r),y=Math.cos(r),x=o.data.get(n.dataId).values;for(let A=0;A<u;A++){let b=A*c*p*d;for(let w=0;w<p;w++){let C=w*(c*d);for(let T=0;T<c;T++){let N=T*d;for(let M=0;M<d;M++){let F=[u,w,T,M],E=F[2],I=F[1],_=(E-h)*y-(I-f)*g,O=(E-h)*g+(I-f)*y;_=Math.round(_+h),O=Math.round(O+f);let B=s;if(typeof s!="number"&&(M===3?B=m:B=s[M]),_>=0&&_<c&&O>=0&&O<p){let U=O*(c*d),G=_*d,q=b+U+G+M;B=x[q]}let $=b+C+N+M;l[$]=B}}}}return{dataId:o.write(l,n.shape,n.dtype),shape:n.shape,dtype:n.dtype}}},hV=ct(wo,e=>{let t=Math.floor(e);return e-t<.5?Math.floor(e):e-t>.5?Math.ceil(e):t%2===0?t:t+1}),fV={kernelName:wo,backendName:"cpu",kernelFunc:hV};function mV(e){let{inputs:t,backend:a,attrs:n}=e,{indices:r,updates:s}=t,{shape:i}=n,{sliceRank:o,numUpdates:l,sliceSize:u,strides:p,outputSize:c}=S.calculateShapes(s,r,i),d=!0,h=a.bufferSync(r),f=a.bufferSync(s),m=Bs(h,f,i,c,u,l,o,p,0,d);return a.makeTensorInfo(i,m.dtype,m.values)}var gV={kernelName:Io,backendName:"cpu",kernelFunc:mV};function yV(e,t){let a=0,n=e.length,r=0;for(;a<n;)r=Math.floor((a+n)/2),e[r]<t?a=r+1:n=r;return n}function xV(e,t){let a=0,n=e.length,r=0;for(;a<n;)r=Math.floor((a+n)/2),e[r]<=t?a=r+1:n=r;return n}function AV(e,t,a,n,r,s){let i=v.getArrayFromDType("int32",a*r);for(let o=0;o<a;++o){let l=e.slice(o*n,(o+1)*n),u=o*r;for(let p=0;p<r;++p)i[u+p]=s==="left"?yV(l,t[p+u]):xV(l,t[p+u])}return i}function bV(e){let{inputs:t,backend:a,attrs:n}=e,{sortedSequence:r,values:s}=t,{side:i}=n,o=a.data.get(r.dataId).values,l=a.data.get(s.dataId).values,u=AV(o,l,r.shape[0],r.shape[1],s.shape[1],i);return a.makeTensorInfo(s.shape,"int32",u)}var vV={kernelName:Co,backendName:"cpu",kernelFunc:bV};function wV(e){let{inputs:t,backend:a}=e,{condition:n,t:r,e:s}=t;Ie([n,r,s],"select");let i=n.shape.length,o=a.data.get(n.dataId).values,l=a.data.get(r.dataId).values,u=a.data.get(s.dataId).values,p=pa(r.dtype,s.dtype),c=v.makeZerosTypedArray(v.sizeFromShape(r.shape),p),d=0,h=i===0||i>1||r.shape.length===1?1:v.sizeFromShape(r.shape.slice(1));for(let f=0;f<o.length;f++)for(let m=0;m<h;m++)o[f]===1?c[d++]=l[f]:c[d++]=u[f];return a.makeTensorInfo(r.shape,p,c)}var kV={kernelName:ku,backendName:"cpu",kernelFunc:wV},IV=S.SELU_SCALEALPHA,SV=S.SELU_SCALE,CV=ct(To,e=>e>=0?SV*e:IV*(Math.exp(e)-1)),TV={kernelName:To,backendName:"cpu",kernelFunc:CV},NV=ct(Ro,e=>e<0?-1:e>0?1:0),RV={kernelName:Ro,backendName:"cpu",kernelFunc:NV},EV=ct(No,e=>Math.sin(e)),MV={kernelName:No,backendName:"cpu",kernelFunc:EV},FV=ct(Su,e=>Math.sinh(e)),_V={kernelName:Su,backendName:"cpu",kernelFunc:FV},$V=11920928955078125e-23,tx=Math.log($V)+2,DV=ct(Mo,e=>{let t=e>-tx,a=e<tx,n=Math.exp(e),r;return a?r=n:t?r=e:r=Math.log(1+n),r}),PV={kernelName:Mo,backendName:"cpu",kernelFunc:DV};function OV(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{blockShape:s,paddings:i}=n;Ie([r],"spaceToBatchND");let o=v.sizeFromShape(s),l=[[0,0]];l.push(...i);for(let g=1+s.length;g<r.shape.length;++g)l.push([0,0]);let u=G6.kernelFunc({inputs:{x:r},backend:a,attrs:{paddings:l,constantValue:0}}),p=S.getReshaped(u.shape,s,o,!1),c=S.getPermuted(p.length,s.length,!1),d=S.getReshapedPermuted(u.shape,s,o,!1),h=At({inputs:{x:u},backend:a,attrs:{shape:p}}),f=Ua({inputs:{x:h},backend:a,attrs:{perm:c}}),m=At({inputs:{x:f},backend:a,attrs:{shape:d}});return a.disposeIntermediateTensorInfo(u),a.disposeIntermediateTensorInfo(h),a.disposeIntermediateTensorInfo(f),m}var zV={kernelName:Cu,backendName:"cpu",kernelFunc:OV};function LV(e){let{inputs:t,backend:a}=e,{indices:n,values:r,denseShape:s,defaultValue:i}=t;if(s.shape.length!==1)throw new Error(`Dense shape must be a vector, saw:
|
|
${s.shape}`);if(n.shape.length!==2)throw new Error(`Indices must be a matrix, saw:
|
|
${n.shape}`);if(r.shape.length!==1)throw new Error(`Values must be a vector, saw:
|
|
${r.shape}`);if(i.shape.length!==0)throw new Error(`Default value must be a scalar, saw:
|
|
${i.shape}`);let o=a.data.get(n.dataId).values,l=a.data.get(r.dataId).values,u=a.data.get(s.dataId).values,p=a.data.get(i.dataId).values[0],[c,d,h,f,m]=b6(o,n.shape,n.dtype,l,r.dtype,u,p);return[a.makeTensorInfo(d,n.dtype,c),a.makeTensorInfo([d[0]],r.dtype,h),a.makeTensorInfo([f.length],"bool",new Uint8Array(f.map(g=>Number(g)))),a.makeTensorInfo([m.length],n.dtype,new Int32Array(m))]}var WV={kernelName:hp,backendName:"cpu",kernelFunc:LV};function BV(e){let{inputs:t,backend:a}=e,{inputIndices:n,inputShape:r,newShape:s}=t;if(n.shape.length!==2)throw new Error(`Input indices should be a matrix but received shape
|
|
${n.shape}`);if(r.shape.length!==1)throw new Error(`Input shape should be a vector but received shape
|
|
${r.shape}`);if(s.shape.length!==1)throw new Error(`Target shape should be a vector but received shape ${s.shape}`);let i=Array.from(a.data.get(r.dataId).values),o=a.data.get(n.dataId).values,l=Array.from(a.data.get(s.dataId).values),[u,p,c]=v6(o,n.shape,n.dtype,i,l);return[a.makeTensorInfo(p,n.dtype,u),a.makeTensorInfo([c.length],s.dtype,new Int32Array(c))]}var VV={kernelName:Nu,backendName:"cpu",kernelFunc:BV};function UV(e){let{inputs:t,backend:a}=e,{data:n,indices:r,segmentIds:s}=t;if(n.shape.length<1)throw new Error("Data should be at least 1 dimensional but received scalar");if(r.shape.length!==1)throw new Error(`Indices should be a vector but received shape
|
|
${r.shape}`);if(s.shape.length!==1)throw new Error(`Segment ids should be a vector but received shape
|
|
${s.shape}`);if(r.shape[0]!==s.shape[0])throw new Error("segmentIds and indices should have same size.");let i=a.data.get(n.dataId).values,o=a.data.get(r.dataId).values,l=a.data.get(s.dataId).values,[u,p]=ng(i,n.shape,n.dtype,o,l,!0);return a.makeTensorInfo(p,n.dtype,u)}var GV={kernelName:fp,backendName:"cpu",kernelFunc:UV};function HV(e){let{inputs:t,backend:a}=e,{data:n,indices:r,segmentIds:s}=t;if(n.shape.length<1)throw new Error("Data should be at least 1 dimensional but received scalar");if(r.shape.length!==1)throw new Error(`Indices should be a vector but received shape
|
|
${r.shape}`);if(s.shape.length!==1)throw new Error(`Segment ids should be a vector but received shape
|
|
${s.shape}`);if(r.shape[0]!==s.shape[0])throw new Error("segmentIds and indices should have same size.");let i=a.data.get(n.dataId).values,o=a.data.get(r.dataId).values,l=a.data.get(s.dataId).values,[u,p]=ng(i,n.shape,n.dtype,o,l);return a.makeTensorInfo(p,n.dtype,u)}var jV={kernelName:mp,backendName:"cpu",kernelFunc:HV};function qV(e){let{inputs:t,backend:a,attrs:n}=e,{sparseIndices:r,sparseValues:s,defaultValue:i}=t,{outputShape:o}=n,{sliceRank:l,numUpdates:u,sliceSize:p,strides:c,outputSize:d}=S.calculateShapes(s,r,o),h=!1,f=a.bufferSync(r),m;switch(s.dtype){case"bool":{let g=a.bufferSync(s),y=!!a.data.get(i.dataId).values[0];m=Bs(f,g,o,d,p,u,l,c,y,h);break}case"float32":{let g=a.bufferSync(s),y=a.data.get(i.dataId).values[0];m=Bs(f,g,o,d,p,u,l,c,y,h);break}case"int32":{let g=a.bufferSync(s),y=a.data.get(i.dataId).values[0];m=Bs(f,g,o,d,p,u,l,c,y,h);break}case"string":{let g=a.bufferSync(s),y=v.decodeString(a.data.get(i.dataId).values[0]);m=Bs(f,g,o,d,p,u,l,c,y,h);break}default:throw new Error(`Unsupported type ${s.dtype}`)}return a.makeTensorInfo(o,m.dtype,m.values)}var XV={kernelName:Do,backendName:"cpu",kernelFunc:qV};function KV(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{numOrSizeSplits:s,axis:i}=n,o=v.parseAxisParam(i,r.shape)[0],l=S.prepareSplitSize(r,s,o),u=new Array(r.shape.length).fill(0),p=r.shape.slice();return l.map(c=>{let d=[...p];d[o]=c;let h=Xs({inputs:{x:r},backend:a,attrs:{begin:u,size:d}});return u[o]+=c,h})}var YV={kernelName:Tu,backendName:"cpu",kernelFunc:KV},ZV={kernelName:gp,backendName:"cpu",kernelFunc:({inputs:e,backend:t})=>{let{x:a}=e,n=t;Ie(a,"square");let r=n.data.get(a.dataId).values,s=new Float32Array(r.length);for(let i=0;i<r.length;++i){let o=r[i];s[i]=o*o}return{dataId:n.write(s,a.shape,a.dtype),shape:a.shape,dtype:a.dtype}}},JV=ct(is,(e,t)=>{let a=t;return isNaN(e)?NaN:e>0?1:a.alpha}),QV={kernelName:is,backendName:"cpu",kernelFunc:JV};function eU(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{begin:s,end:i,strides:o,beginMask:l,endMask:u,ellipsisMask:p,newAxisMask:c,shrinkAxisMask:d}=n;Ie(r,"stridedSlice");let{finalShapeSparse:h,finalShape:f,isIdentity:m,sliceDim0:g,isSimpleSlice:y,begin:x,end:A,strides:b}=Nt.sliceInfo(r.shape,s,i,o,l,u,p,c,d),w;if(m)w=At({inputs:{x:r},backend:a,attrs:{shape:f}});else if(g||y){v.assert(r.shape.length>=1,()=>`Input must have rank at least 1, got: ${r.shape.length}`);let C=Nt.computeOutShape(x,A,b),T=Xs({inputs:{x:r},backend:a,attrs:{begin:x,size:C}});w=At({inputs:{x:T},backend:a,attrs:{shape:f}}),a.disposeIntermediateTensorInfo(T)}else{let C=a.bufferSync(r),T=I6(h,C,b,x);w=a.makeTensorInfo(f,T.dtype,T.values)}return w}var tU={kernelName:Oo,backendName:"cpu",kernelFunc:eU};function aU(e){let{inputs:t,backend:a,attrs:n}=e,{separator:r,nGramWidths:s,leftPad:i,rightPad:o,padWidth:l,preserveShortSequences:u}=n,{data:p,dataSplits:c}=t,d=a.data.get(p.dataId).values,h=a.data.get(c.dataId).values,[f,m]=rg(d,h,r,s,i,o,l,u);return[a.makeTensorInfo([f.length],"string",f),a.makeTensorInfo(c.shape,"int32",m)]}var nU={kernelName:Ru,backendName:"cpu",kernelFunc:aU};function rU(e){let{inputs:t,backend:a,attrs:n}=e,{skipEmpty:r}=n,{input:s,delimiter:i}=t;if(s.dtype!=="string")throw new Error("Input must be of datatype string");if(s.shape.length!==1)throw new Error(`Input must be a vector, got shape: ${s.shape}`);if(i.shape.length!==0)throw new Error(`Delimiter must be a scalar, got shape: ${i.shape}`);let o=a.data.get(s.dataId).values,l=a.data.get(i.dataId).values[0],[u,p,c]=sg(o,l,r),d=p.length;return[a.makeTensorInfo([d,2],"int32",u),a.makeTensorInfo([d],"string",p),a.makeTensorInfo([2],"int32",new Int32Array(c))]}var sU={kernelName:xp,backendName:"cpu",kernelFunc:rU};function iU(e){let{inputs:t,backend:a,attrs:n}=e,{numBuckets:r}=n,{input:s}=t;if(s.dtype!=="string")throw new Error("Input must be of datatype string");if(r<=0)throw new Error("Number of buckets must be at least 1");let i=a.data.get(s.dataId).values,o=ig(i,r);return a.makeTensorInfo(s.shape,"int32",o)}var oU={kernelName:Ap,backendName:"cpu",kernelFunc:iU},lU=ct(Lo,e=>Math.tan(e)),uU={kernelName:Lo,backendName:"cpu",kernelFunc:lU},dU=ct(Wo,e=>Math.tanh(e)),pU={kernelName:Wo,backendName:"cpu",kernelFunc:dU};function cU(e){let{inputs:t,backend:a}=e,{tensor:n,indices:r,updates:s}=t,{sliceRank:i,numUpdates:o,sliceSize:l,strides:u,outputSize:p}=S.calculateShapes(s,r,n.shape),c=!1,d=a.bufferSync(r),h=a.bufferSync(s),f=a.bufferSync(n),m=Bs(d,h,n.shape,p,l,o,i,u,f,c);return a.makeTensorInfo(n.shape,m.dtype,m.values)}var hU={kernelName:So,backendName:"cpu",kernelFunc:cU};function fU(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{reps:s}=n;Ie(r,"tile");let i=C6(a.bufferSync(r),s);return a.makeTensorInfo(i.shape,i.dtype,i.values)}var mU={kernelName:ss,backendName:"cpu",kernelFunc:fU};function gU(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{k:s,sorted:i}=n;Ie(r,"topk");let o=a.data.get(r.dataId).values,[l,u]=N6(o,r.shape,r.dtype,s,i);return[a.makeTensorInfo(l.shape,l.dtype,l.values),a.makeTensorInfo(u.shape,u.dtype,u.values)]}var yU={kernelName:Bo,backendName:"cpu",kernelFunc:gU};function xU(e){let{inputs:t,attrs:a,backend:n}=e,{image:r,transforms:s}=t,{interpolation:i,fillMode:o,fillValue:l,outputShape:u}=a,[p,c,d,h]=r.shape,[f,m]=u!=null?u:[c,d],g=[p,f,m,h],y=v.computeStrides(r.shape),x=y[0],A=y[1],b=y[2],w=v.computeStrides(g),C=w[0],T=w[1],N=w[2],M=v.getTypedArrayFromDType(r.dtype,v.sizeFromShape(g));M.fill(l);let F=n.data.get(r.dataId).values,E=n.data.get(s.dataId).values;for(let I=0;I<p;++I){let _=s.shape[0]===1?E:E.subarray(I*8,I*8+8);for(let O=0;O<f;++O)for(let B=0;B<m;++B)for(let $=0;$<h;++$){let U,G=_[6]*B+_[7]*O+1;if(G===0)continue;let q=(_[0]*B+_[1]*O+_[2])/G,H=(_[3]*B+_[4]*O+_[5])/G,V=ax(q,d,o),Z=ax(H,c,o);switch(i){case"nearest":U=IU(F,c,d,x,A,b,I,Z,V,$,l);break;case"bilinear":U=SU(F,c,d,x,A,b,I,Z,V,$,l);break;default:throw new Error(`Error in Transform: Expect 'nearest' or 'bilinear', but got ${i}`)}let X=I*C+O*T+B*N+$;M[X]=U}return n.makeTensorInfo(g,r.dtype,M)}return{dataId:n.write(M,g,r.dtype),shape:r.shape,dtype:r.dtype}}var AU={kernelName:Vo,backendName:"cpu",kernelFunc:xU};function ax(e,t,a){switch(a){case"reflect":return bU(e,t);case"wrap":return vU(e,t);case"nearest":return kU(e,t);case"constant":default:return wU(e,t)}}function bU(e,t){let a=e;if(a<0)if(t<=1)a=0;else{let n=2*t;a<n&&(a=n*Math.trunc(-a/n)+a),a=a<-t?a+n:-a-1}else if(a>t-1)if(t<=1)a=0;else{let n=2*t;a-=n*Math.trunc(a/n),a>=t&&(a=n-a-1)}return v.clamp(0,a,t-1)}function vU(e,t){let a=e;if(a<0)if(t<=1)a=0;else{let n=t-1;a+=t*(Math.trunc(-a/n)+1)}else if(a>t-1)if(t<=1)a=0;else{let n=t-1;a-=t*Math.trunc(a/n)}return v.clamp(0,a,t-1)}function wU(e,t){return e}function kU(e,t){return v.clamp(0,e,t-1)}function xd(e,t,a,n,r,s,i,o,l,u,p){let c=i*n+o*r+l*s+u;return 0<=o&&o<t&&0<=l&&l<a?e[c]:p}function IU(e,t,a,n,r,s,i,o,l,u,p){let c=Math.round(o),d=Math.round(l);return xd(e,t,a,n,r,s,i,c,d,u,p)}function SU(e,t,a,n,r,s,i,o,l,u,p){let c=Math.floor(o),d=Math.floor(l),h=c+1,f=d+1,m=(f-l)*xd(e,t,a,n,r,s,i,c,d,u,p)+(l-d)*xd(e,t,a,n,r,s,i,c,f,u,p),g=(f-l)*xd(e,t,a,n,r,s,i,h,d,u,p)+(l-d)*xd(e,t,a,n,r,s,i,h,f,u,p);return(h-o)*m+(o-c)*g}function CU(e){let{inputs:t,attrs:a,backend:n}=e,{axis:r}=a,{x:s}=t;Ie(s,"unique");let i=n.data.get(s.dataId).values,{outputValues:o,outputShape:l,indices:u}=lg(i,r,s.shape,s.dtype);return[n.makeTensorInfo(l,s.dtype,o),n.makeTensorInfo([u.length],"int32",u)]}var TU={kernelName:bp,backendName:"cpu",kernelFunc:CU};function NU(e){let{inputs:t,backend:a,attrs:n}=e,{value:r}=t,{axis:s}=n;s<0&&(s+=r.shape.length);let i=r.shape.length,o=r.shape[s],l=new Array(i-1),u=0;for(let h=0;h<i;h++)h!==s&&(l[u++]=r.shape[h]);let p=new Array(i).fill(0),c=r.shape.slice();c[s]=1;let d=new Array(o);for(let h=0;h<d.length;h++){p[s]=h;let f=Xs({inputs:{x:r},backend:a,attrs:{begin:p,size:c}});d[h]=At({inputs:{x:f},backend:a,attrs:{shape:l}}),a.disposeIntermediateTensorInfo(f)}return d}var RU={kernelName:Eu,backendName:"cpu",kernelFunc:NU};function EU(e){let{inputs:t,backend:a,attrs:n}=e,{x:r,segmentIds:s}=t,{numSegments:i}=n;Ie(r,"unsortedSegmentSum");let o=r.shape.length,l=s.shape.length,u=[],p=[],c=o-l,d=s;for(let f=0;f<c;++f){let m=ch({inputs:{input:d},backend:a,attrs:{dim:f+1}});d=m,p.push(m)}for(let f=0;f<i;++f){let m=v.createScalarValue(f,"int32"),g=a.makeTensorInfo([],"int32",m),y=K7({inputs:{a:g,b:d},backend:a}),x=ts({inputs:{x:y},backend:a,attrs:{dtype:"float32"}}),A=Yh({inputs:{a:x,b:r},backend:a}),b=Wp({inputs:{x:A},backend:a,attrs:{axis:0,keepDims:!1}});u.push(b),p.push(g),p.push(y),p.push(x),p.push(A),p.push(b)}let h=U6({inputs:u,backend:a,attrs:{axis:0}});return p.forEach(f=>a.disposeIntermediateTensorInfo(f)),h}var MU={kernelName:vp,backendName:"cpu",kernelFunc:EU},FU=[Tz,oO,Rz,Mz,hO,_z,Dz,Oz,Lz,Bz,Uz,Hz,qz,Yz,Jz,tL,nL,sL,oL,Sz,uL,pL,hL,mO,mL,pO,yO,yL,lO,AL,vL,wL,IL,CL,NL,EL,FL,$L,PL,zL,WL,VL,GL,jL,qL,KL,ZL,QL,eW,tW,aW,sW,xz,oW,xO,mW,AO,gW,vO,wW,kW,SW,kO,SO,TW,RW,MW,_W,TO,RO,uO,DW,bL,OW,LW,BW,Az,MO,_O,UW,DO,HW,XW,YW,QW,tB,nB,rB,OO,iB,lB,dB,cB,fB,gB,xB,LO,bB,kB,TB,BO,UO,EB,_B,PB,HO,zB,WB,BB,G6,HB,vz,XO,qB,KB,ZB,QB,dO,A1,tV,wz,kz,Iz,nV,sV,oV,uV,pV,cV,fV,az,gV,vV,kV,TV,rz,RV,MV,_V,sz,SB,PV,zV,WV,VV,GV,jV,XV,YV,lz,ZV,dz,cz,QV,tU,nU,sU,oU,gz,nW,uU,pU,hU,mU,yU,AU,jO,TU,RU,MU,LB];for(let e of FU)yn(e);var H6={};Ze(H6,{assertNotComplex:()=>Pu,bindCanvasToFramebuffer:()=>UU,bindColorTextureToFramebuffer:()=>Kc,bindTextureToProgramUniformSampler:()=>iv,bindTextureUnit:()=>nv,bindVertexBufferToProgramAttribute:()=>v1,callAndCheck:()=>pe,canBeRepresented:()=>j6,createFragmentShader:()=>K6,createFramebuffer:()=>av,createProgram:()=>Y6,createStaticIndexBuffer:()=>Q6,createStaticVertexBuffer:()=>J6,createTexture:()=>ev,createVertexShader:()=>X6,getBatchDim:()=>Ks,getExtensionOrThrow:()=>Ad,getFramebufferErrorMessage:()=>ov,getMaxTexturesInShader:()=>pv,getNumChannels:()=>BU,getProgramUniformLocation:()=>sv,getProgramUniformLocationOrThrow:()=>rv,getRowsCols:()=>Ys,getShapeAs3D:()=>vd,getTextureShapeFromLogicalShape:()=>uv,getWebGLDisjointQueryTimerVersion:()=>cv,getWebGLErrorMessage:()=>q6,getWebGLMaxTextureSize:()=>dv,hasExtension:()=>mn,isCapableOfRenderingToFloatTexture:()=>hv,isDownloadFloatTextureEnabled:()=>fv,isReshapeFree:()=>Gd,isWebGLFenceEnabled:()=>mv,isWebGLVersionEnabled:()=>k1,linkProgram:()=>Z6,logShaderSourceAndInfoLog:()=>hg,resetMaxTextureSize:()=>GU,resetMaxTexturesInShader:()=>HU,unbindColorTextureFromFramebuffer:()=>w1,unbindTextureUnit:()=>VU,validateFramebuffer:()=>bd,validateProgram:()=>Xc,validateTextureSize:()=>tv});var Ls={},Bc={alpha:!1,antialias:!1,premultipliedAlpha:!1,preserveDrawingBuffer:!1,depth:!1,stencil:!1,failIfMajorPerformanceCaveat:!0};function Zh(e,t){Ls[e]=t}function Wn(e,t){if(!(e in Ls)||t!=null){let n=$U(e,t);if(n!==null)Ls[e]=n;else return console.log("Could not get context for WebGL version",e),null}let a=Ls[e];return a==null||a.isContextLost()?(delete Ls[e],Wn(e)):(a.disable(a.DEPTH_TEST),a.disable(a.STENCIL_TEST),a.disable(a.BLEND),a.disable(a.DITHER),a.disable(a.POLYGON_OFFSET_FILL),a.disable(a.SAMPLE_COVERAGE),a.enable(a.SCISSOR_TEST),a.enable(a.CULL_FACE),a.cullFace(a.BACK),Ls[e])}function _U(e){if(!W().getBool("IS_SAFARI")&&typeof OffscreenCanvas!="undefined"&&e===2)return new OffscreenCanvas(300,150);if(typeof document!="undefined")return document.createElement("canvas");throw new Error("Cannot create a canvas in this context")}function $U(e,t){if(e!==1&&e!==2)throw new Error("Cannot get WebGL rendering context, WebGL is disabled.");let a=t==null?_U(e):t;return a.addEventListener("webglcontextlost",n=>{n.preventDefault(),delete Ls[e]},!1),W().getBool("SOFTWARE_WEBGL_ENABLED")&&(Bc.failIfMajorPerformanceCaveat=!1),e===1?a.getContext("webgl",Bc)||a.getContext("experimental-webgl",Bc):a.getContext("webgl2",Bc)}var Ud;(function(e){e[e.DENSE=0]="DENSE",e[e.SHARED_BATCH=1]="SHARED_BATCH"})(Ud||(Ud={}));var fn;(function(e){e[e.RENDER=0]="RENDER",e[e.UPLOAD=1]="UPLOAD",e[e.PIXELS=2]="PIXELS",e[e.DOWNLOAD=3]="DOWNLOAD"})(fn||(fn={}));var da;(function(e){e[e.UNPACKED_FLOAT16=0]="UNPACKED_FLOAT16",e[e.UNPACKED_FLOAT32=1]="UNPACKED_FLOAT32",e[e.PACKED_4X1_UNSIGNED_BYTE=2]="PACKED_4X1_UNSIGNED_BYTE",e[e.PACKED_2X2_FLOAT32=3]="PACKED_2X2_FLOAT32",e[e.PACKED_2X2_FLOAT16=4]="PACKED_2X2_FLOAT16"})(da||(da={}));function Bp(e,t){return[t,e]}function DU(e,t){return e*t}function Vc(e){let t=v.sizeFromShape(e),a=Math.ceil(t/4);return v.sizeToSquarishShape(a)}function Du(e,t){return[Math.max(1,Math.ceil(t/2)),Math.max(1,Math.ceil(e/2))]}function PU(e,t){let[a,n]=Du(e,t);return a*n*4}function cg(e,t){let a=e,n,r,s,i,o,l,u,p,c,d;return W().getNumber("WEBGL_VERSION")===2?(n=a.R32F,r=a.R16F,s=a.RGBA16F,i=a.RGBA32F,o=a.RED,u=4,p=1,c=a.HALF_FLOAT,d=a.FLOAT,l=a.RGBA8):(n=e.RGBA,r=e.RGBA,s=e.RGBA,i=a.RGBA,o=e.RGBA,u=4,p=4,c=t!=null?t.HALF_FLOAT_OES:null,d=e.FLOAT,l=e.RGBA),{internalFormatFloat:n,internalFormatHalfFloat:r,internalFormatPackedHalfFloat:s,internalFormatPackedFloat:i,textureFormatFloat:o,downloadTextureFormat:l,downloadUnpackNumChannels:u,defaultNumChannels:p,textureTypeHalfFloat:c,textureTypeFloat:d}}function pe(e,t){let a=t();return W().getBool("DEBUG")&&OU(e),a}function OU(e){let t=e.getError();if(t!==e.NO_ERROR)throw new Error("WebGL Error: "+q6(e,t))}var zU=596e-10,LU=65504;function j6(e){return!!(W().getBool("WEBGL_RENDER_FLOAT32_ENABLED")||e===0||zU<Math.abs(e)&&Math.abs(e)<LU)}function q6(e,t){switch(t){case e.NO_ERROR:return"NO_ERROR";case e.INVALID_ENUM:return"INVALID_ENUM";case e.INVALID_VALUE:return"INVALID_VALUE";case e.INVALID_OPERATION:return"INVALID_OPERATION";case e.INVALID_FRAMEBUFFER_OPERATION:return"INVALID_FRAMEBUFFER_OPERATION";case e.OUT_OF_MEMORY:return"OUT_OF_MEMORY";case e.CONTEXT_LOST_WEBGL:return"CONTEXT_LOST_WEBGL";default:return`Unknown error code ${t}`}}function Ad(e,t){return Er(e,()=>e.getExtension(t),'Extension "'+t+'" not supported on this browser.')}function X6(e,t){let a=Er(e,()=>e.createShader(e.VERTEX_SHADER),"Unable to create vertex WebGLShader.");if(pe(e,()=>e.shaderSource(a,t)),pe(e,()=>e.compileShader(a)),e.getShaderParameter(a,e.COMPILE_STATUS)===!1)throw console.log(e.getShaderInfoLog(a)),new Error("Failed to compile vertex shader.");return a}function K6(e,t){let a=Er(e,()=>e.createShader(e.FRAGMENT_SHADER),"Unable to create fragment WebGLShader.");if(pe(e,()=>e.shaderSource(a,t)),pe(e,()=>e.compileShader(a)),W().get("ENGINE_COMPILE_ONLY"))return a;if(e.getShaderParameter(a,e.COMPILE_STATUS)===!1)throw hg(t,e.getShaderInfoLog(a)),new Error("Failed to compile fragment shader.");return a}var WU=/ERROR: [0-9]+:([0-9]+):/g;function hg(e,t){let a=WU.exec(t);if(a==null){console.log(`Couldn't parse line number in error: ${t}`),console.log(e);return}let n=+a[1],r=e.split(`
|
|
`),s=r.length.toString().length+2,i=r.map((c,d)=>v.rightPad((d+1).toString(),s)+c),o=0;for(let c=0;c<i.length;c++)o=Math.max(i[c].length,o);let l=i.slice(0,n-1),u=i.slice(n-1,n),p=i.slice(n);console.log(l.join(`
|
|
`)),console.log(t.split(`
|
|
`)[0]),console.log(`%c ${v.rightPad(u[0],o)}`,"border:1px solid red; background-color:#e3d2d2; color:#a61717"),console.log(p.join(`
|
|
`))}function Y6(e){return Er(e,()=>e.createProgram(),"Unable to create WebGLProgram.")}function Z6(e,t){if(pe(e,()=>e.linkProgram(t)),!W().get("ENGINE_COMPILE_ONLY")&&e.getProgramParameter(t,e.LINK_STATUS)===!1)throw console.log(e.getProgramInfoLog(t)),new Error("Failed to link vertex and fragment shaders.")}function Xc(e,t){if(pe(e,()=>e.validateProgram(t)),e.getProgramParameter(t,e.VALIDATE_STATUS)===!1)throw console.log(e.getProgramInfoLog(t)),new Error("Shader program validation failed.")}function J6(e,t){let a=Er(e,()=>e.createBuffer(),"Unable to create WebGLBuffer");return pe(e,()=>e.bindBuffer(e.ARRAY_BUFFER,a)),pe(e,()=>e.bufferData(e.ARRAY_BUFFER,t,e.STATIC_DRAW)),a}function Q6(e,t){let a=Er(e,()=>e.createBuffer(),"Unable to create WebGLBuffer");return pe(e,()=>e.bindBuffer(e.ELEMENT_ARRAY_BUFFER,a)),pe(e,()=>e.bufferData(e.ELEMENT_ARRAY_BUFFER,t,e.STATIC_DRAW)),a}function BU(){return W().getNumber("WEBGL_VERSION")===2?1:4}function ev(e){return Er(e,()=>e.createTexture(),"Unable to create WebGLTexture.")}function tv(e,t){let a=W().getNumber("WEBGL_MAX_TEXTURE_SIZE");if(e<=0||t<=0){let n=`[${e}x${t}]`;throw new Error("Requested texture size "+n+" is invalid.")}if(e>a||t>a){let n=`[${e}x${t}]`,r=`[${a}x${a}]`;throw new Error("Requested texture size "+n+" greater than WebGL maximum on this browser / GPU "+r+".")}}function av(e){return Er(e,()=>e.createFramebuffer(),"Unable to create WebGLFramebuffer.")}function v1(e,t,a,n,r,s,i){let o=e.getAttribLocation(t,a);return o===-1?!1:(pe(e,()=>e.bindBuffer(e.ARRAY_BUFFER,n)),pe(e,()=>e.vertexAttribPointer(o,r,e.FLOAT,!1,s,i)),pe(e,()=>e.enableVertexAttribArray(o)),!0)}function nv(e,t,a){lv(e,a),pe(e,()=>e.activeTexture(e.TEXTURE0+a)),pe(e,()=>e.bindTexture(e.TEXTURE_2D,t))}function VU(e,t){lv(e,t),pe(e,()=>e.activeTexture(e.TEXTURE0+t)),pe(e,()=>e.bindTexture(e.TEXTURE_2D,null))}function rv(e,t,a){return Er(e,()=>e.getUniformLocation(t,a),'uniform "'+a+'" not present in program.')}function sv(e,t,a){return e.getUniformLocation(t,a)}function iv(e,t,a,n){pe(e,()=>nv(e,t,n)),pe(e,()=>e.uniform1i(a,n))}function UU(e){pe(e,()=>e.bindFramebuffer(e.FRAMEBUFFER,null)),pe(e,()=>e.viewport(0,0,e.canvas.width,e.canvas.height)),pe(e,()=>e.scissor(0,0,e.canvas.width,e.canvas.height))}function Kc(e,t,a){pe(e,()=>e.bindFramebuffer(e.FRAMEBUFFER,a)),pe(e,()=>e.framebufferTexture2D(e.FRAMEBUFFER,e.COLOR_ATTACHMENT0,e.TEXTURE_2D,t,0))}function w1(e,t){pe(e,()=>e.bindFramebuffer(e.FRAMEBUFFER,t)),pe(e,()=>e.framebufferTexture2D(e.FRAMEBUFFER,e.COLOR_ATTACHMENT0,e.TEXTURE_2D,null,0))}function bd(e){let t=e.checkFramebufferStatus(e.FRAMEBUFFER);if(t!==e.FRAMEBUFFER_COMPLETE)throw new Error("Error binding framebuffer: "+ov(e,t))}function ov(e,t){switch(t){case e.FRAMEBUFFER_INCOMPLETE_ATTACHMENT:return"FRAMEBUFFER_INCOMPLETE_ATTACHMENT";case e.FRAMEBUFFER_INCOMPLETE_MISSING_ATTACHMENT:return"FRAMEBUFFER_INCOMPLETE_MISSING_ATTACHMENT";case e.FRAMEBUFFER_INCOMPLETE_DIMENSIONS:return"FRAMEBUFFER_INCOMPLETE_DIMENSIONS";case e.FRAMEBUFFER_UNSUPPORTED:return"FRAMEBUFFER_UNSUPPORTED";default:return`unknown error ${t}`}}function Er(e,t,a){let n=pe(e,()=>t());if(n==null)throw new Error(a);return n}function lv(e,t){let a=e.MAX_COMBINED_TEXTURE_IMAGE_UNITS-1,n=t+e.TEXTURE0;if(n<e.TEXTURE0||n>a){let r=`[gl.TEXTURE0, gl.TEXTURE${a}]`;throw new Error(`textureUnit must be in ${r}.`)}}function Ks(e,t=2){return v.sizeFromShape(e.slice(0,e.length-t))}function Ys(e){if(e.length===0)throw Error("Cannot get rows and columns of an empty shape array.");return[e.length>1?e[e.length-2]:1,e[e.length-1]]}function vd(e){let t=[1,1,1];return e.length===0||e.length===1&&e[0]===1||(t=[Ks(e),...Ys(e)]),t}function uv(e,t=!1){let a=W().getNumber("WEBGL_MAX_TEXTURE_SIZE"),n=W().getNumber("WEBGL_MAX_SIZE_FOR_NARROW_TEXTURE");n===1/0&&W().getBool("WEBGL_AUTO_SQUARIFY_NARROW_TEXTURE_SHAPE")&&(n=a/2),t&&(a=a*2,n=n*2,e=e.map((o,l)=>l>=e.length-2?v.nearestLargerEven(e[l]):e[l]),e.length===1&&(e=[2,e[0]])),e.length!==2&&(e=v.squeezeShape(e).newShape);let r=v.sizeFromShape(e),s=null;e.length<=1&&r<=a?s=[1,r]:e.length===2&&e[0]<=a&&e[1]<=a?s=e:e.length===3&&e[0]*e[1]<=a&&e[2]<=a?s=[e[0]*e[1],e[2]]:e.length===3&&e[0]<=a&&e[1]*e[2]<=a?s=[e[0],e[1]*e[2]]:e.length===4&&e[0]*e[1]*e[2]<=a&&e[3]<=a?s=[e[0]*e[1]*e[2],e[3]]:e.length===4&&e[0]<=a&&e[1]*e[2]*e[3]<=a&&(s=[e[0],e[1]*e[2]*e[3]]);let i=s!=null&&Math.max(...s)>n&&Math.min(...s)<=(t?2:1)&&Math.min(...s)>0;if(s==null||i)if(t){let o=Ks(e),l=2,u=2;e.length&&([l,u]=Ys(e)),r=o*(l/2)*(u/2),s=v.sizeToSquarishShape(r).map(p=>p*2)}else s=v.sizeToSquarishShape(r);return s}function Uc(e){return e%2===0}function Gd(e,t){if(e=e.slice(-2),t=t.slice(-2),v.arraysEqual(e,t)||!e.length||!t.length||e[0]===0||e[1]===0||t[0]===0||t[1]===0)return!0;if(e.length!==t.length){let a=e[e.length-1],n=t[t.length-1];if(a===n||Uc(a)&&Uc(n)&&(e[0]===1||t[0]===1))return!0}return e[1]===t[1]&&Uc(e[0])&&Uc(t[0])}var Yc,Zc;function dv(e){if(Yc==null){let t=Wn(e);Yc=t.getParameter(t.MAX_TEXTURE_SIZE)}return Yc}function GU(){Yc=null}function HU(){Zc=null}function pv(e){if(Zc==null){let t=Wn(e);Zc=t.getParameter(t.MAX_TEXTURE_IMAGE_UNITS)}return Math.min(16,Zc)}function cv(e){if(e===0)return 0;let t,a=Wn(e);return mn(a,"EXT_disjoint_timer_query_webgl2")&&e===2?t=2:mn(a,"EXT_disjoint_timer_query")?t=1:t=0,t}function mn(e,t){return e.getExtension(t)!=null}function k1(e){try{if(Wn(e)!=null)return!0}catch(t){return console.log("Error when getting WebGL context: ",t),!1}return!1}function hv(e){if(e===0)return!1;let t=Wn(e);if(e===1){if(!mn(t,"OES_texture_float"))return!1}else if(!mn(t,"EXT_color_buffer_float"))return!1;return I1(t)}function fv(e){if(e===0)return!1;let t=Wn(e);if(e===1){if(!mn(t,"OES_texture_float")||!mn(t,"WEBGL_color_buffer_float"))return!1}else{if(mn(t,"EXT_color_buffer_float"))return I1(t);let a="EXT_color_buffer_half_float";if(mn(t,a)){let n=t.getExtension(a);return jU(t,n)}return!1}return I1(t)}function I1(e){let t=cg(e),a=e.createTexture();e.bindTexture(e.TEXTURE_2D,a);let n=1,r=1;e.texImage2D(e.TEXTURE_2D,0,t.internalFormatFloat,n,r,0,t.textureFormatFloat,t.textureTypeFloat,null);let s=e.createFramebuffer();e.bindFramebuffer(e.FRAMEBUFFER,s),e.framebufferTexture2D(e.FRAMEBUFFER,e.COLOR_ATTACHMENT0,e.TEXTURE_2D,a,0);let i=e.checkFramebufferStatus(e.FRAMEBUFFER)===e.FRAMEBUFFER_COMPLETE;return e.bindTexture(e.TEXTURE_2D,null),e.bindFramebuffer(e.FRAMEBUFFER,null),e.deleteTexture(a),e.deleteFramebuffer(s),i}function jU(e,t){let a=cg(e,t),n=e.createTexture();e.bindTexture(e.TEXTURE_2D,n);let r=1,s=1;e.texImage2D(e.TEXTURE_2D,0,a.internalFormatHalfFloat,r,s,0,a.textureFormatFloat,a.textureTypeHalfFloat,null);let i=e.createFramebuffer();e.bindFramebuffer(e.FRAMEBUFFER,i),e.framebufferTexture2D(e.FRAMEBUFFER,e.COLOR_ATTACHMENT0,e.TEXTURE_2D,n,0);let o=e.checkFramebufferStatus(e.FRAMEBUFFER)===e.FRAMEBUFFER_COMPLETE;return e.bindTexture(e.TEXTURE_2D,null),e.bindFramebuffer(e.FRAMEBUFFER,null),e.deleteTexture(n),e.deleteFramebuffer(i),o}function mv(e){return e!==2?!1:Wn(e).fenceSync!=null}function Pu(e,t){Array.isArray(e)||(e=[e]),e.forEach(a=>{a!=null&&v.assert(a.dtype!=="complex64",()=>`${t} does not support complex64 tensors in the WebGL backend.`)})}var Se=W();Se.registerFlag("HAS_WEBGL",()=>Se.getNumber("WEBGL_VERSION")>0);Se.registerFlag("WEBGL_VERSION",()=>k1(2)?2:k1(1)?1:0);Se.registerFlag("WEBGL_CHECK_NUMERICAL_PROBLEMS",()=>!1);Se.registerFlag("WEBGL_BUFFER_SUPPORTED",()=>Se.get("WEBGL_VERSION")===2);Se.registerFlag("WEBGL_CPU_FORWARD",()=>!0);Se.registerFlag("WEBGL_FORCE_F16_TEXTURES",()=>!1);Se.registerFlag("WEBGL_PACK",()=>Se.getBool("HAS_WEBGL"));Se.registerFlag("WEBGL_PACK_NORMALIZATION",()=>Se.getBool("WEBGL_PACK"));Se.registerFlag("WEBGL_PACK_CLIP",()=>Se.getBool("WEBGL_PACK"));Se.registerFlag("WEBGL_PACK_DEPTHWISECONV",()=>Se.getBool("WEBGL_PACK"));Se.registerFlag("WEBGL_PACK_BINARY_OPERATIONS",()=>Se.getBool("WEBGL_PACK"));Se.registerFlag("WEBGL_PACK_UNARY_OPERATIONS",()=>Se.getBool("WEBGL_PACK"));Se.registerFlag("WEBGL_PACK_ARRAY_OPERATIONS",()=>Se.getBool("WEBGL_PACK"));Se.registerFlag("WEBGL_PACK_IMAGE_OPERATIONS",()=>Se.getBool("WEBGL_PACK"));Se.registerFlag("WEBGL_PACK_REDUCE",()=>Se.getBool("WEBGL_PACK"));Se.registerFlag("WEBGL_LAZILY_UNPACK",()=>Se.getBool("WEBGL_PACK"));Se.registerFlag("WEBGL_CONV_IM2COL",()=>Se.getBool("WEBGL_PACK"));Se.registerFlag("WEBGL_MAX_TEXTURE_SIZE",()=>dv(Se.getNumber("WEBGL_VERSION")));Se.registerFlag("WEBGL_MAX_TEXTURES_IN_SHADER",()=>pv(Se.getNumber("WEBGL_VERSION")));Se.registerFlag("WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_VERSION",()=>{let e=Se.getNumber("WEBGL_VERSION");return e===0?0:cv(e)});Se.registerFlag("WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_RELIABLE",()=>Se.getNumber("WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_VERSION")>0&&!Sp.isMobile());Se.registerFlag("WEBGL_RENDER_FLOAT32_CAPABLE",()=>hv(Se.getNumber("WEBGL_VERSION")));Se.registerFlag("WEBGL_RENDER_FLOAT32_ENABLED",()=>Se.getBool("WEBGL_FORCE_F16_TEXTURES")?!1:Se.getBool("WEBGL_RENDER_FLOAT32_CAPABLE"));Se.registerFlag("WEBGL_DOWNLOAD_FLOAT_ENABLED",()=>fv(Se.getNumber("WEBGL_VERSION")));Se.registerFlag("WEBGL_FENCE_API_ENABLED",()=>mv(Se.getNumber("WEBGL_VERSION")));Se.registerFlag("WEBGL_SIZE_UPLOAD_UNIFORM",()=>Se.getBool("WEBGL_RENDER_FLOAT32_ENABLED")?4:0);Se.registerFlag("WEBGL_DELETE_TEXTURE_THRESHOLD",()=>-1,e=>{if(e<0&&e!==-1)throw new Error(`WEBGL_DELETE_TEXTURE_THRESHOLD must be -1 (indicating never delete) or at least 0, but got ${e}.`)});Se.registerFlag("WEBGL_FLUSH_THRESHOLD",()=>Sp.isMobile()?1:-1,e=>{if(e<0&&e!==-1)throw new Error(`WEBGL_FLUSH_THRESHOLD must be -1 (indicating never manual flush) or at least 0, but got ${e}.`)});Se.registerFlag("CPU_HANDOFF_SIZE_THRESHOLD",()=>128);Se.registerFlag("WEBGL_USE_SHAPES_UNIFORMS",()=>!1);Se.registerFlag("TOPK_LAST_DIM_CPU_HANDOFF_SIZE_THRESHOLD",()=>1e5);Se.registerFlag("TOPK_K_CPU_HANDOFF_THRESHOLD",()=>128);Se.registerFlag("WEBGL_EXP_CONV",()=>!1);Se.registerFlag("SOFTWARE_WEBGL_ENABLED",()=>Se.getBool("IS_TEST"));Se.registerFlag("WEBGL_MAX_SIZE_FOR_NARROW_TEXTURE",()=>1/0);Se.registerFlag("WEBGL_AUTO_SQUARIFY_NARROW_TEXTURE_SHAPE",()=>!1);Se.registerFlag("WEBGL2_ISNAN_CUSTOM",()=>!1);Se.registerFlag("ENGINE_COMPILE_ONLY",()=>!1);function Ea(){let e,t,a,n,r,s,i,o,l,u;return W().getNumber("WEBGL_VERSION")===2?(e="#version 300 es",t="in",a="out",n="in",r="texture",s="outputColor",i="out vec4 outputColor;",o=W().getBool("WEBGL2_ISNAN_CUSTOM")?`
|
|
bool isnan_custom(float val) {
|
|
uint floatToUint = floatBitsToUint(val);
|
|
return (floatToUint & 0x7fffffffu) > 0x7f800000u;
|
|
}
|
|
|
|
bvec4 isnan_custom(vec4 val) {
|
|
return bvec4(isnan_custom(val.x),
|
|
isnan_custom(val.y), isnan_custom(val.z), isnan_custom(val.w));
|
|
}
|
|
|
|
#define isnan(value) isnan_custom(value)
|
|
`:"",l="",u=`
|
|
#define round(value) newRound(value)
|
|
int newRound(float value) {
|
|
return int(floor(value + 0.5));
|
|
}
|
|
|
|
ivec4 newRound(vec4 value) {
|
|
return ivec4(floor(value + vec4(0.5)));
|
|
}
|
|
`):(e="",t="attribute",a="varying",n="varying",r="texture2D",s="gl_FragColor",i="",o=`
|
|
#define isnan(value) isnan_custom(value)
|
|
bool isnan_custom(float val) {
|
|
return (val > 0. || val < 1. || val == 0.) ? false : true;
|
|
}
|
|
bvec4 isnan_custom(vec4 val) {
|
|
return bvec4(isnan(val.x), isnan(val.y), isnan(val.z), isnan(val.w));
|
|
}
|
|
`,l=`
|
|
uniform float INFINITY;
|
|
|
|
bool isinf(float val) {
|
|
return abs(val) == INFINITY;
|
|
}
|
|
bvec4 isinf(vec4 val) {
|
|
return equal(abs(val), vec4(INFINITY));
|
|
}
|
|
`,u=`
|
|
int round(float value) {
|
|
return int(floor(value + 0.5));
|
|
}
|
|
|
|
ivec4 round(vec4 value) {
|
|
return ivec4(floor(value + vec4(0.5)));
|
|
}
|
|
`),{version:e,attribute:t,varyingVs:a,varyingFs:n,texture2D:r,output:s,defineOutput:i,defineSpecialNaN:o,defineSpecialInf:l,defineRound:u}}function jo(e,t,a="index"){let n=v.computeStrides(t);return n.map((r,s)=>{let i=`int ${e[s]} = ${a} / ${r}`,o=s===n.length-1?`int ${e[s+1]} = ${a} - ${e[s]} * ${r}`:`index -= ${e[s]} * ${r}`;return`${i}; ${o};`}).join("")}function Jh(e,t,a="index"){let n=v.computeStrides(t);return n.map((r,s)=>{let i=`int ${e[s]} = ${a} / outShapeStrides[${s}]`,o=s===n.length-1?`int ${e[s+1]} = ${a} - ${e[s]} * outShapeStrides[${s}]`:`index -= ${e[s]} * outShapeStrides[${s}]`;return`${i}; ${o};`}).join("")}function qU(e,t){let a=e.length,n=e.map(s=>`${t}[${s}]`),r=new Array(a-1);r[a-2]=n[a-1];for(let s=a-3;s>=0;--s)r[s]=`(${r[s+1]} * ${n[s+1]})`;return r}function XU(e,t,a="index"){let n=e.map((s,i)=>i),r=qU(n,t);return r.map((s,i)=>{let o=`int ${e[i]} = ${a} / ${r[i]}`,l=i===r.length-1?`int ${e[i+1]} = ${a} - ${e[i]} * ${r[i]}`:`index -= ${e[i]} * ${r[i]}`;return`${o}; ${l};`}).join("")}function fg(e){let t=v.computeStrides(e).map(a=>a.toString());return`
|
|
int getFlatIndex(ivec3 coords) {
|
|
return coords.x * ${t[0]} + coords.y * ${t[1]} + coords.z;
|
|
}
|
|
`}function mg(){return`
|
|
int getFlatIndex(ivec3 coords) {
|
|
return coords.x * outShapeStrides[0] + coords.y * outShapeStrides[1] + coords.z;
|
|
}
|
|
`}var gv=`
|
|
const float FLOAT_MAX = 1.70141184e38;
|
|
const float FLOAT_MIN = 1.17549435e-38;
|
|
|
|
lowp vec4 encode_float(highp float v) {
|
|
if (isnan(v)) {
|
|
return vec4(255, 255, 255, 255);
|
|
}
|
|
|
|
highp float av = abs(v);
|
|
|
|
if(av < FLOAT_MIN) {
|
|
return vec4(0.0, 0.0, 0.0, 0.0);
|
|
} else if(v > FLOAT_MAX) {
|
|
return vec4(0.0, 0.0, 128.0, 127.0) / 255.0;
|
|
} else if(v < -FLOAT_MAX) {
|
|
return vec4(0.0, 0.0, 128.0, 255.0) / 255.0;
|
|
}
|
|
|
|
highp vec4 c = vec4(0,0,0,0);
|
|
|
|
highp float e = floor(log2(av));
|
|
highp float m = exp2(fract(log2(av))) - 1.0;
|
|
|
|
c[2] = floor(128.0 * m);
|
|
m -= c[2] / 128.0;
|
|
c[1] = floor(32768.0 * m);
|
|
m -= c[1] / 32768.0;
|
|
c[0] = floor(8388608.0 * m);
|
|
|
|
highp float ebias = e + 127.0;
|
|
c[3] = floor(ebias / 2.0);
|
|
ebias -= c[3] * 2.0;
|
|
c[2] += floor(ebias) * 128.0;
|
|
|
|
c[3] += 128.0 * step(0.0, -v);
|
|
|
|
return c / 255.0;
|
|
}
|
|
`,{getBroadcastDims:yv}=S;function KU(e,t,a){let n=[];if(e.forEach(d=>{let h=v.sizeFromShape(d.shapeInfo.logicalShape);if(d.shapeInfo.isUniform?n.push(`uniform float ${d.name}${h>1?`[${h}]`:""};`):(n.push(`uniform sampler2D ${d.name};`),n.push(`uniform int offset${d.name};`)),a.enableShapeUniforms){let{uniformShape:f}=gg(a.packedInputs,d.shapeInfo.logicalShape,d.shapeInfo.texShape);switch(f.length){case 1:n.push(`uniform int ${d.name}Shape;`);break;case 2:n.push(`uniform ivec2 ${d.name}Shape;`);break;case 3:n.push(`uniform ivec3 ${d.name}Shape;`);break;case 4:n.push(`uniform ivec4 ${d.name}Shape;`);break;default:break}n.push(`uniform ivec2 ${d.name}TexShape;`)}}),a.enableShapeUniforms){switch(t.logicalShape.length){case 1:n.push("uniform int outShape;");break;case 2:n.push("uniform ivec2 outShape;"),n.push("uniform int outShapeStrides;");break;case 3:n.push("uniform ivec3 outShape;"),n.push("uniform ivec2 outShapeStrides;");break;case 4:n.push("uniform ivec4 outShape;"),n.push("uniform ivec3 outShapeStrides;");break;default:break}n.push("uniform ivec2 outTexShape;")}a.customUniforms&&a.customUniforms.forEach(d=>{n.push(`uniform ${d.type} ${d.name}${d.arrayIndex?`[${d.arrayIndex}]`:""};`)});let r=n.join(`
|
|
`),s=e.map(d=>YU(d,t,a.packedInputs,a.enableShapeUniforms)).join(`
|
|
`),i=t.texShape,o=Ea(),l=QU(o),u,p,c=aG(o);return t.isPacked?(u=ZU(t.logicalShape,i,a.enableShapeUniforms),p=tG(o)):(u=JU(t.logicalShape,i,a.enableShapeUniforms),p=eG(o)),a.packedInputs&&(c+=iG),[c,l,p,r,u,s,a.userCode].join(`
|
|
`)}function Ou(e,t=!1){let a=e.shapeInfo.logicalShape;switch(a.length){case 0:return xG(e,t);case 1:return bG(e,t);case 2:return wG(e,t);case 3:return IG(e,t);case 4:return CG(e,t);case 5:return TG(e);case 6:return NG(e);default:throw new Error(`${a.length}-D input sampling is not yet supported`)}}function xv(e,t){switch(e.shapeInfo.logicalShape.length){case 0:return yG(e);case 1:return AG(e,t);case 2:return vG(e,t);case 3:return kG(e,t);default:return SG(e,t)}}function YU(e,t,a=!1,n){let r="";a?r+=xv(e,n):r+=Ou(e,n);let s=e.shapeInfo.logicalShape,i=t.logicalShape;return s.length<=i.length&&(a?r+=RG(e,t):r+=EG(e,t)),r}function ZU(e,t,a){switch(e.length){case 0:return Av();case 1:return oG(e,t,a);case 2:return mG(e,t,a);case 3:return uG(e,t,a);default:return pG(e,t,a)}}function JU(e,t,a){switch(e.length){case 0:return Av();case 1:return lG(e,t,a);case 2:return gG(e,t,a);case 3:return dG(e,t,a);case 4:return cG(e,t,a);case 5:return hG(e,t);case 6:return fG(e,t);default:throw new Error(`${e.length}-D output sampling is not yet supported`)}}function QU(e){return`
|
|
float sampleTexture(sampler2D textureSampler, vec2 uv) {
|
|
return ${e.texture2D}(textureSampler, uv).r;
|
|
}
|
|
`}function eG(e){return`
|
|
void setOutput(float val) {
|
|
${e.output} = vec4(val, 0, 0, 0);
|
|
}
|
|
`}function tG(e){return`
|
|
void setOutput(vec4 val) {
|
|
${e.output} = val;
|
|
}
|
|
`}function aG(e){return`${e.version}
|
|
precision highp float;
|
|
precision highp int;
|
|
precision highp sampler2D;
|
|
${e.varyingFs} vec2 resultUV;
|
|
${e.defineOutput}
|
|
const vec2 halfCR = vec2(0.5, 0.5);
|
|
|
|
struct ivec5
|
|
{
|
|
int x;
|
|
int y;
|
|
int z;
|
|
int w;
|
|
int u;
|
|
};
|
|
|
|
struct ivec6
|
|
{
|
|
int x;
|
|
int y;
|
|
int z;
|
|
int w;
|
|
int u;
|
|
int v;
|
|
};
|
|
|
|
uniform float NAN;
|
|
${e.defineSpecialNaN}
|
|
${e.defineSpecialInf}
|
|
${e.defineRound}
|
|
|
|
int imod(int x, int y) {
|
|
return x - y * (x / y);
|
|
}
|
|
|
|
int idiv(int a, int b, float sign) {
|
|
int res = a / b;
|
|
int mod = imod(a, b);
|
|
if (sign < 0. && mod != 0) {
|
|
res -= 1;
|
|
}
|
|
return res;
|
|
}
|
|
|
|
//Based on the work of Dave Hoskins
|
|
//https://www.shadertoy.com/view/4djSRW
|
|
#define HASHSCALE1 443.8975
|
|
float random(float seed){
|
|
vec2 p = resultUV * seed;
|
|
vec3 p3 = fract(vec3(p.xyx) * HASHSCALE1);
|
|
p3 += dot(p3, p3.yzx + 19.19);
|
|
return fract((p3.x + p3.y) * p3.z);
|
|
}
|
|
|
|
${nG}
|
|
${rG}
|
|
${sG}
|
|
`}var nG=`
|
|
vec2 uvFromFlat(int texNumR, int texNumC, int index) {
|
|
int texR = index / texNumC;
|
|
int texC = index - texR * texNumC;
|
|
return (vec2(texC, texR) + halfCR) / vec2(texNumC, texNumR);
|
|
}
|
|
vec2 packedUVfrom1D(int texNumR, int texNumC, int index) {
|
|
int texelIndex = index / 2;
|
|
int texR = texelIndex / texNumC;
|
|
int texC = texelIndex - texR * texNumC;
|
|
return (vec2(texC, texR) + halfCR) / vec2(texNumC, texNumR);
|
|
}
|
|
`,rG=`
|
|
vec2 packedUVfrom2D(int texelsInLogicalRow, int texNumR,
|
|
int texNumC, int row, int col) {
|
|
int texelIndex = (row / 2) * texelsInLogicalRow + (col / 2);
|
|
int texR = texelIndex / texNumC;
|
|
int texC = texelIndex - texR * texNumC;
|
|
return (vec2(texC, texR) + halfCR) / vec2(texNumC, texNumR);
|
|
}
|
|
`,sG=`
|
|
vec2 packedUVfrom3D(int texNumR, int texNumC,
|
|
int texelsInBatch, int texelsInLogicalRow, int b,
|
|
int row, int col) {
|
|
int index = b * texelsInBatch + (row / 2) * texelsInLogicalRow + (col / 2);
|
|
int texR = index / texNumC;
|
|
int texC = index - texR * texNumC;
|
|
return (vec2(texC, texR) + halfCR) / vec2(texNumC, texNumR);
|
|
}
|
|
`,iG=`
|
|
float getChannel(vec4 frag, vec2 innerDims) {
|
|
vec2 modCoord = mod(innerDims, 2.);
|
|
return modCoord.x == 0. ?
|
|
(modCoord.y == 0. ? frag.r : frag.g) :
|
|
(modCoord.y == 0. ? frag.b : frag.a);
|
|
}
|
|
float getChannel(vec4 frag, int dim) {
|
|
float modCoord = mod(float(dim), 2.);
|
|
return modCoord == 0. ? frag.r : frag.g;
|
|
}
|
|
`;function Av(){return`
|
|
int getOutputCoords() {
|
|
return 0;
|
|
}
|
|
`}function oG(e,t,a){let n=[Math.ceil(t[0]/2),Math.ceil(t[1]/2)];return n[0]===1?a?`
|
|
int getOutputCoords() {
|
|
return 2 * int(resultUV.x * ceil(float(outTexShape[1]) / 2.0));
|
|
}
|
|
`:`
|
|
int getOutputCoords() {
|
|
return 2 * int(resultUV.x * ${n[1]}.0);
|
|
}
|
|
`:n[1]===1?a?`
|
|
int getOutputCoords() {
|
|
return 2 * int(resultUV.y * ceil(float(outTexShape[0]) / 2.0));
|
|
}
|
|
`:`
|
|
int getOutputCoords() {
|
|
return 2 * int(resultUV.y * ${n[0]}.0);
|
|
}
|
|
`:a?`
|
|
int getOutputCoords() {
|
|
ivec2 packedTexShape = ivec2(ceil(float(outTexShape[0]) / 2.0), ceil(float(outTexShape[1]) / 2.0));
|
|
ivec2 resTexRC = ivec2(resultUV.yx *
|
|
vec2(packedTexShape[0], packedTexShape[1]));
|
|
return 2 * (resTexRC.x * packedTexShape[1] + resTexRC.y);
|
|
}
|
|
`:`
|
|
int getOutputCoords() {
|
|
ivec2 resTexRC = ivec2(resultUV.yx *
|
|
vec2(${n[0]}, ${n[1]}));
|
|
return 2 * (resTexRC.x * ${n[1]} + resTexRC.y);
|
|
}
|
|
`}function lG(e,t,a){return t[0]===1?a?`
|
|
int getOutputCoords() {
|
|
return int(resultUV.x * float(outTexShape[1]));
|
|
}
|
|
`:`
|
|
int getOutputCoords() {
|
|
return int(resultUV.x * ${t[1]}.0);
|
|
}
|
|
`:t[1]===1?a?`
|
|
int getOutputCoords() {
|
|
return int(resultUV.y * float(outTexShape[0]));
|
|
}
|
|
`:`
|
|
int getOutputCoords() {
|
|
return int(resultUV.y * ${t[0]}.0);
|
|
}
|
|
`:a?`
|
|
int getOutputCoords() {
|
|
ivec2 resTexRC = ivec2(resultUV.yx *
|
|
vec2(outTexShape[0], outTexShape[1]));
|
|
return resTexRC.x * outTexShape[1] + resTexRC.y;
|
|
}
|
|
`:`
|
|
int getOutputCoords() {
|
|
ivec2 resTexRC = ivec2(resultUV.yx *
|
|
vec2(${t[0]}, ${t[1]}));
|
|
return resTexRC.x * ${t[1]} + resTexRC.y;
|
|
}
|
|
`}function uG(e,t,a){if(a)return`
|
|
ivec3 getOutputCoords() {
|
|
ivec2 packedTexShape = ivec2(ceil(float(outTexShape[0]) / 2.0), ceil(float(outTexShape[1]) / 2.0));
|
|
int texelsInLogicalRow = int(ceil(float(outShape[2]) / 2.0));
|
|
int texelsInBatch = texelsInLogicalRow * int(ceil(float(outShape[1]) / 2.0));
|
|
ivec2 resTexRC = ivec2(resultUV.yx *
|
|
vec2(packedTexShape[0], packedTexShape[1]));
|
|
int index = resTexRC.x * packedTexShape[1] + resTexRC.y;
|
|
|
|
int b = index / texelsInBatch;
|
|
index -= b * texelsInBatch;
|
|
|
|
int r = 2 * (index / texelsInLogicalRow);
|
|
int c = imod(index, texelsInLogicalRow) * 2;
|
|
|
|
return ivec3(b, r, c);
|
|
}
|
|
`;let n=[Math.ceil(t[0]/2),Math.ceil(t[1]/2)],r=Math.ceil(e[2]/2),s=r*Math.ceil(e[1]/2);return`
|
|
ivec3 getOutputCoords() {
|
|
ivec2 resTexRC = ivec2(resultUV.yx *
|
|
vec2(${n[0]}, ${n[1]}));
|
|
int index = resTexRC.x * ${n[1]} + resTexRC.y;
|
|
|
|
int b = index / ${s};
|
|
index -= b * ${s};
|
|
|
|
int r = 2 * (index / ${r});
|
|
int c = imod(index, ${r}) * 2;
|
|
|
|
return ivec3(b, r, c);
|
|
}
|
|
`}function dG(e,t,a){if(a)return`
|
|
ivec3 getOutputCoords() {
|
|
ivec2 resTexRC = ivec2(resultUV.yx *
|
|
vec2(outTexShape[0], outTexShape[1]));
|
|
int index = resTexRC.x * outTexShape[1] + resTexRC.y;
|
|
${Jh(["r","c","d"],e)}
|
|
return ivec3(r, c, d);
|
|
}
|
|
`;let n=jo(["r","c","d"],e);return`
|
|
ivec3 getOutputCoords() {
|
|
ivec2 resTexRC = ivec2(resultUV.yx *
|
|
vec2(${t[0]}, ${t[1]}));
|
|
int index = resTexRC.x * ${t[1]} + resTexRC.y;
|
|
${n}
|
|
return ivec3(r, c, d);
|
|
}
|
|
`}function pG(e,t,a){if(a)return`
|
|
ivec4 getOutputCoords() {
|
|
ivec2 packedTexShape = ivec2(ceil(float(outTexShape[0]) / 2.0), ceil(float(outTexShape[1]) / 2.0));
|
|
ivec2 resTexRC = ivec2(resultUV.yx *
|
|
vec2(packedTexShape[0], packedTexShape[1]));
|
|
int index = resTexRC.x * packedTexShape[1] + resTexRC.y;
|
|
|
|
int texelsInLogicalRow = int(ceil(float(outShape[3]) / 2.0));
|
|
int texelsInBatch = texelsInLogicalRow * int(ceil(float(outShape[2]) / 2.0));
|
|
int texelsInBatchN = texelsInBatch * outShape[1];
|
|
|
|
int b2 = index / texelsInBatchN;
|
|
index -= b2 * texelsInBatchN;
|
|
|
|
int b = index / texelsInBatch;
|
|
index -= b * texelsInBatch;
|
|
|
|
int r = 2 * (index / texelsInLogicalRow);
|
|
int c = imod(index, texelsInLogicalRow) * 2;
|
|
|
|
return ivec4(b2, b, r, c);
|
|
}
|
|
`;let n=[Math.ceil(t[0]/2),Math.ceil(t[1]/2)],r=Math.ceil(e[e.length-1]/2),s=r*Math.ceil(e[e.length-2]/2),i=s,o="",l="b, r, c";for(let u=2;u<e.length-1;u++)i*=e[e.length-u-1],o=`
|
|
int b${u} = index / ${i};
|
|
index -= b${u} * ${i};
|
|
`+o,l=`b${u}, `+l;return`
|
|
ivec${e.length} getOutputCoords() {
|
|
ivec2 resTexRC = ivec2(resultUV.yx *
|
|
vec2(${n[0]}, ${n[1]}));
|
|
int index = resTexRC.x * ${n[1]} + resTexRC.y;
|
|
|
|
${o}
|
|
|
|
int b = index / ${s};
|
|
index -= b * ${s};
|
|
|
|
int r = 2 * (index / ${r});
|
|
int c = imod(index, ${r}) * 2;
|
|
|
|
return ivec${e.length}(${l});
|
|
}
|
|
`}function cG(e,t,a){if(a)return`
|
|
ivec4 getOutputCoords() {
|
|
ivec2 resTexRC = ivec2(resultUV.yx *
|
|
vec2(outTexShape[0], outTexShape[1]));
|
|
int index = resTexRC.x * outTexShape[1] + resTexRC.y;
|
|
${Jh(["r","c","d","d2"],e)}
|
|
return ivec4(r, c, d, d2);
|
|
}
|
|
`;let n=jo(["r","c","d","d2"],e);return`
|
|
ivec4 getOutputCoords() {
|
|
ivec2 resTexRC = ivec2(resultUV.yx *
|
|
vec2(${t[0]}, ${t[1]}));
|
|
int index = resTexRC.x * ${t[1]} + resTexRC.y;
|
|
${n}
|
|
return ivec4(r, c, d, d2);
|
|
}
|
|
`}function hG(e,t){let a=jo(["r","c","d","d2","d3"],e);return`
|
|
ivec5 getOutputCoords() {
|
|
ivec2 resTexRC = ivec2(resultUV.yx * vec2(${t[0]},
|
|
${t[1]}));
|
|
|
|
int index = resTexRC.x * ${t[1]} + resTexRC.y;
|
|
|
|
${a}
|
|
|
|
ivec5 outShape = ivec5(r, c, d, d2, d3);
|
|
return outShape;
|
|
}
|
|
`}function fG(e,t){let a=jo(["r","c","d","d2","d3","d4"],e);return`
|
|
ivec6 getOutputCoords() {
|
|
ivec2 resTexRC = ivec2(resultUV.yx *
|
|
vec2(${t[0]}, ${t[1]}));
|
|
int index = resTexRC.x * ${t[1]} + resTexRC.y;
|
|
|
|
${a}
|
|
|
|
ivec6 result = ivec6(r, c, d, d2, d3, d4);
|
|
return result;
|
|
}
|
|
`}function mG(e,t,a){let n=[Math.ceil(t[0]/2),Math.ceil(t[1]/2)];if(v.arraysEqual(e,t))return a?`
|
|
ivec2 getOutputCoords() {
|
|
ivec2 packedTexShape = ivec2(ceil(float(outTexShape[0]) / 2.0), ceil(float(outTexShape[1]) / 2.0));
|
|
return 2 * ivec2(resultUV.yx * vec2(packedTexShape[0], packedTexShape[1]));
|
|
}
|
|
`:`
|
|
ivec2 getOutputCoords() {
|
|
return 2 * ivec2(resultUV.yx * vec2(${n[0]}, ${n[1]}));
|
|
}
|
|
`;let r=Math.ceil(e[1]/2);return a?`
|
|
ivec2 getOutputCoords() {
|
|
ivec2 packedTexShape = ivec2(ceil(float(outTexShape[0]) / 2.0), ceil(float(outTexShape[1]) / 2.0));
|
|
int texelsInLogicalRow = int(ceil(float(outShape[1]) / 2.0));
|
|
ivec2 resTexRC = ivec2(resultUV.yx *
|
|
vec2(packedTexShape[0], packedTexShape[1]));
|
|
|
|
int index = resTexRC.x * packedTexShape[1] + resTexRC.y;
|
|
int r = 2 * (index / texelsInLogicalRow);
|
|
int c = imod(index, texelsInLogicalRow) * 2;
|
|
|
|
return ivec2(r, c);
|
|
}
|
|
`:`
|
|
ivec2 getOutputCoords() {
|
|
ivec2 resTexRC = ivec2(resultUV.yx *
|
|
vec2(${n[0]}, ${n[1]}));
|
|
|
|
int index = resTexRC.x * ${n[1]} + resTexRC.y;
|
|
int r = 2 * (index / ${r});
|
|
int c = imod(index, ${r}) * 2;
|
|
|
|
return ivec2(r, c);
|
|
}
|
|
`}function gG(e,t,a){return v.arraysEqual(e,t)?a?`
|
|
ivec2 getOutputCoords() {
|
|
return ivec2(resultUV.yx * vec2(outTexShape[0], outTexShape[1]));
|
|
}
|
|
`:`
|
|
ivec2 getOutputCoords() {
|
|
return ivec2(resultUV.yx * vec2(${t[0]}, ${t[1]}));
|
|
}
|
|
`:e[1]===1?a?`
|
|
ivec2 getOutputCoords() {
|
|
ivec2 resTexRC = ivec2(resultUV.yx *
|
|
vec2(outTexShape[0], outTexShape[1]));
|
|
int index = resTexRC.x * outTexShape[1] + resTexRC.y;
|
|
return ivec2(index, 0);
|
|
}
|
|
`:`
|
|
ivec2 getOutputCoords() {
|
|
ivec2 resTexRC = ivec2(resultUV.yx *
|
|
vec2(${t[0]}, ${t[1]}));
|
|
int index = resTexRC.x * ${t[1]} + resTexRC.y;
|
|
return ivec2(index, 0);
|
|
}
|
|
`:e[0]===1?a?`
|
|
ivec2 getOutputCoords() {
|
|
ivec2 resTexRC = ivec2(resultUV.yx *
|
|
vec2(outTexShape[0], outTexShape[1]));
|
|
int index = resTexRC.x * outTexShape[1] + resTexRC.y;
|
|
return ivec2(0, index);
|
|
}
|
|
`:`
|
|
ivec2 getOutputCoords() {
|
|
ivec2 resTexRC = ivec2(resultUV.yx *
|
|
vec2(${t[0]}, ${t[1]}));
|
|
int index = resTexRC.x * ${t[1]} + resTexRC.y;
|
|
return ivec2(0, index);
|
|
}
|
|
`:a?`
|
|
ivec2 getOutputCoords() {
|
|
ivec2 resTexRC = ivec2(resultUV.yx *
|
|
vec2(outTexShape[0], outTexShape[1]));
|
|
int index = resTexRC.x * outTexShape[1] + resTexRC.y;
|
|
int r = index / outShape[1];
|
|
int c = index - r * outShape[1];
|
|
return ivec2(r, c);
|
|
}
|
|
`:`
|
|
ivec2 getOutputCoords() {
|
|
ivec2 resTexRC = ivec2(resultUV.yx *
|
|
vec2(${t[0]}, ${t[1]}));
|
|
int index = resTexRC.x * ${t[1]} + resTexRC.y;
|
|
int r = index / ${e[1]};
|
|
int c = index - r * ${e[1]};
|
|
return ivec2(r, c);
|
|
}
|
|
`}function qo(e){return`offset${e}`}function yG(e){let t=e.name,a="get"+t.charAt(0).toUpperCase()+t.slice(1),n=Ea();return`
|
|
vec4 ${a}() {
|
|
return ${n.texture2D}(${t}, halfCR);
|
|
}
|
|
`}function xG(e,t){let a=e.name,n="get"+a.charAt(0).toUpperCase()+a.slice(1);if(e.shapeInfo.isUniform)return`float ${n}() {return ${a};}`;let[r,s]=e.shapeInfo.texShape;if(r===1&&s===1)return`
|
|
float ${n}() {
|
|
return sampleTexture(${a}, halfCR);
|
|
}
|
|
`;let i=qo(a);if(t)return`
|
|
float ${n}() {
|
|
vec2 uv = uvFromFlat(${a}TexShape[0], ${a}TexShape[1], ${i});
|
|
return sampleTexture(${a}, uv);
|
|
}
|
|
`;let[o,l]=e.shapeInfo.texShape;return`
|
|
float ${n}() {
|
|
vec2 uv = uvFromFlat(${o}, ${l}, ${i});
|
|
return sampleTexture(${a}, uv);
|
|
}
|
|
`}function AG(e,t){let a=e.name,n="get"+a.charAt(0).toUpperCase()+a.slice(1),r=e.shapeInfo.texShape,s=Ea();if(t)return`
|
|
vec4 ${n}(int index) {
|
|
ivec2 packedTexShape = ivec2(ceil(float(${a}TexShape[0]) / 2.0), ceil(float(${a}TexShape[1]) / 2.0));
|
|
vec2 uv = packedUVfrom1D(
|
|
packedTexShape[0], packedTexShape[1], index);
|
|
return ${s.texture2D}(${a}, uv);
|
|
}
|
|
`;let i=[Math.ceil(r[0]/2),Math.ceil(r[1]/2)];return`
|
|
vec4 ${n}(int index) {
|
|
vec2 uv = packedUVfrom1D(
|
|
${i[0]}, ${i[1]}, index);
|
|
return ${s.texture2D}(${a}, uv);
|
|
}
|
|
`}function bG(e,t){let a=e.name,n="get"+a.charAt(0).toUpperCase()+a.slice(1);if(e.shapeInfo.isUniform)return`
|
|
float ${n}(int index) {
|
|
${zu(e)}
|
|
}
|
|
`;let r=e.shapeInfo.texShape,s=r[0],i=r[1];if(i===1&&s===1)return`
|
|
float ${n}(int index) {
|
|
return sampleTexture(${a}, halfCR);
|
|
}
|
|
`;let o=qo(a);return i===1?t?`
|
|
float ${n}(int index) {
|
|
vec2 uv = vec2(0.5, (float(index + ${o}) + 0.5) / float(${a}TexShape[0]));
|
|
return sampleTexture(${a}, uv);
|
|
}
|
|
`:`
|
|
float ${n}(int index) {
|
|
vec2 uv = vec2(0.5, (float(index + ${o}) + 0.5) / ${s}.0);
|
|
return sampleTexture(${a}, uv);
|
|
}
|
|
`:s===1?t?`
|
|
float ${n}(int index) {
|
|
vec2 uv = vec2((float(index + ${o}) + 0.5) / float(${a}TexShape[1]), 0.5);
|
|
return sampleTexture(${a}, uv);
|
|
}
|
|
`:`
|
|
float ${n}(int index) {
|
|
vec2 uv = vec2((float(index + ${o}) + 0.5) / ${i}.0, 0.5);
|
|
return sampleTexture(${a}, uv);
|
|
}
|
|
`:t?`
|
|
float ${n}(int index) {
|
|
vec2 uv = uvFromFlat(${a}TexShape[0], ${a}TexShape[1], index + ${o});
|
|
return sampleTexture(${a}, uv);
|
|
}
|
|
`:`
|
|
float ${n}(int index) {
|
|
vec2 uv = uvFromFlat(${s}, ${i}, index + ${o});
|
|
return sampleTexture(${a}, uv);
|
|
}
|
|
`}function vG(e,t){let a=e.shapeInfo.logicalShape,n=e.name,r="get"+n.charAt(0).toUpperCase()+n.slice(1),s=e.shapeInfo.texShape,i=s[0],o=s[1],l=Ea();if(s!=null&&v.arraysEqual(a,s))return t?`
|
|
vec4 ${r}(int row, int col) {
|
|
vec2 uv = (vec2(col, row) + halfCR) / vec2(${n}TexShape[1], ${n}TexShape[0]);
|
|
|
|
return ${l.texture2D}(${n}, uv);
|
|
}
|
|
`:`
|
|
vec4 ${r}(int row, int col) {
|
|
vec2 uv = (vec2(col, row) + halfCR) / vec2(${o}.0, ${i}.0);
|
|
|
|
return ${l.texture2D}(${n}, uv);
|
|
}
|
|
`;if(t)return`
|
|
vec4 ${r}(int row, int col) {
|
|
ivec2 packedTexShape = ivec2(ceil(float(${n}TexShape[0]) / 2.0), ceil(float(${n}TexShape[1]) / 2.0));
|
|
int valuesPerRow = int(ceil(float(${n}Shape[1]) / 2.0));
|
|
vec2 uv = packedUVfrom2D(valuesPerRow, packedTexShape[0], packedTexShape[1], row, col);
|
|
return ${l.texture2D}(${n}, uv);
|
|
}
|
|
`;let u=[Math.ceil(s[0]/2),Math.ceil(s[1]/2)],p=Math.ceil(a[1]/2);return`
|
|
vec4 ${r}(int row, int col) {
|
|
vec2 uv = packedUVfrom2D(${p}, ${u[0]}, ${u[1]}, row, col);
|
|
return ${l.texture2D}(${n}, uv);
|
|
}
|
|
`}function wG(e,t){let a=e.shapeInfo.logicalShape,n=e.name,r="get"+n.charAt(0).toUpperCase()+n.slice(1),s=e.shapeInfo.texShape;if(s!=null&&v.arraysEqual(a,s)){if(t)return`
|
|
float ${r}(int row, int col) {
|
|
vec2 uv = (vec2(col, row) + halfCR) / vec2(${n}TexShape[1], ${n}TexShape[0]);
|
|
return sampleTexture(${n}, uv);
|
|
}
|
|
`;let d=s[0],h=s[1];return`
|
|
float ${r}(int row, int col) {
|
|
vec2 uv = (vec2(col, row) + halfCR) / vec2(${h}.0, ${d}.0);
|
|
return sampleTexture(${n}, uv);
|
|
}
|
|
`}let{newShape:i,keptDims:o}=v.squeezeShape(a),l=i;if(l.length<a.length){let d=Lu(e,l),h=["row","col"];return`
|
|
${Ou(d,t)}
|
|
float ${r}(int row, int col) {
|
|
return ${r}(${Wu(h,o)});
|
|
}
|
|
`}if(e.shapeInfo.isUniform)return`
|
|
float ${r}(int row, int col) {
|
|
int index = round(dot(vec2(row, col), vec2(${a[1]}, 1)));
|
|
${zu(e)}
|
|
}
|
|
`;let u=s[0],p=s[1],c=qo(n);return p===1?t?`
|
|
float ${r}(int row, int col) {
|
|
float index = dot(vec3(row, col, ${c}), vec3(${n}Shape[1], 1, 1));
|
|
vec2 uv = vec2(0.5, (index + 0.5) / float(${n}TexShape[0]));
|
|
return sampleTexture(${n}, uv);
|
|
}
|
|
`:`
|
|
float ${r}(int row, int col) {
|
|
float index = dot(vec3(row, col, ${c}), vec3(${a[1]}, 1, 1));
|
|
vec2 uv = vec2(0.5, (index + 0.5) / ${u}.0);
|
|
return sampleTexture(${n}, uv);
|
|
}
|
|
`:u===1?t?`
|
|
float ${r}(int row, int col) {
|
|
float index = dot(vec3(row, col, ${c}), vec3(${n}Shape[1], 1, 1));
|
|
vec2 uv = vec2((index + 0.5) / float(${n}TexShape[1]), 0.5);
|
|
return sampleTexture(${n}, uv);
|
|
}
|
|
`:`
|
|
float ${r}(int row, int col) {
|
|
float index = dot(vec3(row, col, ${c}), vec3(${a[1]}, 1, 1));
|
|
vec2 uv = vec2((index + 0.5) / ${p}.0, 0.5);
|
|
return sampleTexture(${n}, uv);
|
|
}
|
|
`:t?`
|
|
float ${r}(int row, int col) {
|
|
// Explicitly use integer operations as dot() only works on floats.
|
|
int index = row * ${n}Shape[1] + col + ${c};
|
|
vec2 uv = uvFromFlat(${n}TexShape[0], ${n}TexShape[1], index);
|
|
return sampleTexture(${n}, uv);
|
|
}
|
|
`:`
|
|
float ${r}(int row, int col) {
|
|
// Explicitly use integer operations as dot() only works on floats.
|
|
int index = row * ${a[1]} + col + ${c};
|
|
vec2 uv = uvFromFlat(${u}, ${p}, index);
|
|
return sampleTexture(${n}, uv);
|
|
}
|
|
`}function kG(e,t){let a=e.shapeInfo.logicalShape,n=e.name,r="get"+n.charAt(0).toUpperCase()+n.slice(1),s=e.shapeInfo.texShape,i=[Math.ceil(s[0]/2),Math.ceil(s[1]/2)];if(a[0]===1){let d=a.slice(1),h=[1,2],f=Lu(e,d),m=["b","row","col"];return`
|
|
${xv(f,t)}
|
|
vec4 ${r}(int b, int row, int col) {
|
|
return ${r}(${Wu(m,h)});
|
|
}
|
|
`}let o=Ea();if(t)return`
|
|
vec4 ${r}(int b, int row, int col) {
|
|
ivec2 packedTexShape = ivec2(ceil(float(${n}TexShape[0]) / 2.0), ceil(float(${n}TexShape[1]) / 2.0));
|
|
int valuesPerRow = int(ceil(float(${n}Shape[2]) / 2.0));
|
|
int texelsInBatch = valuesPerRow * int(ceil(float(${n}Shape[1]) / 2.0));
|
|
vec2 uv = packedUVfrom3D(
|
|
packedTexShape[0], packedTexShape[1], texelsInBatch, valuesPerRow, b, row, col);
|
|
return ${o.texture2D}(${n}, uv);
|
|
}
|
|
`;let l=i[0],u=i[1],p=Math.ceil(a[2]/2),c=p*Math.ceil(a[1]/2);return`
|
|
vec4 ${r}(int b, int row, int col) {
|
|
vec2 uv = packedUVfrom3D(
|
|
${l}, ${u}, ${c}, ${p}, b, row, col);
|
|
return ${o.texture2D}(${n}, uv);
|
|
}
|
|
`}function IG(e,t){let a=e.shapeInfo.logicalShape,n=e.name,r="get"+n.charAt(0).toUpperCase()+n.slice(1),s=a[1]*a[2],i=a[2],{newShape:o,keptDims:l}=v.squeezeShape(a),u=o;if(u.length<a.length){let m=Lu(e,u),g=["row","col","depth"];return`
|
|
${Ou(m,t)}
|
|
float ${r}(int row, int col, int depth) {
|
|
return ${r}(${Wu(g,l)});
|
|
}
|
|
`}if(e.shapeInfo.isUniform)return`
|
|
float ${r}(int row, int col, int depth) {
|
|
int index = round(dot(vec3(row, col, depth),
|
|
vec3(${s}, ${i}, 1)));
|
|
${zu(e)}
|
|
}
|
|
`;let p=e.shapeInfo.texShape,c=p[0],d=p[1],h=e.shapeInfo.flatOffset;if(d===s&&h==null)return t?`
|
|
float ${r}(int row, int col, int depth) {
|
|
int stride1 = ${n}Shape[2];
|
|
float texR = float(row);
|
|
float texC = dot(vec2(col, depth), vec2(stride1, 1));
|
|
vec2 uv = (vec2(texC, texR) + halfCR) /
|
|
vec2(${n}TexShape[1], ${n}TexShape[0]);
|
|
return sampleTexture(${n}, uv);
|
|
}
|
|
`:`
|
|
float ${r}(int row, int col, int depth) {
|
|
float texR = float(row);
|
|
float texC = dot(vec2(col, depth), vec2(${i}, 1));
|
|
vec2 uv = (vec2(texC, texR) + halfCR) /
|
|
vec2(${d}.0, ${c}.0);
|
|
return sampleTexture(${n}, uv);
|
|
}
|
|
`;if(d===i&&h==null)return t?`
|
|
float ${r}(int row, int col, int depth) {
|
|
float texR = dot(vec2(row, col), vec2(${n}Shape[1], 1));
|
|
float texC = float(depth);
|
|
vec2 uv = (vec2(texC, texR) + halfCR) / vec2(${n}TexShape[1], ${n}TexShape[0]);
|
|
return sampleTexture(${n}, uv);
|
|
}
|
|
`:`
|
|
float ${r}(int row, int col, int depth) {
|
|
float texR = dot(vec2(row, col), vec2(${a[1]}, 1));
|
|
float texC = float(depth);
|
|
vec2 uv = (vec2(texC, texR) + halfCR) / vec2(${d}.0, ${c}.0);
|
|
return sampleTexture(${n}, uv);
|
|
}
|
|
`;let f=qo(n);return t?`
|
|
float ${r}(int row, int col, int depth) {
|
|
// Explicitly use integer operations as dot() only works on floats.
|
|
int stride0 = ${n}Shape[1] * ${n}Shape[2];
|
|
int stride1 = ${n}Shape[2];
|
|
int index = row * stride0 + col * stride1 + depth + ${f};
|
|
vec2 uv = uvFromFlat(${n}TexShape[0], ${n}TexShape[1], index);
|
|
return sampleTexture(${n}, uv);
|
|
}
|
|
`:`
|
|
float ${r}(int row, int col, int depth) {
|
|
// Explicitly use integer operations as dot() only works on floats.
|
|
int index = row * ${s} + col * ${i} + depth + ${f};
|
|
vec2 uv = uvFromFlat(${c}, ${d}, index);
|
|
return sampleTexture(${n}, uv);
|
|
}
|
|
`}function SG(e,t){let a=e.name,n="get"+a.charAt(0).toUpperCase()+a.slice(1),r=Ea();if(t)return`
|
|
vec4 ${n}(int b2, int b, int row, int col) {
|
|
int valuesPerRow = int(ceil(float(${a}Shape[3]) / 2.0));
|
|
int texelsInBatch = valuesPerRow * int(ceil(float(${a}Shape[2]) / 2.0));
|
|
int index = b * texelsInBatch + (row / 2) * valuesPerRow + (col / 2);
|
|
texelsInBatch *= ${a}Shape[1];
|
|
index = b2 * texelsInBatch + index;
|
|
ivec2 packedTexShape = ivec2(ceil(float(${a}TexShape[0]) / 2.0), ceil(float(${a}TexShape[1]) / 2.0));
|
|
int texR = index / packedTexShape[1];
|
|
int texC = index - texR * packedTexShape[1];
|
|
vec2 uv = (vec2(texC, texR) + halfCR) / vec2(packedTexShape[1], packedTexShape[0]); return ${r.texture2D}(${a}, uv);
|
|
}
|
|
`;let s=e.shapeInfo.logicalShape,i=s.length,o=e.shapeInfo.texShape,l=[Math.ceil(o[0]/2),Math.ceil(o[1]/2)],u=l[0],p=l[1],c=Math.ceil(s[i-1]/2),d=c*Math.ceil(s[i-2]/2),h="int b, int row, int col",f=`b * ${d} + (row / 2) * ${c} + (col / 2)`;for(let m=2;m<i-1;m++)h=`int b${m}, `+h,d*=s[i-m-1],f=`b${m} * ${d} + `+f;return`
|
|
vec4 ${n}(${h}) {
|
|
int index = ${f};
|
|
int texR = index / ${p};
|
|
int texC = index - texR * ${p};
|
|
vec2 uv = (vec2(texC, texR) + halfCR) / vec2(${p}, ${u});
|
|
return ${r.texture2D}(${a}, uv);
|
|
}
|
|
`}function CG(e,t){let a=e.shapeInfo.logicalShape,n=e.name,r="get"+n.charAt(0).toUpperCase()+n.slice(1),s=a[3],i=a[2]*s,o=a[1]*i,{newShape:l,keptDims:u}=v.squeezeShape(a);if(l.length<a.length){let x=Lu(e,l),A=["row","col","depth","depth2"];return`
|
|
${Ou(x,t)}
|
|
float ${r}(int row, int col, int depth, int depth2) {
|
|
return ${r}(${Wu(A,u)});
|
|
}
|
|
`}if(e.shapeInfo.isUniform)return`
|
|
float ${r}(int row, int col, int depth, int depth2) {
|
|
int index = round(dot(vec4(row, col, depth, depth2),
|
|
vec4(${o}, ${i}, ${s}, 1)));
|
|
${zu(e)}
|
|
}
|
|
`;let p=e.shapeInfo.flatOffset,c=e.shapeInfo.texShape,d=c[0],h=c[1],f=`int stride2 = ${n}Shape[3];`,m=`int stride1 = ${n}Shape[2] * stride2;`,g=`int stride0 = ${n}Shape[1] * stride1;`;if(h===o&&p==null)return t?`
|
|
float ${r}(int row, int col, int depth, int depth2) {
|
|
${f}
|
|
${m}
|
|
float texR = float(row);
|
|
float texC =
|
|
dot(vec3(col, depth, depth2),
|
|
vec3(stride1, stride2, 1));
|
|
vec2 uv = (vec2(texC, texR) + halfCR) /
|
|
vec2(${n}TexShape[1], ${n}TexShape[0]);
|
|
return sampleTexture(${n}, uv);
|
|
}
|
|
`:`
|
|
float ${r}(int row, int col, int depth, int depth2) {
|
|
float texR = float(row);
|
|
float texC =
|
|
dot(vec3(col, depth, depth2),
|
|
vec3(${i}, ${s}, 1));
|
|
vec2 uv = (vec2(texC, texR) + halfCR) /
|
|
vec2(${h}.0, ${d}.0);
|
|
return sampleTexture(${n}, uv);
|
|
}
|
|
`;if(h===s&&p==null)return t?`
|
|
float ${r}(int row, int col, int depth, int depth2) {
|
|
float texR = dot(vec3(row, col, depth),
|
|
vec3(${n}Shape[1] * ${n}Shape[2], ${n}Shape[2], 1));
|
|
float texC = float(depth2);
|
|
vec2 uv = (vec2(texC, texR) + halfCR) /
|
|
vec2(${n}TexShape[1], ${n}TexShape[0]);
|
|
return sampleTexture(${n}, uv);
|
|
}
|
|
`:`
|
|
float ${r}(int row, int col, int depth, int depth2) {
|
|
float texR = dot(vec3(row, col, depth),
|
|
vec3(${a[1]*a[2]}, ${a[2]}, 1));
|
|
float texC = float(depth2);
|
|
vec2 uv = (vec2(texC, texR) + halfCR) /
|
|
vec2(${h}.0, ${d}.0);
|
|
return sampleTexture(${n}, uv);
|
|
}
|
|
`;let y=qo(n);return t?`
|
|
float ${r}(int row, int col, int depth, int depth2) {
|
|
// Explicitly use integer operations as dot() only works on floats.
|
|
${f}
|
|
${m}
|
|
${g}
|
|
int index = row * stride0 + col * stride1 +
|
|
depth * stride2 + depth2;
|
|
vec2 uv = uvFromFlat(${n}TexShape[0], ${n}TexShape[1], index + ${y});
|
|
return sampleTexture(${n}, uv);
|
|
}
|
|
`:`
|
|
float ${r}(int row, int col, int depth, int depth2) {
|
|
// Explicitly use integer operations as dot() only works on floats.
|
|
int index = row * ${o} + col * ${i} +
|
|
depth * ${s} + depth2;
|
|
vec2 uv = uvFromFlat(${d}, ${h}, index + ${y});
|
|
return sampleTexture(${n}, uv);
|
|
}
|
|
`}function TG(e){let t=e.shapeInfo.logicalShape,a=e.name,n="get"+a.charAt(0).toUpperCase()+a.slice(1),r=t[4],s=t[3]*r,i=t[2]*s,o=t[1]*i,{newShape:l,keptDims:u}=v.squeezeShape(t);if(l.length<t.length){let m=Lu(e,l),g=["row","col","depth","depth2","depth3"];return`
|
|
${Ou(m)}
|
|
float ${n}(int row, int col, int depth, int depth2, int depth3) {
|
|
return ${n}(${Wu(g,u)});
|
|
}
|
|
`}if(e.shapeInfo.isUniform)return`
|
|
float ${n}(int row, int col, int depth, int depth2, int depth3) {
|
|
float index = dot(
|
|
vec4(row, col, depth, depth2),
|
|
vec4(${o}, ${i}, ${s}, ${r})) +
|
|
depth3;
|
|
${zu(e)}
|
|
}
|
|
`;let p=e.shapeInfo.flatOffset,c=e.shapeInfo.texShape,d=c[0],h=c[1];if(h===o&&p==null)return`
|
|
float ${n}(int row, int col, int depth, int depth2, int depth3) {
|
|
int texR = row;
|
|
float texC = dot(vec4(col, depth, depth2, depth3),
|
|
vec4(${i}, ${s}, ${r}, 1));
|
|
vec2 uv = (vec2(texC, texR) + halfCR) /
|
|
vec2(${h}.0, ${d}.0);
|
|
return sampleTexture(${a}, uv);
|
|
}
|
|
`;if(h===r&&p==null)return`
|
|
float ${n}(int row, int col, int depth, int depth2, int depth3) {
|
|
float texR = dot(
|
|
vec4(row, col, depth, depth2),
|
|
vec4(${t[1]*t[2]*t[3]},
|
|
${t[2]*t[3]}, ${t[3]}, 1));
|
|
int texC = depth3;
|
|
vec2 uv = (vec2(texC, texR) + halfCR) /
|
|
vec2(${h}.0, ${d}.0);
|
|
return sampleTexture(${a}, uv);
|
|
}
|
|
`;let f=qo(a);return`
|
|
float ${n}(int row, int col, int depth, int depth2, int depth3) {
|
|
// Explicitly use integer operations as dot() only works on floats.
|
|
int index = row * ${o} + col * ${i} + depth * ${s} +
|
|
depth2 * ${r} + depth3 + ${f};
|
|
vec2 uv = uvFromFlat(${d}, ${h}, index);
|
|
return sampleTexture(${a}, uv);
|
|
}
|
|
`}function NG(e){let t=e.shapeInfo.logicalShape,a=e.name,n="get"+a.charAt(0).toUpperCase()+a.slice(1),{newShape:r,keptDims:s}=v.squeezeShape(t);if(r.length<t.length){let g=Lu(e,r),y=["row","col","depth","depth2","depth3","depth4"];return`
|
|
${Ou(g)}
|
|
float ${n}(int row, int col, int depth,
|
|
int depth2, int depth3, int depth4) {
|
|
return ${n}(${Wu(y,s)});
|
|
}
|
|
`}let i=t[5],o=t[4]*i,l=t[3]*o,u=t[2]*l,p=t[1]*u;if(e.shapeInfo.isUniform)return`
|
|
float ${n}(int row, int col, int depth,
|
|
int depth2, int depth3, int depth4) {
|
|
int index = round(dot(
|
|
vec4(row, col, depth, depth2),
|
|
vec4(${p}, ${u}, ${l}, ${o})) +
|
|
dot(
|
|
vec2(depth3, depth4),
|
|
vec2(${i}, 1)));
|
|
${zu(e)}
|
|
}
|
|
`;let c=e.shapeInfo.flatOffset,d=e.shapeInfo.texShape,h=d[0],f=d[1];if(f===p&&c==null)return`
|
|
float ${n}(int row, int col, int depth,
|
|
int depth2, int depth3, int depth4) {
|
|
int texR = row;
|
|
float texC = dot(vec4(col, depth, depth2, depth3),
|
|
vec4(${u}, ${l}, ${o}, ${i})) +
|
|
float(depth4);
|
|
vec2 uv = (vec2(texC, texR) + halfCR) /
|
|
vec2(${f}.0, ${h}.0);
|
|
return sampleTexture(${a}, uv);
|
|
}
|
|
`;if(f===i&&c==null)return`
|
|
float ${n}(int row, int col, int depth,
|
|
int depth2, int depth3, int depth4) {
|
|
float texR = dot(vec4(row, col, depth, depth2),
|
|
vec4(${t[1]*t[2]*t[3]*t[4]},
|
|
${t[2]*t[3]*t[4]},
|
|
${t[3]*t[4]},
|
|
${t[4]})) + float(depth3);
|
|
int texC = depth4;
|
|
vec2 uv = (vec2(texC, texR) + halfCR) /
|
|
vec2(${f}.0, ${h}.0);
|
|
return sampleTexture(${a}, uv);
|
|
}
|
|
`;let m=qo(a);return`
|
|
float ${n}(int row, int col, int depth,
|
|
int depth2, int depth3, int depth4) {
|
|
// Explicitly use integer operations as dot() only works on floats.
|
|
int index = row * ${p} + col * ${u} + depth * ${l} +
|
|
depth2 * ${o} + depth3 * ${i} + depth4 + ${m};
|
|
vec2 uv = uvFromFlat(${h}, ${f}, index);
|
|
return sampleTexture(${a}, uv);
|
|
}
|
|
`}function zu(e){let t=e.name,a=v.sizeFromShape(e.shapeInfo.logicalShape);return a<2?`return ${t};`:`
|
|
for (int i = 0; i < ${a}; i++) {
|
|
if (i == index) {
|
|
return ${t}[i];
|
|
}
|
|
}
|
|
`}function RG(e,t){let a=e.name,n=a.charAt(0).toUpperCase()+a.slice(1),r="get"+n+"AtOutCoords",s=e.shapeInfo.logicalShape.length,i=t.logicalShape.length,o=yv(e.shapeInfo.logicalShape,t.logicalShape),l=ft(i),u=i-s,p,c=["x","y","z","w","u","v"];s===0?p="":i<2&&o.length>=1?p="coords = 0;":p=o.map(g=>`coords.${c[g+u]} = 0;`).join(`
|
|
`);let d="";i<2&&s>0?d="coords":d=e.shapeInfo.logicalShape.map((g,y)=>`coords.${c[y+u]}`).join(", ");let h="return outputValue;",f=v.sizeFromShape(e.shapeInfo.logicalShape)===1,m=v.sizeFromShape(t.logicalShape)===1;if(s===1&&!f&&!m)h=`
|
|
return vec4(outputValue.xy, outputValue.xy);
|
|
`;else if(f&&!m)i===1?h=`
|
|
return vec4(outputValue.x, outputValue.x, 0., 0.);
|
|
`:h=`
|
|
return vec4(outputValue.x);
|
|
`;else if(o.length){let g=s-2,y=s-1;o.indexOf(g)>-1&&o.indexOf(y)>-1?h="return vec4(outputValue.x);":o.indexOf(g)>-1?h="return vec4(outputValue.x, outputValue.y, outputValue.x, outputValue.y);":o.indexOf(y)>-1&&(h="return vec4(outputValue.xx, outputValue.zz);")}return`
|
|
vec4 ${r}() {
|
|
${l} coords = getOutputCoords();
|
|
${p}
|
|
vec4 outputValue = get${n}(${d});
|
|
${h}
|
|
}
|
|
`}function EG(e,t){let a=e.name,n=a.charAt(0).toUpperCase()+a.slice(1),r="get"+n+"AtOutCoords",s=t.texShape,i=e.shapeInfo.texShape,o=e.shapeInfo.logicalShape.length,l=t.logicalShape.length;if(!e.shapeInfo.isUniform&&o===l&&e.shapeInfo.flatOffset==null&&v.arraysEqual(i,s))return`
|
|
float ${r}() {
|
|
return sampleTexture(${a}, resultUV);
|
|
}
|
|
`;let u=ft(l),p=yv(e.shapeInfo.logicalShape,t.logicalShape),c=l-o,d,h=["x","y","z","w","u","v"];o===0?d="":l<2&&p.length>=1?d="coords = 0;":d=p.map(m=>`coords.${h[m+c]} = 0;`).join(`
|
|
`);let f="";return l<2&&o>0?f="coords":f=e.shapeInfo.logicalShape.map((m,g)=>`coords.${h[g+c]}`).join(", "),`
|
|
float ${r}() {
|
|
${u} coords = getOutputCoords();
|
|
${d}
|
|
return get${n}(${f});
|
|
}
|
|
`}function ft(e){if(e<=1)return"int";if(e===2)return"ivec2";if(e===3)return"ivec3";if(e===4)return"ivec4";if(e===5)return"ivec5";if(e===6)return"ivec6";throw Error(`GPU for rank ${e} is not yet supported`)}function gg(e,t,a){let{newShape:n,keptDims:r}=v.squeezeShape(t),s=t.length,i=e&&s===3&&t[0]===1,o=i?t.slice(1):n,l=!e&&s>1&&!v.arraysEqual(t,a)&&n.length<s||i;return{useSqueezeShape:l,uniformShape:l?o:t,keptDims:r}}function Lu(e,t){let a=JSON.parse(JSON.stringify(e));return a.shapeInfo.logicalShape=t,a}function Wu(e,t){return t.map(a=>e[a]).join(", ")}function MG(e,t,a,n){let r=a.map((p,c)=>{let d={logicalShape:p.shape,texShape:p.isUniform?null:p.texData.texShape,isUniform:p.isUniform,isPacked:p.isUniform?!1:p.texData.isPacked,flatOffset:null};return p.texData!=null&&p.texData.slice!=null&&p.texData.slice.flatOffset>0&&(d.flatOffset=p.texData.slice.flatOffset),{name:t.variableNames[c],shapeInfo:d}}),s=r.map(p=>p.shapeInfo),i={logicalShape:n.shape,texShape:n.texData.texShape,isUniform:!1,isPacked:n.texData.isPacked,flatOffset:null},o=KU(r,i,t),l=K6(e.gl,o),u=e.createProgram(l);return W().get("ENGINE_COMPILE_ONLY")?{program:t,fragmentShader:l,source:o,webGLProgram:u,inShapeInfos:s,outShapeInfo:i,variablesLocations:null,customUniformLocations:null,infLoc:null,nanLoc:null,outShapeLocation:null,outShapeStridesLocation:null,outTexShapeLocation:null}:(e.buildVao(u),Object.assign({program:t,fragmentShader:l,source:o,webGLProgram:u,inShapeInfos:s,outShapeInfo:i},bv(e,t,u)))}function bv(e,t,a){let n=[],r=[],s,i,o,l=null,u=null;u=e.getUniformLocation(a,"NAN",!1),W().getNumber("WEBGL_VERSION")===1&&(l=e.getUniformLocation(a,"INFINITY",!1));let p=!1;for(let c of t.variableNames){let d={name:c,uniform:e.getUniformLocation(a,c,p),offset:e.getUniformLocation(a,`offset${c}`,p)};t.enableShapeUniforms&&(d.shape=e.getUniformLocation(a,`${c}Shape`,p),d.texShape=e.getUniformLocation(a,`${c}TexShape`,p)),n.push(d)}if(t.enableShapeUniforms&&(s=e.getUniformLocation(a,"outShape",p),o=e.getUniformLocation(a,"outShapeStrides",p),i=e.getUniformLocation(a,"outTexShape",p)),t.customUniforms)for(let c of t.customUniforms)r.push(e.getUniformLocation(a,c.name,p));return{variablesLocations:n,customUniformLocations:r,infLoc:l,nanLoc:u,outShapeLocation:s,outShapeStridesLocation:o,outTexShapeLocation:i}}function nx(e,t){if(e.length!==t.length)throw Error(`Binary was compiled with ${e.length} inputs, but was executed with ${t.length} inputs`);e.forEach((a,n)=>{let r=a.logicalShape,s=t[n],i=s.shape;if(!v.arraysEqual(r,i))throw Error(`Binary was compiled with different shapes than the current args. Shapes ${r} and ${i} must match`);if(a.isUniform&&s.isUniform)return;let o=a.texShape,l=s.isUniform?null:s.texData.texShape;if(!v.arraysEqual(o,l))throw Error(`Binary was compiled with different texture shapes than the current args. Shape ${o} and ${l} must match`)})}function FG(e,t,a,n,r){t.program.enableShapeUniforms||(nx(t.inShapeInfos,a),nx([t.outShapeInfo],[n]));let s=n.texData.texture,i=n.texData.texShape;n.texData.isPacked?e.setOutputPackedMatrixTexture(s.texture,i[0],i[1]):e.setOutputMatrixTexture(s.texture,i[0],i[1]),e.setProgram(t.webGLProgram),e.bindVertexArray(t.webGLProgram.vao),W().getNumber("WEBGL_VERSION")===1&&t.infLoc!==null&&e.gl.uniform1f(t.infLoc,1/0),t.nanLoc!==null&&e.gl.uniform1f(t.nanLoc,NaN);for(let l=0;l<a.length;++l){let u=a[l],{uniform:p,offset:c,shape:d,texShape:h}=t.variablesLocations[l];if(d){let{uniformShape:f}=gg(t.program.packedInputs,u.shape,u.texData.texShape);switch(f.length){case 1:e.gl.uniform1iv(d,new Int32Array(f));break;case 2:e.gl.uniform2iv(d,new Int32Array(f));break;case 3:e.gl.uniform3iv(d,new Int32Array(f));break;case 4:e.gl.uniform4iv(d,new Int32Array(f));break;default:break}}if(h&&e.gl.uniform2i(h,u.texData.texShape[0],u.texData.texShape[1]),p!=null){if(u.isUniform){if(v.sizeFromShape(u.shape)<2)e.gl.uniform1f(p,u.uniformValues[0]);else{let f=u.uniformValues;f instanceof Float32Array||(f=new Float32Array(f)),e.gl.uniform1fv(p,f)}continue}u.texData.slice!=null&&c!=null&&e.gl.uniform1i(c,u.texData.slice.flatOffset),e.setInputMatrixTexture(u.texData.texture.texture,p,l)}}let o=t.outShapeLocation;if(o)switch(n.shape.length){case 1:e.gl.uniform1iv(o,new Int32Array(n.shape));break;case 2:e.gl.uniform2iv(o,new Int32Array(n.shape));break;case 3:e.gl.uniform3iv(o,new Int32Array(n.shape));break;case 4:e.gl.uniform4iv(o,new Int32Array(n.shape));break;default:break}if(t.outShapeStridesLocation){let l=v.computeStrides(n.shape);switch(n.shape.length){case 2:e.gl.uniform1iv(t.outShapeStridesLocation,new Int32Array(l));break;case 3:e.gl.uniform2iv(t.outShapeStridesLocation,new Int32Array(l));break;case 4:e.gl.uniform3iv(t.outShapeStridesLocation,new Int32Array(l));break;default:break}}if(t.outTexShapeLocation&&e.gl.uniform2i(t.outTexShapeLocation,n.texData.texShape[0],n.texData.texShape[1]),t.program.customUniforms&&r)for(let l=0;l<t.program.customUniforms.length;++l){let u=t.program.customUniforms[l],p=t.customUniformLocations[l],c=r[l];if(u.type==="float")e.gl.uniform1fv(p,c);else if(u.type==="vec2")e.gl.uniform2fv(p,c);else if(u.type==="vec3")e.gl.uniform3fv(p,c);else if(u.type==="vec4")e.gl.uniform4fv(p,c);else if(u.type==="int")e.gl.uniform1iv(p,c);else if(u.type==="ivec2")e.gl.uniform2iv(p,c);else if(u.type==="ivec3")e.gl.uniform3iv(p,c);else if(u.type==="ivec4")e.gl.uniform4iv(p,c);else throw Error(`uniform type ${u.type} is not supported yet.`)}e.executeProgram()}function _G(e,t,a){let n="";t.concat(a).forEach(i=>{let o=i.texData!=null&&i.texData.slice!=null&&i.texData.slice.flatOffset>0;if(e.enableShapeUniforms&&!i.isUniform){let l=i.texData.texShape,{useSqueezeShape:u,uniformShape:p,keptDims:c}=gg(e.packedInputs,i.shape,l),d="",h="",f="";if(p.length===1&&e.packedInputs){let w=[Math.ceil(l[0]/2),Math.ceil(l[1]/2)];d=`${w[0]>1}_${w[1]>1}`}else if(p.length===2&&!e.packedInputs)h=`${p[0]>1}_${p[1]>1}`;else if(p.length>2&&!e.packedInputs){let w=v.computeStrides(p);f=`${w[0]===l[1]}_${w[w.length-1]===l[1]}`}let m=i.shape.length,g=p.length===2&&v.arraysEqual(i.shape,l),y=v.sizeFromShape(i.shape)===1,x=S.getBroadcastDims(i.shape,a.shape),A=!e.packedInputs&&m===a.shape.length&&v.arraysEqual(l,a.texData.texShape),b=e.packedInputs||p.length>2?"":`${l[0]>1}_${l[1]>1}`;n+=`${m}_${A}_${u?c:""}_${p.length}_${y}_${x}_${g}_${d}_${h}_${f}_${b}_${o}`}else{let l=i.isUniform?"uniform":i.texData.texShape;n+=`${i.shape}_${l}_${o}`}});let r=e.userCode,s=e.constructor.name;return s+="_"+n+"_"+r+`${W().getNumber("WEBGL_VERSION")}`,s}function ya(e){return W().getBool("WEBGL_USE_SHAPES_UNIFORMS")&&e<=4}var $G=class{constructor(e){this.variableNames=["A"],this.packedInputs=!1,this.packedOutput=!0,this.outPackingScheme=Ud.DENSE,this.customUniforms=[{name:"texShape",type:"ivec2"}];let t=Ea();this.outputShape=e,this.enableShapeUniforms=ya(this.outputShape.length),this.userCode=`
|
|
ivec3 outCoordsFromFlatIndex(int index) {
|
|
${this.enableShapeUniforms?Jh(["r","c","d"],e):jo(["r","c","d"],e)}
|
|
return ivec3(r, c, d);
|
|
}
|
|
|
|
void main() {
|
|
ivec2 resTexRC = ivec2(resultUV.yx * vec2(texShape[0], texShape[1]));
|
|
int index = 4 * (resTexRC.x * texShape[1] + resTexRC.y);
|
|
|
|
vec4 result = vec4(0.);
|
|
|
|
for (int i=0; i<4; i++) {
|
|
int flatIndex = index + i;
|
|
ivec3 rc = outCoordsFromFlatIndex(flatIndex);
|
|
result[i] = getA(rc.x, rc.y, rc.z);
|
|
}
|
|
|
|
${t.output} = result;
|
|
}
|
|
`}},DG=class{constructor(e){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!0,this.outPackingScheme=Ud.DENSE,this.customUniforms=[{name:"texShape",type:"ivec2"}];let t=Ea();this.outputShape=e,this.enableShapeUniforms=ya(this.outputShape.length),this.userCode=`
|
|
ivec3 outCoordsFromFlatIndex(int index) {
|
|
${this.enableShapeUniforms?Jh(["r","c","d"],e):jo(["r","c","d"],e)}
|
|
return ivec3(r, c, d);
|
|
}
|
|
|
|
void main() {
|
|
ivec2 resTexRC = ivec2(resultUV.yx * vec2(texShape[0], texShape[1]));
|
|
int index = 4 * (resTexRC.x * texShape[1] + resTexRC.y);
|
|
|
|
vec4 result = vec4(0.);
|
|
|
|
for (int i=0; i<4; i++) {
|
|
int flatIndex = index + i;
|
|
ivec3 rc = outCoordsFromFlatIndex(flatIndex);
|
|
result[i] = getChannel(getA(rc.x, rc.y, rc.z), vec2(rc.y, rc.z));
|
|
}
|
|
|
|
${t.output} = result;
|
|
}
|
|
`}},PG=class{constructor(e){this.variableNames=["A"],this.outTexUsage=fn.DOWNLOAD;let t=Ea();this.outputShape=e,this.userCode=`
|
|
${gv}
|
|
|
|
void main() {
|
|
float x = getAAtOutCoords();
|
|
${t.output} = encode_float(x);
|
|
}
|
|
`}},OG=class{constructor(e){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!1,this.outTexUsage=fn.DOWNLOAD;let t=Ea();this.outputShape=e,this.userCode=`
|
|
${gv}
|
|
|
|
void main() {
|
|
ivec3 coords = getOutputCoords();
|
|
float x = getChannel(getAAtOutCoords(), vec2(coords.y, coords.z));
|
|
${t.output} = encode_float(x);
|
|
}
|
|
`}},zG={R:0,G:1,B:2,A:3},rx=class{constructor(e,t=!1,a="RGBA"){this.variableNames=["A"],this.customUniforms=[{name:"texShape",type:"ivec2"}];let n=Ea();this.outputShape=e,this.enableShapeUniforms=ya(this.outputShape.length);let r="result";t&&(r="floor(result * 255. + 0.5)");let s="";for(let i=0;i<a.length;i++){let o=a[i];s+=`
|
|
if(offset == ${i}) {
|
|
result = values[${zG[o]}];
|
|
}`}this.userCode=`
|
|
${this.enableShapeUniforms?mg():fg(e)}
|
|
|
|
void main() {
|
|
ivec3 coords = getOutputCoords();
|
|
int flatIndex = getFlatIndex(coords);
|
|
float result = 0.;
|
|
int offset = imod(flatIndex, ${a.length});
|
|
|
|
flatIndex = idiv(flatIndex, ${a.length}, 1.);
|
|
|
|
int r = flatIndex / texShape[1];
|
|
if (r < texShape[0]) {
|
|
int c = imod(flatIndex, texShape[1]);
|
|
vec2 uv = (vec2(c, r) + halfCR) / vec2(texShape[1], texShape[0]);
|
|
vec4 values = ${n.texture2D}(A, uv);
|
|
${s}
|
|
}
|
|
${n.output} = vec4(${r}, 0., 0., 0.);
|
|
}
|
|
`}},LG=class{constructor(e,t=!1){this.variableNames=["A"],this.packedInputs=!1,this.packedOutput=!0,this.customUniforms=[{name:"texShape",type:"ivec2"}];let a=Ea();this.outputShape=e,this.enableShapeUniforms=ya(this.outputShape.length);let n="",r="result";t&&(r="floor(result * 255. + 0.5)");for(let s=0;s<=1;s++)for(let i=0;i<=1;i++){let o=s*2+i;n+=`
|
|
localCoords = coords;
|
|
if(localCoords[2] + ${i} < ${this.enableShapeUniforms?"outShape[2]":`${e[2]}`}) {
|
|
localCoords[2] += ${i};
|
|
if (localCoords[1] + ${s} < ${this.enableShapeUniforms?"outShape[1]":`${e[1]}`}) {
|
|
localCoords[1] += ${s};
|
|
|
|
flatIndex = getFlatIndex(localCoords);
|
|
offset = imod(flatIndex, 4);
|
|
|
|
flatIndex = idiv(flatIndex, 4, 1.);
|
|
|
|
int r = flatIndex / texShape[1];
|
|
int c = imod(flatIndex, texShape[1]);
|
|
vec2 uv = (vec2(c, r) + halfCR) / vec2(texShape[1], texShape[0]);
|
|
values = ${a.texture2D}(A, uv);
|
|
|
|
if (offset == 0) {
|
|
result[${o}] = values[0];
|
|
} else if (offset == 1) {
|
|
result[${o}] = values[1];
|
|
} else if (offset == 2) {
|
|
result[${o}] = values[2];
|
|
} else {
|
|
result[${o}] = values[3];
|
|
}
|
|
}
|
|
}
|
|
`}this.userCode=`
|
|
${this.enableShapeUniforms?mg():fg(e)}
|
|
|
|
void main() {
|
|
ivec3 coords = getOutputCoords();
|
|
|
|
vec4 result = vec4(0.);
|
|
int flatIndex, r, c, offset;
|
|
ivec3 localCoords;
|
|
vec2 uv;
|
|
vec4 values;
|
|
|
|
${n}
|
|
|
|
${a.output} = ${r};
|
|
}
|
|
`}},vv={};Ze(vv,{bindVertexProgramAttributeStreams:()=>Ev,createBufferFromOutputTexture:()=>_v,createFloat16MatrixTexture:()=>Cv,createFloat16PackedMatrixTexture:()=>Rv,createFloat32MatrixTexture:()=>Sv,createIndexBuffer:()=>Iv,createPackedMatrixTexture:()=>Nv,createUnsignedBytesMatrixTexture:()=>Tv,createVertexBuffer:()=>kv,createVertexShader:()=>wv,downloadByteEncodedFloatMatrixFromOutputTexture:()=>Dv,downloadFloat32MatrixFromBuffer:()=>$v,downloadMatrixFromPackedOutputTexture:()=>Ov,downloadPackedMatrixFromBuffer:()=>Pv,getInternalFormatForFloat16MatrixTexture:()=>xg,getInternalFormatForFloat16PackedMatrixTexture:()=>vg,getInternalFormatForFloat32MatrixTexture:()=>yg,getInternalFormatForPackedMatrixTexture:()=>bg,getInternalFormatForUnsignedBytesMatrixTexture:()=>Ag,uploadDenseMatrixToTexture:()=>Mv,uploadPixelDataToTexture:()=>Fv});function wv(e){let t=Ea(),a=`${t.version}
|
|
precision highp float;
|
|
${t.attribute} vec3 clipSpacePos;
|
|
${t.attribute} vec2 uv;
|
|
${t.varyingVs} vec2 resultUV;
|
|
|
|
void main() {
|
|
gl_Position = vec4(clipSpacePos, 1);
|
|
resultUV = uv;
|
|
}`;return X6(e,a)}function kv(e){let t=new Float32Array([-1,1,0,0,1,-1,-1,0,0,0,1,1,0,1,1,1,-1,0,1,0]);return J6(e,t)}function Iv(e){let t=new Uint16Array([0,1,2,2,1,3]);return Q6(e,t)}function Vp(e,t,a,n,r,s){tv(t,a);let i=ev(e),o=e.TEXTURE_2D;return pe(e,()=>e.bindTexture(o,i)),pe(e,()=>e.texParameteri(o,e.TEXTURE_WRAP_S,e.CLAMP_TO_EDGE)),pe(e,()=>e.texParameteri(o,e.TEXTURE_WRAP_T,e.CLAMP_TO_EDGE)),pe(e,()=>e.texParameteri(o,e.TEXTURE_MIN_FILTER,e.NEAREST)),pe(e,()=>e.texParameteri(o,e.TEXTURE_MAG_FILTER,e.NEAREST)),W().getNumber("WEBGL_VERSION")===1?pe(e,()=>e.texImage2D(o,0,n,t,a,0,r,s,null)):pe(e,()=>e.texStorage2D(o,1,n,t,a)),pe(e,()=>e.bindTexture(e.TEXTURE_2D,null)),{texture:i,texShape:[a,t]}}function yg(e){return e.internalFormatFloat}function Sv(e,t,a,n){let[r,s]=Bp(t,a);return Vp(e,r,s,yg(n),n.textureFormatFloat,e.FLOAT)}function xg(e){return e.internalFormatHalfFloat}function Cv(e,t,a,n){let[r,s]=Bp(t,a);return Vp(e,r,s,xg(n),n.textureFormatFloat,n.textureTypeHalfFloat)}function Ag(e){return e.downloadTextureFormat}function Tv(e,t,a,n){let[r,s]=Bp(t,a);return Vp(e,r,s,Ag(n),e.RGBA,e.UNSIGNED_BYTE)}function bg(e){return e.internalFormatPackedFloat}function Nv(e,t,a,n){let[r,s]=Du(t,a);return Vp(e,r,s,bg(n),e.RGBA,e.FLOAT)}function vg(e){return e.internalFormatPackedHalfFloat}function Rv(e,t,a,n){let[r,s]=Du(t,a);return Vp(e,r,s,vg(n),e.RGBA,n.textureTypeHalfFloat)}function Ev(e,t,a){return pe(e,()=>e.bindBuffer(e.ARRAY_BUFFER,a)),v1(e,t,"clipSpacePos",a,3,20,0)&&v1(e,t,"uv",a,2,20,12)}function Mv(e,t,a,n,r,s){pe(e,()=>e.bindTexture(e.TEXTURE_2D,t));let i,o,l;r instanceof Uint8Array?(i=new Uint8Array(a*n*4),o=e.UNSIGNED_BYTE,l=e.RGBA):(i=new Float32Array(a*n*4),o=e.FLOAT,l=s.internalFormatPackedFloat),i.set(r),W().getNumber("WEBGL_VERSION")===2?pe(e,()=>e.texSubImage2D(e.TEXTURE_2D,0,0,0,a,n,e.RGBA,o,i)):pe(e,()=>e.texImage2D(e.TEXTURE_2D,0,l,a,n,0,e.RGBA,o,i)),pe(e,()=>e.bindTexture(e.TEXTURE_2D,null))}function Fv(e,t,a){pe(e,()=>e.bindTexture(e.TEXTURE_2D,t)),a.data instanceof Uint8Array?W().getNumber("WEBGL_VERSION")===2?pe(e,()=>e.texSubImage2D(e.TEXTURE_2D,0,0,0,a.width,a.height,e.RGBA,e.UNSIGNED_BYTE,a.data)):pe(e,()=>e.texImage2D(e.TEXTURE_2D,0,e.RGBA,a.width,a.height,0,e.RGBA,e.UNSIGNED_BYTE,a.data)):W().getNumber("WEBGL_VERSION")===2?pe(e,()=>e.texSubImage2D(e.TEXTURE_2D,0,0,0,e.RGBA,e.UNSIGNED_BYTE,a)):pe(e,()=>e.texImage2D(e.TEXTURE_2D,0,e.RGBA,e.RGBA,e.UNSIGNED_BYTE,a)),pe(e,()=>e.bindTexture(e.TEXTURE_2D,null))}function _v(e,t,a,n){let r=e.createBuffer();pe(e,()=>e.bindBuffer(e.PIXEL_PACK_BUFFER,r));let s=4*4*t*a;return pe(e,()=>e.bufferData(e.PIXEL_PACK_BUFFER,s,e.STREAM_READ)),pe(e,()=>e.readPixels(0,0,a,t,e.RGBA,e.FLOAT,0)),pe(e,()=>e.bindBuffer(e.PIXEL_PACK_BUFFER,null)),r}function $v(e,t,a){let n=e,r=new Float32Array(a);return n.bindBuffer(n.PIXEL_PACK_BUFFER,t),n.getBufferSubData(n.PIXEL_PACK_BUFFER,0,r),n.bindBuffer(n.PIXEL_PACK_BUFFER,null),r}function Dv(e,t,a,n){let[r,s]=Bp(t,a),i=4,o=new Uint8Array(DU(t*a,i));return pe(e,()=>e.readPixels(0,0,r,s,n.downloadTextureFormat,e.UNSIGNED_BYTE,o)),new Float32Array(o.buffer)}function Pv(e,t,a,n,r,s,i,o){let l=e,u=new Float32Array(PU(s,i));return l.bindBuffer(l.PIXEL_PACK_BUFFER,t),l.getBufferSubData(l.PIXEL_PACK_BUFFER,0,u),l.bindBuffer(l.PIXEL_PACK_BUFFER,null),u}function Ov(e,t,a){let n=new Float32Array(t*a*4);return pe(e,()=>e.readPixels(0,0,a,t,e.RGBA,e.FLOAT,n)),n}var _l=class{constructor(e){this.outputTexture=null,this.program=null,this.disposed=!1,this.itemsToPoll=[];let t=W().getNumber("WEBGL_VERSION");if(e!=null?(this.gl=e,Zh(t,e)):this.gl=Wn(t),e=this.gl,W().getNumber("WEBGL_VERSION")===2){let r=e;this.createVertexArray=()=>pe(r,()=>r.createVertexArray()),this.bindVertexArray=s=>pe(r,()=>r.bindVertexArray(s)),this.deleteVertexArray=s=>pe(r,()=>r.deleteVertexArray(s)),this.getVertexArray=()=>pe(r,()=>r.getParameter(r.VERTEX_ARRAY_BINDING))}else if(e!=null){let r=e.getExtension("OES_vertex_array_object");if(r==null)throw new Error("All WebGL1 implementations are expected to offer OES_vertex_array_object.");this.createVertexArray=()=>pe(e,()=>r.createVertexArrayOES()),this.bindVertexArray=s=>pe(e,()=>r.bindVertexArrayOES(s)),this.deleteVertexArray=s=>pe(e,()=>r.deleteVertexArrayOES(s)),this.getVertexArray=()=>pe(e,()=>e.getParameter(r.VERTEX_ARRAY_BINDING_OES))}let a="WEBGL_color_buffer_float",n="EXT_color_buffer_half_float";if(this.parallelCompilationExtension=this.gl.getExtension("KHR_parallel_shader_compile"),W().getNumber("WEBGL_VERSION")===1){let r="OES_texture_float",s="OES_texture_half_float";if(this.textureFloatExtension=Ad(this.gl,r),mn(this.gl,s))this.textureHalfFloatExtension=Ad(this.gl,s);else if(W().get("WEBGL_FORCE_F16_TEXTURES"))throw new Error("GL context does not support half float textures, yet the environment flag WEBGL_FORCE_F16_TEXTURES is set to true.");if(this.colorBufferFloatExtension=this.gl.getExtension(a),mn(this.gl,n))this.colorBufferHalfFloatExtension=Ad(this.gl,n);else if(W().get("WEBGL_FORCE_F16_TEXTURES"))throw new Error("GL context does not support color renderable half floats, yet the environment flag WEBGL_FORCE_F16_TEXTURES is set to true.")}else if(a="EXT_color_buffer_float",mn(this.gl,a))this.colorBufferFloatExtension=this.gl.getExtension(a);else if(mn(this.gl,n))this.colorBufferHalfFloatExtension=this.gl.getExtension(n);else throw new Error("GL context does not support color renderable floats");this.vertexBuffer=kv(this.gl),this.indexBuffer=Iv(this.gl),this.framebuffer=av(this.gl),this.textureConfig=cg(this.gl,this.textureHalfFloatExtension)}get debug(){return W().getBool("DEBUG")}dispose(){if(this.disposed)return;this.program!=null&&console.warn("Disposing a GPGPUContext that still has a bound WebGLProgram. This is probably a resource leak, delete the program with GPGPUContext.deleteProgram before disposing."),this.outputTexture!=null&&console.warn("Disposing a GPGPUContext that still has a bound output matrix texture. This is probably a resource leak, delete the output matrix texture with GPGPUContext.deleteMatrixTexture before disposing.");let e=this.gl;pe(e,()=>e.finish()),pe(e,()=>e.bindFramebuffer(e.FRAMEBUFFER,null)),pe(e,()=>e.deleteFramebuffer(this.framebuffer)),pe(e,()=>e.bindBuffer(e.ARRAY_BUFFER,null)),pe(e,()=>e.bindBuffer(e.ELEMENT_ARRAY_BUFFER,null)),pe(e,()=>e.deleteBuffer(this.indexBuffer)),this.disposed=!0}createFloat32MatrixTexture(e,t){return this.throwIfDisposed(),Sv(this.gl,e,t,this.textureConfig)}createFloat16MatrixTexture(e,t){return this.throwIfDisposed(),Cv(this.gl,e,t,this.textureConfig)}createUnsignedBytesMatrixTexture(e,t){return this.throwIfDisposed(),Tv(this.gl,e,t,this.textureConfig)}uploadPixelDataToTexture(e,t){this.throwIfDisposed(),Fv(this.gl,e,t)}uploadDenseMatrixToTexture(e,t,a,n){this.throwIfDisposed(),Mv(this.gl,e,t,a,n,this.textureConfig)}createFloat16PackedMatrixTexture(e,t){return this.throwIfDisposed(),Rv(this.gl,e,t,this.textureConfig)}createPackedMatrixTexture(e,t){return this.throwIfDisposed(),Nv(this.gl,e,t,this.textureConfig)}deleteMatrixTexture(e){this.throwIfDisposed(),this.outputTexture===e&&(w1(this.gl,this.framebuffer),this.outputTexture=null),pe(this.gl,()=>this.gl.deleteTexture(e))}downloadByteEncodedFloatMatrixFromOutputTexture(e,t,a){return this.downloadMatrixDriver(e,()=>Dv(this.gl,t,a,this.textureConfig))}downloadPackedMatrixFromBuffer(e,t,a,n,r,s){return Pv(this.gl,e,t,a,n,r,s,this.textureConfig)}downloadFloat32MatrixFromBuffer(e,t){return $v(this.gl,e,t)}createBufferFromTexture(e,t,a){this.bindTextureToFrameBuffer(e);let n=_v(this.gl,t,a,this.textureConfig);return this.unbindTextureToFrameBuffer(),n}createAndWaitForFence(){let e=this.createFence(this.gl);return this.pollFence(e)}createFence(e){let t,a;if(W().getBool("WEBGL_FENCE_API_ENABLED")){let n=e,r=n.fenceSync(n.SYNC_GPU_COMMANDS_COMPLETE,0);e.flush(),a=()=>{let s=n.clientWaitSync(r,0,0);return s===n.ALREADY_SIGNALED||s===n.CONDITION_SATISFIED},t=r}else W().getNumber("WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_VERSION")>0?(t=this.beginQuery(),this.endQuery(),a=()=>this.isQueryAvailable(t,W().getNumber("WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_VERSION"))):a=()=>!0;return{query:t,isFencePassed:a}}downloadMatrixFromPackedTexture(e,t,a){return this.downloadMatrixDriver(e,()=>Ov(this.gl,t,a))}createProgram(e){this.throwIfDisposed();let t=this.gl;this.vertexShader==null&&(this.vertexShader=wv(t));let a=Y6(t);pe(t,()=>t.attachShader(a,this.vertexShader)),pe(t,()=>t.attachShader(a,e)),Z6(t,a);let n=Object.assign(a,{vao:this.createVertexArray()});return this.debug&&Xc(t,n),n}buildVao(e){this.setProgram(e),this.bindVertexArray(e.vao);let t=this.gl;pe(t,()=>t.bindBuffer(t.ELEMENT_ARRAY_BUFFER,this.indexBuffer)),Ev(t,e,this.vertexBuffer)}deleteProgram(e){this.throwIfDisposed(),e===this.program&&(this.program=null),e!=null&&(pe(this.gl,()=>this.gl.deleteProgram(e)),this.deleteVertexArray(e.vao))}setProgram(e){this.throwIfDisposed(),this.program=e,this.program!=null&&this.debug&&Xc(this.gl,this.program),pe(this.gl,()=>this.gl.useProgram(e))}getUniformLocation(e,t,a=!0){return this.throwIfDisposed(),a?rv(this.gl,e,t):sv(this.gl,e,t)}getAttributeLocation(e,t){return this.throwIfDisposed(),pe(this.gl,()=>this.gl.getAttribLocation(e,t))}getUniformLocationNoThrow(e,t){return this.throwIfDisposed(),this.gl.getUniformLocation(e,t)}setInputMatrixTexture(e,t,a){this.throwIfDisposed(),this.throwIfNoProgram(),iv(this.gl,e,t,a)}setOutputMatrixTexture(e,t,a){this.setOutputMatrixTextureDriver(e,a,t)}setOutputPackedMatrixTexture(e,t,a){this.throwIfDisposed();let[n,r]=Du(t,a);this.setOutputMatrixTextureDriver(e,n,r)}setOutputMatrixWriteRegion(e,t,a,n){this.setOutputMatrixWriteRegionDriver(a,e,n,t)}setOutputPackedMatrixWriteRegion(e,t,a,n){throw new Error("setOutputPackedMatrixWriteRegion not implemented.")}debugValidate(){this.program!=null&&Xc(this.gl,this.program),bd(this.gl)}executeProgram(){this.throwIfDisposed(),this.throwIfNoProgram();let e=this.gl;if(this.debug){let t=this.getVertexArray();console.assert(t===this.program.vao,"VAO changed between setProgram and executeProgram!"),this.debugValidate()}pe(e,()=>e.drawElements(e.TRIANGLES,6,e.UNSIGNED_SHORT,0))}blockUntilAllProgramsCompleted(){this.throwIfDisposed(),pe(this.gl,()=>this.gl.finish())}getQueryTimerExtension(){return this.disjointQueryTimerExtension==null&&(this.disjointQueryTimerExtension=Ad(this.gl,W().getNumber("WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_VERSION")===2?"EXT_disjoint_timer_query_webgl2":"EXT_disjoint_timer_query")),this.disjointQueryTimerExtension}getQueryTimerExtensionWebGL2(){return this.getQueryTimerExtension()}getQueryTimerExtensionWebGL1(){return this.getQueryTimerExtension()}beginQuery(){if(W().getNumber("WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_VERSION")===2){let a=this.gl,n=this.getQueryTimerExtensionWebGL2(),r=a.createQuery();return a.beginQuery(n.TIME_ELAPSED_EXT,r),r}let e=this.getQueryTimerExtensionWebGL1(),t=e.createQueryEXT();return e.beginQueryEXT(e.TIME_ELAPSED_EXT,t),t}endQuery(){if(W().getNumber("WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_VERSION")===2){let t=this.gl,a=this.getQueryTimerExtensionWebGL2();t.endQuery(a.TIME_ELAPSED_EXT);return}let e=this.getQueryTimerExtensionWebGL1();e.endQueryEXT(e.TIME_ELAPSED_EXT)}async waitForQueryAndGetTime(e){return await v.repeatedTry(()=>this.disposed||this.isQueryAvailable(e,W().getNumber("WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_VERSION"))),this.getQueryTime(e,W().getNumber("WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_VERSION"))}getQueryTime(e,t){if(t===0)return null;if(t===2){let a=this.gl;return a.getQueryParameter(e,a.QUERY_RESULT)/1e6}else{let a=this.getQueryTimerExtensionWebGL1();return a.getQueryObjectEXT(e,a.QUERY_RESULT_EXT)/1e6}}isQueryAvailable(e,t){if(t===0)return!0;if(t===2){let a=this.gl,n=this.getQueryTimerExtensionWebGL2(),r=a.getQueryParameter(e,a.QUERY_RESULT_AVAILABLE);return this.disjoint==null&&(this.disjoint=this.gl.getParameter(n.GPU_DISJOINT_EXT)),r&&!this.disjoint}else{let a=this.getQueryTimerExtensionWebGL1(),n=a.getQueryObjectEXT(e,a.QUERY_RESULT_AVAILABLE_EXT);return this.disjoint==null&&(this.disjoint=this.gl.getParameter(a.GPU_DISJOINT_EXT)),n&&!this.disjoint}}pollFence(e){return new Promise(t=>{this.addItemToPoll(()=>e.isFencePassed(),()=>t())})}pollItems(){let e=WG(this.itemsToPoll.map(t=>t.isDoneFn));for(let t=0;t<=e;++t){let{resolveFn:a}=this.itemsToPoll[t];a()}this.itemsToPoll=this.itemsToPoll.slice(e+1)}addItemToPoll(e,t){if(this.itemsToPoll.push({isDoneFn:e,resolveFn:t}),this.itemsToPoll.length>1)return;let a;"setTimeoutCustom"in W().platform&&(a=W().platform.setTimeoutCustom.bind(W().platform)),v.repeatedTry(()=>(this.pollItems(),this.itemsToPoll.length===0),()=>0,null,a)}bindTextureToFrameBuffer(e){this.throwIfDisposed(),Kc(this.gl,e,this.framebuffer),this.debug&&bd(this.gl)}unbindTextureToFrameBuffer(){this.outputTexture!=null?(Kc(this.gl,this.outputTexture,this.framebuffer),this.debug&&bd(this.gl)):w1(this.gl,this.framebuffer)}downloadMatrixDriver(e,t){this.bindTextureToFrameBuffer(e);let a=t();return this.unbindTextureToFrameBuffer(),a}setOutputMatrixTextureDriver(e,t,a){this.throwIfDisposed();let n=this.gl;Kc(n,e,this.framebuffer),this.debug&&bd(n),this.outputTexture=e,pe(n,()=>n.viewport(0,0,t,a)),pe(n,()=>n.scissor(0,0,t,a))}setOutputMatrixWriteRegionDriver(e,t,a,n){this.throwIfDisposed(),pe(this.gl,()=>this.gl.scissor(e,t,a,n))}throwIfDisposed(){if(this.disposed)throw new Error("Attempted to use disposed GPGPUContext.")}throwIfNoProgram(){if(this.program==null)throw new Error("No GPU program is currently set.")}};function WG(e){let t=0;for(;t<e.length&&e[t]();++t);return t-1}var{addImpl:BG,bincountImpl:zv,bincountReduceImpl:VG,castImpl:UG,ceilImpl:GG,concatImpl:HG,equalImpl:jG,expImpl:qG,expm1Impl:XG,floorImpl:KG,gatherNdImpl:YG,gatherV2Impl:ZG,greaterImpl:JG,greaterEqualImpl:QG,lessImpl:eH,lessEqualImpl:tH,linSpaceImpl:aH,logImpl:nH,maxImpl:rH,maximumImpl:sH,minimumImpl:iH,multiplyImpl:oH,negImpl:lH,notEqualImpl:uH,prodImpl:dH,raggedGatherImpl:pH,raggedRangeImpl:cH,raggedTensorToTensorImpl:hH,rangeImpl:fH,rsqrtImpl:mH,scatterImpl:gH,sigmoidImpl:yH,simpleAbsImpl:Lv,sliceImpl:xH,sparseFillEmptyRowsImpl:AH,sparseReshapeImpl:bH,sparseSegmentReductionImpl:Wv,sqrtImpl:vH,staticRegexReplaceImpl:wH,stridedSliceImpl:kH,stringNGramsImpl:IH,stringSplitImpl:SH,stringToHashBucketFastImpl:CH,subImpl:TH,tileImpl:NH,topKImpl:RH,transposeImpl:wg,uniqueImpl:EH}=Kh;function Bv(e,t){return["x","y","z","w","u","v"].slice(0,t).map(a=>`${e}.${a}`)}function Ia(e,t){return t===1?[e]:Bv(e,t)}function MH(e,t){if(e===1)return"rc";let a="";for(let n=0;n<e;n++)a+=t[n],n<e-1&&(a+=",");return a}var FH=class{constructor(e){if(this.variableNames=["A"],this.packedInputs=!1,this.packedOutput=!0,this.outputShape=e,this.rank=e.length,this.enableShapeUniforms=ya(this.outputShape.length),this.rank===0)this.userCode=`
|
|
void main() {
|
|
setOutput(vec4(getA(), 0., 0., 0.));
|
|
}
|
|
`;else{let t=Ia("rc",this.rank),a=ft(this.rank),n=this.getOutOfBoundsCondition(t),r=this.getSetup(t),s=this.getOutput(t);this.userCode=`
|
|
void main() {
|
|
${a} rc = getOutputCoords();
|
|
|
|
if(${n}) {
|
|
setOutput(vec4(0));
|
|
} else {
|
|
${r}
|
|
|
|
setOutput(vec4(${s}));
|
|
}
|
|
}
|
|
`}}getSourceCoordsArr(e){let t=[];for(let a=0;a<=1;a++)for(let n=0;n<=1;n++){let r=`${a===0?"r":"rp1"}, ${n===0?"c":"cp1"}`;for(let s=2;s<this.rank;s++)r=`${e[e.length-1-s]},`+r;t.push(r)}return t}getOutOfBoundsCondition(e){if(this.rank===1)return`rc > ${this.enableShapeUniforms?"outShape":this.outputShape[0]}`;let t="";for(let a=this.rank-2;a<this.rank;a++)t+=`${e[a]} >= ${this.enableShapeUniforms?`outShape[${a}]`:this.outputShape[a]}`,a<this.rank-1&&(t+="||");return t}getSetup(e){if(this.rank===1)return"";let t=e.slice(-2),a=this.enableShapeUniforms?`outShape[${this.rank} - 1]`:this.outputShape[this.rank-1],n=this.enableShapeUniforms?`outShape[${this.rank} - 2]`:this.outputShape[this.rank-2];return`
|
|
int r = ${t[0]};
|
|
int c = ${t[1]};
|
|
int rp1 = r + 1;
|
|
int cp1 = c + 1;
|
|
|
|
bool cEdge = cp1 >= ${a};
|
|
bool rEdge = rp1 >= ${n};
|
|
`}getOutput(e){let t=this.getSourceCoordsArr(e);return this.rank===1?`getA(rc), (rc + 1 >= ${this.enableShapeUniforms?"outShape":this.outputShape[0]} ? 0. : getA(rc + 1)), 0, 0`:`getA(${t[0]}),
|
|
cEdge ? 0. : getA(${t[1]}),
|
|
rEdge ? 0. : getA(${t[2]}),
|
|
rEdge || cEdge ? 0. : getA(${t[3]})`}},Vv=class{constructor(e,t){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!0,this.customUniforms=[{name:"inputShape",type:"ivec3"}],this.outputShape=e,this.enableShapeUniforms=ya(this.outputShape.length);let a="";for(let n=0;n<4;n++){let r="thisRC = rc;";n%2===1&&(r+="thisRC.z += 1;"),n>1&&(r+="thisRC.y += 1;"),a+=`
|
|
${r}
|
|
${n>0?"if(thisRC.y < rows && thisRC.z < cols){":""}
|
|
int flatIndex = getFlatIndex(thisRC);
|
|
|
|
ivec3 inputRC = inputCoordsFromReshapedOutCoords(flatIndex);
|
|
vec2 inputRCInnerDims = vec2(float(inputRC.y),float(inputRC.z));
|
|
|
|
result[${n}] =
|
|
getChannel(getA(inputRC.x, inputRC.y, inputRC.z), inputRCInnerDims);
|
|
${n>0?"}":""}
|
|
`}this.userCode=`
|
|
${_H(t,this.enableShapeUniforms)}
|
|
${this.enableShapeUniforms?mg():fg(e)}
|
|
|
|
void main() {
|
|
ivec3 rc = getOutputCoords();
|
|
|
|
vec4 result = vec4(0.);
|
|
|
|
ivec3 thisRC;
|
|
int rows = ${this.enableShapeUniforms?"outShape[1]":e[1]};
|
|
int cols = ${this.enableShapeUniforms?"outShape[2]":e[2]};
|
|
|
|
${a}
|
|
|
|
setOutput(result);
|
|
}
|
|
`}};function _H(e,t){return`
|
|
ivec3 inputCoordsFromReshapedOutCoords(int index) {
|
|
${t?XU(["r","c","d"],"inputShape"):jo(["r","c","d"],e)}
|
|
return ivec3(r, c, d);
|
|
}
|
|
`}var $H=class{constructor(e){this.gpgpu=e,this.numUsedTextures=0,this.numFreeTextures=0,this._numBytesAllocated=0,this._numBytesFree=0,this.freeTextures={},this.usedTextures={},this.logEnabled=!1}acquireTexture(e,t,a){let n=ix(t,a),r=ox(e,n,a);r in this.freeTextures||(this.freeTextures[r]=[]),r in this.usedTextures||(this.usedTextures[r]=[]);let s=sx(e,n,this.gpgpu.gl,this.gpgpu.textureConfig,a);if(this.freeTextures[r].length>0){this.numFreeTextures--,this.numUsedTextures++,this._numBytesFree-=s,this.log();let o=this.freeTextures[r].pop();return this.usedTextures[r].push(o),o}let i;return n===da.PACKED_2X2_FLOAT32?i=this.gpgpu.createPackedMatrixTexture(e[0],e[1]):n===da.PACKED_2X2_FLOAT16?i=this.gpgpu.createFloat16PackedMatrixTexture(e[0],e[1]):n===da.UNPACKED_FLOAT32?i=this.gpgpu.createFloat32MatrixTexture(e[0],e[1]):n===da.UNPACKED_FLOAT16?i=this.gpgpu.createFloat16MatrixTexture(e[0],e[1]):n===da.PACKED_4X1_UNSIGNED_BYTE&&(i=this.gpgpu.createUnsignedBytesMatrixTexture(e[0],e[1])),this.usedTextures[r].push(i),this.numUsedTextures++,this._numBytesAllocated+=s,this.log(),i}releaseTexture(e,t,a,n){if(this.freeTextures==null)return;let r=ix(a,n),s=ox(t,r,n);s in this.freeTextures||(this.freeTextures[s]=[]);let i=sx(t,r,this.gpgpu.gl,this.gpgpu.textureConfig,n),o=W().get("WEBGL_DELETE_TEXTURE_THRESHOLD");o!==-1&&this._numBytesAllocated>o?(this.gpgpu.deleteMatrixTexture(e.texture),this._numBytesAllocated-=i):(this.freeTextures[s].push(e),this.numFreeTextures++,this._numBytesFree+=i),this.numUsedTextures--;let l=this.usedTextures[s],u=l&&l.indexOf(e);if(u==null||u<0)throw new Error("Cannot release a texture that was never provided by this texture manager");l[u]=l[l.length-1],l.pop(),this.log()}log(){if(!this.logEnabled)return;let e=this.numFreeTextures+this.numUsedTextures;console.log("Free/Used",`${this.numFreeTextures} / ${this.numUsedTextures}`,`(${e})`);let t=this._numBytesFree/this._numBytesAllocated;console.log(`Bytes allocated: ${this._numBytesAllocated}`),console.log(`Bytes unused: ${this._numBytesFree} (${Math.round(100*t)}%)`)}get numBytesAllocated(){return this._numBytesAllocated}get numBytesFree(){return this._numBytesFree}getNumUsedTextures(){return this.numUsedTextures}getNumFreeTextures(){return this.numFreeTextures}dispose(){if(this.freeTextures!=null){for(let e in this.freeTextures)this.freeTextures[e].forEach(t=>{this.gpgpu.deleteMatrixTexture(t.texture)});for(let e in this.usedTextures)this.usedTextures[e].forEach(t=>{this.gpgpu.deleteMatrixTexture(t.texture)});this.freeTextures=null,this.usedTextures=null,this.numUsedTextures=0,this.numFreeTextures=0,this._numBytesAllocated=0,this._numBytesFree=0}}};function DH(e,t){let a=e;if(t===a.R32F)return 4;if(t===a.R16F)return 2;if(t===a.RGBA32F||t===e.RGBA)return 16;if(t===a.RGBA16F)return 8;if(t===a.RGBA8)return 4;throw new Error(`Unknown internal format ${t}`)}function sx(e,t,a,n,r){let s=PH(t,n),i;if(r){let[l,u]=Du(e[0],e[1]);i=l*u}else{let[l,u]=Bp(e[0],e[1]);i=l*u}let o=DH(a,s);return i*o}function PH(e,t){switch(e){case da.PACKED_2X2_FLOAT32:return bg(t);case da.PACKED_2X2_FLOAT16:return vg(t);case da.UNPACKED_FLOAT32:return yg(t);case da.UNPACKED_FLOAT16:return xg(t);case da.PACKED_4X1_UNSIGNED_BYTE:return Ag(t);default:throw new Error(`Unknown physical texture type ${e}`)}}function OH(e){return W().getBool("WEBGL_RENDER_FLOAT32_ENABLED")?e?da.PACKED_2X2_FLOAT32:da.UNPACKED_FLOAT32:e?da.PACKED_2X2_FLOAT16:da.UNPACKED_FLOAT16}function ix(e,t){if(e===fn.UPLOAD)return da.PACKED_2X2_FLOAT32;if(e===fn.RENDER||e==null)return OH(t);if(e===fn.DOWNLOAD||e===fn.PIXELS)return da.PACKED_4X1_UNSIGNED_BYTE;throw new Error(`Unknown logical texture type ${e}`)}function ox(e,t,a){return`${e[0]}_${e[1]}_${t}_${a}`}var Yn=class{constructor(e,t){this.variableNames=["A"],this.outputShape=e,this.enableShapeUniforms=ya(this.outputShape.length),this.userCode=`
|
|
float unaryOperation(float x) {
|
|
${t}
|
|
}
|
|
|
|
void main() {
|
|
float x = getAAtOutCoords();
|
|
float y = unaryOperation(x);
|
|
|
|
setOutput(y);
|
|
}
|
|
`}},Rn="if (isnan(x)) return x;",zH="return x;",lx="return abs(x);",LH="return (x >= 0.0) ? x : (exp(x) - 1.0);",WH=Rn+`
|
|
return (x < 0.0) ? 0.0 : x;
|
|
`,BH=Rn+`
|
|
return (x < 0.0) ? 0.0 : min(6.0, x);
|
|
`,Lr="return x;",VH="return 1.0 / (1.0 + exp(-1.0 * x));",UH="return x;",GH=`
|
|
vec4 result;
|
|
|
|
result.r = (x.r >= 0.0) ? x.r : (exp(x.r) - 1.0);
|
|
result.g = (x.g >= 0.0) ? x.g : (exp(x.g) - 1.0);
|
|
result.b = (x.b >= 0.0) ? x.b : (exp(x.b) - 1.0);
|
|
result.a = (x.a >= 0.0) ? x.a : (exp(x.a) - 1.0);
|
|
|
|
return result;
|
|
`,HH=`
|
|
vec4 result = x * vec4(greaterThanEqual(x, vec4(0.0)));
|
|
bvec4 isNaN = isnan(x);
|
|
|
|
result.r = isNaN.r ? x.r : result.r;
|
|
result.g = isNaN.g ? x.g : result.g;
|
|
result.b = isNaN.b ? x.b : result.b;
|
|
result.a = isNaN.a ? x.a : result.a;
|
|
|
|
return result;
|
|
`,jH=`
|
|
vec4 result = min(x, vec4(6.)) * vec4(greaterThanEqual(x, vec4(0.0)));
|
|
bvec4 isNaN = isnan(x);
|
|
|
|
result.r = isNaN.r ? x.r : result.r;
|
|
result.g = isNaN.g ? x.g : result.g;
|
|
result.b = isNaN.b ? x.b : result.b;
|
|
result.a = isNaN.a ? x.a : result.a;
|
|
|
|
return result;
|
|
`,qH="return 1.0 / (1.0 + exp(-1.0 * x));",Gr=class{constructor(e,t){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!0,this.outputShape=e,this.enableShapeUniforms=ya(this.outputShape.length),this.userCode=`
|
|
vec4 unaryOperation(vec4 x) {
|
|
${t}
|
|
}
|
|
|
|
void main() {
|
|
vec4 x = getAAtOutCoords();
|
|
vec4 y = unaryOperation(x);
|
|
|
|
setOutput(y);
|
|
}
|
|
`}},XH=class{constructor(e){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!1,this.outputShape=e,this.enableShapeUniforms=ya(this.outputShape.length);let t=e.length,a=Ia("rc",t),n=ft(t),r=MH(t,a),s=a.slice(-2),i=t<=1?"rc":`vec2(${s.join(",")})`;this.userCode=`
|
|
void main() {
|
|
${n} rc = getOutputCoords();
|
|
vec4 packedInput = getA(${r});
|
|
|
|
setOutput(getChannel(packedInput, ${i}));
|
|
}
|
|
`}},KH=Nn.whereImpl,YH=1e-7,ZH=1e-4,W2={};function JH(e){return e in W2||(W2[e]={}),W2[e]}var QH=W().getNumber("CPU_HANDOFF_SIZE_THRESHOLD"),ej=600;function tj(){return W().global.screen==null?1024:W().global.screen.height*W().global.screen.width*window.devicePixelRatio*ej/1024/1024}var Bu=class extends Kl{nextDataId(){return Bu.nextDataId++}constructor(e){if(super(),this.pendingRead=new WeakMap,this.pendingDisposal=new WeakSet,this.dataRefCount=new WeakMap,this.numBytesInGPU=0,this.uploadWaitMs=0,this.downloadWaitMs=0,this.lastGlFlushTime=0,this.warnedAboutMemory=!1,this.pendingDeletes=0,this.disposed=!1,!W().getBool("HAS_WEBGL"))throw new Error("WebGL is not supported on this device");let t;if(e!=null){if(e instanceof _l)t=e;else{let a=Wn(W().getNumber("WEBGL_VERSION"),e);t=new _l(a)}this.binaryCache={},this.gpgpuCreatedLocally=!1}else{let a=Wn(W().getNumber("WEBGL_VERSION"));t=new _l(a),this.binaryCache=JH(W().getNumber("WEBGL_VERSION")),this.gpgpuCreatedLocally=!0}this.gpgpu=t,this.canvas=this.gpgpu.gl.canvas,this.textureManager=new $H(this.gpgpu),this.numMBBeforeWarning=tj(),this.texData=new Zd(this,It())}numDataIds(){return this.texData.numDataIds()-this.pendingDeletes}writeTexture(e,t,a,n,r,s){let i=this.makeTensorInfo(t,a),o=this.texData.get(i.dataId);o.isPacked=!1,o.texture={texture:e,texShape:[n,r]},o.texShape=[n,r];let l=vd(t),u=new rx(l,!1,s),p=this.runWebGLProgram(u,[i],a,[[n,r]]);return p.shape=t,o.texture=null,this.disposeIntermediateTensorInfo(i),p.dataId}write(e,t,a){if((W().getBool("WEBGL_CHECK_NUMERICAL_PROBLEMS")||W().getBool("DEBUG"))&&this.checkNumericalProblems(e),a==="complex64"&&e!=null)throw new Error("Cannot write to a complex64 dtype. Please use tf.complex(real, imag).");let n={id:this.nextDataId()};return this.texData.set(n,{shape:t,dtype:a,values:e,usage:fn.UPLOAD,refCount:1}),n}refCount(e){return this.texData.has(e)?this.texData.get(e).refCount:0}incRef(e){let t=this.texData.get(e);t.refCount++}decRef(e){if(this.texData.has(e)){let t=this.texData.get(e);t.refCount--}}move(e,t,a,n,r){if(W().getBool("DEBUG")&&this.checkNumericalProblems(t),n==="complex64")throw new Error("Cannot write to a complex64 dtype. Please use tf.complex(real, imag).");this.texData.set(e,{shape:a,dtype:n,values:t,usage:fn.UPLOAD,refCount:r})}disposeIntermediateTensorInfo(e){this.disposeData(e.dataId)}readSync(e){let t=this.texData.get(e),{values:a,dtype:n,complexTensorInfos:r,slice:s,shape:i,isPacked:o}=t;if(s!=null){let c;o?c=new Gr(i,Lr):c=new Yn(i,Lr);let d=this.runWebGLProgram(c,[{dataId:e,shape:i,dtype:n}],n),h=this.readSync(d.dataId);return this.disposeIntermediateTensorInfo(d),h}if(a!=null)return this.convertAndCacheOnCPU(e);if(n==="string")return a;let l=this.activeTimers!=null,u;l&&(u=v.now());let p;if(n==="complex64"){let c=this.readSync(r.real.dataId),d=this.readSync(r.imag.dataId);p=S.mergeRealAndImagArrays(c,d)}else p=this.getValuesFromTexture(e);return l&&(this.downloadWaitMs+=v.now()-u),this.convertAndCacheOnCPU(e,p)}async read(e){if(this.pendingRead.has(e)){let h=this.pendingRead.get(e);return new Promise(f=>h.push(f))}let t=this.texData.get(e),{values:a,shape:n,slice:r,dtype:s,complexTensorInfos:i,isPacked:o}=t;if(r!=null){let h;o?h=new Gr(n,Lr):h=new Yn(n,Lr);let f=this.runWebGLProgram(h,[{dataId:e,shape:n,dtype:s}],s),m=this.read(f.dataId);return this.disposeIntermediateTensorInfo(f),m}if(a!=null)return this.convertAndCacheOnCPU(e);if(W().getBool("DEBUG")&&!W().getBool("WEBGL_DOWNLOAD_FLOAT_ENABLED")&&W().getNumber("WEBGL_VERSION")===2)throw new Error("tensor.data() with WEBGL_DOWNLOAD_FLOAT_ENABLED=false and WEBGL_VERSION=2 not yet supported.");let l=null,u;if(s!=="complex64"&&W().get("WEBGL_BUFFER_SUPPORTED")){u=this.decode(e);let h=this.texData.get(u.dataId);l=this.gpgpu.createBufferFromTexture(h.texture.texture,...Vc(n))}this.pendingRead.set(e,[]),s!=="complex64"&&await this.gpgpu.createAndWaitForFence();let p;if(s==="complex64"){let h=await Promise.all([this.read(i.real.dataId),this.read(i.imag.dataId)]),f=h[0],m=h[1];p=S.mergeRealAndImagArrays(f,m)}else if(l==null)p=this.getValuesFromTexture(e);else{let h=v.sizeFromShape(n);p=this.gpgpu.downloadFloat32MatrixFromBuffer(l,h)}if(u!=null&&this.disposeIntermediateTensorInfo(u),l!=null){let h=this.gpgpu.gl;pe(h,()=>h.deleteBuffer(l))}let c=this.convertAndCacheOnCPU(e,p),d=this.pendingRead.get(e);return this.pendingRead.delete(e),d.forEach(h=>h(c)),this.pendingDisposal.has(e)&&(this.pendingDisposal.delete(e),this.disposeData(e)&&It().removeDataId(e,this),this.pendingDeletes--),c}readToGPU(e,t={}){let a=this.texData.get(e),{values:n,shape:r,slice:s,dtype:i,isPacked:o,texture:l}=a;if(i==="complex64")throw new Error("Does not support reading texture for complex64 dtype.");if(s!=null){let d;o?d=new Gr(r,Lr):d=new Yn(r,Lr);let h=this.runWebGLProgram(d,[{dataId:e,shape:r,dtype:i}],i),f=this.readToGPU(h,t);return this.disposeIntermediateTensorInfo(h),f}if(l==null)throw n!=null?new Error("Data is not on GPU but on CPU."):new Error("There is no data on GPU or CPU.");let u=this.decode(e,t.customTexShape),p=It().makeTensorFromTensorInfo(u),c=this.texData.get(u.dataId);return Object.assign({tensorRef:p},c.texture)}bufferSync(e){let t=this.readSync(e.dataId);if(e.dtype==="string")try{let a=t.map(n=>v.decodeString(n));return $e(e.shape,e.dtype,a)}catch(a){throw new Error("Failed to decode encoded string bytes into utf-8")}return $e(e.shape,e.dtype,t)}checkNumericalProblems(e){if(e!=null)for(let t=0;t<e.length;t++){let a=e[t];if(!j6(a))throw W().getBool("WEBGL_RENDER_FLOAT32_CAPABLE")?Error(`The value ${a} cannot be represented with your current settings. Consider enabling float32 rendering: 'tf.env().set('WEBGL_RENDER_FLOAT32_ENABLED', true);'`):Error(`The value ${a} cannot be represented on this device.`)}}getValuesFromTexture(e){let{shape:t,dtype:a,isPacked:n}=this.texData.get(e),r=v.sizeFromShape(t);if(W().getBool("WEBGL_DOWNLOAD_FLOAT_ENABLED")){let c=this.decode(e),d=this.texData.get(c.dataId),h=this.gpgpu.downloadMatrixFromPackedTexture(d.texture.texture,...Vc(t)).subarray(0,r);return this.disposeIntermediateTensorInfo(c),h}let s=W().getBool("WEBGL_PACK")&&n===!0,i=s?vd(t):t,o=s?new OG(i):new PG(i),l=this.runWebGLProgram(o,[{shape:i,dtype:a,dataId:e}],"float32"),u=this.texData.get(l.dataId),p=this.gpgpu.downloadByteEncodedFloatMatrixFromOutputTexture(u.texture.texture,u.texShape[0],u.texShape[1]).subarray(0,r);return this.disposeIntermediateTensorInfo(l),p}timerAvailable(){return W().getNumber("WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_RELIABLE")>0}time(e){let t=this.activeTimers,a=[],n=!1;this.programTimersStack==null?(this.programTimersStack=a,n=!0):this.activeTimers.push(a),this.activeTimers=a,e();let r=v.flatten(this.activeTimers.map(o=>o.query)).filter(o=>o!=null),s=v.flatten(this.activeTimers.map(o=>o.name)).filter(o=>o!=null);this.activeTimers=t,n&&(this.programTimersStack=null);let i={uploadWaitMs:this.uploadWaitMs,downloadWaitMs:this.downloadWaitMs,kernelMs:null,wallMs:null};return(async()=>{if(W().getNumber("WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_RELIABLE")>0){let o=await Promise.all(r);i.kernelMs=v.sum(o),i.getExtraProfileInfo=()=>o.map((l,u)=>({name:s[u],ms:l})).map(l=>`${l.name}: ${l.ms}`).join(", ")}else i.kernelMs={error:"WebGL query timers are not supported in this environment."};return this.uploadWaitMs=0,this.downloadWaitMs=0,i})()}memory(){return{unreliable:!1,numBytesInGPU:this.numBytesInGPU,numBytesInGPUAllocated:this.textureManager.numBytesAllocated,numBytesInGPUFree:this.textureManager.numBytesFree}}startTimer(){return W().getNumber("WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_RELIABLE")>0?this.gpgpu.beginQuery():{startMs:v.now(),endMs:null}}endTimer(e){return W().getNumber("WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_RELIABLE")>0?(this.gpgpu.endQuery(),e):(e.endMs=v.now(),e)}async getQueryTime(e){if(W().getNumber("WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_RELIABLE")>0)return this.gpgpu.waitForQueryAndGetTime(e);let t=e;return t.endMs-t.startMs}disposeData(e,t=!1){if(this.pendingDisposal.has(e))return!1;if(!this.texData.has(e))return!0;if(t?this.texData.get(e).refCount=0:this.texData.get(e).refCount--,!t&&this.texData.get(e).refCount>0)return!1;if(this.pendingRead.has(e))return this.pendingDisposal.add(e),this.pendingDeletes++,!1;this.releaseGPUData(e);let{complexTensorInfos:a}=this.texData.get(e);return a!=null&&(this.disposeData(a.real.dataId,t),this.disposeData(a.imag.dataId,t)),this.texData.delete(e),!0}releaseGPUData(e){let{texture:t,dtype:a,texShape:n,usage:r,isPacked:s,slice:i}=this.texData.get(e),o=i&&i.origDataId||e,l=this.dataRefCount.get(o);l>1?this.dataRefCount.set(o,l-1):(this.dataRefCount.delete(o),t!=null&&(this.numBytesInGPU-=this.computeBytes(n,a),this.textureManager.releaseTexture(t,n,r,s)));let u=this.texData.get(e);u.texture=null,u.texShape=null,u.isPacked=!1,u.slice=null}getTexture(e){return this.uploadToGPU(e),this.texData.get(e).texture.texture}getDataInfo(e){return this.texData.get(e)}shouldExecuteOnCPU(e,t=QH){return W().getBool("WEBGL_CPU_FORWARD")&&e.every(a=>this.texData.get(a.dataId).texture==null&&v.sizeFromShape(a.shape)<t)}getGPGPUContext(){return this.gpgpu}where(e){S.warn("tf.where() in webgl locks the UI thread. Call tf.whereAsync() instead");let t=e.dataSync();return KH(e.shape,t)}packedUnaryOp(e,t,a){let n=new Gr(e.shape,t),r=this.compileAndRun(n,[e],a);return It().makeTensorFromTensorInfo(r)}abs(e){if(this.shouldExecuteOnCPU([e])&&e.dtype!=="complex64"){let n=Lv(this.texData.get(e.dataId).values);return this.makeOutput(e.shape,e.dtype,n)}if(W().getBool("WEBGL_PACK_UNARY_OPERATIONS"))return this.packedUnaryOp(e,lx,e.dtype);let t=new Yn(e.shape,lx),a=this.compileAndRun(t,[e]);return It().makeTensorFromTensorInfo(a)}makeTensorInfo(e,t,a){let n;if(t==="string"&&a!=null&&a.length>0&&v.isString(a[0])){let r=a.map(s=>v.encodeString(s));n=this.write(r,e,t)}else n=this.write(a,e,t);return this.texData.get(n).usage=null,{dataId:n,shape:e,dtype:t}}makeOutput(e,t,a){return It().makeTensorFromTensorInfo(this.makeTensorInfo(e,t,a),this)}unpackTensor(e){let t=new XH(e.shape);return this.runWebGLProgram(t,[e],e.dtype)}packTensor(e){let t=new FH(e.shape),a=!0;return this.runWebGLProgram(t,[e],e.dtype,null,a)}packedReshape(e,t){let a=[Ks(e.shape),...Ys(e.shape)],n={dtype:e.dtype,shape:a,dataId:e.dataId},r=[Ks(t),...Ys(t)],s=new Vv(r,a),i=!0,o=[a],l=this.runWebGLProgram(s,[n],e.dtype,o,i);return{dataId:l.dataId,shape:t,dtype:l.dtype}}decode(e,t){let a=this.texData.get(e),{isPacked:n,shape:r,dtype:s}=a;if(t!=null){let c=v.sizeFromShape(r),d=t[0]*t[1]*4;v.assert(c<=d,()=>"customTexShape is too small. Row * Column * 4 should be equal or larger than the size of the tensor data.")}let i=vd(r),o;n?o=new DG(i):o=new $G(i);let l=!0,u=[t!=null?t:Vc(i)],p=this.runWebGLProgram(o,[{shape:i,dtype:s,dataId:e}],s,u,l,t);return{dtype:s,shape:r,dataId:p.dataId}}runWebGLProgram(e,t,a,n,r=!1,s){let i=this.makeTensorInfo(e.outputShape,a),o=this.texData.get(i.dataId);if(e.packedOutput&&(o.isPacked=!0),e.outPackingScheme===Ud.DENSE){let g=s!=null?s:Vc(e.outputShape);o.texShape=g.map(y=>y*2)}if(e.outTexUsage!=null&&(o.usage=e.outTexUsage),v.sizeFromShape(i.shape)===0)return o.values=v.getTypedArrayFromDType(i.dtype,0),i;let l=[],u=t.map(g=>{if(g.dtype==="complex64")throw new Error("GPGPUProgram does not support complex64 input. For complex64 dtypes, please separate the program into real and imaginary parts.");let y=this.texData.get(g.dataId);if(y.texture==null){if(!e.packedInputs&&v.sizeFromShape(g.shape)<=W().getNumber("WEBGL_SIZE_UPLOAD_UNIFORM"))return{shape:g.shape,texData:null,isUniform:!0,uniformValues:y.values};e.packedInputs&&(y.isPacked=!0,y.shape=g.shape)}if(this.uploadToGPU(g.dataId),!!y.isPacked!=!!e.packedInputs)g=y.isPacked?this.unpackTensor(g):this.packTensor(g),l.push(g),y=this.texData.get(g.dataId);else if(y.isPacked&&!Gd(y.shape,g.shape)){let x=g,A=g.shape;g.shape=y.shape,g=this.packedReshape(g,A),l.push(g),y=this.texData.get(g.dataId),x.shape=A}return{shape:g.shape,texData:y,isUniform:!1}});this.uploadToGPU(i.dataId);let p={shape:i.shape,texData:o,isUniform:!1},c=_G(e,u,p),d=this.getAndSaveBinary(c,()=>MG(this.gpgpu,e,u,p)),h=this.activeTimers!=null,f;h&&(f=this.startTimer()),W().get("ENGINE_COMPILE_ONLY")||FG(this.gpgpu,d,u,p,n),l.forEach(g=>this.disposeIntermediateTensorInfo(g)),h&&(f=this.endTimer(f),this.activeTimers.push({name:e.constructor.name,query:this.getQueryTime(f)}));let m=W().get("WEBGL_FLUSH_THRESHOLD");if(m>0){let g=v.now();g-this.lastGlFlushTime>m&&(this.gpgpu.gl.flush(),this.lastGlFlushTime=g)}if(!W().getBool("WEBGL_LAZILY_UNPACK")&&o.isPacked&&r===!1){let g=this.unpackTensor(i);return this.disposeIntermediateTensorInfo(i),g}return i}compileAndRun(e,t,a,n,r=!1){return a=a||t[0].dtype,this.runWebGLProgram(e,t,a,n,r)}getAndSaveBinary(e,t){return e in this.binaryCache||(this.binaryCache[e]=t()),this.binaryCache[e]}getTextureManager(){return this.textureManager}dispose(){this.disposed||(W().getBool("IS_TEST")||Object.keys(this.binaryCache).forEach(e=>{this.gpgpu.deleteProgram(this.binaryCache[e].webGLProgram),delete this.binaryCache[e]}),this.textureManager.dispose(),this.canvas!=null&&typeof HTMLCanvasElement!="undefined"&&this.canvas instanceof HTMLCanvasElement?this.canvas.remove():this.canvas=null,this.gpgpuCreatedLocally&&(this.gpgpu.program=null,this.gpgpu.dispose()),this.disposed=!0)}floatPrecision(){return this.floatPrecisionValue==null&&(this.floatPrecisionValue=Pe(()=>{if(!W().get("WEBGL_RENDER_FLOAT32_ENABLED")){let e=W().getBool("DEBUG");W().set("DEBUG",!1);let t=this.abs(Ue(1e-8)).dataSync()[0];if(W().set("DEBUG",e),t>0)return 32}return 16})),this.floatPrecisionValue}epsilon(){return this.floatPrecision()===32?YH:ZH}uploadToGPU(e){let t=this.texData.get(e),{shape:a,dtype:n,values:r,texture:s,usage:i,isPacked:o}=t;if(s!=null)return;let l=this.activeTimers!=null,u;l&&(u=v.now());let p=t.texShape;if(p==null&&(p=uv(a,o),t.texShape=p),r!=null){let c=vd(a),d,h=p[1],f=p[0],m=r instanceof Uint8Array||r instanceof Uint8ClampedArray;(o||!m)&&([h,f]=Du(p[0],p[1])),o?d=new LG(c,m):d=new rx(c,m);let g=m?[f,h]:p,y=this.makeTensorInfo(g,n),x=this.texData.get(y.dataId);m?x.usage=fn.PIXELS:x.usage=fn.UPLOAD,x.texShape=g,this.gpgpu.uploadDenseMatrixToTexture(this.getTexture(y.dataId),h,f,r);let A=[[f,h]],b=!0,w=this.runWebGLProgram(d,[y],n,A,b),C=this.texData.get(w.dataId);t.texShape=C.texShape,t.isPacked=C.isPacked,t.usage=C.usage,W().get("ENGINE_COMPILE_ONLY")?this.disposeData(w.dataId):(t.texture=C.texture,t.values=null,this.texData.delete(w.dataId)),this.disposeIntermediateTensorInfo(y),l&&(this.uploadWaitMs+=v.now()-u)}else{let c=this.acquireTexture(p,i,n,o);t.texture=c}}convertAndCacheOnCPU(e,t){let a=this.texData.get(e),{dtype:n}=a;return t!=null&&(a.values=aj(t,n)),a.values}acquireTexture(e,t,a,n){if(this.numBytesInGPU+=this.computeBytes(e,a),!this.warnedAboutMemory&&this.numBytesInGPU>this.numMBBeforeWarning*1024*1024){let r=(this.numBytesInGPU/1024/1024).toFixed(2);this.warnedAboutMemory=!0,console.warn(`High memory usage in GPU: ${r} MB, most likely due to a memory leak`)}return this.textureManager.acquireTexture(e,t,n)}computeBytes(e,t){return e[0]*e[1]*v.bytesPerElement(t)}checkCompileCompletion(){for(let[,e]of Object.entries(this.binaryCache))this.checkCompletion_(e)}async checkCompileCompletionAsync(){let e=[];if(this.gpgpu.parallelCompilationExtension){for(let[,t]of Object.entries(this.binaryCache))e.push(this.checkCompletionAsync_(t));return Promise.all(e)}else{for(let[,t]of Object.entries(this.binaryCache)){let a=new Promise(n=>{try{this.checkCompletion_(t),n(!0)}catch(r){throw r}});e.push(a)}return Promise.all(e)}}async checkCompletionAsync_(e){return this.gpgpu.gl.getProgramParameter(e.webGLProgram,this.gpgpu.parallelCompilationExtension.COMPLETION_STATUS_KHR)?this.checkCompletion_(e):(await x7(),this.checkCompletionAsync_(e))}checkCompletion_(e){if(this.gpgpu.gl.getProgramParameter(e.webGLProgram,this.gpgpu.gl.LINK_STATUS)===!1)throw console.log(this.gpgpu.gl.getProgramInfoLog(e.webGLProgram)),this.gpgpu.gl.getShaderParameter(e.fragmentShader,this.gpgpu.gl.COMPILE_STATUS)===!1?(hg(e.source,this.gpgpu.gl.getShaderInfoLog(e.fragmentShader)),new Error("Failed to compile fragment shader.")):new Error("Failed to link vertex and fragment shaders.");return!0}getUniformLocations(){for(let e of Object.values(this.binaryCache)){this.gpgpu.buildVao(e.webGLProgram);let{variablesLocations:t,customUniformLocations:a,infLoc:n,nanLoc:r,outShapeLocation:s,outShapeStridesLocation:i,outTexShapeLocation:o}=bv(this.gpgpu,e.program,e.webGLProgram);e.variablesLocations=t,e.customUniformLocations=a,e.infLoc=n,e.nanLoc=r,e.outShapeLocation=s,e.outShapeStridesLocation=i,e.outTexShapeLocation=o}}createTensorFromGPUData(e,t,a){e.channels=e.channels||"RGBA";let{texture:n,height:r,width:s,channels:i}=e,o=It().backend;if(!o.gpgpu.gl.isTexture(n))throw new Error("The texture is invalid. Also, please make sure the texture and the TFJS WebGL backend are using the same canvas. If you want to use your own custom canvas, you have to create and use the custom TFJS WebGL backend created from the canvas through 'new tf.MathBackendWebGL(customCanvas)'.");let l=o.writeTexture(n,t,a,r,s,i);return It().makeTensorFromDataId(l,t,a,o)}};Bu.nextDataId=0;function aj(e,t){if(t==="float32"||t==="complex64")return e;if(t==="int32"||t==="bool"){let a=t==="int32"?new Int32Array(e.length):new Uint8Array(e.length);for(let n=0;n<a.length;++n)a[n]=Math.round(e[n]);return a}else throw new Error(`Unknown dtype ${t}`)}var nj="4.5.0";function Uv(){W().set("WEBGL_FORCE_F16_TEXTURES",!0)}Sp.isBrowser()&&Go("webgl",()=>new Bu,2);var rj={forceHalfFloat:Uv},kg=`
|
|
if (isnan(a)) return a;
|
|
if (isnan(b)) return b;
|
|
`,ql=class{constructor(e,t,a){this.variableNames=["A","B"],this.outputShape=S.assertAndGetBroadcastShape(t,a),this.enableShapeUniforms=ya(this.outputShape.length),this.userCode=`
|
|
float binaryOperation(float a, float b) {
|
|
${e}
|
|
}
|
|
|
|
void main() {
|
|
float a = getAAtOutCoords();
|
|
float b = getBAtOutCoords();
|
|
setOutput(binaryOperation(a, b));
|
|
}
|
|
`}},Xo=`
|
|
result.r = isNaN.r ? NAN : result.r;
|
|
result.g = isNaN.g ? NAN : result.g;
|
|
result.b = isNaN.b ? NAN : result.b;
|
|
result.a = isNaN.a ? NAN : result.a;
|
|
`,Up=class{constructor(e,t,a,n=!1){this.variableNames=["A","B"],this.supportsBroadcasting=!0,this.packedInputs=!0,this.packedOutput=!0,this.outputShape=S.assertAndGetBroadcastShape(t,a);let r=this.outputShape.length;this.enableShapeUniforms=ya(r);let s="";if(n)if(r===0||v.sizeFromShape(this.outputShape)===1)s=`
|
|
result.y = 0.;
|
|
result.z = 0.;
|
|
result.w = 0.;
|
|
`;else if(s=`
|
|
${ft(r)} coords = getOutputCoords();
|
|
`,r===1)this.enableShapeUniforms?s+=`
|
|
result.y = (coords + 1) >= outShape ? 0. : result.y;
|
|
result.z = 0.;
|
|
result.w = 0.;
|
|
`:s+=`
|
|
result.y = (coords + 1) >= ${this.outputShape[0]} ? 0. : result.y;
|
|
result.z = 0.;
|
|
result.w = 0.;
|
|
`;else{let i=Ia("coords",r);this.enableShapeUniforms?s+=`
|
|
bool nextRowOutOfBounds =
|
|
(${i[r-2]} + 1) >= outShape[${r} - 2];
|
|
bool nextColOutOfBounds =
|
|
(${i[r-1]} + 1) >= outShape[${r} - 1];
|
|
result.y = nextColOutOfBounds ? 0. : result.y;
|
|
result.z = nextRowOutOfBounds ? 0. : result.z;
|
|
result.w = nextColOutOfBounds || nextRowOutOfBounds ? 0. : result.w;
|
|
`:s+=`
|
|
bool nextRowOutOfBounds =
|
|
(${i[r-2]} + 1) >= ${this.outputShape[r-2]};
|
|
bool nextColOutOfBounds =
|
|
(${i[r-1]} + 1) >= ${this.outputShape[r-1]};
|
|
result.y = nextColOutOfBounds ? 0. : result.y;
|
|
result.z = nextRowOutOfBounds ? 0. : result.z;
|
|
result.w = nextColOutOfBounds || nextRowOutOfBounds ? 0. : result.w;
|
|
`}this.userCode=`
|
|
vec4 binaryOperation(vec4 a, vec4 b) {
|
|
${e}
|
|
}
|
|
|
|
void main() {
|
|
vec4 a = getAAtOutCoords();
|
|
vec4 b = getBAtOutCoords();
|
|
|
|
vec4 result = binaryOperation(a, b);
|
|
${s}
|
|
|
|
setOutput(result);
|
|
}
|
|
`}};function en(e){let{inputs:t,backend:a}=e,{x:n}=t;return a.incRef(n.dataId),{dataId:n.dataId,shape:n.shape,dtype:n.dtype}}var sj={kernelName:zi,backendName:"webgl",kernelFunc:en};function ds(e){let{inputs:t,backend:a}=e,{real:n,imag:r}=t,s=a.makeTensorInfo(n.shape,"complex64"),i=a.texData.get(s.dataId),o=en({inputs:{x:n},backend:a}),l=en({inputs:{x:r},backend:a});return i.complexTensorInfos={real:o,imag:l},s}var ij={kernelName:tp,backendName:"webgl",kernelFunc:ds},Gv="return (a < 0.) ? b * a : a;",Hv=`
|
|
vec4 aLessThanZero = vec4(lessThan(a, vec4(0.)));
|
|
return (aLessThanZero * (b * a)) + ((vec4(1.0) - aLessThanZero) * a);
|
|
`;function oj(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{alpha:s}=n,i=a.makeTensorInfo([],"float32",v.createScalarValue(s,"float32")),o=W().getBool("WEBGL_PACK_BINARY_OPERATIONS")?new Up(Hv,r.shape,i.shape):new ql(Gv,r.shape,i.shape),l=a.runWebGLProgram(o,[r,i],"float32");return a.disposeIntermediateTensorInfo(i),l}var lj={kernelName:Vi,backendName:"webgl",kernelFunc:oj},jv="return (a < 0.) ? b * a : a;",qv=`
|
|
vec4 aLessThanZero = vec4(lessThan(a, vec4(0.)));
|
|
return (aLessThanZero * (b * a)) + ((vec4(1.0) - aLessThanZero) * a);
|
|
`;function uj(e){let{inputs:t,backend:a}=e,{x:n,alpha:r}=t,s=W().getBool("WEBGL_PACK_BINARY_OPERATIONS")?new Up(qv,n.shape,r.shape):new ql(jv,n.shape,r.shape);return a.runWebGLProgram(s,[n,r],"float32")}var dj={kernelName:fo,backendName:"webgl",kernelFunc:uj},Vu="if (isnan(x)) return x;";function et({opSnippet:e,packedOpSnippet:t,cpuKernelImpl:a,dtype:n}){return({inputs:r,backend:s})=>{let{x:i}=r,o=s,l=n||i.dtype;if(o.shouldExecuteOnCPU([i])&&a!=null){let c=o.texData.get(i.dataId),d=a(c.values,l);return o.makeTensorInfo(i.shape,l,d)}let u=W().getBool("WEBGL_PACK_UNARY_OPERATIONS")&&t!=null,p;return u?p=new Gr(i.shape,t):p=new Yn(i.shape,e),o.runWebGLProgram(p,[i],l)}}function fa({opSnippet:e,packedOpSnippet:t,checkOutOfBounds:a=!1,supportsComplex:n=!1,cpuKernelImpl:r,dtype:s}){return({inputs:i,backend:o})=>{let{a:l,b:u}=i,p=o;if(n&&l.dtype==="complex64"){let f=p.texData.get(l.dataId),m=p.texData.get(u.dataId),[g,y]=[[f.complexTensorInfos.real,m.complexTensorInfos.real],[f.complexTensorInfos.imag,m.complexTensorInfos.imag]].map(A=>{let[b,w]=A,C={dataId:b.dataId,dtype:b.dtype,shape:l.shape},T={dataId:w.dataId,dtype:w.dtype,shape:u.shape},N=new ql(e,l.shape,u.shape);return p.runWebGLProgram(N,[C,T],pa(b.dtype,w.dtype))}),x=ds({inputs:{real:g,imag:y},backend:p});return p.disposeIntermediateTensorInfo(g),p.disposeIntermediateTensorInfo(y),x}let c=s||pa(l.dtype,u.dtype);if((l.dtype==="string"||u.dtype==="string"||p.shouldExecuteOnCPU([l,u]))&&r!=null){let f=p.texData.get(l.dataId).values,m=p.texData.get(u.dataId).values,g=l.dtype==="string"?S.fromUint8ToStringArray(f):f,y=l.dtype==="string"?S.fromUint8ToStringArray(m):m,[x,A]=r(l.shape,u.shape,g,y,c),b=p.makeTensorInfo(A,c),w=p.texData.get(b.dataId);return w.values=x,b}let d=W().getBool("WEBGL_PACK_BINARY_OPERATIONS")&&t!=null,h;return d?h=new Up(t,l.shape,u.shape,a):h=new ql(e,l.shape,u.shape),p.runWebGLProgram(h,[l,u],c)}}function Hd(e,t=!1){if(e==="linear")return t?UH:zH;if(e==="relu")return t?HH:WH;if(e==="elu")return t?GH:LH;if(e==="relu6")return t?jH:BH;if(e==="prelu")return t?qv:jv;if(e==="leakyrelu")return t?Hv:Gv;if(e==="sigmoid")return t?qH:VH;throw new Error(`Activation ${e} has not been implemented for the WebGL backend.`)}var Xv=class{constructor(e,t,a,n=!1,r=!1,s=!1,i=null,o=!1,l=!1){this.variableNames=["matrixA","matrixB"],this.packedInputs=!0,this.packedOutput=!0,this.outputShape=a,this.enableShapeUniforms=ya(this.outputShape.length);let u=n?e[1]:e[2],p=Math.ceil(u/2),c=n?"i * 2, rc.y":"rc.y, i * 2",d=r?"rc.z, i * 2":"i * 2, rc.z",h=n?["a.xxyy","a.zzww"]:["a.xxzz","a.yyww"],f=r?["b.xzxz","b.ywyw"]:["b.xyxy","b.zwzw"],m="",g="";i&&(o?m=`vec4 activation(vec4 a) {
|
|
vec4 b = getPreluActivationWeightsAtOutCoords();
|
|
${i}
|
|
}`:l?m=`vec4 activation(vec4 a) {
|
|
vec4 b = getLeakyreluAlphaAtOutCoords();
|
|
${i}
|
|
}`:m=`vec4 activation(vec4 x) {
|
|
${i}
|
|
}`,g="result = activation(result);");let y=s?"result += getBiasAtOutCoords();":"";s&&this.variableNames.push("bias"),o&&this.variableNames.push("preluActivationWeights"),l&&this.variableNames.push("leakyreluAlpha");let x="rc.x",A="rc.x";e[0]<t[0]?x=`imod(rc.x, ${e[0]})`:t[0]<e[0]&&(A=`imod(rc.x, ${t[0]})`),this.userCode=`
|
|
${m}
|
|
// Don't use uniform for sharedDimensionPacked for performance.
|
|
const float sharedDimension = ${p}.0;
|
|
|
|
vec4 dot2x2ARowBCol(ivec3 rc) {
|
|
vec4 result = vec4(0);
|
|
int batchA = ${x};
|
|
int batchB = ${A};
|
|
for (int i = 0; i < ${p}; i++) {
|
|
vec4 a = getMatrixA(batchA, ${c});
|
|
vec4 b = getMatrixB(batchB, ${d});
|
|
|
|
// These swizzled products need to be separately added.
|
|
// See: https://github.com/tensorflow/tfjs/issues/1735
|
|
result += (${h[0]} * ${f[0]});
|
|
result += (${h[1]} * ${f[1]});
|
|
}
|
|
return result;
|
|
}
|
|
|
|
void main() {
|
|
ivec3 rc = getOutputCoords();
|
|
vec4 result = dot2x2ARowBCol(rc);
|
|
|
|
${y}
|
|
|
|
${g}
|
|
|
|
setOutput(result);
|
|
}
|
|
`}},ux={REAL:"return areal * breal - aimag * bimag;",IMAG:"return areal * bimag + aimag * breal;"},dx=class{constructor(e,t,a){this.variableNames=["AReal","AImag","BReal","BImag"],this.outputShape=S.assertAndGetBroadcastShape(t,a),this.userCode=`
|
|
float binaryOpComplex(
|
|
float areal, float aimag, float breal, float bimag) {
|
|
${e}
|
|
}
|
|
|
|
void main() {
|
|
float areal = getARealAtOutCoords();
|
|
float aimag = getAImagAtOutCoords();
|
|
float breal = getBRealAtOutCoords();
|
|
float bimag = getBImagAtOutCoords();
|
|
setOutput(binaryOpComplex(areal, aimag, breal, bimag));
|
|
}
|
|
`}},px="return a * b;";function Ig(e){let{inputs:t,backend:a}=e,{a:n,b:r}=t,s=S.upcastType(n.dtype,r.dtype);if(n.dtype==="complex64"){let o=a.texData.get(n.dataId),l=a.texData.get(r.dataId),u=new dx(ux.REAL,n.shape,r.shape),p=new dx(ux.IMAG,n.shape,r.shape),c=[{dataId:o.complexTensorInfos.real.dataId,dtype:o.complexTensorInfos.real.dtype,shape:n.shape},{dataId:o.complexTensorInfos.imag.dataId,dtype:o.complexTensorInfos.imag.dtype,shape:n.shape},{dataId:l.complexTensorInfos.real.dataId,dtype:l.complexTensorInfos.real.dtype,shape:r.shape},{dataId:l.complexTensorInfos.imag.dataId,dtype:l.complexTensorInfos.imag.dtype,shape:r.shape}],d=a.runWebGLProgram(u,c,"float32"),h=a.runWebGLProgram(p,c,"float32"),f=ds({inputs:{real:d,imag:h},backend:a});return a.disposeIntermediateTensorInfo(d),a.disposeIntermediateTensorInfo(h),f}if(a.shouldExecuteOnCPU([n,r])){let o=a.texData.get(n.dataId),l=a.texData.get(r.dataId),[u,p]=oH(n.shape,r.shape,o.values,l.values,s),c=a.makeTensorInfo(p,s),d=a.texData.get(c.dataId);return d.values=u,c}let i;return W().getBool("WEBGL_PACK_BINARY_OPERATIONS")?i=new Up(px,n.shape,r.shape):i=new ql(px,n.shape,r.shape),a.runWebGLProgram(i,[n,r],s)}var pj={kernelName:io,backendName:"webgl",kernelFunc:Ig};function cj(e,t,a){let n=[Ks(e.shape),...Ys(e.shape)],r={dtype:e.dtype,shape:n,dataId:e.dataId},s=[Ks(t),...Ys(t)],i=new Vv(s,n),o=!0,l=[n],u=a.runWebGLProgram(i,[r],e.dtype,l,o);return{dataId:u.dataId,shape:t,dtype:u.dtype}}function ue(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{shape:s}=n,i=a,o=v.sizeFromShape(r.shape),l=v.inferFromImplicitShape(s,o),u=v.sizeFromShape(l);v.assert(o===u,()=>`The new shape (${l}) has ${u} elements and the old shape (${r.shape}) has ${o} elements. The new shape and old shape must have the same number of elements.`);let p=i.texData.get(r.dataId);return p.isPacked&&!Gd(r.shape,l)&&!(p.texture!==null&&Gd(p.shape,l))?cj(r,l,i):(i.incRef(r.dataId),{dataId:r.dataId,shape:l,dtype:r.dtype})}var hj={kernelName:bu,backendName:"webgl",kernelFunc:ue},cx=class{constructor(e,t){this.variableNames=["x"];let{windowSize:a,batchSize:n,inSize:r,outSize:s}=e;this.outputShape=[n,s];let i=Math.floor(a/4)*4,o=a%4,l="sumValue += dot(values, ones);";if(t!=null){let p=1/t;l=`sumValue += dot(values * ${v.isInt(p)?p.toPrecision(2):p}, ones);`}let u="";r%a>0&&(u=`
|
|
if (inIdx < 0 || inIdx >= ${r}) {
|
|
return 0.0;
|
|
}
|
|
`),this.userCode=`
|
|
const vec4 ones = vec4(1.0, 1.0, 1.0, 1.0);
|
|
|
|
float getValue(int batch, int inIdx) {
|
|
${u}
|
|
return getX(batch, inIdx);
|
|
}
|
|
|
|
void main() {
|
|
ivec2 coords = getOutputCoords();
|
|
int batch = coords[0];
|
|
int outIdx = coords[1];
|
|
int inOffset = outIdx * ${a};
|
|
|
|
float sumValue = 0.0;
|
|
|
|
for (int i = 0; i < ${i}; i += 4) {
|
|
int inIdx = inOffset + i;
|
|
vec4 values = vec4(
|
|
getValue(batch, inIdx),
|
|
getValue(batch, inIdx + 1),
|
|
getValue(batch, inIdx + 2),
|
|
getValue(batch, inIdx + 3)
|
|
);
|
|
|
|
${l}
|
|
}
|
|
|
|
int inIdx = inOffset + ${i};
|
|
if (${o===1}) {
|
|
vec4 values = vec4(getValue(batch, inIdx), 0.0, 0.0, 0.0);
|
|
|
|
${l}
|
|
} else if (${o===2}) {
|
|
vec4 values = vec4(
|
|
getValue(batch, inIdx),
|
|
getValue(batch, inIdx + 1), 0.0, 0.0);
|
|
|
|
${l}
|
|
} else if (${o===3}) {
|
|
vec4 values = vec4(
|
|
getValue(batch, inIdx),
|
|
getValue(batch, inIdx + 1),
|
|
getValue(batch, inIdx + 2), 0.0);
|
|
|
|
${l}
|
|
}
|
|
setOutput(sumValue);
|
|
}
|
|
`}},fj=class{constructor(e,t){this.variableNames=["x"];let{windowSize:a,batchSize:n,inSize:r,outSize:s}=e;this.outputShape=[n,s];let i="0.0",o="";t==="prod"?i="1.0":t==="min"?(i="1.0 / 1e-20",o="min"):t==="max"&&(i="-1.0 / 1e-20",o="max");let l=`${t}(${t}(${t}(minMaxValue[0], minMaxValue[1]), minMaxValue[2]), minMaxValue[3])`;t==="sum"?l="sumValue":t==="prod"?l="prodValue":t==="all"?l="allValue":t==="any"&&(l="anyValue");let u=Math.floor(a/4)*4,p=a%4,c=`
|
|
if (${t==="sum"}) {
|
|
sumValue += dot(values, ones);
|
|
} else if (${t==="prod"}) {
|
|
vec2 tmp = vec2(values[0], values[1]) * vec2(values[2], values[3]);
|
|
prodValue *= tmp[0] * tmp[1];
|
|
} else {
|
|
minMaxValue = ${o}(values, minMaxValue);
|
|
if (${t==="min"} || ${t==="max"}) {
|
|
minMaxValue = ${o}(values, minMaxValue);
|
|
bvec4 isNaN = isnan(values);
|
|
if (isNaN.r || isNaN.g || isNaN.b || isNaN.a) {
|
|
minMaxValue = vec4(NAN);
|
|
}
|
|
}
|
|
}
|
|
`,d="vec4";t==="all"?(i="1.0",c=`
|
|
bool reducedAllValue = all(values);
|
|
float floatedReducedAllValue = float(reducedAllValue);
|
|
allValue = float(allValue >= 1.0 && floatedReducedAllValue >= 1.0);
|
|
`,d="bvec4"):t==="any"&&(i="0.0",c=`
|
|
bool reducedAnyValue = any(values);
|
|
float floatedReducedAnyValue = float(reducedAnyValue);
|
|
anyValue = float(anyValue >= 1.0 || floatedReducedAnyValue >= 1.0);
|
|
`,d="bvec4");let h="";r%a>0&&(h=`
|
|
if (inIdx < 0 || inIdx >= ${r}) {
|
|
return initializationValue;
|
|
}
|
|
`),this.userCode=`
|
|
const float initializationValue = ${i};
|
|
const vec4 ones = vec4(1.0, 1.0, 1.0, 1.0);
|
|
|
|
float getValue(int batch, int inIdx) {
|
|
${h}
|
|
return getX(batch, inIdx);
|
|
}
|
|
|
|
void main() {
|
|
ivec2 coords = getOutputCoords();
|
|
int batch = coords[0];
|
|
int outIdx = coords[1];
|
|
int inOffset = outIdx * ${a};
|
|
|
|
vec4 minMaxValue = vec4(${i});
|
|
float prodValue = 1.0;
|
|
float sumValue = 0.0;
|
|
float allValue = 1.0;
|
|
float anyValue = 0.0;
|
|
|
|
for (int i = 0; i < ${u}; i += 4) {
|
|
int inIdx = inOffset + i;
|
|
${d} values = ${d}(
|
|
getValue(batch, inIdx),
|
|
getValue(batch, inIdx + 1),
|
|
getValue(batch, inIdx + 2),
|
|
getValue(batch, inIdx + 3)
|
|
);
|
|
|
|
${c}
|
|
}
|
|
|
|
int inIdx = inOffset + ${u};
|
|
if (${p===1}) {
|
|
${d} values = ${d}(
|
|
getValue(batch, inIdx),
|
|
initializationValue,
|
|
initializationValue,
|
|
initializationValue
|
|
);
|
|
|
|
${c}
|
|
} else if (${p===2}) {
|
|
${d} values = ${d}(
|
|
getValue(batch, inIdx),
|
|
getValue(batch, inIdx + 1),
|
|
initializationValue,
|
|
initializationValue
|
|
);
|
|
|
|
${c}
|
|
} else if (${p===3}) {
|
|
${d} values = ${d}(
|
|
getValue(batch, inIdx),
|
|
getValue(batch, inIdx + 1),
|
|
getValue(batch, inIdx + 2),
|
|
initializationValue
|
|
);
|
|
|
|
${c}
|
|
}
|
|
setOutput(${l});
|
|
}
|
|
`}};function mj(e){let t=[];for(;t.length===0||t[t.length-1].outSize!==1;){let a=t.length?t[t.length-1].outSize:e[1],n=S.computeOptimalWindowSize(a);t.push({inSize:a,windowSize:n,outSize:Math.ceil(a/n)})}return t}function Ko(e,t,a,n){let r=mj(e.shape),s=e;for(let i=0;i<r.length;i++){let{inSize:o,windowSize:l,outSize:u}=r[i],p,c;a==="mean"?p=i===0?new cx({windowSize:l,inSize:o,batchSize:e.shape[0],outSize:u},o):new cx({windowSize:l,inSize:o,batchSize:e.shape[0],outSize:u}):p=new fj({windowSize:l,inSize:o,batchSize:e.shape[0],outSize:u},a),c=s,s=n.runWebGLProgram(p,[s],t),c.dataId!==e.dataId&&n.disposeIntermediateTensorInfo(c)}return s}var gj=class{constructor(e,t){this.variableNames=["A"];let a=new Array(e.length);for(let s=0;s<a.length;s++)a[s]=e[t[s]];this.outputShape=a,this.rank=a.length;let n=ft(this.rank),r=yj(t);this.userCode=`
|
|
void main() {
|
|
${n} resRC = getOutputCoords();
|
|
setOutput(getA(${r}));
|
|
}
|
|
`}};function yj(e){let t=e.length;if(t>6)throw Error(`Transpose for rank ${t} is not yet supported`);let a=["resRC.x","resRC.y","resRC.z","resRC.w","resRC.u","resRC.v"],n=new Array(t);for(let r=0;r<e.length;r++)n[e[r]]=a[r];return n.join()}var xj=class{constructor(e,t){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!0;let a=new Array(e.length);for(let u=0;u<a.length;u++)a[u]=e[t[u]];if(this.outputShape=a,this.rank=a.length,this.rank>6)throw Error(`Packed transpose for rank ${this.rank} is not yet supported.`);let n=ft(this.rank),r=Bv("rc",this.rank),s=new Array(this.rank);for(let u=0;u<t.length;u++)s[t[u]]=r[u];let i=`vec2(${s.slice(-2).join()})`,o=`++${r[this.rank-1]} < ${a[this.rank-1]}`,l=`getChannel(getA(${s.join()}), ${i})`;this.userCode=`
|
|
void main() {
|
|
${n} rc = getOutputCoords();
|
|
vec4 result = vec4(0.);
|
|
result[0] = ${l};
|
|
if(${o}) {
|
|
result[1] = ${l};
|
|
}
|
|
--${r[this.rank-1]};
|
|
if(++${r[this.rank-2]} < ${a[this.rank-2]}) {
|
|
result[2] = ${l};
|
|
if(${o}) {
|
|
result[3] = ${l};
|
|
}
|
|
}
|
|
setOutput(result);
|
|
}
|
|
`}};function Qh(e,t,a){let n=W().getBool("WEBGL_PACK_ARRAY_OPERATIONS")?new xj(e.shape,t):new gj(e.shape,t);return a.runWebGLProgram(n,[e],e.dtype)}function Aj(e,t,a,n){let r=t,s=e.shape.length,i=v.parseAxisParam(r,e.shape),o=i,l=S.getAxesPermutation(o,s),u=l!=null,p=e;u&&(p=Qh(e,l,n),o=S.getInnerMostAxes(o.length,s)),S.assertAxesAreInnerMostDims("sum",o,s);let[c,d]=S.computeOutAndReduceShapes(p.shape,o),h=c;a&&(h=S.expandShapeToKeepDim(c,i));let f=v.sizeFromShape(d),m=v.sizeFromShape(e.shape)/f,g=ue({inputs:{x:p},attrs:{shape:[m,f]},backend:n}),y=Ip(e.dtype),x=Ko(g,y,"sum",n),A=ue({inputs:{x},attrs:{shape:h},backend:n});return n.disposeIntermediateTensorInfo(g),n.disposeIntermediateTensorInfo(x),u&&n.disposeIntermediateTensorInfo(p),A}function e0(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{axis:s,keepDims:i}=n;return Aj(r,s,i,a)}var bj={kernelName:_o,backendName:"webgl",kernelFunc:e0};function Ta(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{perm:s}=n,i=a,o=r.shape.length,l=new Array(o);for(let p=0;p<l.length;p++)l[p]=r.shape[s[p]];let u;if(i.shouldExecuteOnCPU([r])){let p=i.texData.get(r.dataId).values,c=wg(p,r.shape,r.dtype,s,l);u=i.makeTensorInfo(l,r.dtype);let d=i.texData.get(u.dataId);d.values=c}else u=Qh(r,s,i);return u}var vj={kernelName:wr,backendName:"webgl",kernelFunc:Ta},Kv=1e3;function fh({a:e,b:t,transposeA:a,transposeB:n,backend:r,bias:s=null,preluActivationWeights:i=null,leakyreluAlpha:o=0,activation:l=null}){let u=e.shape.length,p=t.shape.length,c=a?e.shape[u-2]:e.shape[u-1],d=n?t.shape[p-1]:t.shape[p-2],h=a?e.shape[u-1]:e.shape[u-2],f=n?t.shape[p-2]:t.shape[p-1],m=e.shape.slice(0,-2),g=t.shape.slice(0,-2),y=v.sizeFromShape(m),x=v.sizeFromShape(g),A=Ho.assertAndGetBroadcastShape(e.shape.slice(0,-2),t.shape.slice(0,-2)).concat([h,f]);v.assert(c===d,()=>`Error in matMul: inner shapes (${c}) and (${d}) of Tensors with shapes ${e.shape} and ${t.shape} and transposeA=${a} and transposeB=${n} must match.`);let b=a?[y,c,h]:[y,h,c],w=n?[x,f,d]:[x,d,f],C=ue({inputs:{x:e},backend:r,attrs:{shape:b}}),T=ue({inputs:{x:t},backend:r,attrs:{shape:w}}),N=[C,T],M=Math.max(y,x),F=a?C.shape[1]:C.shape[2],E=s!=null,I=i!=null,_=l==="leakyrelu",O=l!=null?Hd(l,!0):null,B=E||I||_||O!=null,$;if((h===1||f===1)&&F>Kv&&B===!1){let G=C,q=T;a&&(G=Ta({inputs:{x:C},backend:r,attrs:{perm:[0,2,1]}}),N.push(G)),n&&(q=Ta({inputs:{x:T},backend:r,attrs:{perm:[0,2,1]}}),N.push(q));let H=f!==1,V=f===1,Z=G;H&&(Z=ue({inputs:{x:G},backend:r,attrs:{shape:[M,F,1]}}),N.push(Z));let X=f===1?2:1,re=q;V&&(re=ue({inputs:{x:q},backend:r,attrs:{shape:[M,1,F]}}),N.push(re));let ee=Ig({inputs:{a:Z,b:re},backend:r});$=e0({inputs:{x:ee},backend:r,attrs:{axis:X,keepDims:!0}}),N.push(ee)}else{let G=pa(e.dtype,t.dtype),q=new Xv(b,w,[M,h,f],a,n,E,O,I,_),H=[C,T];if(s!=null&&H.push(s),I&&H.push(i),_){let V=r.makeTensorInfo([],"float32",v.createScalarValue(o,"float32"));H.push(V),N.push(V)}$=r.runWebGLProgram(q,H,G)}let U=ue({inputs:{x:$},backend:r,attrs:{shape:A}});N.push($);for(let G of N)r.disposeIntermediateTensorInfo(G);return U}function wj(e){let{inputs:t,backend:a,attrs:n}=e,{a:r,b:s,bias:i,preluActivationWeights:o}=t,{transposeA:l,transposeB:u,activation:p,leakyreluAlpha:c}=n;return fh({a:r,b:s,transposeA:l,transposeB:u,backend:a,bias:i,preluActivationWeights:o,leakyreluAlpha:c,activation:p})}var kj={kernelName:Xr,backendName:"webgl",kernelFunc:wj},hx="return abs(x);";function Ij(e){let{inputs:t,backend:a}=e,{x:n}=t;if(a.shouldExecuteOnCPU([n])&&n.dtype!=="complex64"){let s=a.texData.get(n.dataId),i=Lv(s.values);return a.makeTensorInfo(n.shape,n.dtype,i)}let r;return W().getBool("WEBGL_PACK_UNARY_OPERATIONS")?r=new Gr(n.shape,hx):r=new Yn(n.shape,hx),a.runWebGLProgram(r,[n],n.dtype)}var Sj={kernelName:Zl,backendName:"webgl",kernelFunc:Ij},Cj=Rn+`
|
|
if (abs(x) > 1.) {
|
|
return NAN;
|
|
}
|
|
return acos(x);
|
|
`,Tj=et({opSnippet:Cj}),Nj={kernelName:Qs,backendName:"webgl",kernelFunc:Tj},Rj=Rn+`
|
|
if (x < 1.0) return NAN;
|
|
return log(x + sqrt(x * x - 1.0));`,Ej=et({opSnippet:Rj}),Mj={kernelName:ei,backendName:"webgl",kernelFunc:Ej},fx="return a + b;",Fj=fa({opSnippet:fx,packedOpSnippet:fx,supportsComplex:!0,cpuKernelImpl:BG}),_j={kernelName:ns,backendName:"webgl",kernelFunc:Fj},$j=class{constructor(e,t){this.outputShape=[],this.outputShape=e,this.variableNames=t.map((r,s)=>`T${s}`);let a=[];this.variableNames.forEach(r=>{a.push(`float v${r} = get${r}AtOutCoords();`)});let n=this.variableNames.map(r=>`v${r}`).join(" + ");this.userCode=`
|
|
void main() {
|
|
${a.join(`
|
|
`)}
|
|
|
|
float result = ${n};
|
|
setOutput(result);
|
|
}
|
|
`}},Dj=class{constructor(e,t){this.outputShape=[],this.packedInputs=!0,this.packedOutput=!0,this.outputShape=e,this.variableNames=t.map((r,s)=>`T${s}`);let a=[];this.variableNames.forEach(r=>{a.push(`vec4 v${r} = get${r}AtOutCoords();`)});let n=this.variableNames.map(r=>`v${r}`).join(" + ");this.userCode=`
|
|
void main() {
|
|
${a.join(`
|
|
`)}
|
|
|
|
vec4 result = ${n};
|
|
setOutput(result);
|
|
}
|
|
`}};function Jc(e){let{inputs:t,backend:a}=e,n=t;if(n.length===1)return en({inputs:{x:n[0]},backend:a});if(n.length>W().get("WEBGL_MAX_TEXTURES_IN_SHADER")){let o=Math.floor(n.length/2),l=Jc({inputs:n.slice(0,o),backend:a}),u=Jc({inputs:n.slice(o),backend:a});return Jc({inputs:[l,u],backend:a})}let r=n.map(o=>o.dtype).reduce((o,l)=>pa(o,l)),s=n.map(o=>o.shape),i=W().getBool("WEBGL_PACK")?new Dj(n[0].shape,s):new $j(n[0].shape,s);return a.runWebGLProgram(i,n,r)}var Pj={kernelName:ti,backendName:"webgl",kernelFunc:Jc};function Oj(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{axis:s,keepDims:i}=n,o=r.shape.length,l=v.parseAxisParam(s,r.shape),u=l,p=S.getAxesPermutation(u,o),c=r;p!=null&&(c=Ta({inputs:{x:r},backend:a,attrs:{perm:p}}),u=S.getInnerMostAxes(u.length,o)),S.assertAxesAreInnerMostDims("all",u,o);let[d,h]=S.computeOutAndReduceShapes(c.shape,u),f=v.sizeFromShape(h),m=ue({inputs:{x:c},backend:a,attrs:{shape:[-1,f]}}),g=Ko(m,m.dtype,"all",a),y;if(i){let x=S.expandShapeToKeepDim(d,l);y=ue({inputs:{x:g},backend:a,attrs:{shape:x}})}else y=ue({inputs:{x:g},backend:a,attrs:{shape:d}});return a.disposeIntermediateTensorInfo(m),a.disposeIntermediateTensorInfo(g),p!=null&&a.disposeIntermediateTensorInfo(c),y}var zj={kernelName:ai,backendName:"webgl",kernelFunc:Oj};function Lj(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{axis:s,keepDims:i}=n,o=r.shape.length,l=v.parseAxisParam(s,r.shape),u=l,p=S.getAxesPermutation(u,o),c=r;p!=null&&(c=Ta({inputs:{x:r},backend:a,attrs:{perm:p}}),u=S.getInnerMostAxes(u.length,o)),S.assertAxesAreInnerMostDims("any",u,o);let[d,h]=S.computeOutAndReduceShapes(c.shape,u),f=v.sizeFromShape(h),m=ue({inputs:{x:c},backend:a,attrs:{shape:[-1,f]}}),g=Ko(m,m.dtype,"any",a),y;if(i){let x=S.expandShapeToKeepDim(d,l);y=ue({inputs:{x:g},backend:a,attrs:{shape:x}})}else y=ue({inputs:{x:g},backend:a,attrs:{shape:d}});return a.disposeIntermediateTensorInfo(m),a.disposeIntermediateTensorInfo(g),p!=null&&a.disposeIntermediateTensorInfo(c),y}var Wj={kernelName:ni,backendName:"webgl",kernelFunc:Lj},Bj=class{constructor(e,t,a){this.variableNames=["A"];let{windowSize:n,batchSize:r,outSize:s}=e;a||this.variableNames.push("bestIndicesA"),this.outputShape=[r,s];let i=t==="max"?">":"<",o=a?"inOffset + i;":"round(getBestIndicesA(batch, inOffset + i));";this.userCode=`
|
|
void main() {
|
|
ivec2 coords = getOutputCoords();
|
|
int batch = coords[0];
|
|
int outIdx = coords[1];
|
|
int inOffset = outIdx * ${n};
|
|
|
|
int bestIndex = inOffset;
|
|
float bestValue = getA(batch, bestIndex);
|
|
|
|
for (int i = 0; i < ${n}; i++) {
|
|
int inIdx = ${o};
|
|
float candidate = getA(batch, inIdx);
|
|
if (candidate ${i} bestValue) {
|
|
bestValue = candidate;
|
|
bestIndex = inIdx;
|
|
}
|
|
}
|
|
setOutput(float(bestIndex));
|
|
}
|
|
`}},Vj=class{constructor(e,t,a,n){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!0,v.assert(e.length>2,()=>`Packed arg${a.charAt(0).toUpperCase()+a.slice(1)} supports only inputs with rank above 2.`);let r=e[e.length-1],s=Math.ceil(r/t);this.outputShape=e.slice(0,-1),s>1&&this.outputShape.push(s),n||this.variableNames.push("bestIndicesA");let i=this.outputShape,o=i.length,l=ft(o),u=Ia("coords",o),p,c;if(s===1){c=o+1;let T=ft(c);p=`
|
|
${T} sourceLocR = ${T}(${u.join()}, 0);
|
|
++${u[o-1]};
|
|
${T} sourceLocG = ${T}(${u.join()}, 0);
|
|
++${u[o-2]};
|
|
${T} sourceLocA = ${T}(${u.join()}, 0);
|
|
--${u[o-1]};
|
|
${T} sourceLocB = ${T}(${u.join()}, 0);
|
|
--${u[o-2]};`}else c=o,p=`
|
|
${l} sourceLocR = coords;
|
|
++${u[o-1]};
|
|
${l} sourceLocG = coords;
|
|
++${u[o-2]};
|
|
${l} sourceLocA = coords;
|
|
--${u[o-1]};
|
|
${l} sourceLocB = coords;
|
|
--${u[o-2]};`;let d=["x","y","z","w","u","v"].slice(0,c),h="."+d[c-1],f=d.map(T=>"int "+T),m=Ia("sourceLocR",c-1).concat("inIdx.r"),g=Ia("sourceLocG",c-1).concat("inIdx.g"),y=Ia("sourceLocB",c-1).concat("inIdx.b"),x=Ia("sourceLocA",c-1).concat("inIdx.a"),A=a==="max"?"greaterThan":"lessThan",b=n?"":`
|
|
inIdx = round(vec4(getBestIndicesAChannel(${m.join()}),
|
|
getBestIndicesAChannel(${g.join()}),
|
|
getBestIndicesAChannel(${y.join()}),
|
|
getBestIndicesAChannel(${x.join()})));`,w=`vec4(
|
|
getAChannel(${m.join()}),
|
|
hasNextCol ? getAChannel(${g.join()}) : 0.,
|
|
hasNextRow ? getAChannel(${y.join()}) : 0.,
|
|
hasNextRow && hasNextCol ? getAChannel(${x.join()}) : 0.)`,C=n?"":`
|
|
float getBestIndicesAChannel(${f.join()}) {
|
|
return getChannel(getBestIndicesA(${d.join()}),
|
|
vec2(${d.slice(-2).join()}));
|
|
}`;this.userCode=`
|
|
float getAChannel(${f.join()}) {
|
|
return getChannel(getA(${d.join()}),
|
|
vec2(${d.slice(-2).join()}));
|
|
}
|
|
${C}
|
|
void main() {
|
|
${l} coords = getOutputCoords();
|
|
bool hasNextCol = ${u[o-1]} < ${i[o-1]-1};
|
|
bool hasNextRow = ${u[o-2]} < ${i[o-2]-1};
|
|
${p}
|
|
ivec4 srcIdx = ivec4(sourceLocR${h}, sourceLocG${h},
|
|
sourceLocB${h}, sourceLocA${h}) * ${t};
|
|
ivec4 inIdx = srcIdx;
|
|
vec4 bestIndex = vec4(inIdx);
|
|
vec4 bestValue = ${w};
|
|
|
|
for (int i = 0; i < ${t}; i++) {
|
|
inIdx = srcIdx;
|
|
${b}
|
|
vec4 candidate = ${w};
|
|
bvec4 nan = isnan(candidate);
|
|
bvec4 replace = bvec4(
|
|
vec4(${A}(candidate, bestValue)) * (vec4(1.0) - vec4(nan)));
|
|
|
|
bestValue = vec4(replace.x ? candidate.x : bestValue.x,
|
|
replace.y ? candidate.y : bestValue.y,
|
|
replace.z ? candidate.z : bestValue.z,
|
|
replace.w ? candidate.w : bestValue.w);
|
|
bestIndex = mix(bestIndex, vec4(inIdx), vec4(replace));
|
|
srcIdx++;
|
|
}
|
|
setOutput(bestIndex);
|
|
}
|
|
`}};function Yv(e,t,a,n=null){let r=t.shape[0],s=t.shape[1];n!=null&&(r=n.shape[0],s=n.shape[1]);let i=S.computeOptimalWindowSize(s),o={windowSize:i,inSize:s,batchSize:r,outSize:Math.ceil(s/i)},l=new Bj(o,a,n==null),u=[t];n!=null&&u.push(n);let p=e.runWebGLProgram(l,u,"int32");if(p.shape[1]===1)return p;let c=Yv(e,t,a,p);return e.disposeIntermediateTensorInfo(p),c}function Zv(e,t,a,n=null){let r=n!=null?n.shape:t.shape,s=r[r.length-1],i=S.computeOptimalWindowSize(s),o=new Vj(r,i,a,n==null),l=n==null?[t]:[t,n],u=e.runWebGLProgram(o,l,"int32");if(u.shape.length===t.shape.length){let p=Zv(e,t,a,u);return e.disposeIntermediateTensorInfo(u),p}return u}function Jv(e,t,a,n){let r=[a];if(S.assertAxesAreInnerMostDims("arg"+n.charAt(0).toUpperCase()+n.slice(1),r,t.shape.length),!W().getBool("WEBGL_PACK_REDUCE")||t.shape.length<=2){let s=[],i=e.texData.get(t.dataId),o=i!==null&&i.isPacked,l=t;o&&(l=e.unpackTensor(t),s.push(l));let[u,p]=S.computeOutAndReduceShapes(l.shape,r),c=v.sizeFromShape(p),d=ue({inputs:{x:l},backend:e,attrs:{shape:[-1,c]}});s.push(d);let h=Yv(e,d,n);s.push(h);let f=ue({inputs:{x:h},backend:e,attrs:{shape:u}});return s.forEach(m=>e.disposeIntermediateTensorInfo(m)),f}return Zv(e,t,n)}function Uj(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{axis:s}=n,i=v.parseAxisParam(s,r.shape),o=S.getAxesPermutation(i,r.shape.length),l=r,u=[];o!=null&&(l=Ta({inputs:{x:r},backend:a,attrs:{perm:o}}),u.push(l),i=S.getInnerMostAxes(i.length,l.shape.length)),S.assertAxesAreInnerMostDims("argMax",[i[0]],l.shape.length);let p=Jv(a,l,i[0],"max");return u.forEach(c=>a.disposeIntermediateTensorInfo(c)),p}var Gj={kernelName:Jl,backendName:"webgl",kernelFunc:Uj};function Hj(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{axis:s}=n,i=v.parseAxisParam(s,r.shape),o=S.getAxesPermutation(i,r.shape.length),l=r,u=[];o!=null&&(l=Ta({inputs:{x:r},backend:a,attrs:{perm:o}}),u.push(l),i=S.getInnerMostAxes(i.length,l.shape.length)),S.assertAxesAreInnerMostDims("argMin",[i[0]],l.shape.length);let p=Jv(a,l,i[0],"min");return u.forEach(c=>a.disposeIntermediateTensorInfo(c)),p}var jj={kernelName:Ql,backendName:"webgl",kernelFunc:Hj},qj=Rn+`
|
|
if (abs(x) > 1.) {
|
|
return NAN;
|
|
}
|
|
return asin(x);
|
|
`,Xj=et({opSnippet:qj}),Kj={kernelName:ri,backendName:"webgl",kernelFunc:Xj},Yj=Rn+"return log(x + sqrt(x * x + 1.0));",Zj=et({opSnippet:Yj}),Jj={kernelName:si,backendName:"webgl",kernelFunc:Zj},Qj=Rn+`
|
|
return atan(x);
|
|
`,eq=et({opSnippet:Qj}),tq={kernelName:ii,backendName:"webgl",kernelFunc:eq},aq=kg+`
|
|
return atan(a, b);
|
|
`,nq=`
|
|
vec4 result = atan(a, b);
|
|
bvec4 isNaNA = isnan(a);
|
|
bvec4 isNaNB = isnan(b);
|
|
bvec4 isNaN = bvec4(isNaNA.x || isNaNB.x, isNaNA.y || isNaNB.y, isNaNA.z || isNaNB.z, isNaNA.w || isNaNB.w);
|
|
`+Xo+`
|
|
return result;
|
|
`,rq=fa({opSnippet:aq,packedOpSnippet:nq}),sq={kernelName:li,backendName:"webgl",kernelFunc:rq},iq=Rn+`
|
|
if ((x < -1.0) || (x > 1.0)) return NAN;
|
|
return (log(1.0 + x) - log(1.0 - x)) / 2.0;`,oq=et({opSnippet:iq}),lq={kernelName:oi,backendName:"webgl",kernelFunc:oq},jd=class{constructor(e,t,a,n=!1,r=!1){if(this.variableNames=["x"],t==="avg"&&a)throw new Error("Cannot compute positions for average pool.");let s=e.filterWidth,i=e.strideHeight,o=e.strideWidth,l=e.dilationHeight,u=e.dilationWidth,p=e.effectiveFilterHeight,c=e.effectiveFilterWidth,d=e.padInfo.top,h=e.padInfo.left;this.outputShape=e.outShape;let f=t==="avg",m=`((batch * ${e.inHeight} + xR) * ${e.inWidth} + xC) * ${e.inChannels} + d`,g=`(xR * ${e.inWidth} + xC) * ${e.inChannels} + d`,y="0.0";if(f||(y="-1.0 / 1e-20"),a){let T=">=";this.userCode=`
|
|
const ivec2 strides = ivec2(${i}, ${o});
|
|
const ivec2 pads = ivec2(${d}, ${h});
|
|
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int batch = coords[0];
|
|
int d = coords[3];
|
|
|
|
ivec2 xRCCorner = coords.yz * strides - pads;
|
|
int xRCorner = xRCCorner.x;
|
|
int xCCorner = xRCCorner.y;
|
|
|
|
// max/min x(?, ?, d) to get y(yR, yC, d).
|
|
// ? = to be determined
|
|
float minMaxValue = 0.0;
|
|
float minMaxValueFound = 0.0;
|
|
int minMaxPosition = 0;
|
|
float avgValue = 0.0;
|
|
|
|
for (int wR = 0; wR < ${p};
|
|
wR += ${l}) {
|
|
int xR = xRCorner + wR;
|
|
|
|
if (xR < 0 || xR >= ${e.inHeight}) {
|
|
continue;
|
|
}
|
|
|
|
for (int wC = 0; wC < ${c};
|
|
wC += ${u}) {
|
|
int xC = xCCorner + wC;
|
|
|
|
if (xC < 0 || xC >= ${e.inWidth}) {
|
|
continue;
|
|
}
|
|
|
|
float value = getX(batch, xR, xC, d);
|
|
|
|
// If a min / max value has already been found, use it. If not,
|
|
// use the current value.
|
|
float currMinMaxValue = mix(
|
|
value, minMaxValue, minMaxValueFound);
|
|
if (value ${T} currMinMaxValue) {
|
|
minMaxValue = value;
|
|
minMaxValueFound = 1.0;
|
|
minMaxPosition = ${n?r?m:g:`wR * ${c} + wC`};
|
|
}
|
|
}
|
|
}
|
|
setOutput(float(minMaxPosition));
|
|
}
|
|
`;return}let x="max",A=`${t}(${t}(${t}(minMaxValue[0], minMaxValue[1]), minMaxValue[2]), minMaxValue[3])`;t==="avg"&&(A="avgValue / max(count, 1.0)");let b=Math.floor(s/4)*4,w=s%4,C=`
|
|
if (${f}) {
|
|
avgValue += dot(values, ones);
|
|
} else {
|
|
minMaxValue = ${x}(values, minMaxValue);
|
|
}
|
|
`;this.userCode=`
|
|
const ivec2 strides = ivec2(${i}, ${o});
|
|
const ivec2 pads = ivec2(${d}, ${h});
|
|
const float initializationValue = ${y};
|
|
const vec4 ones = vec4(1.0, 1.0, 1.0, 1.0);
|
|
|
|
float count = 0.0;
|
|
|
|
float getValue(int batch, int xR, int xC, int d) {
|
|
if (xC < 0 || xC >= ${e.inWidth}) {
|
|
return initializationValue;
|
|
}
|
|
count += 1.0;
|
|
return getX(batch, xR, xC, d);
|
|
}
|
|
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int batch = coords[0];
|
|
int d = coords[3];
|
|
|
|
ivec2 xRCCorner = coords.yz * strides - pads;
|
|
int xRCorner = xRCCorner.x;
|
|
int xCCorner = xRCCorner.y;
|
|
|
|
// max/min x(?, ?, d) to get y(yR, yC, d).
|
|
// ? = to be determined
|
|
vec4 minMaxValue = vec4(${y});
|
|
float avgValue = 0.0;
|
|
count = 0.0;
|
|
|
|
for (int wR = 0; wR < ${p};
|
|
wR += ${l}) {
|
|
int xR = xRCorner + wR;
|
|
|
|
if (xR < 0 || xR >= ${e.inHeight}) {
|
|
continue;
|
|
}
|
|
|
|
for (int wC = 0; wC < ${b}; wC += 4) {
|
|
int xC = xCCorner + wC * ${u};
|
|
|
|
vec4 values = vec4(
|
|
getValue(batch, xR, xC, d),
|
|
getValue(batch, xR, xC + ${u}, d),
|
|
getValue(batch, xR, xC + 2 * ${u}, d),
|
|
getValue(batch, xR, xC + 3 * ${u}, d)
|
|
);
|
|
|
|
${C}
|
|
}
|
|
|
|
int xC = xCCorner + ${b};
|
|
if (${w===1}) {
|
|
vec4 values = vec4(
|
|
getValue(batch, xR, xC, d),
|
|
initializationValue,
|
|
initializationValue,
|
|
initializationValue
|
|
);
|
|
|
|
${C}
|
|
} else if (${w===2}) {
|
|
vec4 values = vec4(
|
|
getValue(batch, xR, xC, d),
|
|
getValue(batch, xR, xC + ${u}, d),
|
|
initializationValue,
|
|
initializationValue
|
|
);
|
|
|
|
${C}
|
|
} else if (${w===3}) {
|
|
vec4 values = vec4(
|
|
getValue(batch, xR, xC, d),
|
|
getValue(batch, xR, xC + ${u}, d),
|
|
getValue(batch, xR, xC + 2 * ${u}, d),
|
|
initializationValue
|
|
);
|
|
|
|
${C}
|
|
}
|
|
}
|
|
setOutput(${A});
|
|
}
|
|
`}},Sg=class{constructor(e,t,a,n=!1,r=!1){if(this.variableNames=["x"],t==="avg"&&a)throw new Error("Cannot compute positions for average pool.");let s=e.filterWidth,i=e.strideDepth,o=e.strideHeight,l=e.strideWidth,u=e.dilationDepth,p=e.dilationHeight,c=e.dilationWidth,d=e.effectiveFilterDepth,h=e.effectiveFilterHeight,f=e.effectiveFilterWidth,m=e.padInfo.front,g=e.padInfo.top,y=e.padInfo.left;this.outputShape=e.outShape;let x=t==="avg",A="0.0";if(x||(A="-1.0 / 1e-20"),a){let M=">=";this.userCode=`
|
|
const ivec3 strides =
|
|
ivec3(${i}, ${o}, ${l});
|
|
const ivec3 pads = ivec3(${m}, ${g}, ${y});
|
|
|
|
void main() {
|
|
ivec5 coords = getOutputCoords();
|
|
int batch = coords.x;
|
|
int ch = coords.u;
|
|
|
|
ivec3 xCorner = ivec3(coords.y, coords.z, coords.w) * strides - pads;
|
|
int xDCorner = xCorner.x;
|
|
int xRCorner = xCorner.y;
|
|
int xCCorner = xCorner.z;
|
|
|
|
// max/min x(?, ?, ?, ch) to get y(yD, yR, yC, ch).
|
|
// ? = to be determined
|
|
float minMaxValue = 0.0;
|
|
float minMaxValueFound = 0.0;
|
|
int minMaxPosition = 0;
|
|
|
|
for (int wD = 0; wD < ${d};
|
|
wD += ${u}) {
|
|
int xD = xDCorner + wD;
|
|
|
|
if (xD < 0 || xD >= ${e.inDepth}) {
|
|
continue;
|
|
}
|
|
|
|
for (int wR = 0; wR < ${h};
|
|
wR += ${p}) {
|
|
int xR = xRCorner + wR;
|
|
|
|
if (xR < 0 || xR >= ${e.inHeight}) {
|
|
continue;
|
|
}
|
|
|
|
for (int wC = 0; wC < ${f};
|
|
wC += ${c}) {
|
|
int xC = xCCorner + wC;
|
|
|
|
if (xC < 0 || xC >= ${e.inWidth}) {
|
|
continue;
|
|
}
|
|
|
|
float value = getX(batch, xD, xR, xC, ch);
|
|
|
|
// If a min / max value has already been found, use it. If not,
|
|
// use the current value.
|
|
float currMinMaxValue = mix(
|
|
value, minMaxValue, minMaxValueFound);
|
|
if (value ${M} currMinMaxValue) {
|
|
minMaxValue = value;
|
|
minMaxValueFound = 1.0;
|
|
minMaxPosition = ${n?r?`(((batch * ${e.inDepth} + xD) * ${e.inHeight} + xR) * ${e.inWidth} + xC) * ${e.inChannels} + ch`:`((xD * ${e.inHeight} + xR) * ${e.inWidth} + xC) * ${e.inChannels} + ch`:`wD * ${h} * ${f} +
|
|
wR * ${f} + wC`};
|
|
}
|
|
}
|
|
}
|
|
}
|
|
setOutput(float(minMaxPosition));
|
|
}
|
|
`;return}let b="max",w=`${t}(${t}(${t}(minMaxValue[0], minMaxValue[1]), minMaxValue[2]), minMaxValue[3])`;t==="avg"&&(w="avgValue / max(count, 1.0)");let C=Math.floor(s/4)*4,T=s%4,N=`
|
|
if (${x}) {
|
|
avgValue += dot(values, ones);
|
|
} else {
|
|
minMaxValue = ${b}(values, minMaxValue);
|
|
}
|
|
`;this.userCode=`
|
|
const ivec3 strides =
|
|
ivec3(${i}, ${o}, ${l});
|
|
const ivec3 pads = ivec3(${m}, ${g}, ${y});
|
|
const float initializationValue = ${A};
|
|
const vec4 ones = vec4(1.0, 1.0, 1.0, 1.0);
|
|
|
|
float count = 0.0;
|
|
|
|
float getValue(int batch, int xD, int xR, int xC, int ch) {
|
|
if (xC < 0 || xC >= ${e.inWidth}) {
|
|
return initializationValue;
|
|
}
|
|
count += 1.0;
|
|
return getX(batch, xD, xR, xC, ch);
|
|
}
|
|
|
|
void main() {
|
|
ivec5 coords = getOutputCoords();
|
|
int batch = coords.x;
|
|
int ch = coords.u;
|
|
|
|
ivec3 xCorner = ivec3(coords.y, coords.z, coords.w) * strides - pads;
|
|
int xDCorner = xCorner.x;
|
|
int xRCorner = xCorner.y;
|
|
int xCCorner = xCorner.z;
|
|
|
|
// max/min x(?, ?, ?, d) to get y(yD, yR, yC, ch).
|
|
// ? = to be determined
|
|
vec4 minMaxValue = vec4(${A});
|
|
float avgValue = 0.0;
|
|
count = 0.0;
|
|
|
|
for (int wD = 0; wD < ${d};
|
|
wD += ${u}) {
|
|
int xD = xDCorner + wD;
|
|
|
|
if (xD < 0 || xD >= ${e.inDepth}) {
|
|
continue;
|
|
}
|
|
|
|
for (int wR = 0; wR < ${h};
|
|
wR += ${p}) {
|
|
int xR = xRCorner + wR;
|
|
|
|
if (xR < 0 || xR >= ${e.inHeight}) {
|
|
continue;
|
|
}
|
|
|
|
for (int wC = 0; wC < ${C}; wC += 4) {
|
|
int xC = xCCorner + wC * ${c};
|
|
|
|
vec4 values = vec4(
|
|
getValue(batch, xD, xR, xC, ch),
|
|
getValue(batch, xD, xR, xC + ${c}, ch),
|
|
getValue(batch, xD, xR, xC + 2 * ${c}, ch),
|
|
getValue(batch, xD, xR, xC + 3 * ${c}, ch)
|
|
);
|
|
|
|
${N}
|
|
}
|
|
|
|
int xC = xCCorner + ${C};
|
|
if (${T===1}) {
|
|
vec4 values = vec4(
|
|
getValue(batch, xD, xR, xC, ch),
|
|
initializationValue,
|
|
initializationValue,
|
|
initializationValue
|
|
);
|
|
|
|
${N}
|
|
} else if (${T===2}) {
|
|
vec4 values = vec4(
|
|
getValue(batch, xD, xR, xC, ch),
|
|
getValue(batch, xD, xR, xC + ${c}, ch),
|
|
initializationValue,
|
|
initializationValue
|
|
);
|
|
|
|
${N}
|
|
} else if (${T===3}) {
|
|
vec4 values = vec4(
|
|
getValue(batch, xD, xR, xC, ch),
|
|
getValue(batch, xD, xR, xC + ${c}, ch),
|
|
getValue(batch, xD, xR, xC + 2 * ${c}, ch),
|
|
initializationValue
|
|
);
|
|
|
|
${N}
|
|
}
|
|
}
|
|
}
|
|
setOutput(${w});
|
|
}
|
|
`}};function uq(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t;Pu(r,"avgPool");let{filterSize:s,strides:i,pad:o,dimRoundingMode:l}=n,u=1;v.assert(S.eitherStridesOrDilationsAreOne(i,u),()=>`Error in avgPool: Either strides or dilations must be 1. Got strides ${i} and dilations '${u}'`);let p=S.computePool2DInfo(r.shape,s,i,u,o,l);if(p.filterWidth===1&&p.filterHeight===1&&v.arraysEqual(p.inShape,p.outShape))return en({inputs:{x:r},backend:a});let c=new jd(p,"avg",!1);return a.runWebGLProgram(c,[r],"float32")}var dq={kernelName:ui,backendName:"webgl",kernelFunc:uq};function pq(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{filterSize:s,strides:i,pad:o,dimRoundingMode:l,dataFormat:u}=n,p=[1,1,1],c=S.computePool3DInfo(r.shape,s,i,p,o,l,u),d=new Sg(c,"avg",!1);return a.runWebGLProgram(d,[r],"float32")}var cq={kernelName:eu,backendName:"webgl",kernelFunc:pq},hq=class{constructor(e){this.variableNames=["dy"],this.outputShape=e.inShape;let t=e.filterHeight,a=e.filterWidth,n=e.strideHeight,r=e.strideWidth,s=e.dilationHeight,i=e.dilationWidth,o=e.effectiveFilterHeight,l=e.effectiveFilterWidth,u=o-1-e.padInfo.top,p=l-1-e.padInfo.left,c=1/(t*a);this.userCode=`
|
|
const ivec2 pads = ivec2(${u}, ${p});
|
|
const float avgMultiplier = float(${c});
|
|
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int b = coords[0];
|
|
int d = coords[3];
|
|
|
|
ivec2 dyRCCorner = coords.yz - pads;
|
|
int dyRCorner = dyRCCorner.x;
|
|
int dyCCorner = dyRCCorner.y;
|
|
|
|
// Convolve dy(?, ?, d) with pos mask(:, :, d) to get dx(xR, xC, d).
|
|
// ? = to be determined. : = across all values in that axis.
|
|
float dotProd = 0.0;
|
|
for (int wR = 0; wR < ${o};
|
|
wR += ${s}) {
|
|
float dyR = float(dyRCorner + wR) / ${n}.0;
|
|
|
|
if (dyR < 0.0 || dyR >= ${e.outHeight}.0 || fract(dyR) > 0.0) {
|
|
continue;
|
|
}
|
|
int idyR = int(dyR);
|
|
|
|
for (int wC = 0; wC < ${l};
|
|
wC+= ${i}) {
|
|
float dyC = float(dyCCorner + wC) / ${r}.0;
|
|
|
|
if (dyC < 0.0 || dyC >= ${e.outWidth}.0 ||
|
|
fract(dyC) > 0.0) {
|
|
continue;
|
|
}
|
|
int idyC = int(dyC);
|
|
|
|
float dyValue = getDy(b, idyR, idyC, d);
|
|
|
|
dotProd += dyValue * avgMultiplier;
|
|
}
|
|
}
|
|
setOutput(dotProd);
|
|
}
|
|
`}},fq=class{constructor(e){this.variableNames=["dy"],this.outputShape=e.inShape;let t=e.filterDepth,a=e.filterHeight,n=e.filterWidth,r=e.strideDepth,s=e.strideHeight,i=e.strideWidth,o=e.dilationDepth,l=e.dilationHeight,u=e.dilationWidth,p=e.effectiveFilterDepth,c=e.effectiveFilterHeight,d=e.effectiveFilterWidth,h=p-1-e.padInfo.front,f=c-1-e.padInfo.top,m=d-1-e.padInfo.left,g=1/(t*a*n);this.userCode=`
|
|
const ivec3 pads = ivec3(${h}, ${f}, ${m});
|
|
const float avgMultiplier = float(${g});
|
|
|
|
void main() {
|
|
ivec5 coords = getOutputCoords();
|
|
int batch = coords.x;
|
|
int ch = coords.u;
|
|
|
|
ivec3 dyCorner = ivec3(coords.y, coords.z, coords.w) - pads;
|
|
int dyDCorner = dyCorner.x;
|
|
int dyRCorner = dyCorner.y;
|
|
int dyCCorner = dyCorner.z;
|
|
|
|
// Convolve dy(?, ?, ?, d) with pos mask(:, :, :, ch) to get
|
|
// dx(xD, xR, xC, ch).
|
|
// ? = to be determined. : = across all values in that axis.
|
|
float dotProd = 0.0;
|
|
|
|
for (int wD = 0; wD < ${p};
|
|
wD += ${o}) {
|
|
float dyD = float(dyDCorner + wD) / ${r}.0;
|
|
|
|
if (dyD < 0.0 || dyD >= ${e.outDepth}.0 || fract(dyD) > 0.0) {
|
|
continue;
|
|
}
|
|
int idyD = int(dyD);
|
|
|
|
for (int wR = 0; wR < ${c};
|
|
wR += ${l}) {
|
|
float dyR = float(dyRCorner + wR) / ${s}.0;
|
|
|
|
if (dyR < 0.0 || dyR >= ${e.outHeight}.0 ||
|
|
fract(dyR) > 0.0) {
|
|
continue;
|
|
}
|
|
int idyR = int(dyR);
|
|
|
|
for (int wC = 0; wC < ${d};
|
|
wC += ${u}) {
|
|
float dyC = float(dyCCorner + wC) / ${i}.0;
|
|
|
|
if (dyC < 0.0 || dyC >= ${e.outWidth}.0 ||
|
|
fract(dyC) > 0.0) {
|
|
continue;
|
|
}
|
|
int idyC = int(dyC);
|
|
|
|
float dyValue = getDy(batch, idyD, idyR, idyC, ch);
|
|
|
|
dotProd += dyValue * avgMultiplier;
|
|
}
|
|
}
|
|
}
|
|
setOutput(dotProd);
|
|
}
|
|
`}};function mq(e){let{inputs:t,backend:a,attrs:n}=e,{dy:r,input:s}=t,i=s,{filterSize:o,strides:l,pad:u,dimRoundingMode:p}=n,c=[1,1,1],d=S.computePool3DInfo(i.shape,o,l,c,u,p),h=new fq(d);return a.runWebGLProgram(h,[r],i.dtype)}var gq={kernelName:ep,backendName:"webgl",kernelFunc:mq};function yq(e){let{inputs:t,backend:a,attrs:n}=e,{dy:r,input:s}=t,i=s;Pu([r,s],"avgPoolGrad");let{filterSize:o,strides:l,pad:u}=n,p=S.computePool2DInfo(i.shape,o,l,1,u),c=new hq(p);return a.runWebGLProgram(c,[r],i.dtype)}var xq={kernelName:kh,backendName:"webgl",kernelFunc:yq};function Aq(e){let{inputs:t,backend:a,attrs:n}=e,{a:r,b:s}=t,{transposeA:i,transposeB:o}=n;return fh({a:r,b:s,transposeA:i,transposeB:o,backend:a})}var bq={kernelName:di,backendName:"webgl",kernelFunc:Aq},vq=class{constructor(e,t,a,n,r,s){this.outputShape=[],this.variableNames=["x","mean","variance"],S.assertAndGetBroadcastShape(e,t),S.assertAndGetBroadcastShape(e,a);let i="0.0";n!=null&&(S.assertAndGetBroadcastShape(e,n),this.variableNames.push("offset"),i="getOffsetAtOutCoords()");let o="1.0";r!=null&&(S.assertAndGetBroadcastShape(e,r),this.variableNames.push("scale"),o="getScaleAtOutCoords()"),this.outputShape=e,this.userCode=`
|
|
void main() {
|
|
float x = getXAtOutCoords();
|
|
float mean = getMeanAtOutCoords();
|
|
float variance = getVarianceAtOutCoords();
|
|
float offset = ${i};
|
|
float scale = ${o};
|
|
float inv = scale * inversesqrt(variance + float(${s}));
|
|
setOutput(dot(vec3(x, -mean, offset), vec3(inv, inv, 1)));
|
|
}
|
|
`}},wq=class{constructor(e,t,a,n,r,s){this.packedInputs=!0,this.packedOutput=!0,this.variableNames=["x","mean","variance"],S.assertAndGetBroadcastShape(e,t),S.assertAndGetBroadcastShape(e,a);let i="vec4(0.0)";n!=null&&(S.assertAndGetBroadcastShape(e,n),this.variableNames.push("offset"),i="getOffsetAtOutCoords()");let o="vec4(1.0)";r!=null&&(S.assertAndGetBroadcastShape(e,r),this.variableNames.push("scale"),o="getScaleAtOutCoords()"),this.outputShape=e,this.userCode=`
|
|
void main() {
|
|
vec4 offset = ${i};
|
|
vec4 scale = ${o};
|
|
|
|
vec4 x = getXAtOutCoords();
|
|
vec4 mean = getMeanAtOutCoords();
|
|
vec4 variance = getVarianceAtOutCoords();
|
|
|
|
vec4 inv = scale * inversesqrt(variance + vec4(${s}));
|
|
|
|
setOutput((x - mean) * inv + offset);
|
|
}
|
|
`}},kq=({inputs:e,backend:t,attrs:a})=>{let{x:n,mean:r,variance:s,offset:i,scale:o}=e;v.assert(r.shape.length===s.shape.length,()=>"Batch normalization gradient requires mean and variance to have equal ranks."),v.assert(i==null||r.shape.length===i.shape.length,()=>"Batch normalization gradient requires mean and offset to have equal ranks."),v.assert(o==null||r.shape.length===o.shape.length,()=>"Batch normalization gradient requires mean and scale to have equal ranks.");let{varianceEpsilon:l}=a;l==null&&(l=.001);let u=[n,r,s],p=null;i!=null&&(p=i.shape,u.push(i));let c=null;o!=null&&(c=o.shape,u.push(o));let d=W().getBool("WEBGL_PACK_NORMALIZATION")?new wq(n.shape,r.shape,s.shape,p,c,l):new vq(n.shape,r.shape,s.shape,p,c,l);return t.runWebGLProgram(d,u,u[0].dtype)},Iq={kernelName:$i,backendName:"webgl",kernelFunc:kq},Sq=class{constructor(e){this.variableNames=["source"],this.outputShape=e,this.rank=e.length;let t=ft(this.rank);this.customUniforms=[{name:"start",arrayIndex:this.rank,type:"int"}];let a=Cq(this.rank),n,r=e.map((s,i)=>`sourceLoc.${S1[i]} = start[${i}] + coords.${S1[i]};`);n=`
|
|
${t} sourceLoc;
|
|
${t} coords = getOutputCoords();
|
|
${r.join(`
|
|
`)}
|
|
`,this.userCode=`
|
|
void main() {
|
|
${n}
|
|
setOutput(getSource(${a}));
|
|
}
|
|
`}},S1=["x","y","z","w","u","v"];function Cq(e){if(e===1)return"sourceLoc";if(e<=6)return S1.slice(0,e).map(t=>"sourceLoc."+t).join(",");throw Error(`Slicing for rank ${e} is not yet supported`)}var Tq=class{constructor(e){this.variableNames=["source"],this.packedInputs=!0,this.packedOutput=!0,this.outputShape=e,this.rank=e.length,this.customUniforms=[{name:"start",arrayIndex:this.rank,type:"int"}];let t=ft(this.rank),a=Ia("coords",this.rank),n=Ia("sourceLoc",this.rank),r=this.rank===1?"sourceLoc":`vec2(${n.slice(-2).join()})`,s=`getChannel(getSource(${n.join()}), ${r})`,i=`
|
|
result.x = ${s};
|
|
if (++${a[this.rank-1]} < ${e[this.rank-1]}) {
|
|
++${n[this.rank-1]};
|
|
result.y = ${s};
|
|
--${n[this.rank-1]};
|
|
}
|
|
`,o=this.rank===1?"":`
|
|
--${a[this.rank-1]};
|
|
if (++${a[this.rank-2]} < ${e[this.rank-2]}) {
|
|
++${n[this.rank-2]};
|
|
result.z = ${s};
|
|
if (++${a[this.rank-1]} < ${e[this.rank-1]}) {
|
|
++${n[this.rank-1]};
|
|
result.w = ${s};
|
|
}
|
|
}
|
|
`,l=this.rank<=4?`sourceLoc = coords +
|
|
${t}(${e.map((u,p)=>`start[${p}]`).join()});`:e.map((u,p)=>`${n[p]} = ${a[p]} + start[${p}];`).join(`
|
|
`);this.userCode=`
|
|
void main() {
|
|
${t} coords = getOutputCoords();
|
|
${t} sourceLoc;
|
|
${l}
|
|
vec4 result = vec4(0.);
|
|
${i}
|
|
${o}
|
|
setOutput(result);
|
|
}
|
|
`}};function Nq(e,t,a,n){let r=n.texData.get(e.dataId),s=n.makeTensorInfo(a,e.dtype),i=n.texData.get(s.dataId);Object.assign(i,r),i.refCount=1,i.shape=a,i.dtype=e.dtype;let o=Nt.computeFlatOffset(t,v.computeStrides(e.shape));r.slice&&(o+=r.slice.flatOffset),i.slice={flatOffset:o,origDataId:r.slice&&r.slice.origDataId||e.dataId};let l=n.dataRefCount.get(i.slice.origDataId)||1;return n.dataRefCount.set(i.slice.origDataId,l+1),s}function Uu(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{begin:s,size:i}=n,[o,l]=Nt.parseSliceParams(r,s,i);if(Nt.assertParamsValid(r,o,l),v.sizeFromShape(l)===0)return a.makeTensorInfo(l,r.dtype,[]);if(a.shouldExecuteOnCPU([r])||r.dtype==="string"){let c=a.texData.get(r.dataId),d=xH(c.values,o,l,r.shape,r.dtype);return a.makeTensorInfo(l,r.dtype,d)}let{isPacked:u}=a.texData.get(r.dataId),p=Nt.isSliceContinous(r.shape,o,l);if(u||!p){let c=W().getBool("WEBGL_PACK_ARRAY_OPERATIONS")?new Tq(l):new Sq(l),d=[o];return a.runWebGLProgram(c,[r],r.dtype,d)}return a.uploadToGPU(r.dataId),Nq(r,o,l,a)}var Rq={kernelName:Iu,backendName:"webgl",kernelFunc:Uu},Eq=e=>{let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{blockShape:s,crops:i}=n;v.assert(r.shape.length<=4,()=>"batchToSpaceND for rank > 4 with a WebGL backend not implemented yet");let o=s.reduce((x,A)=>x*A),l=S.getReshaped(r.shape,s,o),u=S.getPermuted(l.length,s.length),p=S.getReshapedPermuted(r.shape,s,o),c=S.getSliceBeginCoords(i,s.length),d=S.getSliceSize(p,i,s.length),h=[],f=ue({inputs:{x:r},backend:a,attrs:{shape:l}}),m=Ta({inputs:{x:f},backend:a,attrs:{perm:u}}),g=ue({inputs:{x:m},backend:a,attrs:{shape:p}}),y=Uu({inputs:{x:g},backend:a,attrs:{begin:c,size:d}});return h.push(f),h.push(m),h.push(g),h.forEach(x=>a.disposeIntermediateTensorInfo(x)),y},Mq={kernelName:tu,backendName:"webgl",kernelFunc:Eq};function Fq(e){let{inputs:t,backend:a,attrs:n}=e,{x:r,weights:s}=t,{size:i}=n,o=a.readSync(r.dataId),l=a.readSync(s.dataId),u=zv(o,l,s.dtype,s.shape,i);return a.makeTensorInfo([i],s.dtype,u)}var _q={kernelName:pi,backendName:"webgl",kernelFunc:Fq};function $q(e){let{inputs:t,backend:a}=e,{s0:n,s1:r}=t,s=a.readSync(n.dataId),i=a.readSync(r.dataId),o=S.assertAndGetBroadcastShape(Array.from(s),Array.from(i));return a.makeTensorInfo([o.length],"int32",Int32Array.from(o))}var Dq={kernelName:au,backendName:"webgl",kernelFunc:$q},Pq="return float(a != b);",Qv=fa({opSnippet:Pq,cpuKernelImpl:uH,dtype:"bool"}),Oq={kernelName:oo,backendName:"webgl",kernelFunc:Qv};function Gp(e){let{inputs:t,backend:a}=e,{input:n}=t,r=a.texData.get(n.dataId);return en({inputs:{x:r.complexTensorInfos.real},backend:a})}var zq={kernelName:cp,backendName:"webgl",kernelFunc:Gp},Lq="return float(int(x));";function Wq(e,t){let a=new Yn(e.shape,Lq),n=t.runWebGLProgram(a,[e],"int32");return{dataId:n.dataId,shape:n.shape,dtype:n.dtype}}function C1(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{dtype:s}=n;if(s==="complex64"){if(r.dtype==="complex64")return en({inputs:{x:r},backend:a});let i=gn(r.shape),o=C1({inputs:{x:r},backend:a,attrs:{dtype:"float32"}}),l=ds({inputs:{real:o,imag:i},backend:a});return i.dispose(),a.disposeIntermediateTensorInfo(o),l}if(r.dtype==="complex64"){let i=Gp({inputs:{input:r},backend:a}),o=C1({inputs:{x:i},backend:a,attrs:{dtype:s}});return a.disposeIntermediateTensorInfo(i),o}if(!v.hasEncodingLoss(r.dtype,s)){let i=en({inputs:{x:r},backend:a});return{dataId:i.dataId,shape:i.shape,dtype:s}}if(a.shouldExecuteOnCPU([r])){let i=a.texData.get(r.dataId).values,[o,l,u]=UG(i,r.shape,r.dtype,s);return a.makeTensorInfo(o,l,u)}if(s==="int32")return Wq(r,a);if(s==="bool"){let i=a.makeTensorInfo([],"bool",v.getTypedArrayFromDType("bool",1)),o=Qv({inputs:{a:r,b:i},backend:a});return a.disposeIntermediateTensorInfo(i),o}throw new Error(`Error in Cast: failed to cast ${r.dtype} to ${s}`)}var Bq={kernelName:ci,backendName:"webgl",kernelFunc:C1},mx="return ceil(x);",Vq=et({opSnippet:mx,packedOpSnippet:mx,cpuKernelImpl:GG}),Uq={kernelName:hi,backendName:"webgl",kernelFunc:Vq},Gq=class{constructor(e){this.variableNames=["A"],this.customUniforms=[{name:"minVal",type:"float"},{name:"maxVal",type:"float"}],this.outputShape=e,this.userCode=`
|
|
|
|
void main() {
|
|
float value = getAAtOutCoords();
|
|
if (isnan(value)) {
|
|
setOutput(value);
|
|
return;
|
|
}
|
|
|
|
setOutput(clamp(value, minVal, maxVal));
|
|
}
|
|
`}},Hq=class{constructor(e){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!0,this.customUniforms=[{name:"minVal",type:"float"},{name:"maxVal",type:"float"}],this.outputShape=e,this.userCode=`
|
|
void main() {
|
|
vec4 value = getAAtOutCoords();
|
|
|
|
if (any(isnan(value))) {
|
|
setOutput(value);
|
|
return;
|
|
}
|
|
|
|
setOutput(clamp(value, vec4(minVal), vec4(maxVal)));
|
|
}
|
|
`}};function jq(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{clipValueMin:s,clipValueMax:i}=n,o;W().getBool("WEBGL_PACK_CLIP")?o=new Hq(r.shape):o=new Gq(r.shape);let l=[[s],[i]];return a.runWebGLProgram(o,[r],r.dtype,l)}var qq={kernelName:rs,backendName:"webgl",kernelFunc:jq},Xq=class{constructor(e){this.variableNames=["real","imag"],this.outputShape=e,this.userCode=`
|
|
void main() {
|
|
float re = abs(getRealAtOutCoords());
|
|
float im = abs(getImagAtOutCoords());
|
|
float mx = max(re, im);
|
|
|
|
// sadly the length function in glsl is not underflow-safe
|
|
// (at least not on Intel GPUs). So the safe solution is
|
|
// to ensure underflow-safety in all cases.
|
|
setOutput(
|
|
mx == 0.0 ? 0.0 : mx * length(vec2(1, min(re, im)/mx))
|
|
);
|
|
}
|
|
`}};function gx(e,t){return{dataId:t.dataId,dtype:t.dtype,shape:e.shape}}function Kq(e){let{inputs:t,backend:a}=e,{x:n}=t,r=a.texData.get(n.dataId),s=new Xq(n.shape),i=[gx(n,r.complexTensorInfos.real),gx(n,r.complexTensorInfos.imag)];return a.runWebGLProgram(s,i,i[0].dtype)}var Yq={kernelName:ap,backendName:"webgl",kernelFunc:Kq},Zq=class{constructor(e){this.outputShape=[],this.outputShape=S.computeOutShape(e,1),this.variableNames=e.map((s,i)=>`T${i}`);let t=new Array(e.length-1);t[0]=e[0][1];for(let s=1;s<t.length;s++)t[s]=t[s-1]+e[s][1];let a=[`if (yC < ${t[0]}) setOutput(getT0(yR, yC));`];for(let s=1;s<t.length;s++){let i=t[s-1];a.push(`else if (yC < ${t[s]}) setOutput(getT${s}(yR, yC-${i}));`)}let n=t.length,r=t[t.length-1];a.push(`else setOutput(getT${n}(yR, yC-${r}));`),this.userCode=`
|
|
void main() {
|
|
ivec2 coords = getOutputCoords();
|
|
int yR = coords.x;
|
|
int yC = coords.y;
|
|
|
|
${a.join(`
|
|
`)}
|
|
}
|
|
`}},Jq=class{constructor(e,t){this.packedInputs=!0,this.packedOutput=!0,this.outputShape=[],this.outputShape=S.computeOutShape(e,t);let a=this.outputShape,n=a.length,r=ft(n),s=Ia("coords",n),i=["x","y","z","w","u","v"].slice(0,n);this.variableNames=e.map((f,m)=>`T${m}`);let o=new Array(e.length-1);o[0]=e[0][t];for(let f=1;f<o.length;f++)o[f]=o[f-1]+e[f][t];let l=i[t],u=i.slice(-2),p=i.join(),c=`if (${l} < ${o[0]}) {
|
|
return getChannel(
|
|
getT0(${p}), vec2(${u.join()}));
|
|
}`;for(let f=1;f<o.length;f++){let m=o[f-1];c+=`
|
|
if (${l} < ${o[f]} && ${l} >= ${o[f-1]}) {
|
|
return getChannel(
|
|
getT${f}(${Gc(i,l,m)}),
|
|
vec2(${Gc(u,l,m)}));
|
|
}`}let d=o.length,h=o[o.length-1];c+=`
|
|
return getChannel(
|
|
getT${d}(${Gc(i,l,h)}),
|
|
vec2(${Gc(u,l,h)}));`,this.userCode=`
|
|
float getValue(${i.map(f=>"int "+f)}) {
|
|
${c}
|
|
}
|
|
|
|
void main() {
|
|
${r} coords = getOutputCoords();
|
|
vec4 result = vec4(getValue(${s}), 0., 0., 0.);
|
|
|
|
${s[n-1]} = ${s[n-1]} + 1;
|
|
if (${s[n-1]} < ${a[n-1]}) {
|
|
result.g = getValue(${s});
|
|
}
|
|
|
|
${s[n-2]} = ${s[n-2]} + 1;
|
|
if (${s[n-2]} < ${a[n-2]}) {
|
|
result.a = getValue(${s});
|
|
}
|
|
|
|
${s[n-1]} = ${s[n-1]} - 1;
|
|
if (${s[n-2]} < ${a[n-2]} &&
|
|
${s[n-1]} < ${a[n-1]}) {
|
|
result.b = getValue(${s});
|
|
}
|
|
setOutput(result);
|
|
}
|
|
`}};function Gc(e,t,a){let n=e.indexOf(t);return e.map((r,s)=>s===n?`${r} - ${a}`:r).join()}function t0(e){let{inputs:t,backend:a}=e,{input:n}=t,r=a.texData.get(n.dataId);return en({inputs:{x:r.complexTensorInfos.imag},backend:a})}var Qq={kernelName:up,backendName:"webgl",kernelFunc:t0};function wd(e,t,a){let n=e[0].dtype;if(n==="complex64"){let h=e.map(x=>Gp({inputs:{input:x},backend:a})),f=e.map(x=>t0({inputs:{input:x},backend:a})),m=wd(h,t,a),g=wd(f,t,a),y=ds({inputs:{real:m,imag:g},backend:a});return h.forEach(x=>a.disposeIntermediateTensorInfo(x)),f.forEach(x=>a.disposeIntermediateTensorInfo(x)),a.disposeIntermediateTensorInfo(m),a.disposeIntermediateTensorInfo(g),y}let r=a.shouldExecuteOnCPU(e);if(n==="string"&&(r=!0),r){let h=e.map(b=>{let w=[-1,v.sizeFromShape(b.shape.slice(t))];return ue({inputs:{x:b},backend:a,attrs:{shape:w}})}),f=h.map(b=>({vals:a.readSync(b.dataId),shape:b.shape})),m=S.computeOutShape(h.map(b=>b.shape),1),g=h[0].shape[0]===1,y=HG(f,m,n,g),x=S.computeOutShape(e.map(b=>b.shape),t),A=a.makeTensorInfo(x,n,y);return h.forEach(b=>a.disposeIntermediateTensorInfo(b)),A}let s=e.filter(h=>v.sizeFromShape(h.shape)>0),i=W().getBool("WEBGL_PACK_ARRAY_OPERATIONS")&&s[0].shape.length>1;if(s.length===1){let h=i?new Yn(e[0].shape,Lr):new Gr(e[0].shape,Lr);return a.runWebGLProgram(h,e,n)}let o=W().getNumber("WEBGL_MAX_TEXTURES_IN_SHADER");if(s.length>o){let h=[];for(let m=0;m<s.length;m+=o){let g=s.slice(m,m+o);h.push(wd(g,t,a))}let f=wd(h,t,a);for(let m of h)a.disposeIntermediateTensorInfo(m);return f}if(i){let h=new Jq(s.map(f=>f.shape),t);return a.runWebGLProgram(h,s,n)}let{tensors2D:l,outShape:u}=eX(s,t,a),p=new Zq(l.map(h=>h.shape)),c=a.runWebGLProgram(p,l,n);l.forEach(h=>a.disposeIntermediateTensorInfo(h));let d=ue({inputs:{x:c},attrs:{shape:u},backend:a});return a.disposeIntermediateTensorInfo(c),d}function eX(e,t,a){let n=S.computeOutShape(e.map(r=>r.shape),t);return{tensors2D:e.map(r=>ue({inputs:{x:r},attrs:{shape:[-1,v.sizeFromShape(r.shape.slice(t))]},backend:a})),outShape:n}}function e8(e){let{inputs:t,backend:a,attrs:n}=e,{axis:r}=n,s=v.parseAxisParam(r,t[0].shape)[0],i=t.map(u=>u.shape);S.assertParamsConsistent(i,s);let o=S.computeOutShape(t.map(u=>u.shape),s);if(v.sizeFromShape(o)===0)return a.makeTensorInfo(o,t[0].dtype,[]);let l=t.filter(u=>v.sizeFromShape(u.shape)>0);return l.length===1?en({inputs:{x:l[0]},backend:a}):wd(l,s,a)}var tX={kernelName:nu,backendName:"webgl",kernelFunc:e8},t8=class{constructor(e,t=!1,a=null,n=!1,r=!1){this.variableNames=["x","W"],this.outputShape=e.outShape;let s=e.padInfo.top,i=e.padInfo.left,o=e.strideHeight,l=e.strideWidth,u=e.dilationHeight,p=e.dilationWidth,c=e.filterHeight,d=e.filterWidth,h=Math.floor(e.inChannels/4)*4,f=e.inChannels%4,m=e.dataFormat==="channelsLast",g=m?1:2,y=m?2:3,x=m?3:1,A="",b="";a&&(n?A=`float activation(float a) {
|
|
float b = getPreluActivationWeightsAtOutCoords();
|
|
${a}
|
|
}`:r?A=`float activation(float a) {
|
|
float b = getLeakyreluAlphaAtOutCoords();
|
|
${a}
|
|
}`:A=`
|
|
float activation(float x) {
|
|
${a}
|
|
}
|
|
`,b="result = activation(result);");let w=t?"result += getBiasAtOutCoords();":"";t&&this.variableNames.push("bias"),n&&this.variableNames.push("preluActivationWeights"),r&&this.variableNames.push("leakyreluAlpha"),this.userCode=`
|
|
${A}
|
|
|
|
const ivec2 strides = ivec2(${o}, ${l});
|
|
const ivec2 pads = ivec2(${s}, ${i});
|
|
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int batch = coords[0];
|
|
int d2 = coords[${x}];
|
|
|
|
ivec2 xRCCorner =
|
|
ivec2(coords[${g}], coords[${y}]) * strides - pads;
|
|
int xRCorner = xRCCorner.x;
|
|
int xCCorner = xRCCorner.y;
|
|
|
|
// Convolve x(?, ?, d1) with w(:, :, d1, d2) to get y(yR, yC, d2).
|
|
// ? = to be determined. : = across all values in that axis.
|
|
float dotProd = 0.0;
|
|
for (int wR = 0; wR < ${c}; wR++) {
|
|
int xR = xRCorner + wR * ${u};
|
|
|
|
if (xR < 0 || xR >= ${e.inHeight}) {
|
|
continue;
|
|
}
|
|
|
|
for (int wC = 0; wC < ${d}; wC++) {
|
|
int xC = xCCorner + wC * ${p};
|
|
|
|
if (xC < 0 || xC >= ${e.inWidth}) {
|
|
continue;
|
|
}
|
|
|
|
for (int d1 = 0; d1 < ${h}; d1 += 4) {
|
|
vec4 wValues = vec4(
|
|
getW(wR, wC, d1, d2),
|
|
getW(wR, wC, d1 + 1, d2),
|
|
getW(wR, wC, d1 + 2, d2),
|
|
getW(wR, wC, d1 + 3, d2)
|
|
);
|
|
|
|
if (${m}) {
|
|
vec4 xValues = vec4(
|
|
getX(batch, xR, xC, d1),
|
|
getX(batch, xR, xC, d1 + 1),
|
|
getX(batch, xR, xC, d1 + 2),
|
|
getX(batch, xR, xC, d1 + 3)
|
|
);
|
|
dotProd += dot(xValues, wValues);
|
|
} else {
|
|
vec4 xValues = vec4(
|
|
getX(batch, d1, xR, xC),
|
|
getX(batch, d1 + 1, xR, xC),
|
|
getX(batch, d1 + 2, xR, xC),
|
|
getX(batch, d1 + 3, xR, xC)
|
|
);
|
|
dotProd += dot(xValues, wValues);
|
|
}
|
|
}
|
|
|
|
if (${f===1}) {
|
|
|
|
if (${m}) {
|
|
dotProd +=
|
|
getX(batch, xR, xC, ${h}) *
|
|
getW(wR, wC, ${h}, d2);
|
|
} else {
|
|
dotProd +=
|
|
getX(batch, ${h}, xR, xC) *
|
|
getW(wR, wC, ${h}, d2);
|
|
}
|
|
|
|
} else if (${f===2}) {
|
|
vec2 wValues = vec2(
|
|
getW(wR, wC, ${h}, d2),
|
|
getW(wR, wC, ${h} + 1, d2)
|
|
);
|
|
|
|
if (${m}) {
|
|
vec2 xValues = vec2(
|
|
getX(batch, xR, xC, ${h}),
|
|
getX(batch, xR, xC, ${h} + 1)
|
|
);
|
|
dotProd += dot(xValues, wValues);
|
|
} else {
|
|
vec2 xValues = vec2(
|
|
getX(batch, ${h}, xR, xC),
|
|
getX(batch, ${h} + 1, xR, xC)
|
|
);
|
|
dotProd += dot(xValues, wValues);
|
|
}
|
|
|
|
} else if (${f===3}) {
|
|
vec3 wValues = vec3(
|
|
getW(wR, wC, ${h}, d2),
|
|
getW(wR, wC, ${h} + 1, d2),
|
|
getW(wR, wC, ${h} + 2, d2)
|
|
);
|
|
|
|
if (${m}) {
|
|
vec3 xValues = vec3(
|
|
getX(batch, xR, xC, ${h}),
|
|
getX(batch, xR, xC, ${h} + 1),
|
|
getX(batch, xR, xC, ${h} + 2)
|
|
);
|
|
dotProd += dot(xValues, wValues);
|
|
} else {
|
|
vec3 xValues = vec3(
|
|
getX(batch, ${h}, xR, xC),
|
|
getX(batch, ${h} + 1, xR, xC),
|
|
getX(batch, ${h} + 2, xR, xC)
|
|
);
|
|
dotProd += dot(xValues, wValues);
|
|
}
|
|
|
|
}
|
|
}
|
|
}
|
|
|
|
float result = dotProd;
|
|
${w}
|
|
${b}
|
|
setOutput(result);
|
|
}
|
|
`}},aX=class{constructor(e){this.variableNames=["x","W"],this.outputShape=e.outShape;let t=e.padInfo.front,a=e.padInfo.top,n=e.padInfo.left,r=e.strideDepth,s=e.strideHeight,i=e.strideWidth,o=e.dilationDepth,l=e.dilationHeight,u=e.dilationWidth,p=e.filterDepth,c=e.filterHeight,d=e.filterWidth,h=Math.floor(e.inChannels/4)*4,f=e.inChannels%4;this.userCode=`
|
|
const ivec3 strides = ivec3(${r}, ${s}, ${i});
|
|
const ivec3 pads = ivec3(${t}, ${a}, ${n});
|
|
|
|
void main() {
|
|
ivec5 coords = getOutputCoords();
|
|
int batch = coords.x;
|
|
int d2 = coords.u;
|
|
|
|
ivec3 xFRCCorner = ivec3(coords.y, coords.z, coords.w) * strides - pads;
|
|
int xFCorner = xFRCCorner.x;
|
|
int xRCorner = xFRCCorner.y;
|
|
int xCCorner = xFRCCorner.z;
|
|
|
|
// Convolve x(?, ?, ?, d1) with w(:, :, :, d1, d2) to get
|
|
// y(yF, yR, yC, d2). ? = to be determined. : = across all
|
|
// values in that axis.
|
|
float dotProd = 0.0;
|
|
for (int wF = 0; wF < ${p}; wF++) {
|
|
int xF = xFCorner + wF * ${o};
|
|
|
|
if (xF < 0 || xF >= ${e.inDepth}) {
|
|
continue;
|
|
}
|
|
|
|
for (int wR = 0; wR < ${c}; wR++) {
|
|
int xR = xRCorner + wR * ${l};
|
|
|
|
if (xR < 0 || xR >= ${e.inHeight}) {
|
|
continue;
|
|
}
|
|
|
|
for (int wC = 0; wC < ${d}; wC++) {
|
|
int xC = xCCorner + wC * ${u};
|
|
|
|
if (xC < 0 || xC >= ${e.inWidth}) {
|
|
continue;
|
|
}
|
|
|
|
for (int d1 = 0; d1 < ${h}; d1 += 4) {
|
|
vec4 xValues = vec4(
|
|
getX(batch, xF, xR, xC, d1),
|
|
getX(batch, xF, xR, xC, d1 + 1),
|
|
getX(batch, xF, xR, xC, d1 + 2),
|
|
getX(batch, xF, xR, xC, d1 + 3)
|
|
);
|
|
vec4 wValues = vec4(
|
|
getW(wF, wR, wC, d1, d2),
|
|
getW(wF, wR, wC, d1 + 1, d2),
|
|
getW(wF, wR, wC, d1 + 2, d2),
|
|
getW(wF, wR, wC, d1 + 3, d2)
|
|
);
|
|
|
|
dotProd += dot(xValues, wValues);
|
|
}
|
|
|
|
if (${f===1}) {
|
|
dotProd +=
|
|
getX(batch, xF, xR, xC, ${h}) *
|
|
getW(wF, wR, wC, ${h}, d2);
|
|
} else if (${f===2}) {
|
|
vec2 xValues = vec2(
|
|
getX(batch, xF, xR, xC, ${h}),
|
|
getX(batch, xF, xR, xC, ${h} + 1)
|
|
);
|
|
vec2 wValues = vec2(
|
|
getW(wF, wR, wC, ${h}, d2),
|
|
getW(wF, wR, wC, ${h} + 1, d2)
|
|
);
|
|
dotProd += dot(xValues, wValues);
|
|
} else if (${f===3}) {
|
|
vec3 xValues = vec3(
|
|
getX(batch, xF, xR, xC, ${h}),
|
|
getX(batch, xF, xR, xC, ${h} + 1),
|
|
getX(batch, xF, xR, xC, ${h} + 2)
|
|
);
|
|
vec3 wValues = vec3(
|
|
getW(wF, wR, wC, ${h}, d2),
|
|
getW(wF, wR, wC, ${h} + 1, d2),
|
|
getW(wF, wR, wC, ${h} + 2, d2)
|
|
);
|
|
dotProd += dot(xValues, wValues);
|
|
}
|
|
}
|
|
}
|
|
}
|
|
setOutput(dotProd);
|
|
}
|
|
`}},a8=class{constructor(e,t=!1,a=null,n=!1,r=!1){this.variableNames=["x","W"],this.packedInputs=!0,this.packedOutput=!0,this.customUniforms=[{name:"pads",type:"ivec2"},{name:"strides",type:"ivec2"},{name:"dilations",type:"ivec2"},{name:"inDims",type:"ivec2"}],this.outputShape=e.outShape,this.enableShapeUniforms=ya(this.outputShape.length);let s=e.padInfo.left,i=e.strideWidth,o=e.dilationWidth,l=e.filterHeight,u=e.filterWidth,p=u,c=`
|
|
int xR; int xC; int xCOffset;
|
|
vec4 wTexel; vec4 previous; vec4 final;`;for(let m=0;m<u;m++)c+=`
|
|
vec4 xTexelC${m*2};
|
|
int xTexelC${m*2}Ready;
|
|
vec4 xTexelC${m*2+1};
|
|
int xTexelC${m*2+1}Ready;
|
|
vec4 xC${m};`;c+=`
|
|
for (int r = 0; r < ${l}; r++) {
|
|
for (int d1 = 0; d1 < ${e.inChannels}; d1 += 2) {
|
|
`;for(let m=0;m<u;m++)c+=`
|
|
xTexelC${m*2} = vec4(0.0);
|
|
xTexelC${m*2}Ready = 0;
|
|
xTexelC${m*2+1} = vec4(0.0);
|
|
xTexelC${m*2+1}Ready = 0;
|
|
xC${m} = vec4(0.0);`;c+=`
|
|
xR = xRCorner + r * dilations[0];
|
|
if (xR >=0 && xR < inDims[0]) {
|
|
`;for(let m=0;m<(p+1)/2;m++){let g=m*2;if(c+=`
|
|
xC = xCCorner + ${g*o};
|
|
`,i===1){if(g<u&&(s%2===1?(c+=`
|
|
xCOffset = xC + 1;
|
|
if (xCOffset >= 0 && xCOffset < inDims[1] && xTexelC${g}Ready == 0) {
|
|
xTexelC${g} = getX(batch, xR, xCOffset, d1);
|
|
|
|
// Need to manually clear unused channels in case
|
|
// we're reading from recycled texture.
|
|
if (xCOffset + 1 >= inDims[1]) {
|
|
xTexelC${g}.zw = vec2(0.0);
|
|
}
|
|
xTexelC${g}Ready = 1;
|
|
}
|
|
`,o===1&&g>0?c+=`
|
|
xC${g} = vec4(xTexelC${g-2}.zw, xTexelC${g}.xy);
|
|
`:c+=`
|
|
xCOffset = xC + 1 - 2;
|
|
|
|
if (xCOffset >= 0 && xCOffset < inDims[1]) {
|
|
previous = getX(batch, xR, xCOffset, d1);
|
|
|
|
// Need to manually clear unused channels in case
|
|
// we're reading from recycled texture.
|
|
if (xCOffset + 1 >= inDims[1]) {
|
|
previous.zw = vec2(0.0);
|
|
}
|
|
|
|
xC${g} = vec4(previous.zw, xTexelC${g}.xy);
|
|
} else {
|
|
xC${g} = vec4(0.0, 0.0, xTexelC${g}.xy);
|
|
}
|
|
`):c+=`
|
|
if (xC >= 0 && xC < inDims[1] && xTexelC${g}Ready == 0) {
|
|
xTexelC${g} = getX(batch, xR, xC, d1);
|
|
if (xC + 1 >= inDims[1]) {
|
|
xTexelC${g}.zw = vec2(0.0);
|
|
}
|
|
xTexelC${g}Ready = 1;
|
|
}
|
|
|
|
xC${g} = xTexelC${g};
|
|
`,g+1<u)){let y=s%2===0?v.nearestLargerEven(o):o;o%2===0&&s%2===1||o%2!==0&&s%2!==1?(c+=`
|
|
xCOffset = xC + imod(pads[1], 2) + ${y};
|
|
|
|
if (xCOffset >= 0 && xCOffset < inDims[1] && xTexelC${g+1}Ready == 0) {
|
|
xTexelC${g+1} = getX(batch, xR, xCOffset, d1);
|
|
|
|
// Need to manually clear unused channels in case
|
|
// we're reading from recycled texture.
|
|
if (xCOffset + 1 >= inDims[1]) {
|
|
xTexelC${g+1}.zw = vec2(0.0);
|
|
}
|
|
xTexelC${g+1}Ready = 1;
|
|
}
|
|
`,o>1?c+=`
|
|
xCOffset -= 2;
|
|
if (xCOffset >= 0 && xCOffset < inDims[1]) {
|
|
previous = getX(batch, xR, xCOffset, d1);
|
|
xC${g+1} = vec4(previous.zw, xTexelC${g+1}.xy);
|
|
} else {
|
|
xC${g+1} = vec4(0.0, 0.0, xTexelC${g+1}.xy);
|
|
}
|
|
`:c+=`
|
|
xC${g+1} = vec4(xTexelC${g}.zw, xTexelC${g+1}.xy);
|
|
`):y===1?c+=`
|
|
xC${g+1} = xTexelC${g};
|
|
`:c+=`
|
|
xCOffset = xC + ${y};
|
|
|
|
if (xCOffset >= 0 && xCOffset < inDims[1] && xTexelC${g+1}Ready == 0) {
|
|
xTexelC${g+1} = getX(batch, xR, xCOffset, d1);
|
|
if (xCOffset + 1 >= inDims[1]) {
|
|
xTexelC${g+1}.zw = vec2(0.0);
|
|
}
|
|
xTexelC${g+1}Ready = 1;
|
|
}
|
|
|
|
xC${g+1} = xTexelC${g+1};
|
|
`}}else g<u&&(s%2===1?(c+=`
|
|
xCOffset = xC + 1 - strides[1];
|
|
if(xCOffset >= 0 && xCOffset < inDims[1] && xTexelC${g}Ready == 0) {
|
|
xTexelC${g} = getX(batch, xR, xCOffset, d1);
|
|
// Need to manually clear unused channels in case
|
|
// we're reading from recycled texture.
|
|
if (xCOffset + 1 >= inDims[1]) {
|
|
xTexelC${g}.zw = vec2(0.0);
|
|
}
|
|
xTexelC${g}Ready = 1;
|
|
}
|
|
|
|
if(xC + 1 >= 0 && xC + 1 < inDims[1] && xTexelC${g+1}Ready == 0) {
|
|
xTexelC${g+1} = getX(batch, xR, xC + 1, d1);
|
|
// Need to manually clear unused channels in case
|
|
// we're reading from recycled texture.
|
|
if (xC + 2 >= inDims[1]) {
|
|
xTexelC${g+1}.zw = vec2(0.0);
|
|
}
|
|
xTexelC${g+1}Ready = 1;
|
|
}
|
|
|
|
xC${g} = vec4(xTexelC${g}.zw, xTexelC${g+1}.zw);
|
|
`,g+1<u&&(c+=`
|
|
final = vec4(0.0);
|
|
xCOffset = xC + 1 + strides[1];
|
|
if(xCOffset >= 0 && xCOffset < inDims[1]) {
|
|
final = getX(batch, xR, xCOffset, d1);
|
|
}
|
|
xC${g+1} = vec4(xTexelC${g+1}.xy, final.xy);
|
|
`)):(c+=`
|
|
if(xC >= 0 && xC < inDims[1] && xTexelC${g}Ready == 0) {
|
|
xTexelC${g} = getX(batch, xR, xC, d1);
|
|
if (xC + 1 >= inDims[1]) {
|
|
xTexelC${g}.zw = vec2(0.0);
|
|
}
|
|
xTexelC${g}Ready = 1;
|
|
}
|
|
|
|
xCOffset = xC + strides[1];
|
|
if(xCOffset >= 0 && xCOffset < inDims[1] && xTexelC${g+1}Ready == 0) {
|
|
xTexelC${g+1} = getX(batch, xR, xCOffset, d1);
|
|
if (xCOffset + 1 >= inDims[1]) {
|
|
xTexelC${g+1}.zw = vec2(0.);
|
|
}
|
|
xTexelC${g+1}Ready = 1;
|
|
}
|
|
|
|
xC${g} = vec4(
|
|
xTexelC${g}.xy, xTexelC${g+1}.xy);
|
|
`,g+1<u&&(c+=`
|
|
xC${g+1} = vec4(xTexelC${g}.zw, xTexelC${g+1}.zw);
|
|
`)));g<u&&(c+=`
|
|
wTexel = getW(r, ${g}, d1, d2);
|
|
dotProd += xC${g}.xxzz * vec4(wTexel.xy, wTexel.xy);
|
|
if(d1 + 1 < ${e.inChannels}) {
|
|
dotProd += xC${g}.yyww * vec4(wTexel.zw, wTexel.zw);
|
|
}
|
|
`,g+1<u&&(c+=`
|
|
wTexel = getW(r, ${g+1}, d1, d2);
|
|
dotProd += xC${g+1}.xxzz * vec4(wTexel.xy, wTexel.xy);
|
|
if(d1 + 1 < ${e.inChannels}) {
|
|
dotProd += xC${g+1}.yyww * vec4(wTexel.zw, wTexel.zw);
|
|
}
|
|
`))}c+=`
|
|
}
|
|
`,c+=`
|
|
}
|
|
`,c+=`
|
|
}
|
|
`;let d="",h="";a&&(n?d=`vec4 activation(vec4 a) {
|
|
vec4 b = getPreluActivationWeightsAtOutCoords();
|
|
${a}
|
|
}`:r?d=`vec4 activation(vec4 a) {
|
|
vec4 b = getLeakyreluAlphaAtOutCoords();
|
|
${a}
|
|
}`:d=`vec4 activation(vec4 x) {
|
|
${a}
|
|
}`,h="result = activation(result);");let f=t?"result += getBiasAtOutCoords();":"";t&&this.variableNames.push("bias"),n&&this.variableNames.push("preluActivationWeights"),r&&this.variableNames.push("leakyreluAlpha"),this.userCode=`
|
|
${d}
|
|
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int batch = coords.x;
|
|
ivec2 xRCCorner = coords.yz * strides - pads;
|
|
int d2 = coords.w;
|
|
int xRCorner = xRCCorner.x;
|
|
int xCCorner = xRCCorner.y;
|
|
|
|
//intialize dotProd with a small epsilon seems to reduce GPU accuracy loss.
|
|
vec4 dotProd = vec4(0.000000000000001);
|
|
|
|
${c}
|
|
|
|
vec4 result = dotProd - vec4(0.000000000000001);
|
|
${f}
|
|
${h}
|
|
setOutput(result);
|
|
}
|
|
`}},nX=class{constructor(e,t){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!0,this.customUniforms=[{name:"inputShape",type:"ivec4"},{name:"pad",type:"ivec2"},{name:"stride",type:"ivec2"},{name:"dilation",type:"ivec2"},{name:"inChannels",type:"int"},{name:"itemsPerBlockRow",type:"int"},{name:"outWidth",type:"int"}],this.outputShape=e,this.enableShapeUniforms=ya(this.outputShape.length);let{dataFormat:a}=t,n=Ea(),r=a==="channelsLast",s=r?1:2,i=r?2:3,o=this.enableShapeUniforms?"if(blockIndex < outShape[2] && pos < outShape[1]) {":`if(blockIndex < ${e[2]} && pos < ${e[1]}) {`,l="";for(let u=0;u<=1;u++)for(let p=0;p<=1;p++)l+=`
|
|
blockIndex = rc.z + ${p};
|
|
pos = rc.y + ${u};
|
|
|
|
${o}
|
|
offsetY = int(blockIndex / outWidth) * stride[0] - pad[0];
|
|
d0 = offsetY + dilation[0] * (pos / itemsPerBlockRow);
|
|
|
|
if(d0 < inputShape[${s}] && d0 >= 0) {
|
|
// Use custom imod instead mod. On Intel GPU, mod may generate
|
|
// unexpected value.
|
|
// https://github.com/tensorflow/tfjs/issues/5447
|
|
offsetX = imod(blockIndex, outWidth) * stride[1] - pad[1];
|
|
d1 = offsetX + dilation[1] * (imod(pos, itemsPerBlockRow) /
|
|
inChannels);
|
|
|
|
if(d1 < inputShape[${i}] && d1 >= 0) {
|
|
|
|
ch = imod(pos, inChannels);
|
|
|
|
if (${r}) {
|
|
innerDims = vec2(d1, ch);
|
|
result[${u*2+p}] = getChannel(
|
|
getA(rc.x, d0, int(innerDims.x),
|
|
int(innerDims.y)), innerDims);
|
|
} else {
|
|
innerDims = vec2(d0, d1);
|
|
result[${u*2+p}] = getChannel(
|
|
getA(rc.x, ch, int(innerDims.x),
|
|
int(innerDims.y)), innerDims);
|
|
}
|
|
}
|
|
}
|
|
}
|
|
`;this.userCode=`
|
|
void main() {
|
|
ivec3 rc = getOutputCoords();
|
|
|
|
vec4 result = vec4(0);
|
|
|
|
int blockIndex, pos, offsetY, d0, offsetX, d1, ch;
|
|
vec2 innerDims;
|
|
|
|
${l}
|
|
|
|
${n.output} = result;
|
|
}
|
|
`}};function mh(e,t){let a=e.length;return a>=3?t?[...e.slice(0,-3),e[a-3]*e[a-2],e[a-1]]:[...e.slice(0,-3),e[a-3],e[a-2]*e[a-1]]:!t&&a===1&&e[0]>1?[e[0],1]:null}function n8({x:e,filter:t,convInfo:a,backend:n,bias:r=null,preluActivationWeights:s=null,leakyreluAlpha:i=0,activation:o=null}){let l=e.shape,u=n.texData.get(e.dataId),p=a.inChannels,c=l[0]*l[1]*l[2],d=a.outChannels,h=a.dataFormat==="channelsLast",f=!1,m=!1,g,y=[];if(s!=null){let x=mh(s.shape,h);x!=null&&(s=ue({inputs:{x:s},backend:n,attrs:{shape:x}}),y.push(s))}if(r!=null){let x=mh(r.shape,h);x!=null&&(r=ue({inputs:{x:r},backend:n,attrs:{shape:x}}),y.push(r))}if(!((c===1||d===1)&&p>Kv)&&u.isPacked&&h&&u.texture!=null&&l[2]%2!==0&&v.arraysEqual(u.shape.slice(-3),l.slice(-3))){let x=l[0]*l[1]*(l[2]+1),A={dataId:e.dataId,shape:[1,x,a.inChannels],dtype:e.dtype},b=u.shape;u.shape=u.shape.slice(),u.shape[u.shape.length-2]++,v.assert(Gd(u.shape,A.shape),()=>`packed reshape ${u.shape} to ${A.shape} isn't free`);let w=ue({inputs:{x:t},backend:n,attrs:{shape:[1,a.inChannels,a.outChannels]}});y.push(w);let C=fh({a:A,b:w,backend:n,transposeA:f,transposeB:m,bias:r,activation:o,preluActivationWeights:s,leakyreluAlpha:i}),T=n.texData.get(C.dataId);v.assert(T.isPacked,()=>"batchMatMul result is expected to be packed"),u.shape=b,T.shape=a.outShape,g=en({inputs:{x:C},backend:n}),g.shape=a.outShape,y.push(C)}else{let x=a.outHeight*a.outWidth,A=ue({inputs:{x:e},backend:n,attrs:{shape:h?[a.batchSize,x,a.inChannels]:[a.batchSize,a.inChannels,x]}}),b=ue({inputs:{x:t},backend:n,attrs:{shape:[1,a.inChannels,a.outChannels]}}),w=fh({a:h?A:b,b:h?b:A,transposeA:!h,transposeB:m,backend:n,bias:r,activation:o,preluActivationWeights:s,leakyreluAlpha:i});g=ue({inputs:{x:w},backend:n,attrs:{shape:a.outShape}}),y.push(A),y.push(b),y.push(w)}for(let x of y)n.disposeIntermediateTensorInfo(x);return g}function r8({x:e,filter:t,convInfo:a,backend:n,bias:r=null,preluActivationWeights:s=null,leakyreluAlpha:i=0,activation:o=null}){let{filterWidth:l,filterHeight:u,inChannels:p,outWidth:c,outHeight:d,dataFormat:h}=a,f=h==="channelsLast",m=l*u*p,g=d*c,y=[a.batchSize,m,g],x=!0,A=!1,b=[];if(s!=null){let G=mh(s.shape,f);G!=null&&(s=ue({inputs:{x:s},backend:n,attrs:{shape:G}}),b.push(s))}if(r!=null){let G=mh(r.shape,f);G!=null&&(r=ue({inputs:{x:r},backend:n,attrs:{shape:G}}),b.push(r))}let w=ue({inputs:{x:t},backend:n,attrs:{shape:[1,m,v.sizeFromShape(t.shape)/m]}});b.push(w);let C=new nX(y,a),T=[e.shape,[a.padInfo.top,a.padInfo.left],[a.strideHeight,a.strideWidth],[a.dilationHeight,a.dilationWidth],[a.inChannels],[a.filterWidth*a.inChannels],[a.outWidth]],N=n.runWebGLProgram(C,[e],"float32",T),M=ue({inputs:{x:N},backend:n,attrs:{shape:y}});b.push(N),b.push(M);let F=r!=null,E=s!=null,I=o==="leakyrelu",_=o?Hd(o,!0):null,O=new Xv(f?M.shape:w.shape,f?w.shape:M.shape,f?[a.batchSize,g,a.outChannels]:[a.batchSize,a.outChannels,g],x,A,F,_,E,I),B=f?[M,w]:[w,M];if(r&&B.push(r),E&&B.push(s),I){let G=n.makeTensorInfo([],"float32",v.createScalarValue(i,"float32"));B.push(G),b.push(G)}let $=n.runWebGLProgram(O,B,"float32"),U=ue({inputs:{x:$},backend:n,attrs:{shape:a.outShape}});b.push($);for(let G of b)n.disposeIntermediateTensorInfo(G);return U}function rX(e){let{inputs:t,backend:a,attrs:n}=e,{x:r,filter:s}=t,{strides:i,pad:o,dataFormat:l,dilations:u,dimRoundingMode:p}=n,c=S.convertConv2DDataFormat(l),d=S.computeConv2DInfo(r.shape,s.shape,i,u,o,p,!1,c),h;if(d.filterHeight===1&&d.filterWidth===1&&d.dilationHeight===1&&d.dilationWidth===1&&d.strideHeight===1&&d.strideWidth===1&&(d.padInfo.type==="SAME"||d.padInfo.type==="VALID"))h=n8({x:r,filter:s,convInfo:d,backend:a});else if(d.strideWidth<=2&&c==="channelsLast"&&W().getBool("WEBGL_EXP_CONV")){let m=new a8(d),g=[[d.padInfo.top,d.padInfo.left],[d.strideHeight,d.strideWidth],[d.dilationHeight,d.dilationWidth],[d.inHeight,d.inWidth]];h=a.runWebGLProgram(m,[r,s],"float32",g)}else if(W().getBool("WEBGL_CONV_IM2COL"))h=r8({x:r,filter:s,convInfo:d,backend:a});else{let m=new t8(d);h=a.runWebGLProgram(m,[r,s],"float32")}let f=ue({inputs:{x:h},backend:a,attrs:{shape:d.outShape}});return a.disposeIntermediateTensorInfo(h),f}var sX={kernelName:fi,backendName:"webgl",kernelFunc:rX},iX=class{constructor(e){this.variableNames=["x","dy"],this.outputShape=e.filterShape;let t=e.strideHeight,a=e.strideWidth,n=e.padInfo.top,r=e.padInfo.left,s=e.dataFormat==="channelsLast";this.userCode=`
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int wR = coords.x;
|
|
int wC = coords.y;
|
|
int d1 = coords.z;
|
|
int d2 = coords.w;
|
|
|
|
// Convolve x(?, ?, d1) with dy(:, :, d2) to get dw(wR, wC, d1, d2).
|
|
// ? = to be determined. : = across all values in that axis.
|
|
float dotProd = 0.0;
|
|
|
|
for (int b = 0; b < ${e.batchSize}; b++) {
|
|
for (int yR = 0; yR < ${e.outHeight}; yR++) {
|
|
int xR = wR + yR * ${t} - ${n};
|
|
|
|
if (xR < 0 || xR >= ${e.inHeight}) {
|
|
continue;
|
|
}
|
|
|
|
for (int yC = 0; yC < ${e.outWidth}; yC++) {
|
|
int xC = wC + yC * ${a} - ${r};
|
|
|
|
if (xC < 0 || xC >= ${e.inWidth}) {
|
|
continue;
|
|
}
|
|
|
|
${s?`float dyValue = getDy(b, yR, yC, d2);
|
|
float xValue = getX(b, xR, xC, d1);
|
|
dotProd += (xValue * dyValue);`:`float dyValue = getDy(b, d2, yR, yC);
|
|
float xValue = getX(b, d1, xR, xC);
|
|
dotProd += (xValue * dyValue);`}
|
|
}
|
|
}
|
|
}
|
|
setOutput(dotProd);
|
|
}
|
|
`}},oX=class{constructor(e){this.variableNames=["dy","W"],this.outputShape=e.inShape;let t=e.filterHeight,a=e.filterWidth,n=e.strideHeight,r=e.strideWidth,s=e.dataFormat==="channelsLast",i=t-1-e.padInfo.top,o=a-1-e.padInfo.left,l=s?1:2,u=s?2:3,p=s?3:1;this.userCode=`
|
|
const ivec2 pads = ivec2(${i}, ${o});
|
|
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int batch = coords[0];
|
|
int d1 = coords[${p}];
|
|
|
|
ivec2 dyCorner = ivec2(coords[${l}], coords[${u}]) - pads;
|
|
int dyRCorner = dyCorner.x;
|
|
int dyCCorner = dyCorner.y;
|
|
|
|
// Convolve dy(?, ?, d2) with w(:, :, d1, d2) to compute dx(xR, xC, d1).
|
|
// ? = to be determined. : = across all values in that axis.
|
|
float dotProd = 0.0;
|
|
for (int wR = 0; wR < ${t}; wR++) {
|
|
float dyR = float(dyRCorner + wR) / ${n}.0;
|
|
|
|
if (dyR < 0.0 || dyR >= ${e.outHeight}.0 || fract(dyR) > 0.0) {
|
|
continue;
|
|
}
|
|
int idyR = int(dyR);
|
|
|
|
int wRPerm = ${t} - 1 - wR;
|
|
|
|
for (int wC = 0; wC < ${a}; wC++) {
|
|
float dyC = float(dyCCorner + wC) / ${r}.0;
|
|
|
|
if (dyC < 0.0 || dyC >= ${e.outWidth}.0 ||
|
|
fract(dyC) > 0.0) {
|
|
continue;
|
|
}
|
|
int idyC = int(dyC);
|
|
|
|
int wCPerm = ${a} - 1 - wC;
|
|
|
|
for (int d2 = 0; d2 < ${e.outChannels}; d2++) {
|
|
|
|
if (${s}) {
|
|
float xValue = getDy(batch, idyR, idyC, d2);
|
|
float wValue = getW(wRPerm, wCPerm, d1, d2);
|
|
dotProd += xValue * wValue;
|
|
} else {
|
|
float xValue = getDy(batch, d2, idyR, idyC);
|
|
float wValue = getW(wRPerm, wCPerm, d1, d2);
|
|
dotProd += xValue * wValue;
|
|
}
|
|
|
|
}
|
|
}
|
|
}
|
|
setOutput(dotProd);
|
|
}
|
|
`}},lX=class{constructor(e){this.variableNames=["x","dy"],this.outputShape=e.filterShape;let t=e.strideDepth,a=e.strideHeight,n=e.strideWidth,r=e.padInfo.front,s=e.padInfo.top,i=e.padInfo.left;this.userCode=`
|
|
void main() {
|
|
ivec5 coords = getOutputCoords();
|
|
int wF = coords.x;
|
|
int wR = coords.y;
|
|
int wC = coords.z;
|
|
int d1 = coords.w;
|
|
int d2 = coords.u;
|
|
|
|
float dotProd = 0.0;
|
|
|
|
for (int b = 0; b < ${e.batchSize}; b++) {
|
|
for (int yF = 0; yF < ${e.outDepth}; yF++) {
|
|
int xF = wF + yF * ${t} - ${r};
|
|
|
|
if (xF < 0 || xF >= ${e.inDepth}) {
|
|
continue;
|
|
}
|
|
|
|
for (int yR = 0; yR < ${e.outHeight}; yR++) {
|
|
int xR = wR + yR * ${a} - ${s};
|
|
|
|
if (xR < 0 || xR >= ${e.inHeight}) {
|
|
continue;
|
|
}
|
|
|
|
for (int yC = 0; yC < ${e.outWidth}; yC++) {
|
|
int xC = wC + yC * ${n} - ${i};
|
|
|
|
if (xC < 0 || xC >= ${e.inWidth}) {
|
|
continue;
|
|
}
|
|
|
|
float dyValue = getDy(b, yF, yR, yC, d2);
|
|
float xValue = getX(b, xF, xR, xC, d1);
|
|
dotProd += (xValue * dyValue);
|
|
}
|
|
}
|
|
}
|
|
}
|
|
setOutput(dotProd);
|
|
}
|
|
`}},uX=class{constructor(e){this.variableNames=["dy","W"],this.outputShape=e.inShape;let t=e.filterDepth,a=e.filterHeight,n=e.filterWidth,r=e.strideDepth,s=e.strideHeight,i=e.strideWidth,o=t-1-e.padInfo.front,l=a-1-e.padInfo.top,u=n-1-e.padInfo.left;this.userCode=`
|
|
const ivec3 pads = ivec3(${o}, ${l}, ${u});
|
|
|
|
void main() {
|
|
ivec5 coords = getOutputCoords();
|
|
int batch = coords.x;
|
|
int d1 = coords.u;
|
|
|
|
|
|
ivec3 dyCorner = ivec3(coords.y, coords.z, coords.w) - pads;
|
|
int dyFCorner = dyCorner.x;
|
|
int dyRCorner = dyCorner.y;
|
|
int dyCCorner = dyCorner.z;
|
|
|
|
float dotProd = 0.0;
|
|
for (int wF = 0; wF < ${t}; wF++) {
|
|
float dyF = float(dyFCorner + wF) / ${r}.0;
|
|
|
|
if (dyF < 0.0 || dyF >= ${e.outDepth}.0 || fract(dyF) > 0.0) {
|
|
continue;
|
|
}
|
|
int idyF = int(dyF);
|
|
|
|
int wFPerm = ${t} - 1 - wF;
|
|
|
|
for (int wR = 0; wR < ${a}; wR++) {
|
|
float dyR = float(dyRCorner + wR) / ${s}.0;
|
|
|
|
if (dyR < 0.0 || dyR >= ${e.outHeight}.0 ||
|
|
fract(dyR) > 0.0) {
|
|
continue;
|
|
}
|
|
int idyR = int(dyR);
|
|
|
|
int wRPerm = ${a} - 1 - wR;
|
|
|
|
for (int wC = 0; wC < ${n}; wC++) {
|
|
float dyC = float(dyCCorner + wC) / ${i}.0;
|
|
|
|
if (dyC < 0.0 || dyC >= ${e.outWidth}.0 ||
|
|
fract(dyC) > 0.0) {
|
|
continue;
|
|
}
|
|
int idyC = int(dyC);
|
|
|
|
int wCPerm = ${n} - 1 - wC;
|
|
|
|
for (int d2 = 0; d2 < ${e.outChannels}; d2++) {
|
|
float xValue = getDy(batch, idyF, idyR, idyC, d2);
|
|
float wValue = getW(wFPerm, wRPerm, wCPerm, d1, d2);
|
|
dotProd += xValue * wValue;
|
|
}
|
|
}
|
|
}
|
|
}
|
|
setOutput(dotProd);
|
|
}
|
|
`}};function dX(e){let{inputs:t,backend:a,attrs:n}=e,{x:r,dy:s}=t,{strides:i,pad:o,dataFormat:l,dimRoundingMode:u,filterShape:p}=n,c=S.convertConv2DDataFormat(l),d=S.computeConv2DInfo(r.shape,p,i,1,o,u,!1,c),h=new iX(d);return a.runWebGLProgram(h,[r,s],"float32")}var pX={kernelName:np,backendName:"webgl",kernelFunc:dX},cX=class{constructor(e){this.variableNames=["dy","W"],this.packedInputs=!0,this.packedOutput=!0,this.customUniforms=[{name:"strides",type:"vec2"}],this.outputShape=e.inShape,this.enableShapeUniforms=ya(this.outputShape.length);let t=e.filterHeight,a=e.filterWidth,n=t-1-e.padInfo.top,r=a-1-e.padInfo.left;this.userCode=`
|
|
const ivec2 pads = ivec2(${n}, ${r});
|
|
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int batch = coords[0];
|
|
int d1 = coords[3];
|
|
|
|
ivec2 dyCorner = ivec2(coords[1], coords[2]) - pads;
|
|
int dyRCorner = dyCorner.x;
|
|
int dyCCorner = dyCorner.y;
|
|
|
|
vec4 result = vec4(0.);
|
|
for (int wR = 0; wR < ${t}; wR++) {
|
|
float dyR = float(dyRCorner + wR) / strides[0];
|
|
if (dyR < 0.0 || dyR >= ${e.outHeight}.0 || fract(dyR) > 0.0) {
|
|
continue;
|
|
}
|
|
int idyR = int(dyR);
|
|
int wRPerm = ${t} - 1 - wR;
|
|
|
|
for (int wC = 0; wC < ${a}; wC++) {
|
|
int wCPerm = ${a} - 1 - wC;
|
|
|
|
float dyC = float(dyCCorner + wC) / strides[1];
|
|
bool idyCVal = (dyC >= 0.0) && (dyC < ${e.outWidth}.0)
|
|
&& (fract(dyC) == 0.0);
|
|
int idyC = int(dyC);
|
|
|
|
float dyC2 = float(dyCCorner + wC + 1) / strides[1];
|
|
bool idyCVal2 = (dyC2 >= 0.0) && (dyC2 < ${e.outWidth}.0)
|
|
&& (fract(dyC2) == 0.0);
|
|
int idyC2 = int(dyC2);
|
|
|
|
if (idyCVal && idyCVal2) {
|
|
for (int d2 = 0; d2 < ${e.outChannels}; d2 += 2) {
|
|
vec4 wValue = getW(wRPerm, wCPerm, d1, d2);
|
|
vec4 dySample = getDy(batch, idyR, idyC, d2);
|
|
vec4 dySample2 = (idyC / 2 == idyC2 / 2) ?
|
|
dySample : getDy(batch, idyR, idyC2, d2);
|
|
|
|
vec2 dyValue = mod(float(idyC), 2.) == 0. ?
|
|
dySample.xy : dySample.zw;
|
|
result.xy += vec2(dot(dyValue, wValue.xy),
|
|
dot(dyValue, wValue.zw));
|
|
|
|
dyValue = mod(float(idyC2), 2.) == 0. ?
|
|
dySample2.xy : dySample2.zw;
|
|
result.zw += vec2(dot(dyValue, wValue.xy),
|
|
dot(dyValue, wValue.zw));
|
|
}
|
|
} else if (idyCVal) {
|
|
for (int d2 = 0; d2 < ${e.outChannels}; d2 += 2) {
|
|
vec4 wValue = getW(wRPerm, wCPerm, d1, d2);
|
|
vec4 dySample = getDy(batch, idyR, idyC, d2);
|
|
vec2 dyValue = mod(float(idyC), 2.) == 0. ?
|
|
dySample.xy : dySample.zw;
|
|
result.xy += vec2(dot(dyValue, wValue.xy),
|
|
dot(dyValue, wValue.zw));
|
|
}
|
|
} else if (idyCVal2) {
|
|
for (int d2 = 0; d2 < ${e.outChannels}; d2 += 2) {
|
|
vec4 wValue = getW(wRPerm, wCPerm, d1, d2);
|
|
vec4 dySample = getDy(batch, idyR, idyC2, d2);
|
|
vec2 dyValue = mod(float(idyC2), 2.) == 0. ?
|
|
dySample.xy : dySample.zw;
|
|
result.zw += vec2(dot(dyValue, wValue.xy),
|
|
dot(dyValue, wValue.zw));
|
|
}
|
|
}
|
|
}
|
|
}
|
|
setOutput(result);
|
|
}
|
|
`}};function hX(e){let{inputs:t,backend:a,attrs:n}=e,{dy:r,filter:s}=t,{inputShape:i,strides:o,pad:l,dataFormat:u,dimRoundingMode:p}=n,c=S.convertConv2DDataFormat(u),d=S.computeConv2DInfo(i,s.shape,o,1,l,p,!1,c);if(W().getBool("WEBGL_PACK")&&c==="channelsLast"){let h=[[d.strideHeight,d.strideWidth]],f=new cX(d);return a.runWebGLProgram(f,[r,s],"float32",h)}else{let h=new oX(d);return a.runWebGLProgram(h,[r,s],"float32")}}var fX={kernelName:mi,backendName:"webgl",kernelFunc:hX};function mX(e){let{inputs:t,backend:a,attrs:n}=e,{x:r,filter:s}=t,{strides:i,pad:o,dilations:l}=n,u=S.computeConv3DInfo(r.shape,s.shape,i,l,o),p=new aX(u);return a.runWebGLProgram(p,[r,s],"float32")}var gX={kernelName:gi,backendName:"webgl",kernelFunc:mX};function yX(e){let{inputs:t,backend:a,attrs:n}=e,{x:r,dy:s}=t,{strides:i,pad:o,filterShape:l}=n,u=S.computeConv3DInfo(r.shape,l,i,1,o),p=new lX(u);return a.runWebGLProgram(p,[r,s],"float32")}var xX={kernelName:ru,backendName:"webgl",kernelFunc:yX};function AX(e){let{inputs:t,backend:a,attrs:n}=e,{dy:r,filter:s}=t,{pad:i,strides:o,inputShape:l}=n,u=S.computeConv3DInfo(l,s.shape,o,1,i),p=new uX(u);return a.runWebGLProgram(p,[r,s],"float32")}var bX={kernelName:yi,backendName:"webgl",kernelFunc:AX},vX=Vu+`
|
|
return cos(x);
|
|
`,wX=`
|
|
vec4 result = cos(x);
|
|
bvec4 isNaN = isnan(x);
|
|
${Xo}
|
|
return result;
|
|
`,kX=et({opSnippet:vX,packedOpSnippet:wX}),IX={kernelName:xi,backendName:"webgl",kernelFunc:kX},SX=`
|
|
float e2x = exp(-x);
|
|
return (e2x + 1.0 / e2x) / 2.0;
|
|
`,CX=et({opSnippet:SX}),TX={kernelName:Ai,backendName:"webgl",kernelFunc:CX},NX=class{constructor(e,t,a,n,r){this.variableNames=["Image","Boxes","BoxInd"],this.outputShape=[];let[s,i,o,l]=e,[u]=t,[p,c]=a;this.outputShape=[u,p,c,l];let d=n==="bilinear"?1:0,[h,f]=[`${i-1}.0`,`${o-1}.0`],[m,g,y]=p>1?[`${(i-1)/(p-1)}`,"(y2-y1) * height_ratio",`y1*${h} + float(y)*(height_scale)`]:["0.0","0.0",`0.5 * (y1+y2) * ${h}`],[x,A,b]=c>1?[`${(o-1)/(c-1)}`,"(x2-x1) * width_ratio",`x1*${f} + float(x)*(width_scale)`]:["0.0","0.0",`0.5 * (x1+x2) * ${f}`];this.userCode=`
|
|
const float height_ratio = float(${m});
|
|
const float width_ratio = float(${x});
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int b = coords[0];
|
|
int y = coords[1];
|
|
int x = coords[2];
|
|
int d = coords[3];
|
|
|
|
// get box vals
|
|
float y1 = getBoxes(b,0);
|
|
float x1 = getBoxes(b,1);
|
|
float y2 = getBoxes(b,2);
|
|
float x2 = getBoxes(b,3);
|
|
|
|
// get image in batch index
|
|
int bInd = round(getBoxInd(b));
|
|
if(bInd < 0 || bInd >= ${s}) {
|
|
return;
|
|
}
|
|
|
|
float height_scale = ${g};
|
|
float width_scale = ${A};
|
|
|
|
float in_y = ${y};
|
|
if( in_y < 0.0 || in_y > ${h} ) {
|
|
setOutput(float(${r}));
|
|
return;
|
|
}
|
|
float in_x = ${b};
|
|
if( in_x < 0.0 || in_x > ${f} ) {
|
|
setOutput(float(${r}));
|
|
return;
|
|
}
|
|
|
|
vec2 sourceFracIndexCR = vec2(in_x,in_y);
|
|
if(${d} == 1) {
|
|
// Compute the four integer indices.
|
|
ivec2 sourceFloorCR = ivec2(sourceFracIndexCR);
|
|
ivec2 sourceCeilCR = ivec2(ceil(sourceFracIndexCR));
|
|
|
|
float topLeft = getImage(b, sourceFloorCR.y, sourceFloorCR.x, d);
|
|
float bottomLeft = getImage(b, sourceCeilCR.y, sourceFloorCR.x, d);
|
|
float topRight = getImage(b, sourceFloorCR.y, sourceCeilCR.x, d);
|
|
float bottomRight = getImage(b, sourceCeilCR.y, sourceCeilCR.x, d);
|
|
|
|
vec2 fracCR = sourceFracIndexCR - vec2(sourceFloorCR);
|
|
|
|
float top = topLeft + (topRight - topLeft) * fracCR.x;
|
|
float bottom = bottomLeft + (bottomRight - bottomLeft) * fracCR.x;
|
|
float newValue = top + (bottom - top) * fracCR.y;
|
|
setOutput(newValue);
|
|
} else {
|
|
// Compute the coordinators of nearest neighbor point.
|
|
ivec2 sourceNearestCR = ivec2(floor(
|
|
sourceFracIndexCR + vec2(0.5,0.5)));
|
|
float newValue = getImage(b, sourceNearestCR.y, sourceNearestCR.x, d);
|
|
setOutput(newValue);
|
|
}
|
|
}
|
|
`}},RX=e=>{let{inputs:t,backend:a,attrs:n}=e,{image:r,boxes:s,boxInd:i}=t,{cropSize:o,method:l,extrapolationValue:u}=n,p=new NX(r.shape,s.shape,o,l,u);return a.runWebGLProgram(p,[r,s,i],"float32")},EX={kernelName:wi,backendName:"webgl",kernelFunc:RX},qd;(function(e){e.Prod="*",e.Sum="+"})(qd||(qd={}));var yx=class{constructor(e,t,a,n){this.op=e,this.outputShape=t,this.variableNames=["x"],this.customUniforms=[{name:"index",type:"float"}];let r=this.outputShape.length,s=this.op===qd.Prod?"1.0":"0.0",i=a?s:`getX(${xx(r,"coords",this.op)})`,o=this.outputShape[this.outputShape.length-1],l="",u="";a?(l=n?`end != ${o-1}`:"end != 0",u=n?"end + 1":"end - 1"):(l=n?`end + pow2 < ${o}`:"end >= pow2",u=n?"end + pow2":"end - pow2"),this.userCode=`
|
|
void main() {
|
|
${ft(r)} coords = getOutputCoords();
|
|
int end = ${Ax(r,"coords",this.op)};
|
|
float val = ${i};
|
|
int pow2 = int(pow(2.0, index));
|
|
if (${l}) {
|
|
int idx = ${u};
|
|
${Ax(r,"coords",this.op)} = idx;
|
|
val ${this.op}= getX(${xx(r,"coords",this.op)});
|
|
}
|
|
setOutput(val);
|
|
}
|
|
`}};function xx(e,t,a){if(e===1)return`${t}`;if(e===2)return`${t}.x, ${t}.y`;if(e===3)return`${t}.x, ${t}.y, ${t}.z`;if(e===4)return`${t}.x, ${t}.y, ${t}.z, ${t}.w`;throw new Error(`Cumulative ${a} for rank ${e} is not yet supported`)}function Ax(e,t,a){if(e===1)return`${t}`;if(e===2)return`${t}.y`;if(e===3)return`${t}.z`;if(e===4)return`${t}.w`;throw new Error(`Cumulative ${a} for rank ${e} is not yet supported`)}function s8(e,t,a,n,r,s){let i=t.shape.length,o=S.getAxesPermutation([n],i),l=t;o!=null&&(l=Ta({inputs:{x:t},backend:a,attrs:{perm:o}}));let u=S.getInnerMostAxes(1,i)[0];if(u!==i-1)throw new Error(`WebGL cumprod shader expects an inner-most axis=${t.shape.length-1} but got axis=${n}`);let p=l.shape[u],c=en({inputs:{x:l},backend:a});for(let d=0;d<=Math.ceil(Math.log2(p))-1;d++){let h=new yx(e,l.shape,!1,s),f=[[d]],m=c;c=a.runWebGLProgram(h,[c],c.dtype,f),a.disposeIntermediateTensorInfo(m)}if(r){let d=new yx(e,l.shape,r,s),h=c;c=a.runWebGLProgram(d,[c],c.dtype),a.disposeIntermediateTensorInfo(h)}if(o!=null){let d=S.getUndoAxesPermutation(o),h=Ta({inputs:{x:c},backend:a,attrs:{perm:d}});return a.disposeIntermediateTensorInfo(c),a.disposeIntermediateTensorInfo(l),h}return c}function MX(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{axis:s,exclusive:i,reverse:o}=n;return s8(qd.Prod,r,a,s,i,o)}var FX={kernelName:bi,backendName:"webgl",kernelFunc:MX};function _X(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{axis:s,exclusive:i,reverse:o}=n;return s8(qd.Sum,r,a,s,i,o)}var $X={kernelName:vi,backendName:"webgl",kernelFunc:_X};function DX(e){let{inputs:t,backend:a,attrs:n}=e,{x:r,weights:s}=t,{size:i,binaryOutput:o}=n;if(r.shape.length===1){let l=a.readSync(r.dataId),u=a.readSync(s.dataId),p=zv(l,u,s.dtype,s.shape,i);return a.makeTensorInfo([i],s.dtype,p)}else if(r.shape.length===2){let l=a.bufferSync(r),u=a.bufferSync(s),p=VG(l,u,i,o);return a.makeTensorInfo(p.shape,s.dtype,p.values)}throw new Error(`Error in denseBincount: input must be at most rank 2, but got rank${r.shape.length}.`)}var PX={kernelName:su,backendName:"webgl",kernelFunc:DX},OX=class{constructor(e,t,a){this.variableNames=["x"],this.outputShape=[],this.outputShape=e,this.blockSize=t,this.dataFormat=a,this.userCode=`
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int b = coords[0];
|
|
int h = ${this.getHeightCoordString()};
|
|
int w = ${this.getWidthCoordString()};
|
|
int d = ${this.getDepthCoordString()};
|
|
|
|
int in_h = h / ${t};
|
|
int offset_h = imod(h, ${t});
|
|
int in_w = w / ${t};
|
|
int offset_w = imod(w, ${t});
|
|
int offset_d = (offset_h * ${t} + offset_w) *
|
|
${this.getOutputDepthSize()};
|
|
int in_d = d + offset_d;
|
|
|
|
float result = ${this.getInputSamplingString()};
|
|
setOutput(result);
|
|
}
|
|
`}getHeightCoordString(){return this.dataFormat==="NHWC"?"coords[1]":"coords[2]"}getWidthCoordString(){return this.dataFormat==="NHWC"?"coords[2]":"coords[3]"}getDepthCoordString(){return this.dataFormat==="NHWC"?"coords[3]":"coords[1]"}getOutputDepthSize(){return this.dataFormat==="NHWC"?this.outputShape[3]:this.outputShape[1]}getInputSamplingString(){return this.dataFormat==="NHWC"?"getX(b, in_h, in_w, in_d)":"getX(b, in_d, in_h, in_w)"}};function zX(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{blockSize:s,dataFormat:i}=n,o=r.shape[0],l=i==="NHWC"?r.shape[1]:r.shape[2],u=i==="NHWC"?r.shape[2]:r.shape[3],p=i==="NHWC"?r.shape[3]:r.shape[1],c=l*s,d=u*s,h=p/(s*s),f=i==="NHWC"?[o,c,d,h]:[o,h,c,d],m=new OX(f,s,i);return a.runWebGLProgram(m,[r],r.dtype)}var LX={kernelName:ki,backendName:"webgl",kernelFunc:zX},i8=class{constructor(e,t=!1,a=null,n=!1,r=!1){this.variableNames=["x","W"],this.customUniforms=[{name:"pads",type:"ivec2"},{name:"strides",type:"ivec2"},{name:"dilations",type:"ivec2"},{name:"inDims",type:"ivec2"}],this.outputShape=e.outShape,this.enableShapeUniforms=ya(this.outputShape.length);let s=e.filterHeight,i=e.filterWidth,o=e.outChannels/e.inChannels,l="",u="";a&&(n?l=`float activation(float a) {
|
|
float b = getPreluActivationWeightsAtOutCoords();
|
|
${a}
|
|
}`:r?l=`float activation(float a) {
|
|
float b = getLeakyreluAlphaAtOutCoords();
|
|
${a}
|
|
}`:l=`
|
|
float activation(float x) {
|
|
${a}
|
|
}
|
|
`,u="result = activation(result);");let p=t?"result += getBiasAtOutCoords();":"";t&&this.variableNames.push("bias"),n&&this.variableNames.push("preluActivationWeights"),r&&this.variableNames.push("leakyreluAlpha"),this.userCode=`
|
|
${l}
|
|
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int batch = coords.x;
|
|
ivec2 xRCCorner = coords.yz * strides - pads;
|
|
int d2 = coords.w;
|
|
int d1 = d2 / ${o};
|
|
int q = d2 - d1 * ${o};
|
|
|
|
int xRCorner = xRCCorner.x;
|
|
int xCCorner = xRCCorner.y;
|
|
|
|
// Convolve x(?, ?, d1) with w(:, :, d1, q) to get y(yR, yC, d2).
|
|
// ? = to be determined. : = across all values in that axis.
|
|
float dotProd = 0.0;
|
|
// TO DO(dsmilkov): Flatten the two for loops and vec4 the operations.
|
|
for (int wR = 0; wR < ${s}; wR++) {
|
|
int xR = xRCorner + wR * dilations[0];
|
|
|
|
if (xR < 0 || xR >= inDims[0]) {
|
|
continue;
|
|
}
|
|
|
|
for (int wC = 0; wC < ${i}; wC++) {
|
|
int xC = xCCorner + wC * dilations[1];
|
|
|
|
if (xC < 0 || xC >= inDims[1]) {
|
|
continue;
|
|
}
|
|
|
|
float xVal = getX(batch, xR, xC, d1);
|
|
float wVal = getW(wR, wC, d1, q);
|
|
dotProd += xVal * wVal;
|
|
}
|
|
}
|
|
|
|
float result = dotProd;
|
|
${p}
|
|
${u}
|
|
setOutput(result);
|
|
}
|
|
`}},o8=class{constructor(e,t=!1,a=null,n=!1,r=!1){this.variableNames=["x","W"],this.packedInputs=!0,this.packedOutput=!0,this.customUniforms=[{name:"pads",type:"ivec2"},{name:"strides",type:"ivec2"},{name:"dilations",type:"ivec2"},{name:"inDims",type:"ivec2"}],this.outputShape=e.outShape,this.enableShapeUniforms=ya(this.outputShape.length);let s=e.outChannels/e.inChannels,i=e.padInfo.left,o=e.strideWidth,l=e.dilationWidth,u=e.filterHeight,p=e.filterWidth,c=p,d=`
|
|
int xR; int xC; int xCOffset;
|
|
vec4 wTexel; vec4 previous; vec4 final;`;for(let g=0;g<p;g++)d+=`
|
|
vec4 xTexelC${g*2};
|
|
int xTexelC${g*2}Ready;
|
|
vec4 xTexelC${g*2+1};
|
|
int xTexelC${g*2+1}Ready;
|
|
vec4 xC${g};`;d+=`
|
|
for (int r = 0; r < ${u}; r++) {
|
|
`;for(let g=0;g<p;g++)d+=`
|
|
xTexelC${g*2} = vec4(0.0);
|
|
xTexelC${g*2}Ready = 0;
|
|
xTexelC${g*2+1} = vec4(0.0);
|
|
xTexelC${g*2+1}Ready = 0;
|
|
xC${g} = vec4(0.0);`;d+=`
|
|
xR = xRCorner + r * dilations[0];
|
|
if (xR >=0 && xR < inDims[0]) {
|
|
`;for(let g=0;g<(c+1)/2;g++){let y=g*2;if(d+=`
|
|
xC = xCCorner + ${y*l};
|
|
`,o===1){if(y<p&&(i%2===1?(d+=`
|
|
xCOffset = xC + 1;
|
|
if (xCOffset >= 0 && xCOffset < inDims[1] && xTexelC${y}Ready == 0) {
|
|
xTexelC${y} = getX(batch, xR, xCOffset, d1);
|
|
|
|
// Need to manually clear unused channels in case
|
|
// we're reading from recycled texture.
|
|
if (xCOffset + 1 >= inDims[1]) {
|
|
xTexelC${y}.zw = vec2(0.0);
|
|
}
|
|
xTexelC${y}Ready = 1;
|
|
}
|
|
`,l===1&&y>0?d+=`
|
|
xC${y} = vec4(xTexelC${y-2}.zw, xTexelC${y}.xy);
|
|
`:d+=`
|
|
xCOffset = xC + 1 - 2;
|
|
|
|
if (xCOffset >= 0 && xCOffset < inDims[1]) {
|
|
previous = getX(batch, xR, xCOffset, d1);
|
|
|
|
// Need to manually clear unused channels in case
|
|
// we're reading from recycled texture.
|
|
if (xCOffset + 1 >= inDims[1]) {
|
|
previous.zw = vec2(0.0);
|
|
}
|
|
|
|
xC${y} = vec4(previous.zw, xTexelC${y}.xy);
|
|
} else {
|
|
xC${y} = vec4(0.0, 0.0, xTexelC${y}.xy);
|
|
}
|
|
`):d+=`
|
|
if (xC >= 0 && xC < inDims[1] && xTexelC${y}Ready == 0) {
|
|
xTexelC${y} = getX(batch, xR, xC, d1);
|
|
if (xC + 1 >= inDims[1]) {
|
|
xTexelC${y}.zw = vec2(0.0);
|
|
}
|
|
xTexelC${y}Ready = 1;
|
|
}
|
|
|
|
xC${y} = xTexelC${y};
|
|
`,y+1<p)){let x=i%2===0?v.nearestLargerEven(l):l;l%2===0&&i%2===1||l%2!==0&&i%2!==1?(d+=`
|
|
xCOffset = xC + imod(pads[1], 2) + ${x};
|
|
|
|
if (xCOffset >= 0 && xCOffset < inDims[1] && xTexelC${y+1}Ready == 0) {
|
|
xTexelC${y+1} = getX(batch, xR, xCOffset, d1);
|
|
|
|
// Need to manually clear unused channels in case
|
|
// we're reading from recycled texture.
|
|
if (xCOffset + 1 >= inDims[1]) {
|
|
xTexelC${y+1}.zw = vec2(0.0);
|
|
}
|
|
xTexelC${y+1}Ready = 1;
|
|
}
|
|
`,l>1?d+=`
|
|
xCOffset -= 2;
|
|
if (xCOffset >= 0 && xCOffset < inDims[1]) {
|
|
previous = getX(batch, xR, xCOffset, d1);
|
|
xC${y+1} = vec4(previous.zw, xTexelC${y+1}.xy);
|
|
} else {
|
|
xC${y+1} = vec4(0.0, 0.0, xTexelC${y+1}.xy);
|
|
}
|
|
`:d+=`
|
|
xC${y+1} = vec4(xTexelC${y}.zw, xTexelC${y+1}.xy);
|
|
`):x===1?d+=`
|
|
xC${y+1} = xTexelC${y};
|
|
`:d+=`
|
|
xCOffset = xC + ${x};
|
|
|
|
if (xCOffset >= 0 && xCOffset < inDims[1] && xTexelC${y+1}Ready == 0) {
|
|
xTexelC${y+1} = getX(batch, xR, xCOffset, d1);
|
|
if (xCOffset + 1 >= inDims[1]) {
|
|
xTexelC${y+1}.zw = vec2(0.0);
|
|
}
|
|
xTexelC${y+1}Ready = 1;
|
|
}
|
|
|
|
xC${y+1} = xTexelC${y+1};
|
|
`}}else y<p&&(i%2===1?(d+=`
|
|
xCOffset = xC + 1 - strides[1];
|
|
if(xCOffset >= 0 && xCOffset < inDims[1] && xTexelC${y}Ready == 0) {
|
|
xTexelC${y} = getX(batch, xR, xCOffset, d1);
|
|
// Need to manually clear unused channels in case
|
|
// we're reading from recycled texture.
|
|
if (xCOffset + 1 >= inDims[1]) {
|
|
xTexelC${y}.zw = vec2(0.0);
|
|
}
|
|
xTexelC${y}Ready = 1;
|
|
}
|
|
|
|
if(xC + 1 >= 0 && xC + 1 < inDims[1] && xTexelC${y+1}Ready == 0) {
|
|
xTexelC${y+1} = getX(batch, xR, xC + 1, d1);
|
|
// Need to manually clear unused channels in case
|
|
// we're reading from recycled texture.
|
|
if (xC + 2 >= inDims[1]) {
|
|
xTexelC${y+1}.zw = vec2(0.0);
|
|
}
|
|
xTexelC${y+1}Ready = 1;
|
|
}
|
|
|
|
xC${y} = vec4(xTexelC${y}.zw, xTexelC${y+1}.zw);
|
|
`,y+1<p&&(d+=`
|
|
final = vec4(0.0);
|
|
xCOffset = xC + 1 + strides[1];
|
|
if(xCOffset >= 0 && xCOffset < inDims[1]) {
|
|
final = getX(batch, xR, xCOffset, d1);
|
|
}
|
|
xC${y+1} = vec4(xTexelC${y+1}.xy, final.xy);
|
|
`)):(d+=`
|
|
if(xC >= 0 && xC < inDims[1] && xTexelC${y}Ready == 0) {
|
|
xTexelC${y} = getX(batch, xR, xC, d1);
|
|
if (xC + 1 >= inDims[1]) {
|
|
xTexelC${y}.zw = vec2(0.0);
|
|
}
|
|
xTexelC${y}Ready = 1;
|
|
}
|
|
|
|
xCOffset = xC + strides[1];
|
|
if(xCOffset >= 0 && xCOffset < inDims[1] && xTexelC${y+1}Ready == 0) {
|
|
xTexelC${y+1} = getX(batch, xR, xCOffset, d1);
|
|
if (xCOffset + 1 >= inDims[1]) {
|
|
xTexelC${y+1}.zw = vec2(0.);
|
|
}
|
|
xTexelC${y+1}Ready = 1;
|
|
}
|
|
|
|
xC${y} = vec4(
|
|
xTexelC${y}.xy, xTexelC${y+1}.xy);
|
|
`,y+1<p&&(d+=`
|
|
xC${y+1} = vec4(xTexelC${y}.zw, xTexelC${y+1}.zw);
|
|
`)));y<p&&(d+=`
|
|
wTexel = getW(r, ${y}, d1, q);
|
|
dotProd += xC${y} * vec4(wTexel.xz, wTexel.xz);
|
|
`,y+1<p&&(d+=`
|
|
wTexel = getW(r, ${y+1}, d1, q);
|
|
dotProd += xC${y+1} * vec4(wTexel.xz, wTexel.xz);
|
|
`))}d+=`
|
|
}
|
|
`,d+=`
|
|
}
|
|
`;let h="",f="";a&&(n?h=`vec4 activation(vec4 a) {
|
|
vec4 b = getPreluActivationWeightsAtOutCoords();
|
|
${a}
|
|
}`:r?h=`vec4 activation(vec4 a) {
|
|
vec4 b = getLeakyreluAlphaAtOutCoords();
|
|
${a}
|
|
}`:h=`vec4 activation(vec4 x) {
|
|
${a}
|
|
}`,f="result = activation(result);");let m=t?"result += getBiasAtOutCoords();":"";t&&this.variableNames.push("bias"),n&&this.variableNames.push("preluActivationWeights"),r&&this.variableNames.push("leakyreluAlpha"),this.userCode=`
|
|
${h}
|
|
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int batch = coords.x;
|
|
ivec2 xRCCorner = coords.yz * strides - pads;
|
|
int d2 = coords.w;
|
|
int d1 = d2 / ${s};
|
|
int q = d2 - d1 * ${s};
|
|
int xRCorner = xRCCorner.x;
|
|
int xCCorner = xRCCorner.y;
|
|
|
|
//intialize dotProd with a small epsilon seems to reduce GPU accuracy loss.
|
|
vec4 dotProd = vec4(0.000000000000001);
|
|
|
|
${d}
|
|
|
|
vec4 result = dotProd - vec4(0.000000000000001);
|
|
${m}
|
|
${f}
|
|
setOutput(result);
|
|
}
|
|
`}};function WX(e){let{inputs:t,backend:a,attrs:n}=e,{x:r,filter:s}=t,{strides:i,pad:o,dilations:l,dimRoundingMode:u}=n,p=l;p==null&&(p=[1,1]),v.assert(S.eitherStridesOrDilationsAreOne(i,p),()=>`Error in depthwiseConv2d: Either strides or dilations must be 1. Got strides ${i} and dilations '${p}'`);let c=S.computeConv2DInfo(r.shape,s.shape,i,p,o,u,!0),d;W().getBool("WEBGL_PACK_DEPTHWISECONV")&&c.strideWidth<=2&&c.outChannels/c.inChannels===1?d=new o8(c):d=new i8(c);let h=[[c.padInfo.top,c.padInfo.left],[c.strideHeight,c.strideWidth],[c.dilationHeight,c.dilationWidth],[c.inHeight,c.inWidth]];return a.runWebGLProgram(d,[r,s],"float32",h)}var BX={kernelName:Ii,backendName:"webgl",kernelFunc:WX},VX=class{constructor(e){this.variableNames=["x","dy"],this.outputShape=e.filterShape;let t=e.strideHeight,a=e.strideWidth,n=e.padInfo.top,r=e.padInfo.left,s=e.outChannels/e.inChannels;this.userCode=`
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int wR = coords.x;
|
|
int wC = coords.y;
|
|
int d1 = coords.z;
|
|
int dm = coords.w;
|
|
int d2 = d1 * ${s} + dm;
|
|
|
|
float dotProd = 0.0;
|
|
|
|
// TO DO: Vec4 over the batch size
|
|
for (int b = 0; b < ${e.batchSize}; b++) {
|
|
for (int yR = 0; yR < ${e.outHeight}; yR++) {
|
|
int xR = wR + yR * ${t} - ${n};
|
|
|
|
if (xR < 0 || xR >= ${e.inHeight}) {
|
|
continue;
|
|
}
|
|
|
|
for (int yC = 0; yC < ${e.outWidth}; yC++) {
|
|
int xC = wC + yC * ${a} - ${r};
|
|
|
|
if (xC < 0 || xC >= ${e.inWidth}) {
|
|
continue;
|
|
}
|
|
|
|
float dyValue = getDy(b, yR, yC, d2);
|
|
float xValue = getX(b, xR, xC, d1);
|
|
dotProd += (xValue * dyValue);
|
|
}
|
|
}
|
|
}
|
|
setOutput(dotProd);
|
|
}
|
|
`}},UX=class{constructor(e){this.variableNames=["dy","W"],this.outputShape=e.inShape;let t=e.filterHeight,a=e.filterWidth,n=e.strideHeight,r=e.strideWidth,s=t-1-e.padInfo.top,i=a-1-e.padInfo.left,o=e.outChannels/e.inChannels;this.userCode=`
|
|
const ivec2 pads = ivec2(${s}, ${i});
|
|
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int batch = coords[0];
|
|
int d1 = coords[3];
|
|
ivec2 dyCorner = coords.yz - pads;
|
|
int dyRCorner = dyCorner.x;
|
|
int dyCCorner = dyCorner.y;
|
|
|
|
float dotProd = 0.0;
|
|
|
|
for (int wR = 0; wR < ${t}; wR++) {
|
|
float dyR = float(dyRCorner + wR) / ${n}.0;
|
|
|
|
if (dyR < 0.0 || dyR >= ${e.outHeight}.0 || fract(dyR) > 0.0) {
|
|
continue;
|
|
}
|
|
int idyR = int(dyR);
|
|
|
|
int wRPerm = ${t} - 1 - wR;
|
|
|
|
for (int wC = 0; wC < ${a}; wC++) {
|
|
float dyC = float(dyCCorner + wC) / ${r}.0;
|
|
|
|
if (dyC < 0.0 || dyC >= ${e.outWidth}.0 ||
|
|
fract(dyC) > 0.0) {
|
|
continue;
|
|
}
|
|
int idyC = int(dyC);
|
|
|
|
int wCPerm = ${a} - 1 - wC;
|
|
|
|
// TO DO: Vec4 over the channelMul
|
|
for (int dm = 0; dm < ${o}; dm++) {
|
|
int d2 = d1 * ${o} + dm;
|
|
float xValue = getDy(batch, idyR, idyC, d2);
|
|
float wValue = getW(wRPerm, wCPerm, d1, dm);
|
|
dotProd += xValue * wValue;
|
|
}
|
|
}
|
|
}
|
|
setOutput(dotProd);
|
|
}
|
|
`}};function GX(e){let{inputs:t,backend:a,attrs:n}=e,{x:r,dy:s}=t,{strides:i,dilations:o,pad:l,dimRoundingMode:u,filterShape:p}=n,c=S.computeConv2DInfo(r.shape,p,i,o,l,u,!0),d=new VX(c);return a.runWebGLProgram(d,[r,s],"float32")}var HX={kernelName:rp,backendName:"webgl",kernelFunc:GX};function jX(e){let{inputs:t,backend:a,attrs:n}=e,{dy:r,filter:s}=t,{strides:i,dilations:o,pad:l,dimRoundingMode:u,inputShape:p}=n,c=S.computeConv2DInfo(p,s.shape,i,o,l,u,!0),d=new UX(c);return a.runWebGLProgram(d,[r,s],"float32")}var qX={kernelName:sp,backendName:"webgl",kernelFunc:jX},XX=class{constructor(e){this.variableNames=["X"],this.outputShape=[e,e],this.userCode=`
|
|
void main() {
|
|
ivec2 coords = getOutputCoords();
|
|
float val = coords[0] == coords[1] ? getX(coords[0]) : 0.0;
|
|
setOutput(val);
|
|
}
|
|
`}};function KX(e){let{inputs:t,backend:a}=e,{x:n}=t,r=[...n.shape,...n.shape],s=v.sizeFromShape(n.shape),i=ue({inputs:{x:n},backend:a,attrs:{shape:[s]}}),o=new XX(s),l=a.runWebGLProgram(o,[i],i.dtype),u=ue({inputs:{x:l},backend:a,attrs:{shape:r}});return a.disposeIntermediateTensorInfo(i),a.disposeIntermediateTensorInfo(l),u}var YX={kernelName:iu,backendName:"webgl",kernelFunc:KX},ZX=class{constructor(e){this.variableNames=["x","W"],this.outputShape=e.outShape;let{inHeight:t,inWidth:a,padInfo:n,strideHeight:r,strideWidth:s,filterHeight:i,filterWidth:o,dilationHeight:l,dilationWidth:u}=e,{top:p,left:c}=n;this.userCode=`
|
|
const ivec2 strides = ivec2(${r}, ${s});
|
|
const ivec2 pads = ivec2(${p}, ${c});
|
|
const float neg_infinity = -3.4e38;
|
|
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int batch = coords.x;
|
|
int d1 = coords.w;
|
|
ivec2 outTopLeftCorner =
|
|
coords.yz * strides - pads;
|
|
int hBeg = outTopLeftCorner.x;
|
|
int wBeg = outTopLeftCorner.y;
|
|
|
|
float curVal = neg_infinity;
|
|
for (int h = 0; h < ${i}; h++) {
|
|
int hIn = hBeg + h * ${l};
|
|
|
|
if (hIn >= 0 && hIn < ${t}) {
|
|
for (int w = 0; w < ${o}; w++) {
|
|
int wIn = wBeg + w * ${u};
|
|
|
|
if (wIn >= 0 && wIn < ${a}) {
|
|
float xVal = getX(batch, hIn, wIn, d1);
|
|
float wVal = getW(h, w, d1);
|
|
|
|
float val = xVal + wVal;
|
|
if (val > curVal) {
|
|
curVal = val;
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
float result = curVal;
|
|
setOutput(result);
|
|
}
|
|
`}};function JX(e){let{inputs:t,backend:a,attrs:n}=e,{x:r,filter:s}=t,{strides:i,pad:o,dilations:l}=n,u=S.computeDilation2DInfo(r.shape,s.shape,i,o,"NHWC",l),p,c=new ZX(u);p=a.runWebGLProgram(c,[r,s],"float32");let d=ue({inputs:{x:p},backend:a,attrs:{shape:u.outShape}});return a.disposeIntermediateTensorInfo(p),d}var QX={kernelName:Si,backendName:"webgl",kernelFunc:JX};function eK(e){let{inputs:t,backend:a,attrs:n}=e,{equation:r}=n,s=t,{allDims:i,summedDims:o,idDims:l}=S.decodeEinsumEquation(r,s.length);S.checkEinsumDimSizes(i.length,l,s);let{path:u,steps:p}=S.getEinsumComputePath(o,l),c=p.length,d=null,h=i.length,f=[];for(let m=0;m<c;++m){for(let g of p[m]){let{permutationIndices:y,expandDims:x}=S.getEinsumPermutation(h,l[g]),A;S.isIdentityPermutation(y)?A=s[g]:(A=Ta({inputs:{x:s[g]},backend:a,attrs:{perm:y}}),f.push(A));let b=A.shape.slice();for(let w=0;w<x.length;++w)b.splice(x[w],0,1);v.arraysEqual(A.shape,b)||(A=ue({inputs:{x:A},backend:a,attrs:{shape:b}}),f.push(A)),d===null?d=A:(d=Ig({inputs:{a:A,b:d},backend:a}),f.push(d))}m<c-1&&(u[m]>=0&&(d=e0({inputs:{x:d},backend:a,attrs:{axis:u[m]-(i.length-h),keepDims:!1}}),f.push(d)),h--)}for(let m of f)m!==d&&a.disposeIntermediateTensorInfo(m);return d}var tK={kernelName:ip,backendName:"webgl",kernelFunc:eK},aK="return (x >= 0.0) ? x : (exp(x) - 1.0);",nK=`
|
|
vec4 result;
|
|
|
|
result.r = (x.r >= 0.0) ? x.r : (exp(x.r) - 1.0);
|
|
result.g = (x.g >= 0.0) ? x.g : (exp(x.g) - 1.0);
|
|
result.b = (x.b >= 0.0) ? x.b : (exp(x.b) - 1.0);
|
|
result.a = (x.a >= 0.0) ? x.a : (exp(x.a) - 1.0);
|
|
|
|
return result;
|
|
`,rK=et({opSnippet:aK,packedOpSnippet:nK}),sK={kernelName:Ti,backendName:"webgl",kernelFunc:rK},iK="return (b >= 0.0) ? a : a * (b + 1.0);",oK=`
|
|
vec4 bGTEZero = vec4(greaterThanEqual(b, vec4(0.)));
|
|
return (bGTEZero * a) + ((vec4(1.0) - bGTEZero) * (a * (b + vec4(1.0))));
|
|
`,lK=e=>{let{inputs:t,backend:a}=e,{dy:n,y:r}=t,s=W().getBool("WEBGL_PACK_BINARY_OPERATIONS")?new Up(oK,n.shape,r.shape):new ql(iK,n.shape,r.shape);return a.runWebGLProgram(s,[n,r],n.dtype)},uK={kernelName:ou,backendName:"webgl",kernelFunc:lK},dK=`
|
|
return vec4(equal(a, b));
|
|
`,pK="return float(a == b);",cK=fa({opSnippet:pK,packedOpSnippet:dK,dtype:"bool",cpuKernelImpl:jG}),hK={kernelName:Ni,backendName:"webgl",kernelFunc:cK},fK=`
|
|
// Error function is calculated approximately with elementary function.
|
|
// See "Handbook of Mathematical Functions with Formulas,
|
|
// Graphs, and Mathematical Tables", Abramowitz and Stegun.
|
|
float p = ${S.ERF_P};
|
|
float a1 = ${S.ERF_A1};
|
|
float a2 = ${S.ERF_A2};
|
|
float a3 = ${S.ERF_A3};
|
|
float a4 = ${S.ERF_A4};
|
|
float a5 = ${S.ERF_A5};
|
|
|
|
float sign = sign(x);
|
|
x = abs(x);
|
|
float t = 1.0 / (1.0 + p * x);
|
|
return sign * (1.0 - (((((a5*t + a4)*t) + a3)*t + a2)*t + a1)*t*exp(-x*x));
|
|
`,mK=et({opSnippet:fK}),gK={kernelName:lu,backendName:"webgl",kernelFunc:mK},yK=Vu+`
|
|
return exp(x);
|
|
`,xK=`
|
|
vec4 result = exp(x);
|
|
bvec4 isNaN = isnan(x);
|
|
result.r = isNaN.r ? x.r : result.r;
|
|
result.g = isNaN.g ? x.g : result.g;
|
|
result.b = isNaN.b ? x.b : result.b;
|
|
result.a = isNaN.a ? x.a : result.a;
|
|
|
|
return result;
|
|
`,l8=et({opSnippet:yK,packedOpSnippet:xK,cpuKernelImpl:qG,dtype:"float32"}),AK={kernelName:Ri,backendName:"webgl",kernelFunc:l8};function T1(e){let{inputs:t,attrs:a,backend:n}=e,{dim:r}=a,{input:s}=t,i=s.shape.length,o=s.shape.slice(),l=r;return r<0&&(v.assert(-(i+1)<=r,()=>`Axis must be in the interval [${-(i+1)}, ${i}]`),l=i+r+1),o.splice(l,0,1),ue({inputs:{x:s},backend:n,attrs:{shape:o}})}var bK={kernelName:uu,backendName:"webgl",kernelFunc:T1},bx="return exp(x) - 1.0;",vK=et({opSnippet:bx,packedOpSnippet:bx,cpuKernelImpl:XG}),wK={kernelName:Ei,backendName:"webgl",kernelFunc:vK},vx=class{constructor(e,t,a){this.variableNames=["real","imag"];let n=t[1];this.outputShape=t;let r=a?`2.0 * ${Math.PI}`:`-2.0 * ${Math.PI}`,s=a?`${n}.0`:"1.0",i;if(e==="real")i="return real * expR - imag * expI;";else if(e==="imag")i="return real * expI + imag * expR;";else throw new Error(`FFT component must be either "real" or "imag", got ${e}.`);this.userCode=`
|
|
const float exponentMultiplier = ${r};
|
|
|
|
float unaryOpComplex(float real, float expR, float imag, float expI) {
|
|
${i}
|
|
}
|
|
|
|
float mulMatDFT(int batch, int index) {
|
|
float indexRatio = float(index) / float(${n});
|
|
float exponentMultiplierTimesIndexRatio =
|
|
exponentMultiplier * indexRatio;
|
|
|
|
float result = 0.0;
|
|
|
|
for (int i = 0; i < ${n}; i++) {
|
|
// x = (-2|2 * PI / N) * index * i;
|
|
float x = exponentMultiplierTimesIndexRatio * float(i);
|
|
float expR = cos(x);
|
|
float expI = sin(x);
|
|
float real = getReal(batch, i);
|
|
float imag = getImag(batch, i);
|
|
|
|
result +=
|
|
unaryOpComplex(real, expR, imag, expI) / ${s};
|
|
}
|
|
|
|
return result;
|
|
}
|
|
|
|
void main() {
|
|
ivec2 coords = getOutputCoords();
|
|
setOutput(mulMatDFT(coords[0], coords[1]));
|
|
}
|
|
`}};function u8(e,t,a){let n=a.texData.get(e.dataId),r=v.sizeFromShape(e.shape),s=e.shape[e.shape.length-1],i=r/s,o=ue({inputs:{x:e},backend:a,attrs:{shape:[i,s]}}),l=o.shape,u=new vx("real",l,t),p=new vx("imag",l,t),c=[{dataId:n.complexTensorInfos.real.dataId,dtype:n.complexTensorInfos.real.dtype,shape:l},{dataId:n.complexTensorInfos.imag.dataId,dtype:n.complexTensorInfos.imag.dtype,shape:l}],d=a.runWebGLProgram(u,c,"float32"),h=a.runWebGLProgram(p,c,"float32"),f=ds({inputs:{real:d,imag:h},backend:a});a.disposeIntermediateTensorInfo(d),a.disposeIntermediateTensorInfo(h);let m=ue({inputs:{x:f},backend:a,attrs:{shape:e.shape}});return a.disposeIntermediateTensorInfo(o),a.disposeIntermediateTensorInfo(f),m}function kK(e){let{inputs:t,backend:a}=e,{input:n}=t;return u8(n,!1,a)}var IK={kernelName:op,backendName:"webgl",kernelFunc:kK},SK=class{constructor(e,t){this.outputShape=[],this.customUniforms=[{name:"value",type:"float"}],this.variableNames=["x"],this.outputShape=e,this.userCode=`
|
|
void main() {
|
|
// Input can be obtained from uniform value.
|
|
setOutput(value);
|
|
}
|
|
`}};function Hp(e){let{backend:t,attrs:a}=e,{shape:n,value:r}=a,{dtype:s}=a;if(s=s||v.inferDtype(r),s==="string"){let i=v.getArrayFromDType(s,v.sizeFromShape(n));return i.fill(r),t.makeTensorInfo(n,s,i)}else{let i=new SK(n,r),o=[[r]];return t.runWebGLProgram(i,[],s,o)}}var CK={kernelName:du,backendName:"webgl",kernelFunc:Hp},TK=class{constructor(e){this.variableNames=["Image"],this.outputShape=[];let t=e[2];this.outputShape=e,this.userCode=`
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int x = coords[2];
|
|
|
|
int coordX = ${t} - x - 1;
|
|
float outputValue;
|
|
if(coordX >= 0 && coordX < ${t}) {
|
|
outputValue = getImage(coords[0], coords[1], coordX, coords[3]);
|
|
} else {
|
|
outputValue = getImage(coords[0], coords[1], coords[2], coords[3]);
|
|
}
|
|
setOutput(outputValue);
|
|
}
|
|
`}},NK={kernelName:Mi,backendName:"webgl",kernelFunc:({inputs:e,backend:t})=>{let{image:a}=e,n=t,r=new TK(a.shape);return n.runWebGLProgram(r,[a],a.dtype)}},wx="return floor(x);",RK=et({opSnippet:wx,packedOpSnippet:wx,cpuKernelImpl:KG}),EK={kernelName:Fi,backendName:"webgl",kernelFunc:RK},MK=`
|
|
float s = sign(a) * sign(b);
|
|
int ia = round(a);
|
|
int ib = round(b);
|
|
if (ib != 0) {
|
|
// Windows (D3D) wants guaranteed non-zero int division at compile-time.
|
|
return float(idiv(ia, ib, s));
|
|
} else {
|
|
return NAN;
|
|
}
|
|
`,FK=`
|
|
ivec4 ia = round(a);
|
|
ivec4 ib = round(b);
|
|
bvec4 cond = notEqual(ib, ivec4(0));
|
|
ivec4 result = ivec4(0);
|
|
vec4 s = sign(a) * sign(b);
|
|
|
|
// Windows (D3D) wants guaranteed non-zero int division at compile-time.
|
|
if (cond[0]) {
|
|
result[0] = idiv(ia[0], ib[0], s[0]);
|
|
}
|
|
if (cond[1]) {
|
|
result[1] = idiv(ia[1], ib[1], s[1]);
|
|
}
|
|
if (cond[2]) {
|
|
result[2] = idiv(ia[2], ib[2], s[2]);
|
|
}
|
|
if (cond[3]) {
|
|
result[3] = idiv(ia[3], ib[3], s[3]);
|
|
}
|
|
return vec4(result);
|
|
`,_K=fa({opSnippet:MK,packedOpSnippet:FK,dtype:"int32"}),$K={kernelName:_i,backendName:"webgl",kernelFunc:_K},DK=class{constructor(e){this.variableNames=["A"];let t=Ea(),[a,n]=e;this.outputShape=e,this.userCode=`
|
|
void main() {
|
|
ivec3 coords = getOutputCoords();
|
|
int texR = coords[0];
|
|
int texC = coords[1];
|
|
int depth = coords[2];
|
|
vec2 uv = (vec2(texC, texR) + halfCR) / vec2(${n}.0, ${a}.0);
|
|
|
|
vec4 values = ${t.texture2D}(A, uv);
|
|
float value;
|
|
if (depth == 0) {
|
|
value = values.r;
|
|
} else if (depth == 1) {
|
|
value = values.g;
|
|
} else if (depth == 2) {
|
|
value = values.b;
|
|
} else if (depth == 3) {
|
|
value = values.a;
|
|
}
|
|
|
|
setOutput(floor(value * 255.0 + 0.5));
|
|
}
|
|
`}},PK=class{constructor(e){this.variableNames=["A"],this.packedInputs=!1,this.packedOutput=!0;let t=Ea(),[a,n]=e;this.outputShape=e,this.userCode=`
|
|
void main() {
|
|
ivec3 coords = getOutputCoords();
|
|
int texR = coords[0];
|
|
int texC = coords[1];
|
|
int depth = coords[2];
|
|
|
|
vec4 result = vec4(0.);
|
|
|
|
for(int row=0; row<=1; row++) {
|
|
for(int col=0; col<=1; col++) {
|
|
texC = coords[1] + row;
|
|
depth = coords[2] + col;
|
|
|
|
vec2 uv = (vec2(texC, texR) + halfCR) /
|
|
vec2(${n}.0, ${a}.0);
|
|
vec4 values = ${t.texture2D}(A, uv);
|
|
float value;
|
|
if (depth == 0) {
|
|
value = values.r;
|
|
} else if (depth == 1) {
|
|
value = values.g;
|
|
} else if (depth == 2) {
|
|
value = values.b;
|
|
} else if (depth == 3) {
|
|
value = values.a;
|
|
}
|
|
|
|
result[row * 2 + col] = floor(value * 255.0 + 0.5);
|
|
}
|
|
}
|
|
|
|
${t.output} = result;
|
|
}
|
|
`}},OK={kernelName:Ed,backendName:"webgl",kernelFunc:zK},kl,B2=W().getBool("CANVAS2D_WILL_READ_FREQUENTLY_FOR_GPU");function zK(e){let{inputs:t,backend:a,attrs:n}=e,{pixels:r}=t,{numChannels:s}=n,i=typeof HTMLVideoElement!="undefined"&&r instanceof HTMLVideoElement,o=typeof HTMLImageElement!="undefined"&&r instanceof HTMLImageElement,[l,u]=i?[r.videoWidth,r.videoHeight]:[r.width,r.height],p=[u,l],c=[u,l,s];if(o||i){let m=W().getBool("CANVAS2D_WILL_READ_FREQUENTLY_FOR_GPU");(kl==null||m!==B2)&&(B2=m,kl=document.createElement("canvas").getContext("2d",{willReadFrequently:B2})),kl.canvas.width=l,kl.canvas.height=u,kl.drawImage(r,0,0,l,u),r=kl.canvas}let d=a.makeTensorInfo(p,"int32");a.texData.get(d.dataId).usage=fn.PIXELS,a.gpgpu.uploadPixelDataToTexture(a.getTexture(d.dataId),r);let h=W().getBool("WEBGL_PACK")?new PK(c):new DK(c),f=a.runWebGLProgram(h,[d],"int32");return a.disposeData(d.dataId),f}function LK(e){let{inputs:t,backend:a,attrs:n}=e,{x:r,filter:s,bias:i,preluActivationWeights:o}=t,{strides:l,pad:u,dataFormat:p,dilations:c,dimRoundingMode:d,activation:h,leakyreluAlpha:f}=n,m=S.convertConv2DDataFormat(p),g=S.computeConv2DInfo(r.shape,s.shape,l,c,u,d,!1,m),y,x=[],A=i!=null,b=o!=null,w=h==="leakyrelu",C=()=>{let N=[r,s],M=(F,E)=>{if(E==="NCHW"&&F.shape.length===1&&F.shape[0]!==1){let I=ue({inputs:{x:F},backend:a,attrs:{shape:[F.shape[0],1,1]}});return x.push(I),I}return F};if(A&&N.push(M(i,p)),b&&N.push(M(o,p)),w){let F=a.makeTensorInfo([],"float32",v.createScalarValue(f,"float32"));N.push(F),x.push(F)}return N};if(g.filterHeight===1&&g.filterWidth===1&&g.dilationHeight===1&&g.dilationWidth===1&&g.strideHeight===1&&g.strideWidth===1&&(g.padInfo.type==="SAME"||g.padInfo.type==="VALID"))y=n8({x:r,filter:s,convInfo:g,backend:a,bias:i,activation:h,preluActivationWeights:o,leakyreluAlpha:f});else if(g.strideWidth<=2&&m==="channelsLast"&&W().getBool("WEBGL_EXP_CONV")){let N=h?Hd(h,!0):null,M=new a8(g,A,N,b,w),F=[[g.padInfo.top,g.padInfo.left],[g.strideHeight,g.strideWidth],[g.dilationHeight,g.dilationWidth],[g.inHeight,g.inWidth]],E=C();y=a.runWebGLProgram(M,E,"float32",F)}else if(W().getBool("WEBGL_CONV_IM2COL"))y=r8({x:r,filter:s,convInfo:g,backend:a,bias:i,activation:h,preluActivationWeights:o,leakyreluAlpha:f});else{let N=h?Hd(h,!1):null,M=new t8(g,A,N,b,w),F=C();y=a.runWebGLProgram(M,F,"float32")}let T=ue({inputs:{x:y},backend:a,attrs:{shape:g.outShape}});return x.push(y),x.forEach(N=>a.disposeIntermediateTensorInfo(N)),T}var WK={kernelName:Kr,backendName:"webgl",kernelFunc:LK};function BK(e){let{inputs:t,backend:a,attrs:n}=e,{x:r,filter:s,bias:i,preluActivationWeights:o}=t,{strides:l,pad:u,dilations:p,dimRoundingMode:c,activation:d,leakyreluAlpha:h}=n,f=[],m=p;m==null&&(m=[1,1]),v.assert(S.eitherStridesOrDilationsAreOne(l,m),()=>`Error in depthwiseConv2d: Either strides or dilations must be 1. Got strides ${l} and dilations '${m}'`);let g=S.computeConv2DInfo(r.shape,s.shape,l,m,u,c,!0),y=W().getBool("WEBGL_PACK_DEPTHWISECONV")&&g.strideWidth<=2&&g.outChannels/g.inChannels===1,x=d?Hd(d,y):null,A=[r,s],b=i!=null,w=o!=null,C=d==="leakyrelu";if(b&&A.push(i),w&&A.push(o),C){let F=a.makeTensorInfo([],"float32",v.createScalarValue(h,"float32"));A.push(F),f.push(F)}let T;y?T=new o8(g,b,x,w,C):T=new i8(g,b,x,w,C);let N=[[g.padInfo.top,g.padInfo.left],[g.strideHeight,g.strideWidth],[g.dilationHeight,g.dilationWidth],[g.inHeight,g.inWidth]],M=a.runWebGLProgram(T,A,"float32",N);return f.forEach(F=>a.disposeIntermediateTensorInfo(F)),M}var VK={kernelName:Yr,backendName:"webgl",kernelFunc:BK},UK=class{constructor(e,t,a,n){this.sliceDim=e,this.strides=t,this.paramsShape=n,this.variableNames=["x","indices"],this.outputShape=a;let r=ft(a.length),s=`
|
|
int index;`;for(let i=0;i<this.sliceDim;i++)s+=`
|
|
index = round(getIndices(coords[0], ${i}));
|
|
out_of_bounds = out_of_bounds || index < 0;
|
|
out_of_bounds = out_of_bounds || index >= ${this.paramsShape[i]};
|
|
flattenIndex += index * ${this.strides[i]};`;this.userCode=`
|
|
void main() {
|
|
${r} coords = getOutputCoords();
|
|
int flattenIndex = 0;
|
|
bool out_of_bounds = false;
|
|
|
|
${s}
|
|
|
|
setOutput(out_of_bounds ? 0.0 : getX(flattenIndex, coords[1]));
|
|
}
|
|
`}};function GK(e){let{inputs:t,backend:a}=e,{params:n,indices:r}=t,s=r.shape,i=s[s.length-1],o=v.sizeFromShape(n.shape),[l,u,p,c]=S.prepareAndValidate(n,r),d=ue({inputs:{x:r},backend:a,attrs:{shape:[u,i]}}),h=ue({inputs:{x:n},backend:a,attrs:{shape:[v.sizeFromShape(n.shape)/p,p]}});if(a.shouldExecuteOnCPU([n,r])||n.dtype==="string"){let y=a.readSync(r.dataId),x=a.bufferSync(n),A=YG(y,x,n.dtype,u,i,p,c,n.shape,o);return a.makeTensorInfo(l,n.dtype,A.values)}let f=new UK(i,c,[u,p],n.shape),m=a.runWebGLProgram(f,[h,d],h.dtype),g=ue({inputs:{x:m},backend:a,attrs:{shape:l}});return a.disposeIntermediateTensorInfo(d),a.disposeIntermediateTensorInfo(h),a.disposeIntermediateTensorInfo(m),g}var HK={kernelName:Di,backendName:"webgl",kernelFunc:GK},jK=class{constructor(e,t){this.variableNames=["A","indices"],this.outputShape=t,this.rank=t.length;let a=ft(this.rank),n=qK(e,2);this.userCode=`
|
|
void main() {
|
|
${a} resRC = getOutputCoords();
|
|
int index = int(getIndices(resRC.x, resRC.z));
|
|
float inBounds = (index >= 0) && (index < ${e[2]}) ? 1.0 : 0.0;
|
|
setOutput(inBounds * getA(${n}));
|
|
}
|
|
`}};function qK(e,t){let a=["resRC.x","resRC.y","resRC.z","resRC.w"],n=[];for(let r=0;r<e.length;r++)r===2?n.push("index"):n.push(`${a[r]}`);return n.join()}function d8(e){let{inputs:t,backend:a,attrs:n}=e,{x:r,indices:s}=t,{axis:i,batchDims:o}=n,l=v.parseAxisParam(i,r.shape)[0];if(W().get("DEBUG")){let x=a.readSync(s.dataId),A=r.shape[l];for(let b=0;b<x.length;++b){let w=x[b];v.assert(w<=A-1&&w>=0,()=>`GatherV2: the index value ${w} is not in [0, ${A-1}]`)}}let u=S.segment_util.collectGatherOpShapeInfo(r,s,l,o),p=v.sizeFromShape(s.shape),c=[],d=ue({inputs:{x:r},backend:a,attrs:{shape:[u.batchSize,u.outerSize,u.dimSize,u.sliceSize]}}),h=ue({inputs:{x:s},backend:a,attrs:{shape:[u.batchSize,p/u.batchSize]}});c.push(d),c.push(h);let f=[u.batchSize,u.outerSize,p/u.batchSize,u.sliceSize];if(a.shouldExecuteOnCPU([r,s])||r.dtype==="string"){let x=a.bufferSync(h),A=a.bufferSync(d),b=ZG(A,x,f);return c.forEach(w=>a.disposeIntermediateTensorInfo(w)),a.makeTensorInfo(u.outputShape,b.dtype,b.values)}let m=new jK(d.shape,f),g=a.runWebGLProgram(m,[d,h],d.dtype);c.push(g);let y=ue({inputs:{x:g},backend:a,attrs:{shape:u.outputShape}});return c.forEach(x=>a.disposeIntermediateTensorInfo(x)),y}var XK={kernelName:pu,backendName:"webgl",kernelFunc:d8},KK="return float(a > b);",YK=`
|
|
return vec4(greaterThan(a, b));
|
|
`,ZK=fa({opSnippet:KK,packedOpSnippet:YK,cpuKernelImpl:JG,dtype:"bool"}),JK={kernelName:Pi,backendName:"webgl",kernelFunc:ZK},QK="return float(a >= b);",eY=`
|
|
return vec4(greaterThanEqual(a, b));
|
|
`,tY=fa({opSnippet:QK,packedOpSnippet:eY,dtype:"bool",cpuKernelImpl:QG}),aY={kernelName:Oi,backendName:"webgl",kernelFunc:tY};function nY(e){let{inputs:t,backend:a}=e,{input:n}=t;return u8(n,!0,a)}var rY={kernelName:lp,backendName:"webgl",kernelFunc:nY},sY="return float(!isnan(x) && !isinf(x));",iY=et({opSnippet:sY,dtype:"bool"}),oY={kernelName:Li,backendName:"webgl",kernelFunc:iY},lY="return float(isinf(x));",uY=et({opSnippet:lY,dtype:"bool"}),dY={kernelName:Wi,backendName:"webgl",kernelFunc:uY},pY="return float(isnan(x));",cY=et({opSnippet:pY,dtype:"bool"}),hY={kernelName:Bi,backendName:"webgl",kernelFunc:cY},fY="return float(a < b);",mY=`
|
|
return vec4(lessThan(a, b));
|
|
`,gY=fa({opSnippet:fY,packedOpSnippet:mY,cpuKernelImpl:eH,dtype:"bool"}),yY={kernelName:Ui,backendName:"webgl",kernelFunc:gY},xY="return float(a <= b);",AY=`
|
|
return vec4(lessThanEqual(a, b));
|
|
`,bY=fa({opSnippet:xY,packedOpSnippet:AY,cpuKernelImpl:tH,dtype:"bool"}),vY={kernelName:Gi,backendName:"webgl",kernelFunc:bY};function wY(e){let{backend:t,attrs:a}=e,{start:n,stop:r,num:s}=a,i=aH(n,r,s);return t.makeTensorInfo([i.length],"float32",i)}var kY={kernelName:Hi,backendName:"webgl",kernelFunc:wY},IY=Vu+`
|
|
return x < 0.0 ? 0./0. : log(x);
|
|
`,SY=`
|
|
vec4 result = log(x);
|
|
bvec4 isNaN = isnan(x);
|
|
result.r = isNaN.r ? x.r : (x.r < 0.0 ? 0./0. : result.r);
|
|
result.g = isNaN.g ? x.g : (x.g < 0.0 ? 0./0. : result.g);
|
|
result.b = isNaN.b ? x.b : (x.b < 0.0 ? 0./0. : result.b);
|
|
result.a = isNaN.a ? x.a : (x.a < 0.0 ? 0./0. : result.a);
|
|
return result;
|
|
`,CY=et({opSnippet:IY,packedOpSnippet:SY,cpuKernelImpl:nH}),TY={kernelName:ji,backendName:"webgl",kernelFunc:CY},NY=Vu+`
|
|
return log(1.0 + x);
|
|
`,RY=et({opSnippet:NY}),EY={kernelName:qi,backendName:"webgl",kernelFunc:RY},MY="return float(a >= 1.0 && b >= 1.0);",FY=`
|
|
return vec4(
|
|
vec4(greaterThanEqual(a, vec4(1.0))) *
|
|
vec4(greaterThanEqual(b, vec4(1.0))));
|
|
`,_Y=fa({opSnippet:MY,packedOpSnippet:FY,dtype:"bool"}),$Y={kernelName:Xi,backendName:"webgl",kernelFunc:_Y},DY="return float(!(x >= 1.0));",PY=et({opSnippet:DY}),OY={kernelName:Ki,backendName:"webgl",kernelFunc:PY},zY="return float(a >= 1.0 || b >= 1.0);",LY=`
|
|
return min(
|
|
vec4(greaterThanEqual(a, vec4(1.0))) +
|
|
vec4(greaterThanEqual(b, vec4(1.0))),
|
|
vec4(1.0));
|
|
`,WY=fa({opSnippet:zY,packedOpSnippet:LY,dtype:"bool"}),BY={kernelName:Yi,backendName:"webgl",kernelFunc:WY},VY=class{constructor(e,t,a,n,r){this.variableNames=["x"],this.outputShape=[];let s=t,i=e[3]-1;this.outputShape=e;let o,l=`float(${a}) + float(${n}) * sum`;r===.5?o=`inversesqrt(${l})`:r===1?o=`1.0/(${l})`:o=`exp(log(${l}) * float(-${r}));`,this.userCode=`
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int b = coords[0];
|
|
int r = coords[1];
|
|
int c = coords[2];
|
|
int d = coords[3];
|
|
float x = getX(b, r, c, d);
|
|
float sum = 0.0;
|
|
for (int j = -${s}; j <= ${s}; j++) {
|
|
int idx = d + j;
|
|
if (idx >= 0 && idx <= ${i}) {
|
|
float z = getX(b, r, c, idx);
|
|
sum += z * z;
|
|
}
|
|
}
|
|
float val = x * ${o};
|
|
setOutput(val);
|
|
}
|
|
`}},UY=class{constructor(e,t,a,n,r){this.variableNames=["x"],this.outputShape=[],this.packedInputs=!0,this.packedOutput=!0;let s=t,i=e[3]-1;this.outputShape=e;let o,l=`float(${a}) + float(${n}) * sum`;r===.5?o=`inversesqrt(${l})`:r===1?o=`1.0/(${l})`:o=`exp(log(${l}) * float(-${r}));`,this.userCode=`
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int b = coords.x;
|
|
int r = coords.y;
|
|
int c = coords.z;
|
|
int d = coords.w;
|
|
|
|
bool hasNextCol = d < ${this.outputShape[3]};
|
|
bool hasNextRow = c < ${this.outputShape[2]};
|
|
|
|
vec4 sum = vec4(0.);
|
|
vec4 xFragAtOutputCoords = getX(b, r, c, d);
|
|
|
|
vec4 xAtOutputCoords = vec4(
|
|
getChannel(xFragAtOutputCoords, vec2(c, d)),
|
|
hasNextCol ?
|
|
getChannel(xFragAtOutputCoords, vec2(c, d + 1)) : 0.0,
|
|
hasNextRow ?
|
|
getChannel(xFragAtOutputCoords , vec2(c + 1, d)) : 0.0,
|
|
(hasNextRow && hasNextCol) ?
|
|
getChannel(xFragAtOutputCoords, vec2(c + 1, d + 1)) : 0.0
|
|
);
|
|
|
|
int firstChannel = d - ${s};
|
|
vec2 cache = vec2(0.);
|
|
if(firstChannel >= 0){
|
|
vec4 firstChannelFrag = getX(b, r, c, firstChannel);
|
|
cache.x = getChannel(firstChannelFrag, vec2(c, firstChannel));
|
|
if(hasNextRow){
|
|
cache.y = getChannel(firstChannelFrag, vec2(c + 1, firstChannel));
|
|
}
|
|
}
|
|
|
|
ivec2 depth = ivec2(d, d + 1);
|
|
for (int j = - ${s}; j <= ${s}; j++) {
|
|
ivec2 idx = depth + j;
|
|
bvec2 aboveLowerBound = greaterThanEqual(idx, ivec2(0));
|
|
bvec2 belowUpperBound = lessThanEqual(idx, ivec2(${i}));
|
|
|
|
bool depthInRange = aboveLowerBound.x && belowUpperBound.x;
|
|
bool depthPlusOneInRange = aboveLowerBound.y && belowUpperBound.y;
|
|
|
|
if(depthInRange || depthPlusOneInRange){
|
|
vec4 z = vec4(0.);
|
|
vec4 xFragAtCurrentDepth;
|
|
z.xz = cache.xy;
|
|
if(depthPlusOneInRange && hasNextCol){
|
|
xFragAtCurrentDepth = idx.y != d ?
|
|
getX(b, r, c, idx.y) : xFragAtOutputCoords;
|
|
z.y = getChannel(xFragAtCurrentDepth, vec2(c, idx.y));
|
|
if(hasNextRow){
|
|
z.w = getChannel(xFragAtCurrentDepth, vec2(c + 1, idx.y));
|
|
}
|
|
}
|
|
cache.xy = z.yw;
|
|
sum += z * z;
|
|
}
|
|
}
|
|
vec4 result = xAtOutputCoords * ${o};
|
|
setOutput(result);
|
|
}
|
|
`}},GY=e=>{let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{depthRadius:s,bias:i,alpha:o,beta:l}=n,u=W().getBool("WEBGL_PACK_NORMALIZATION")?new UY(r.shape,s,i,o,l):new VY(r.shape,s,i,o,l);return a.runWebGLProgram(u,[r],r.dtype)},HY={kernelName:Zi,backendName:"webgl",kernelFunc:GY},jY=class{constructor(e,t,a,n,r){this.variableNames=["inputImage","outputImage","dy"],this.outputShape=[],this.outputShape=e,this.depth=e[3],this.depthRadius=t,this.bias=a,this.alpha=n,this.beta=r,this.userCode=`
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int b = coords[0];
|
|
int r = coords[1];
|
|
int c = coords[2];
|
|
|
|
float result = 0.0;
|
|
for (int d = 0; d < ${this.depth}; ++d) {
|
|
int depthBegin = int(max(0.0, float(d - ${t})));
|
|
int depthEnd = int(min(float(${this.depth}),
|
|
float(d + ${t} + 1)));
|
|
|
|
const int MIN_DEPTH_BEGIN = 0;
|
|
const int MAX_DEPTH_END = ${this.depth};
|
|
|
|
float norm = 0.0;
|
|
for (int k = MIN_DEPTH_BEGIN; k < MAX_DEPTH_END; ++k) {
|
|
if (k < depthBegin){
|
|
continue;
|
|
}
|
|
else if (k >= depthBegin && k < depthEnd) {
|
|
norm += getInputImage(b, r, c, k) * getInputImage(b, r, c, k);
|
|
}
|
|
else {
|
|
break;
|
|
}
|
|
}
|
|
|
|
norm = float(${n}) * norm + float(${a});
|
|
|
|
for(int k = MIN_DEPTH_BEGIN; k < MAX_DEPTH_END; ++k){
|
|
if (k < depthBegin){
|
|
continue;
|
|
}
|
|
else if (k >= depthBegin && k < depthEnd){
|
|
float dyi = -2.0 * float(${n})
|
|
* float(${r})
|
|
* getInputImage(b, r, c, k) * getOutputImage(b, r, c, d)
|
|
/ norm;
|
|
if (k == d) {
|
|
dyi += pow(norm, -1.0 * ${r});
|
|
}
|
|
if (k == coords[3]) {
|
|
dyi *= getDy(b, r, c, d);
|
|
result += dyi;
|
|
}
|
|
}
|
|
else {
|
|
break;
|
|
}
|
|
}
|
|
}
|
|
setOutput(result);
|
|
}
|
|
`}},qY=e=>{let{inputs:t,backend:a,attrs:n}=e,{x:r,y:s,dy:i}=t,{depthRadius:o,bias:l,alpha:u,beta:p}=n,c=new jY(r.shape,o,l,u,p);return a.runWebGLProgram(c,[r,s,i],r.dtype)},XY={kernelName:cu,backendName:"webgl",kernelFunc:qY};function KY(e,t,a,n){let r=v.sizeFromShape(t),s=v.sizeFromShape(e.shape)/r,i=ue({inputs:{x:e},attrs:{shape:[s,r]},backend:n}),o=Ko(i,e.dtype,"max",n),l=ue({inputs:{x:o},attrs:{shape:a},backend:n});return n.disposeIntermediateTensorInfo(i),n.disposeIntermediateTensorInfo(o),l}function p8(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{reductionIndices:s,keepDims:i}=n,o=r.shape.length,l=v.parseAxisParam(s,r.shape),u=l,p=S.getAxesPermutation(u,o),c=p!=null,d=a.shouldExecuteOnCPU([r]),h=r;if(c){if(d){let x=a.texData.get(h.dataId).values,A=new Array(o);for(let C=0;C<A.length;C++)A[C]=r.shape[p[C]];let b=wg(x,r.shape,r.dtype,p,A);h=a.makeTensorInfo(A,r.dtype);let w=a.texData.get(h.dataId);w.values=b}else h=Qh(r,p,a);u=S.getInnerMostAxes(u.length,o)}S.assertAxesAreInnerMostDims("max",u,o);let[f,m]=S.computeOutAndReduceShapes(h.shape,u),g=f;i&&(g=S.expandShapeToKeepDim(f,l));let y;if(d){let x=a.texData.get(h.dataId).values,A=rH(x,v.sizeFromShape(m),g,r.dtype);y=a.makeTensorInfo(g,r.dtype);let b=a.texData.get(y.dataId);b.values=A}else y=KY(h,m,g,a);return c&&a.disposeIntermediateTensorInfo(h),y}var YY={kernelName:Ji,backendName:"webgl",kernelFunc:p8},ZY=kg+`
|
|
return max(a, b);
|
|
`,JY=`
|
|
vec4 result = vec4(max(a, b));
|
|
bvec4 isNaNA = isnan(a);
|
|
bvec4 isNaNB = isnan(b);
|
|
bvec4 isNaN = bvec4(isNaNA.x || isNaNB.x, isNaNA.y || isNaNB.y, isNaNA.z || isNaNB.z, isNaNA.w || isNaNB.w);
|
|
`+Xo+`
|
|
return result;
|
|
`,QY=fa({opSnippet:ZY,packedOpSnippet:JY,cpuKernelImpl:sH}),eZ={kernelName:Qi,backendName:"webgl",kernelFunc:QY};function tZ(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t;Pu(r,"maxPool");let{filterSize:s,strides:i,pad:o,dimRoundingMode:l}=n,u=1;v.assert(S.eitherStridesOrDilationsAreOne(i,u),()=>`Error in maxPool: Either strides or dilations must be 1. Got strides ${i} and dilations '${u}'`);let p=S.computePool2DInfo(r.shape,s,i,u,o,l);if(p.filterWidth===1&&p.filterHeight===1&&v.arraysEqual(p.inShape,p.outShape))return en({inputs:{x:r},backend:a});let c=new jd(p,"max",!1);return a.runWebGLProgram(c,[r],r.dtype)}var aZ={kernelName:eo,backendName:"webgl",kernelFunc:tZ};function nZ(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{filterSize:s,strides:i,pad:o,dataFormat:l,dimRoundingMode:u}=n,p=[1,1,1],c=S.computePool3DInfo(r.shape,s,i,p,o,u,l),d=new Sg(c,"max",!1);return a.runWebGLProgram(d,[r],r.dtype)}var rZ={kernelName:hu,backendName:"webgl",kernelFunc:nZ},sZ=class{constructor(e){this.variableNames=["dy","maxPos"],this.outputShape=e.inShape;let t=e.strideHeight,a=e.strideWidth,n=e.dilationHeight,r=e.effectiveFilterHeight,s=e.effectiveFilterWidth,i=r-1-e.padInfo.top,o=s-1-e.padInfo.left,l=r*s-1;this.userCode=`
|
|
const ivec2 pads = ivec2(${i}, ${o});
|
|
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int b = coords[0];
|
|
int d = coords[3];
|
|
|
|
ivec2 dyRCCorner = coords.yz - pads;
|
|
int dyRCorner = dyRCCorner.x;
|
|
int dyCCorner = dyRCCorner.y;
|
|
|
|
// Convolve dy(?, ?, d) with pos mask(:, :, d) to get dx(xR, xC, d).
|
|
// ? = to be determined. : = across all values in that axis.
|
|
float dotProd = 0.0;
|
|
for (int wR = 0; wR < ${r};
|
|
wR += ${n}) {
|
|
float dyR = float(dyRCorner + wR) / ${t}.0;
|
|
|
|
if (dyR < 0.0 || dyR >= ${e.outHeight}.0 || fract(dyR) > 0.0) {
|
|
continue;
|
|
}
|
|
int idyR = int(dyR);
|
|
|
|
for (int wC = 0; wC < ${s}; wC++) {
|
|
float dyC = float(dyCCorner + wC) / ${a}.0;
|
|
|
|
if (dyC < 0.0 || dyC >= ${e.outWidth}.0 ||
|
|
fract(dyC) > 0.0) {
|
|
continue;
|
|
}
|
|
int idyC = int(dyC);
|
|
|
|
float dyValue = getDy(b, idyR, idyC, d);
|
|
int maxPosValue = ${l} - int(getMaxPos(b, idyR, idyC, d));
|
|
|
|
// Get the current value, check it against the value from the
|
|
// position matrix.
|
|
int curPosValue = wR * ${s} + wC;
|
|
float mask = float(maxPosValue == curPosValue ? 1.0 : 0.0);
|
|
|
|
dotProd += dyValue * mask;
|
|
}
|
|
}
|
|
setOutput(dotProd);
|
|
}
|
|
`}},iZ=class{constructor(e){this.variableNames=["dy","maxPos"],this.outputShape=e.inShape;let t=e.strideDepth,a=e.strideHeight,n=e.strideWidth,r=e.dilationDepth,s=e.dilationHeight,i=e.dilationWidth,o=e.effectiveFilterDepth,l=e.effectiveFilterHeight,u=e.effectiveFilterWidth,p=o-1-e.padInfo.front,c=l-1-e.padInfo.top,d=u-1-e.padInfo.left,h=o*l*u-1;this.userCode=`
|
|
const ivec3 pads = ivec3(${p}, ${c}, ${d});
|
|
|
|
void main() {
|
|
ivec5 coords = getOutputCoords();
|
|
int batch = coords.x;
|
|
int ch = coords.u;
|
|
|
|
ivec3 dyCorner = ivec3(coords.y, coords.z, coords.w) - pads;
|
|
int dyDCorner = dyCorner.x;
|
|
int dyRCorner = dyCorner.y;
|
|
int dyCCorner = dyCorner.z;
|
|
|
|
// Convolve dy(?, ?, ?, ch) with pos mask(:, :, :, d) to get
|
|
// dx(xD, xR, xC, ch).
|
|
// ? = to be determined. : = across all values in that axis.
|
|
float dotProd = 0.0;
|
|
|
|
for (int wD = 0; wD < ${o};
|
|
wD += ${r}) {
|
|
float dyD = float(dyDCorner + wD) / ${t}.0;
|
|
|
|
if (dyD < 0.0 || dyD >= ${e.outDepth}.0 || fract(dyD) > 0.0) {
|
|
continue;
|
|
}
|
|
int idyD = int(dyD);
|
|
|
|
for (int wR = 0; wR < ${l};
|
|
wR += ${s}) {
|
|
float dyR = float(dyRCorner + wR) / ${a}.0;
|
|
|
|
if (dyR < 0.0 || dyR >= ${e.outHeight}.0 ||
|
|
fract(dyR) > 0.0) {
|
|
continue;
|
|
}
|
|
int idyR = int(dyR);
|
|
|
|
for (int wC = 0; wC < ${u};
|
|
wC += ${i}) {
|
|
float dyC = float(dyCCorner + wC) / ${n}.0;
|
|
|
|
if (dyC < 0.0 || dyC >= ${e.outWidth}.0 ||
|
|
fract(dyC) > 0.0) {
|
|
continue;
|
|
}
|
|
int idyC = int(dyC);
|
|
|
|
float dyValue = getDy(batch, idyD, idyR, idyC, ch);
|
|
int maxPosValue = ${h} -
|
|
int(getMaxPos(batch, idyD, idyR, idyC, ch));
|
|
|
|
// Get the current value, check it against the value from the
|
|
// position matrix.
|
|
int curPosValue =
|
|
wD * ${l} * ${u} +
|
|
wR * ${u} + wC;
|
|
float mask = float(maxPosValue == curPosValue ? 1.0 : 0.0);
|
|
|
|
dotProd += dyValue * mask;
|
|
}
|
|
}
|
|
}
|
|
setOutput(dotProd);
|
|
}
|
|
`}};function oZ(e){let{inputs:t,backend:a,attrs:n}=e,{dy:r,input:s}=t,i=s,{filterSize:o,strides:l,pad:u,dimRoundingMode:p}=n,c=[1,1,1],d=S.computePool3DInfo(i.shape,o,l,c,u,p),h=new Sg(d,"max",!0),f=a.runWebGLProgram(h,[i],i.dtype),m=new iZ(d),g=a.runWebGLProgram(m,[r,f],i.dtype);return a.disposeIntermediateTensorInfo(f),g}var lZ={kernelName:dp,backendName:"webgl",kernelFunc:oZ};function uZ(e){let{inputs:t,backend:a,attrs:n}=e,{dy:r,input:s,output:i}=t,o=s;Pu([s,i],"maxPoolGrad");let{filterSize:l,strides:u,pad:p,dimRoundingMode:c}=n,d=S.computePool2DInfo(o.shape,l,u,1,p,c),h=!0,f=new jd(d,"max",h),m=a.runWebGLProgram(f,[o],o.dtype),g=new sZ(d),y=a.runWebGLProgram(g,[r,m],o.dtype);return a.disposeIntermediateTensorInfo(m),y}var dZ={kernelName:Sh,backendName:"webgl",kernelFunc:uZ};function pZ(e,t,a,n){let r=new jd(a,"max",!1),s=n.runWebGLProgram(r,[e],"float32");r=new jd(a,"max",!0,!0,t);let i=n.runWebGLProgram(r,[e],"float32");return[s,i]}var cZ={kernelName:pp,backendName:"webgl",kernelFunc:({inputs:e,attrs:t,backend:a})=>{let{x:n}=e,{filterSize:r,strides:s,pad:i,includeBatchInIndex:o}=t,l=a;v.assert(n.shape.length===4,()=>`Error in maxPool: input must be rank 4 but got rank ${n.shape.length}.`);let u=[1,1];v.assert(S.eitherStridesOrDilationsAreOne(s,u),()=>`Error in maxPool: Either strides or dilations must be 1. Got strides ${s} and dilations '${u}'`);let p=S.computePool2DInfo(n.shape,r,s,u,i),[c,d]=pZ(n,o,p,l);return[c,d]}};function hZ(e,t,a,n){let r=v.sizeFromShape(t),s=v.sizeFromShape(e.shape)/r,i=ue({inputs:{x:e},attrs:{shape:[s,r]},backend:n}),o=Ko(i,"float32","mean",n),l=ue({inputs:{x:o},attrs:{shape:a},backend:n});return n.disposeIntermediateTensorInfo(i),n.disposeIntermediateTensorInfo(o),l}var fZ={kernelName:to,backendName:"webgl",kernelFunc:({inputs:e,attrs:t,backend:a})=>{let{x:n}=e,{keepDims:r,axis:s}=t,i=a,o=n.shape.length,l=v.parseAxisParam(s,n.shape),u=l,p=S.getAxesPermutation(u,o),c=p!=null,d=i.shouldExecuteOnCPU([n]),h=[],f=n;if(c){if(d){let A=i.texData.get(f.dataId).values,b=new Array(o);for(let T=0;T<b.length;T++)b[T]=n.shape[p[T]];let w=wg(A,n.shape,n.dtype,p,b);f=i.makeTensorInfo(b,n.dtype);let C=i.texData.get(f.dataId);C.values=w}else f=Qh(n,p,i);h.push(f),u=S.getInnerMostAxes(u.length,o)}S.assertAxesAreInnerMostDims("sum",u,o);let[m,g]=S.computeOutAndReduceShapes(f.shape,u),y=m;r&&(y=S.expandShapeToKeepDim(m,l));let x=hZ(f,g,y,i);for(let A of h)i.disposeIntermediateTensorInfo(A);return x}};function mZ(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{axis:s,keepDims:i}=n,o=r.shape.length,l=v.parseAxisParam(s,r.shape),u=l,p=S.getAxesPermutation(u,o),c=r;p!=null&&(c=Ta({inputs:{x:r},backend:a,attrs:{perm:p}}),u=S.getInnerMostAxes(u.length,r.shape.length)),S.assertAxesAreInnerMostDims("min",u,o);let[d,h]=S.computeOutAndReduceShapes(c.shape,u),f=v.sizeFromShape(h),m=ue({inputs:{x:c},backend:a,attrs:{shape:[-1,f]}}),g=Ko(m,m.dtype,"min",a),y;if(i){let x=S.expandShapeToKeepDim(d,l);y=ue({inputs:{x:g},backend:a,attrs:{shape:x}})}else y=ue({inputs:{x:g},backend:a,attrs:{shape:d}});return a.disposeIntermediateTensorInfo(m),a.disposeIntermediateTensorInfo(g),p!=null&&a.disposeIntermediateTensorInfo(c),y}var gZ={kernelName:ao,backendName:"webgl",kernelFunc:mZ},yZ=kg+`
|
|
return min(a, b);
|
|
`,xZ=`
|
|
vec4 result = vec4(min(a, b));
|
|
bvec4 isNaNA = isnan(a);
|
|
bvec4 isNaNB = isnan(b);
|
|
bvec4 isNaN = bvec4(isNaNA.x || isNaNB.x, isNaNA.y || isNaNB.y, isNaNA.z || isNaNB.z, isNaNA.w || isNaNB.w);
|
|
`+Xo+`
|
|
return result;
|
|
`,AZ=fa({opSnippet:yZ,packedOpSnippet:xZ,cpuKernelImpl:iH}),bZ={kernelName:no,backendName:"webgl",kernelFunc:AZ},vZ=class{constructor(e,t,a){this.variableNames=["x"],this.outputShape=t.map((u,p)=>u[0]+e[p]+u[1]);let n=e.length,r=ft(n),s=t.map(u=>u[0]).join(","),i=t.map((u,p)=>u[0]+e[p]).join(","),o=["coords[0]","coords[1]","coords[2]","coords[3]"].slice(0,n),l=a==="reflect"?0:1;if(n===1){this.userCode=`
|
|
int start = ${s};
|
|
int end = ${i};
|
|
|
|
void main() {
|
|
int outC = getOutputCoords();
|
|
if (outC < start) {
|
|
outC = start * 2 - outC - ${l};
|
|
} else if(outC >= end) {
|
|
outC = (end - 1) * 2 - outC + ${l};
|
|
}
|
|
setOutput(getX(outC - start));
|
|
}
|
|
`;return}this.userCode=`
|
|
${r} start = ${r}(${s});
|
|
${r} end = ${r}(${i});
|
|
|
|
void main() {
|
|
${r} outC = getOutputCoords();
|
|
for (int i = 0; i < ${n}; i++) {
|
|
if (outC[i] < start[i]) {
|
|
outC[i] = start[i] * 2 - outC[i] - ${l};
|
|
} else if(outC[i] >= end[i]) {
|
|
outC[i] = (end[i] - 1) * 2 - outC[i] + ${l};
|
|
}
|
|
}
|
|
${r} coords = outC - start;
|
|
setOutput(getX(${o}));
|
|
}
|
|
`}},wZ=class{constructor(e,t,a){this.variableNames=["x"],this.packedInputs=!0,this.packedOutput=!0,this.outputShape=t.map((h,f)=>h[0]+e[f]+h[1]);let n=e.length,r=ft(n),s=t.map(h=>h[0]).join(","),i=t.map((h,f)=>h[0]+e[f]).join(","),o=Ia("rc",n),l=Ia("source",n),u=`${o[n-1]} < ${this.outputShape[n-1]}`,p=n===1?"source":`vec2(${l.slice(-2).join()})`,c=a==="reflect"?0:1,d="";if(n===1){let h=`
|
|
${r} source = rc;
|
|
if (source < start) {
|
|
source = start * 2 - source - ${c};
|
|
} else if (source >= end) {
|
|
source = (end - 1) * 2 - source + ${c};
|
|
}
|
|
source -= start;
|
|
`;d=`
|
|
${r} rc = outputLoc;
|
|
${h}
|
|
result[0] = getChannel(getX(${l.join()}), ${p});
|
|
${o[n-1]} += 1;
|
|
if(${u}) {
|
|
${h}
|
|
result[1] = getChannel(getX(${l.join()}), ${p});
|
|
}
|
|
`}else{let h=`
|
|
${r} source = rc;
|
|
${r} lt = ${r}(lessThan(source, start));
|
|
${r} gte = ${r}(greaterThanEqual(source, end));
|
|
${r} orig = 1 - (lt + gte);
|
|
source = orig * source +
|
|
lt * (start * 2 - source - ${c}) +
|
|
gte * ((end - 1) * 2 - source + ${c});
|
|
source -= start;
|
|
`;d=`
|
|
${r} rc = outputLoc;
|
|
${h}
|
|
result[0] = getChannel(getX(${l.join()}), ${p});
|
|
${o[n-1]} += 1;
|
|
if(${u}) {
|
|
${h}
|
|
result[1] = getChannel(getX(${l.join()}), ${p});
|
|
}
|
|
rc = outputLoc;
|
|
${o[n-2]} += 1;
|
|
if(${o[n-2]} < ${this.outputShape[n-2]}) {
|
|
${h}
|
|
result[2] = getChannel(getX(${l.join()}), ${p});
|
|
${o[n-1]} += 1;
|
|
if(${u}) {
|
|
${h}
|
|
result[3] = getChannel(getX(${l.join()}), ${p});
|
|
}
|
|
}
|
|
`}this.userCode=`
|
|
const ${r} start = ${r}(${s});
|
|
const ${r} end = ${r}(${i});
|
|
|
|
void main() {
|
|
${r} outputLoc = getOutputCoords();
|
|
vec4 result = vec4(0.);
|
|
${d}
|
|
setOutput(result);
|
|
}
|
|
`}},kZ=({inputs:e,backend:t,attrs:a})=>{let{x:n}=e,{paddings:r,mode:s}=a,i=W().getBool("WEBGL_PACK_ARRAY_OPERATIONS")?new wZ(n.shape,r,s):new vZ(n.shape,r,s);return t.runWebGLProgram(i,[n],n.dtype)},IZ={kernelName:ro,backendName:"webgl",kernelFunc:kZ},SZ=`if (b == 0.0) return NAN;
|
|
return mod(a, b);`,CZ=`
|
|
vec4 result = mod(a, b);
|
|
bvec4 isNaN = equal(b, vec4(0.0));
|
|
`+Xo+`
|
|
return result;
|
|
`,TZ=fa({opSnippet:SZ,packedOpSnippet:CZ}),NZ={kernelName:fu,backendName:"webgl",kernelFunc:TZ},RZ=class{constructor(e,t,a){this.variableNames=["probs"],this.customUniforms=[{name:"seed",type:"float"}],this.outputShape=[e,a],this.userCode=`
|
|
void main() {
|
|
ivec2 coords = getOutputCoords();
|
|
int batch = coords[0];
|
|
|
|
float r = random(seed);
|
|
float cdf = 0.0;
|
|
|
|
for (int i = 0; i < ${t-1}; i++) {
|
|
cdf += getProbs(batch, i);
|
|
|
|
if (r < cdf) {
|
|
setOutput(float(i));
|
|
return;
|
|
}
|
|
}
|
|
|
|
// If no other event happened, last event happened.
|
|
setOutput(float(${t-1}));
|
|
}
|
|
`}},EZ=`
|
|
if (a == b) {
|
|
return 1.0;
|
|
};
|
|
return a / b;`,MZ=`
|
|
// vec4 one = vec4(equal(a, b));
|
|
// return one + (vec4(1.0) - one) * a / b;
|
|
vec4 result = a / b;
|
|
if(a.x == b.x) {
|
|
result.x = 1.;
|
|
}
|
|
if(a.y == b.y) {
|
|
result.y = 1.;
|
|
}
|
|
if(a.z == b.z) {
|
|
result.z = 1.;
|
|
}
|
|
if(a.w == b.w) {
|
|
result.w = 1.;
|
|
}
|
|
|
|
return result;
|
|
`,c8=fa({opSnippet:EZ,packedOpSnippet:MZ,checkOutOfBounds:!0}),FZ={kernelName:Ci,backendName:"webgl",kernelFunc:c8},kx="return a - b;",h8=fa({opSnippet:kx,packedOpSnippet:kx,supportsComplex:!0,cpuKernelImpl:TH}),_Z={kernelName:zo,backendName:"webgl",kernelFunc:h8};function f8(e){let{inputs:t,backend:a,attrs:n}=e,{logits:r}=t,{dim:s}=n,i=v.parseAxisParam([s],r.shape),o=p8({inputs:{x:r},backend:a,attrs:{reductionIndices:i,keepDims:!1}}),l=S.expandShapeToKeepDim(o.shape,i),u=ue({inputs:{x:o},backend:a,attrs:{shape:l}}),p=h8({inputs:{a:r,b:u},backend:a}),c=l8({inputs:{x:p},backend:a}),d=e0({inputs:{x:c},backend:a,attrs:{axis:i,keepDims:!1}}),h=ue({inputs:{x:d},backend:a,attrs:{shape:l}}),f=c8({inputs:{a:c,b:h},backend:a});return a.disposeIntermediateTensorInfo(o),a.disposeIntermediateTensorInfo(u),a.disposeIntermediateTensorInfo(p),a.disposeIntermediateTensorInfo(c),a.disposeIntermediateTensorInfo(d),a.disposeIntermediateTensorInfo(h),f}var $Z={kernelName:$o,backendName:"webgl",kernelFunc:f8};function DZ(e){let{inputs:t,backend:a,attrs:n}=e,{logits:r}=t,{numSamples:s,seed:i,normalized:o}=n,l=o?r:f8({inputs:{logits:r},backend:a,attrs:{dim:r.shape.length-1}}),u=l.shape[0],p=l.shape[1],c=new RZ(u,p,s),d=[[i]],h=a.runWebGLProgram(c,[l],"int32",d);return o||a.disposeIntermediateTensorInfo(l),h}var PZ={kernelName:so,backendName:"webgl",kernelFunc:DZ},OZ=Rn+`
|
|
return -x;
|
|
`,zZ=`
|
|
vec4 result = -x;
|
|
bvec4 isNaN = isnan(x);
|
|
|
|
result.r = isNaN.r ? x.r : result.r;
|
|
result.g = isNaN.g ? x.g : result.g;
|
|
result.b = isNaN.b ? x.b : result.b;
|
|
result.a = isNaN.a ? x.a : result.a;
|
|
|
|
return result;
|
|
`;function LZ(e){let{inputs:t,backend:a}=e,{x:n}=t;if(a.shouldExecuteOnCPU([n])){let s=a.texData.get(n.dataId),[i,o]=lH(s.values,n.shape,n.dtype);return a.makeTensorInfo(o,n.dtype,i)}let r;return W().getBool("WEBGL_PACK_UNARY_OPERATIONS")?r=new Gr(n.shape,zZ):r=new Yn(n.shape,OZ),a.runWebGLProgram(r,[n],n.dtype)}var WZ={kernelName:mu,backendName:"webgl",kernelFunc:LZ},BZ=Nn.nonMaxSuppressionV3Impl;function VZ(e){S.warn("tf.nonMaxSuppression() in webgl locks the UI thread. Call tf.nonMaxSuppressionAsync() instead");let{inputs:t,backend:a,attrs:n}=e,{boxes:r,scores:s}=t,{maxOutputSize:i,iouThreshold:o,scoreThreshold:l}=n,u=a.readSync(r.dataId),p=a.readSync(s.dataId),{selectedIndices:c}=BZ(u,p,i,o,l);return a.makeTensorInfo([c.length],"int32",new Int32Array(c))}var UZ={kernelName:lo,backendName:"webgl",kernelFunc:VZ},GZ=Nn.nonMaxSuppressionV4Impl;function HZ(e){S.warn("tf.nonMaxSuppression() in webgl locks the UI thread. Call tf.nonMaxSuppressionAsync() instead");let{inputs:t,backend:a,attrs:n}=e,{boxes:r,scores:s}=t,{maxOutputSize:i,iouThreshold:o,scoreThreshold:l,padToMaxOutputSize:u}=n,p=a.readSync(r.dataId),c=a.readSync(s.dataId),{selectedIndices:d,validOutputs:h}=GZ(p,c,i,o,l,u);return[a.makeTensorInfo([d.length],"int32",new Int32Array(d)),a.makeTensorInfo([],"int32",new Int32Array([h]))]}var jZ={kernelName:gu,backendName:"webgl",kernelFunc:HZ},qZ=Nn.nonMaxSuppressionV5Impl;function XZ(e){S.warn("tf.nonMaxSuppression() in webgl locks the UI thread. Call tf.nonMaxSuppressionAsync() instead");let{inputs:t,backend:a,attrs:n}=e,{boxes:r,scores:s}=t,{maxOutputSize:i,iouThreshold:o,scoreThreshold:l,softNmsSigma:u}=n,p=a.readSync(r.dataId),c=a.readSync(s.dataId),d=i,h=o,f=l,m=u,{selectedIndices:g,selectedScores:y}=qZ(p,c,d,h,f,m);return[a.makeTensorInfo([g.length],"int32",new Int32Array(g)),a.makeTensorInfo([y.length],"float32",new Float32Array(y))]}var KZ={kernelName:uo,backendName:"webgl",kernelFunc:XZ},YZ=class{constructor(e,t,a,n){this.variableNames=["indices"],this.outputShape=[e,t],this.userCode=`
|
|
void main() {
|
|
ivec2 coords = getOutputCoords();
|
|
int index = round(getIndices(coords.x));
|
|
setOutput(mix(float(${n}), float(${a}),
|
|
float(index == coords.y)));
|
|
}
|
|
`}},ZZ=e=>{let{inputs:t,backend:a,attrs:n}=e,{indices:r}=t,{dtype:s,depth:i,onValue:o,offValue:l}=n,u=v.sizeFromShape(r.shape),p=new YZ(u,i,o,l),c=ue({inputs:{x:r},backend:a,attrs:{shape:[u]}}),d=a.runWebGLProgram(p,[c],s);a.disposeIntermediateTensorInfo(c);let h=[...r.shape,i],f=ue({inputs:{x:d},backend:a,attrs:{shape:h}});return a.disposeIntermediateTensorInfo(d),f},JZ={kernelName:po,backendName:"webgl",kernelFunc:ZZ};function gh(e){let{inputs:t,backend:a}=e,{x:n}=t;if(n.dtype==="complex64"){let r=Gp({inputs:{input:n},backend:a}),s=gh({inputs:{x:r},backend:a}),i=t0({inputs:{input:n},backend:a}),o=gh({inputs:{x:i},backend:a}),l=ds({inputs:{real:s,imag:o},backend:a});return a.disposeIntermediateTensorInfo(r),a.disposeIntermediateTensorInfo(s),a.disposeIntermediateTensorInfo(i),a.disposeIntermediateTensorInfo(o),l}else return Hp({attrs:{shape:n.shape,dtype:n.dtype,value:n.dtype==="string"?"":0},backend:a})}var QZ={kernelName:Mu,backendName:"webgl",kernelFunc:gh};function m8(e){let{inputs:t,backend:a}=e,{x:n}=t;if(n.dtype==="string")throw new Error("onesLike is not supported under string dtype");if(n.dtype==="complex64"){let r=Gp({inputs:{input:n},backend:a}),s=m8({inputs:{x:r},backend:a}),i=t0({inputs:{input:n},backend:a}),o=gh({inputs:{x:i},backend:a}),l=ds({inputs:{real:s,imag:o},backend:a});return a.disposeIntermediateTensorInfo(r),a.disposeIntermediateTensorInfo(s),a.disposeIntermediateTensorInfo(i),a.disposeIntermediateTensorInfo(o),l}else return Hp({attrs:{shape:n.shape,dtype:n.dtype,value:1},backend:a})}var eJ={kernelName:yu,backendName:"webgl",kernelFunc:m8};function tJ(e){let{inputs:t,backend:a,attrs:n}=e,{axis:r}=n;if(t.length===1)return T1({inputs:{input:t[0]},backend:a,attrs:{dim:r}});let s=t[0].shape,i=t[0].dtype;t.forEach(p=>{v.assertShapesMatch(s,p.shape,"All tensors passed to stack must have matching shapes"),v.assert(i===p.dtype,()=>"All tensors passed to stack must have matching dtypes")});let o=[],l=t.map(p=>{let c=T1({inputs:{input:p},backend:a,attrs:{dim:r}});return o.push(c),c}),u=e8({inputs:l,backend:a,attrs:{axis:r}});return o.forEach(p=>a.disposeIntermediateTensorInfo(p)),u}var aJ={kernelName:xu,backendName:"webgl",kernelFunc:tJ},nJ=class{constructor(e,t,a){this.variableNames=["x"],this.customUniforms=[{name:"value",type:"float"}],this.outputShape=t.map((l,u)=>l[0]+e[u]+l[1]);let n=e.length,r=ft(n),s=t.map(l=>l[0]).join(","),i=t.map((l,u)=>l[0]+e[u]).join(","),o=["coords[0]","coords[1]","coords[2]","coords[3]"].slice(0,n);if(n===1){this.userCode=`
|
|
int start = ${s};
|
|
int end = ${i};
|
|
|
|
void main() {
|
|
int outC = getOutputCoords();
|
|
if (outC < start || outC >= end) {
|
|
setOutput(value);
|
|
} else {
|
|
setOutput(getX(outC - start));
|
|
}
|
|
}
|
|
`;return}this.userCode=`
|
|
${r} start = ${r}(${s});
|
|
${r} end = ${r}(${i});
|
|
|
|
void main() {
|
|
${r} outC = getOutputCoords();
|
|
if (any(lessThan(outC, start)) || any(greaterThanEqual(outC, end))) {
|
|
setOutput(value);
|
|
} else {
|
|
${r} coords = outC - start;
|
|
setOutput(getX(${o}));
|
|
}
|
|
}
|
|
`}},rJ=class{constructor(e,t,a){this.variableNames=["x"],this.packedInputs=!0,this.packedOutput=!0,this.customUniforms=[{name:"value",type:"float"}],this.outputShape=t.map((f,m)=>f[0]+e[m]+f[1]);let n=e.length,r=ft(n),s=t.map(f=>f[0]).join(","),i=t.map((f,m)=>f[0]+e[m]).join(","),o=Ia("rc",n),l=Ia("source",n),u=`${o[n-1]} < ${this.outputShape[n-1]}`,p=n===1?"source":`vec2(${l.slice(-2).join()})`,c=[`${r} rc = outputLoc;`,`${o[n-1]} += 1;
|
|
if(${u}) {
|
|
`,n===1?"":`}
|
|
rc = outputLoc;
|
|
${o[n-2]} += 1;
|
|
if(${o[n-2]} < ${this.outputShape[n-2]}) {`,n===1?"":` ${o[n-1]} += 1;
|
|
if(${u}) {`],d=n===1?"rc < start || rc >= end":"any(lessThan(rc, start)) || any(greaterThanEqual(rc, end))",h="";for(let f=0,m=n===1?2:4;f<m;f++)h+=`
|
|
${c[f]}
|
|
if (${d}) {
|
|
result[${f}] = float(value);
|
|
} else {
|
|
${r} source = rc - start;
|
|
result[${f}] = getChannel(getX(${l.join()}), ${p});
|
|
}
|
|
`;h+=n===1?"} ":"}}",this.userCode=`
|
|
const ${r} start = ${r}(${s});
|
|
const ${r} end = ${r}(${i});
|
|
|
|
void main() {
|
|
${r} outputLoc = getOutputCoords();
|
|
vec4 result = vec4(0.);
|
|
${h}
|
|
setOutput(result);
|
|
}
|
|
`}},g8=e=>{let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{paddings:s,constantValue:i}=n;if(v.sizeFromShape(r.shape)===0){let u=s.map((p,c)=>p[0]+r.shape[c]+p[1]);return Hp({backend:a,attrs:{shape:u,value:i,dtype:r.dtype}})}let o=W().getBool("WEBGL_PACK_ARRAY_OPERATIONS")?new rJ(r.shape,s,i):new nJ(r.shape,s,i),l=[[i]];return a.runWebGLProgram(o,[r],r.dtype,l)},sJ={kernelName:co,backendName:"webgl",kernelFunc:g8},iJ=`
|
|
if(a < 0.0 && floor(b) < b){
|
|
return NAN;
|
|
}
|
|
if (b == 0.0) {
|
|
return 1.0;
|
|
}
|
|
return (round(mod(b, 2.0)) != 1) ?
|
|
pow(abs(a), b) : sign(a) * pow(abs(a), b);
|
|
`,oJ=`
|
|
// isModRound1 has 1 for components with round(mod(b, 2.0)) == 1, 0 otherwise.
|
|
vec4 isModRound1 = vec4(equal(round(mod(b, 2.0)), ivec4(1)));
|
|
vec4 multiplier = sign(a) * isModRound1 + (vec4(1.0) - isModRound1);
|
|
vec4 result = multiplier * pow(abs(a), b);
|
|
|
|
// Ensure that a^0 = 1, including 0^0 = 1 as this correspond to TF and JS
|
|
bvec4 isExpZero = equal(b, vec4(0.0));
|
|
result.r = isExpZero.r ? 1.0 : result.r;
|
|
result.g = isExpZero.g ? 1.0 : result.g;
|
|
result.b = isExpZero.b ? 1.0 : result.b;
|
|
result.a = isExpZero.a ? 1.0 : result.a;
|
|
|
|
bvec4 isNaN1 = lessThan(a, vec4(0.0));
|
|
bvec4 isNaN2 = lessThan(floor(b), b);
|
|
bvec4 isNaN = bvec4(isNaN1.x && isNaN2.x, isNaN1.y && isNaN2.y, isNaN1.z && isNaN2.z, isNaN1.w && isNaN2.w);
|
|
`+Xo+`
|
|
return result;
|
|
`,lJ=fa({opSnippet:iJ,packedOpSnippet:oJ}),uJ={kernelName:ho,backendName:"webgl",kernelFunc:lJ};function dJ(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{axis:s,keepDims:i}=n,o=r.shape.length,l=[],u=v.parseAxisParam(s,r.shape),p=u,c=S.getAxesPermutation(p,o),d=r;c!=null&&(d=Ta({inputs:{x:r},backend:a,attrs:{perm:c}}),p=S.getInnerMostAxes(p.length,o),l.push(d)),S.assertAxesAreInnerMostDims("prod",p,o);let h;if(a.shouldExecuteOnCPU([d])){let f=a.texData.get(d.dataId).values,{outVals:m,outShape:g,outDtype:y}=dH(d.shape,d.dtype,f,p);h=a.makeTensorInfo(g,y,m)}else{let[f,m]=S.computeOutAndReduceShapes(d.shape,p),g=v.sizeFromShape(m),y=ue({inputs:{x:d},backend:a,attrs:{shape:[-1,g]}}),x=Ip(r.dtype),A=Ko(y,x,"prod",a);h=ue({inputs:{x:A},backend:a,attrs:{shape:f}}),l.push(y),l.push(A)}if(i){l.push(h);let f=S.expandShapeToKeepDim(h.shape,u);h=ue({inputs:{x:h},backend:a,attrs:{shape:f}})}return l.forEach(f=>a.disposeIntermediateTensorInfo(f)),h}var pJ={kernelName:mo,backendName:"webgl",kernelFunc:dJ};function cJ(e){let{inputs:t,backend:a,attrs:n}=e,{paramsNestedSplits:r,paramsDenseValues:s,indices:i}=t,{outputRaggedRank:o}=n,l=r.map(y=>a.readSync(y.dataId)),u=r.map(y=>y.shape),p=a.readSync(s.dataId),c=a.readSync(i.dataId),[d,h,f]=pH(l,u,p,s.shape,s.dtype,c,i.shape,o),m=d.map(y=>a.makeTensorInfo([y.length],"int32",y)),g=a.makeTensorInfo(f,s.dtype,h);return m.concat([g])}var hJ={kernelName:Ch,backendName:"webgl",kernelFunc:cJ};function fJ(e){let{inputs:t,backend:a}=e,{starts:n,limits:r,deltas:s}=t,i=a.readSync(n.dataId),o=a.readSync(r.dataId),l=a.readSync(s.dataId),[u,p]=cH(i,n.shape,n.dtype,o,r.shape,l,s.shape),c=a.makeTensorInfo([u.length],"int32",u),d=a.makeTensorInfo([p.length],n.dtype,p);return[c,d]}var mJ={kernelName:Th,backendName:"webgl",kernelFunc:fJ};function gJ(e){let{inputs:t,backend:a,attrs:n}=e,{shape:r,values:s,defaultValue:i,rowPartitionTensors:o}=t,{rowPartitionTypes:l}=n,u=a.readSync(r.dataId),p=a.readSync(s.dataId),c=a.readSync(i.dataId),d=o.map(g=>a.readSync(g.dataId)),h=o.map(g=>g.shape),[f,m]=hH(u,r.shape,p,s.shape,s.dtype,c,i.shape,d,h,l);return a.makeTensorInfo(f,s.dtype,m)}var yJ={kernelName:Nh,backendName:"webgl",kernelFunc:gJ},y8=e=>{let{backend:t,attrs:a}=e,{start:n,stop:r,step:s,dtype:i}=a,o=fH(n,r,s,i);return t.makeTensorInfo([o.length],i,o)},xJ={kernelName:Au,backendName:"webgl",kernelFunc:y8},AJ="return 1.0 / x;",bJ=et({opSnippet:AJ}),vJ={kernelName:go,backendName:"webgl",kernelFunc:bJ},wJ=Rn+`
|
|
return (x < 0.0) ? 0.0 : x;
|
|
`,kJ=`
|
|
vec4 result = x * vec4(greaterThanEqual(x, vec4(0.0)));
|
|
bvec4 isNaN = isnan(x);
|
|
|
|
result.r = isNaN.r ? x.r : result.r;
|
|
result.g = isNaN.g ? x.g : result.g;
|
|
result.b = isNaN.b ? x.b : result.b;
|
|
result.a = isNaN.a ? x.a : result.a;
|
|
|
|
return result;
|
|
`,IJ=et({opSnippet:wJ,packedOpSnippet:kJ}),SJ={kernelName:yo,backendName:"webgl",kernelFunc:IJ},CJ=Rn+`
|
|
return (x < 0.0) ? 0.0 : min(6.0, x);
|
|
`,TJ=`
|
|
vec4 result = min(x, vec4(6.)) * vec4(greaterThanEqual(x, vec4(0.0)));
|
|
bvec4 isNaN = isnan(x);
|
|
|
|
result.r = isNaN.r ? x.r : result.r;
|
|
result.g = isNaN.g ? x.g : result.g;
|
|
result.b = isNaN.b ? x.b : result.b;
|
|
result.a = isNaN.a ? x.a : result.a;
|
|
|
|
return result;
|
|
`,NJ=et({opSnippet:CJ,packedOpSnippet:TJ}),RJ={kernelName:bo,backendName:"webgl",kernelFunc:NJ},EJ=class{constructor(e,t,a,n,r){this.variableNames=["A"],this.outputShape=[];let[s,i,o,l]=e;this.outputShape=[s,t,a,l];let u=[n&&t>1?i-1:i,n&&a>1?o-1:o],p=[n&&t>1?t-1:t,n&&a>1?a-1:a],c;r?c="(vec2(yRC) + vec2(0.5)) * effectiveInputOverOutputRatioRC - vec2(0.5)":c="vec2(yRC) * effectiveInputOverOutputRatioRC",this.userCode=`
|
|
const vec2 effectiveInputOverOutputRatioRC = vec2(
|
|
${u[0]/p[0]},
|
|
${u[1]/p[1]});
|
|
const vec2 inputShapeRC = vec2(${i}.0, ${o}.0);
|
|
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int b = coords[0];
|
|
int d = coords[3];
|
|
ivec2 yRC = coords.yz;
|
|
|
|
// Fractional source index.
|
|
vec2 sourceFracIndexRC = ${c};
|
|
|
|
// Compute the four integer indices.
|
|
ivec2 sourceFloorRC = ivec2(max(sourceFracIndexRC, vec2(0.0)));
|
|
ivec2 sourceCeilRC = ivec2(
|
|
min(inputShapeRC - 1.0, ceil(sourceFracIndexRC)));
|
|
|
|
float topLeft = getA(b, sourceFloorRC.x, sourceFloorRC.y, d);
|
|
float bottomLeft = getA(b, sourceCeilRC.x, sourceFloorRC.y, d);
|
|
float topRight = getA(b, sourceFloorRC.x, sourceCeilRC.y, d);
|
|
float bottomRight = getA(b, sourceCeilRC.x, sourceCeilRC.y, d);
|
|
|
|
vec2 fracRC = sourceFracIndexRC - vec2(sourceFloorRC);
|
|
|
|
float top = topLeft + (topRight - topLeft) * fracRC.y;
|
|
float bottom = bottomLeft + (bottomRight - bottomLeft) * fracRC.y;
|
|
float newValue = top + (bottom - top) * fracRC.x;
|
|
|
|
setOutput(newValue);
|
|
}
|
|
`}},MJ=class{constructor(e,t,a,n,r){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!0,this.outputShape=[];let[s,i,o,l]=e;this.outputShape=[s,t,a,l];let u=[n&&t>1?i-1:i,n&&a>1?o-1:o],p=[n&&t>1?t-1:t,n&&a>1?a-1:a],c;r?c="(vec3(yRC) + vec3(0.5)) * effectiveInputOverOutputRatioRC - vec3(0.5)":c="vec3(yRC) * effectiveInputOverOutputRatioRC",this.userCode=`
|
|
const vec3 effectiveInputOverOutputRatioRC = vec3(
|
|
${u[0]/p[0]},
|
|
${u[1]/p[1]},
|
|
${u[1]/p[1]});
|
|
const vec3 inputShapeRC = vec3(${i}.0, ${o}.0,
|
|
${o}.0);
|
|
|
|
float getAValue(int b, int r, int c, int d) {
|
|
return getChannel(getA(b, r, c, d), vec2(c, d));
|
|
}
|
|
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int b = coords[0];
|
|
int d = coords[3];
|
|
// Calculate values for next column in yRC.z.
|
|
ivec3 yRC = coords.yzz + ivec3(0, 0, 1);
|
|
|
|
// Fractional source index.
|
|
vec3 sourceFracIndexRC = ${c};
|
|
|
|
// Compute the four integer indices.
|
|
ivec3 sourceFloorRC = ivec3(max(sourceFracIndexRC, vec3(0.0)));
|
|
ivec3 sourceCeilRC = ivec3(
|
|
min(inputShapeRC - 1.0, ceil(sourceFracIndexRC)));
|
|
|
|
// Should we calculate next column and row elements in 2x2 packed cell.
|
|
bool hasNextCol = d < ${l-1};
|
|
bool hasNextRow = coords.z < ${a-1};
|
|
|
|
// In parallel, construct four corners for all four components in
|
|
// packed 2x2 cell.
|
|
vec4 topLeft = vec4(
|
|
getAValue(b, sourceFloorRC.x, sourceFloorRC.y, d),
|
|
hasNextCol ? getAValue(b, sourceFloorRC.x, sourceFloorRC.y, d + 1)
|
|
: 0.0,
|
|
hasNextRow ? getAValue(b, sourceFloorRC.x, sourceFloorRC.z, d)
|
|
: 0.0,
|
|
(hasNextRow && hasNextCol) ?
|
|
getAValue(b, sourceFloorRC.x, sourceFloorRC.z, d + 1) : 0.0);
|
|
|
|
vec4 bottomLeft = vec4(
|
|
getAValue(b, sourceCeilRC.x, sourceFloorRC.y, d),
|
|
hasNextCol ? getAValue(b, sourceCeilRC.x, sourceFloorRC.y, d + 1)
|
|
: 0.0,
|
|
hasNextRow ? getAValue(b, sourceCeilRC.x, sourceFloorRC.z, d)
|
|
: 0.0,
|
|
(hasNextRow && hasNextCol) ?
|
|
getAValue(b, sourceCeilRC.x, sourceFloorRC.z, d + 1) : 0.0);
|
|
|
|
vec4 topRight = vec4(
|
|
getAValue(b, sourceFloorRC.x, sourceCeilRC.y, d),
|
|
hasNextCol ? getAValue(b, sourceFloorRC.x, sourceCeilRC.y, d + 1)
|
|
: 0.0,
|
|
hasNextRow ? getAValue(b, sourceFloorRC.x, sourceCeilRC.z, d)
|
|
: 0.0,
|
|
(hasNextRow && hasNextCol) ?
|
|
getAValue(b, sourceFloorRC.x, sourceCeilRC.z, d + 1) : 0.0);
|
|
|
|
vec4 bottomRight = vec4(
|
|
getAValue(b, sourceCeilRC.x, sourceCeilRC.y, d),
|
|
hasNextCol ? getAValue(b, sourceCeilRC.x, sourceCeilRC.y, d + 1)
|
|
: 0.0,
|
|
hasNextRow ? getAValue(b, sourceCeilRC.x, sourceCeilRC.z, d)
|
|
: 0.0,
|
|
(hasNextRow && hasNextCol) ?
|
|
getAValue(b, sourceCeilRC.x, sourceCeilRC.z, d + 1) : 0.0);
|
|
|
|
vec3 fracRC = sourceFracIndexRC - vec3(sourceFloorRC);
|
|
|
|
vec4 top = mix(topLeft, topRight, fracRC.yyzz);
|
|
vec4 bottom = mix(bottomLeft, bottomRight, fracRC.yyzz);
|
|
vec4 newValue = mix(top, bottom, fracRC.x);
|
|
|
|
setOutput(newValue);
|
|
}
|
|
`}};function FJ(e){let{inputs:t,backend:a,attrs:n}=e,{images:r}=t,{alignCorners:s,halfPixelCenters:i,size:o}=n,[l,u]=o,p=W().getBool("WEBGL_PACK_IMAGE_OPERATIONS")?new MJ(r.shape,l,u,s,i):new EJ(r.shape,l,u,s,i);return a.runWebGLProgram(p,[r],"float32")}var _J={kernelName:Ao,backendName:"webgl",kernelFunc:FJ},$J=class{constructor(e,t,a){this.variableNames=["dy"],this.outputShape=[],this.outputShape=t;let[,n,r]=t,[,s,i]=e,o=[a&&s>1?n-1:n,a&&i>1?r-1:r],l=[a&&s>1?s-1:s,a&&i>1?i-1:i],u=o[0]/l[0],p=o[1]/l[1],c=1/u,d=1/p,h=Math.ceil(c)*2+2,f=Math.ceil(d)*2+2;this.userCode=`
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int b = coords[0];
|
|
int d = coords[3];
|
|
int r = coords[1];
|
|
int c = coords[2];
|
|
|
|
float accumulator = 0.0;
|
|
|
|
const float heightScale = float(${u});
|
|
const float widthScale = float(${p});
|
|
|
|
const float invHeightScale = float(${c});
|
|
const float invWidthScale = float(${d});
|
|
|
|
const int winHeight = int(${h});
|
|
const int winWidth = int(${f});
|
|
|
|
// Compute bounds for where in dy we will look
|
|
float startRLerp = floor(float(r) * invHeightScale);
|
|
int startDyR = int(startRLerp - float(winHeight / 2));
|
|
|
|
float startCLerp = floor(float(c) * invWidthScale);
|
|
int startDyC = int(startCLerp - float(winWidth / 2));
|
|
|
|
// Loop over dy
|
|
for (int dyROffset = 0; dyROffset < winHeight; dyROffset++) {
|
|
int dyR = dyROffset + startDyR;
|
|
|
|
// Guard against the window exceeding the bounds of dy
|
|
if (dyR < 0 || dyR >= ${s}) {
|
|
continue;
|
|
}
|
|
|
|
for (int dyCOffset = 0; dyCOffset < winWidth; dyCOffset++) {
|
|
int dyC = dyCOffset + startDyC;
|
|
|
|
// Guard against the window exceeding the bounds of dy
|
|
if (dyC < 0 || dyC >= ${i}) {
|
|
continue;
|
|
}
|
|
|
|
float dxR = float(dyR) * heightScale;
|
|
int topDxRIndex = int(floor(dxR));
|
|
int bottomDxRIndex = int(min(ceil(dxR), ${n-1}.0));
|
|
float dxRLerp = dxR - float(topDxRIndex);
|
|
float inverseDxRLerp = 1.0 - dxRLerp;
|
|
|
|
float dxC = float(dyC) * widthScale;
|
|
int leftDxCIndex = int(floor(dxC));
|
|
int rightDxCIndex = int(min(ceil(dxC), ${r-1}.0));
|
|
float dxCLerp = dxC - float(leftDxCIndex);
|
|
float inverseDxCLerp = 1.0 - dxCLerp;
|
|
|
|
if (r == topDxRIndex && c == leftDxCIndex) {
|
|
// topLeft
|
|
accumulator +=
|
|
getDy(b, dyR, dyC, d) * inverseDxRLerp * inverseDxCLerp;
|
|
}
|
|
|
|
if (r == topDxRIndex && c == rightDxCIndex) {
|
|
// topRight
|
|
accumulator += getDy(b, dyR, dyC, d) * inverseDxRLerp * dxCLerp;
|
|
}
|
|
|
|
if (r == bottomDxRIndex && c == leftDxCIndex) {
|
|
// bottomLeft
|
|
accumulator += getDy(b, dyR, dyC, d) * dxRLerp * inverseDxCLerp;
|
|
}
|
|
|
|
if (r == bottomDxRIndex && c == rightDxCIndex) {
|
|
// bottomRight
|
|
accumulator += getDy(b, dyR, dyC, d) * dxRLerp * dxCLerp;
|
|
}
|
|
}
|
|
}
|
|
// End loop over dy
|
|
|
|
setOutput(accumulator);
|
|
}
|
|
`}};function DJ(e){let{inputs:t,backend:a,attrs:n}=e,{images:r,dy:s}=t,{alignCorners:i}=n,o=new $J(s.shape,r.shape,i);return a.runWebGLProgram(o,[s],s.dtype)}var PJ={kernelName:wu,backendName:"webgl",kernelFunc:DJ},OJ=class{constructor(e,t,a,n,r){this.variableNames=["A"],this.outputShape=[];let[s,i,o,l]=e;this.outputShape=[s,t,a,l];let u=[n&&t>1?i-1:i,n&&a>1?o-1:o],p=[n&&t>1?t-1:t,n&&a>1?a-1:a],c=n?"0.5":"0.0",d;r?d="max((vec2(yRC) + vec2(0.5)) * effectiveInputOverOutputRatioRC, vec2(0.0))":d="vec2(yRC) * effectiveInputOverOutputRatioRC",this.userCode=`
|
|
const vec2 effectiveInputOverOutputRatioRC = vec2(
|
|
${u[0]/p[0]},
|
|
${u[1]/p[1]});
|
|
const vec2 inputShapeRC = vec2(${i}.0, ${o}.0);
|
|
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int b = coords[0];
|
|
int d = coords[3];
|
|
ivec2 yRC = coords.yz;
|
|
|
|
// Fractional source index.
|
|
vec2 sourceFracIndexRC = ${d};
|
|
|
|
// Compute the coordinators of nearest neighbor point.
|
|
ivec2 sourceNearestRC = ivec2(
|
|
min(inputShapeRC - 1.0, floor(sourceFracIndexRC + ${c})));
|
|
float newValue = getA(b, sourceNearestRC.x, sourceNearestRC.y, d);
|
|
|
|
setOutput(newValue);
|
|
}
|
|
`}},zJ=class{constructor(e,t,a,n,r){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!0,this.outputShape=[];let[s,i,o,l]=e;this.outputShape=[s,t,a,l];let u=[n&&t>1?i-1:i,n&&a>1?o-1:o],p=[n&&t>1?t-1:t,n&&a>1?a-1:a],c=n?"0.5":"0.0",d;r?d="max((vec3(yRC) + vec3(0.5)) * effectiveInputOverOutputRatioRC, vec3(0.0))":d="vec3(yRC) * effectiveInputOverOutputRatioRC",this.userCode=`
|
|
const vec3 effectiveInputOverOutputRatioRC = vec3(
|
|
${u[0]/p[0]},
|
|
${u[1]/p[1]},
|
|
${u[1]/p[1]});
|
|
const vec3 inputShapeRC = vec3(${i}.0, ${o}.0,
|
|
${o}.0);
|
|
|
|
float getAValue(int b, int r, int c, int d) {
|
|
return getChannel(getA(b, r, c, d), vec2(c, d));
|
|
}
|
|
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int b = coords[0];
|
|
int d = coords[3];
|
|
// Calculate values for next column in yRC.z.
|
|
ivec3 yRC = coords.yzz + ivec3(0, 0, 1);
|
|
|
|
// Fractional source index.
|
|
vec3 sourceFracIndexRC = ${d};
|
|
|
|
// Compute the coordinators of nearest neighbor point.
|
|
ivec3 sourceNearestRC = ivec3(
|
|
min(inputShapeRC - 1.0, floor(sourceFracIndexRC + ${c})));
|
|
|
|
// Should we calculate next column and row elements in 2x2 packed cell.
|
|
bool hasNextCol = d < ${l-1};
|
|
bool hasNextRow = coords.z < ${a-1};
|
|
|
|
vec4 newValue = vec4(
|
|
getAValue(b, sourceNearestRC.x, sourceNearestRC.y, d),
|
|
hasNextCol ? getAValue(b, sourceNearestRC.x, sourceNearestRC.y, d + 1)
|
|
: 0.0,
|
|
hasNextRow ? getAValue(b, sourceNearestRC.x, sourceNearestRC.z, d)
|
|
: 0.0,
|
|
(hasNextRow && hasNextCol) ?
|
|
getAValue(b, sourceNearestRC.x, sourceNearestRC.z, d + 1) : 0.0);
|
|
|
|
setOutput(newValue);
|
|
}
|
|
`}};function LJ(e){let{inputs:t,backend:a,attrs:n}=e,{images:r}=t,{alignCorners:s,halfPixelCenters:i,size:o}=n,[l,u]=o,p=W().getBool("WEBGL_PACK_IMAGE_OPERATIONS")?new zJ(r.shape,l,u,s,i):new OJ(r.shape,l,u,s,i);return a.runWebGLProgram(p,[r],r.dtype)}var WJ={kernelName:xo,backendName:"webgl",kernelFunc:LJ},BJ=class{constructor(e,t,a){this.variableNames=["dy"],this.outputShape=[],this.outputShape=t;let[,n,r]=t,[,s,i]=e,o=[a&&s>1?n-1:n,a&&i>1?r-1:r],l=[a&&s>1?s-1:s,a&&i>1?i-1:i],u=o[0]/l[0],p=o[1]/l[1],c=1/u,d=1/p,h=Math.ceil(c)*2+2,f=Math.ceil(d)*2+2;this.userCode=`
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int b = coords[0];
|
|
int d = coords[3];
|
|
int r = coords[1];
|
|
int c = coords[2];
|
|
|
|
float accumulator = 0.0;
|
|
|
|
const float heightScale = float(${u});
|
|
const float widthScale = float(${p});
|
|
|
|
const float invHeightScale = float(${c});
|
|
const float invWidthScale = float(${d});
|
|
|
|
const int winHeight = int(${h});
|
|
const int winWidth = int(${f});
|
|
|
|
// Compute bounds for where in dy we will look
|
|
float startRLerp = floor(float(r) * invHeightScale);
|
|
int startDyR = int(floor(startRLerp - float(winHeight / 2)));
|
|
|
|
float startCLerp = floor(float(c) * invWidthScale);
|
|
int startDyC = int(floor(startCLerp - float(winWidth / 2)));
|
|
|
|
// Loop over dy
|
|
for (int dyROffset = 0; dyROffset < winHeight; dyROffset++) {
|
|
int dyR = dyROffset + startDyR;
|
|
|
|
// Guard against the window exceeding the bounds of dy
|
|
if (dyR < 0 || dyR >= ${s}) {
|
|
continue;
|
|
}
|
|
|
|
for (int dyCOffset = 0; dyCOffset < winWidth; dyCOffset++) {
|
|
int dyC = dyCOffset + startDyC;
|
|
|
|
// Guard against the window exceeding the bounds of dy
|
|
if (dyC < 0 || dyC >= ${i}) {
|
|
continue;
|
|
}
|
|
|
|
float sourceFracRow =
|
|
float(${o[0]}) *
|
|
(float(dyR) / float(${l[0]}));
|
|
|
|
float sourceFracCol =
|
|
float(${o[1]}) *
|
|
(float(dyC) / float(${l[1]}));
|
|
|
|
int sourceNearestRow = int(min(
|
|
float(int(${n}) - 1),
|
|
${a} ? float(round(sourceFracRow)) :
|
|
float(floor(sourceFracRow))));
|
|
|
|
int sourceNearestCol = int(min(
|
|
float(int(${r}) - 1),
|
|
${a} ? float(round(sourceFracCol)) :
|
|
float(floor(sourceFracCol))));
|
|
|
|
if (r == sourceNearestRow && c == sourceNearestCol) {
|
|
accumulator += getDy(b, dyR, dyC, d);
|
|
}
|
|
}
|
|
}
|
|
// End loop over dy
|
|
|
|
setOutput(accumulator);
|
|
}
|
|
`}};function VJ(e){let{inputs:t,backend:a,attrs:n}=e,{images:r,dy:s}=t,{alignCorners:i}=n,o=new BJ(s.shape,r.shape,i);return a.runWebGLProgram(o,[s],s.dtype)}var UJ={kernelName:vu,backendName:"webgl",kernelFunc:VJ},GJ=class{constructor(e,t){this.variableNames=["x"];let a=e.length;if(a>4)throw new Error(`WebGL backend: Reverse of rank-${a} tensor is not yet supported`);if(this.outputShape=e,a===1){this.userCode=`
|
|
void main() {
|
|
int coord = getOutputCoords();
|
|
setOutput(getX(${e[0]} - coord - 1));
|
|
}
|
|
`;return}let n=i=>t.indexOf(i)!==-1&&e[i]!==1?`${e[i]} - coords[${i}] - 1`:`coords[${i}]`,r=e.map((i,o)=>n(o)).join(","),s=ft(a);this.userCode=`
|
|
void main() {
|
|
${s} coords = getOutputCoords();
|
|
setOutput(getX(${r}));
|
|
}
|
|
`}},HJ=class{constructor(e,t){this.variableNames=["x"],this.packedInputs=!0,this.packedOutput=!0;let a=e.length;if(a>4)throw new Error(`WebGL backend: Reverse of rank-${a} tensor is not yet supported`);this.outputShape=e;let n=Ia("rc",a),r=`${n[a-1]} + 1 < ${this.outputShape[a-1]}`,s=`${n[a-2]} + 1 < ${this.outputShape[a-2]}`,i=ft(a);a===1?this.userCode=`
|
|
void main(){
|
|
int rc = getOutputCoords();
|
|
vec4 result = vec4(0.);
|
|
result.r = getChannel(getX(${e[0]} - rc - 1),
|
|
${e[0]} - rc - 1);
|
|
if(${r}){
|
|
result.g = getChannel(getX(${e[0]} - (rc + 1) - 1),
|
|
${e[0]} - (rc + 1) - 1);
|
|
}
|
|
setOutput(result);
|
|
}
|
|
`:this.userCode=`
|
|
void main() {
|
|
${i} rc = getOutputCoords();
|
|
vec4 result = vec4(0.);
|
|
result.r = ${o(n.slice())};
|
|
if(${r}){
|
|
result.g = ${l(n.slice())};
|
|
}
|
|
if(${s}) {
|
|
result.b = ${u(n.slice())};
|
|
if(${r}) {
|
|
result.a = ${p(n.slice())};
|
|
}
|
|
}
|
|
setOutput(result);
|
|
}
|
|
`;function o(h){return c(h)}function l(h){return h[a-1]="("+h[a-1]+" + 1)",c(h)}function u(h){return h[a-2]="("+h[a-2]+" + 1)",c(h)}function p(h){return h[a-1]="("+h[a-1]+" + 1)",h[a-2]="("+h[a-2]+" + 1)",c(h)}function c(h){let f=e.map((y,x)=>d(x,h)),m=f.join(","),g=f.slice(-2).join(",");return`getChannel(getX(${m}), vec2(${g}))`}function d(h,f){return t.indexOf(h)!==-1&&e[h]!==1?`${e[h]} - ${f[h]} - 1`:`${f[h]}`}}};function jJ(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{dims:s}=n,i=r.shape.length,o=v.parseAxisParam(s,r.shape);if(i===0)return en({inputs:{x:r},backend:a});let l=W().getBool("WEBGL_PACK_ARRAY_OPERATIONS")?new HJ(r.shape,o):new GJ(r.shape,o);return a.runWebGLProgram(l,[r],r.dtype)}var qJ={kernelName:vo,backendName:"webgl",kernelFunc:jJ},XJ=class{constructor(e,t){this.variableNames=["Image"],this.outputShape=[],this.customUniforms=[{name:"params",type:"vec4"}];let a=e[1],n=e[2];this.outputShape=e;let r="";typeof t=="number"?r=`float outputValue = ${t.toFixed(2)};`:r=`
|
|
vec3 fill = vec3(${t.join(",")});
|
|
float outputValue = fill[coords[3]];`,this.userCode=`
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int x = coords[2];
|
|
int y = coords[1];
|
|
float coordXFloat = (float(x) - params[0]) * params[3] -
|
|
(float(y) - params[1]) * params[2];
|
|
float coordYFloat = (float(x) - params[0]) * params[2] +
|
|
(float(y) - params[1]) * params[3];
|
|
int coordX = int(round(coordXFloat + params[0]));
|
|
int coordY = int(round(coordYFloat + params[1]));
|
|
${r}
|
|
if(coordX >= 0 && coordX < ${n} && coordY >= 0 && coordY < ${a}) {
|
|
outputValue = getImage(coords[0], coordY, coordX, coords[3]);
|
|
}
|
|
setOutput(outputValue);
|
|
}
|
|
`}},KJ={kernelName:Uo,backendName:"webgl",kernelFunc:({inputs:e,attrs:t,backend:a})=>{let{image:n}=e,{radians:r,fillValue:s,center:i}=t,o=a,l=new XJ(n.shape,s),[u,p]=S.getImageCenter(i,n.shape[1],n.shape[2]),c=[[u,p,Math.sin(r),Math.cos(r)]];return o.runWebGLProgram(l,[n],n.dtype,c)}},YJ=`
|
|
// OpenGL ES does not support round function.
|
|
// The algorithm is based on banker's rounding.
|
|
float base = floor(x);
|
|
if ((x - base) < 0.5) {
|
|
return floor(x);
|
|
} else if ((x - base) > 0.5) {
|
|
return ceil(x);
|
|
} else {
|
|
if (mod(base, 2.0) == 0.0) {
|
|
return base;
|
|
} else {
|
|
return base + 1.0;
|
|
}
|
|
}
|
|
`,ZJ=et({opSnippet:YJ}),JJ={kernelName:wo,backendName:"webgl",kernelFunc:ZJ},QJ="return inversesqrt(x);",eQ=et({opSnippet:QJ,cpuKernelImpl:mH}),tQ={kernelName:ko,backendName:"webgl",kernelFunc:eQ},Cg=class{constructor(e,t,a,n,r,s,i=!0,o=!1){this.variableNames=["updates","indices","defaultValue"],this.outputShape=s;let l=ft(r.length),u=ft(s.length),p="";a===1?p="i":a===2&&(p="i, j");let c=`getIndices(${p})`,d="";n===1?d="i":n===2&&(d="i, coords[1]");let h=`getUpdates(${d})`,f="";o&&(f="coords[0], coords[1]");let m=`getDefaultValue(${f})`,g=t>1?"strides[j]":"strides";this.userCode=`
|
|
${l} strides = ${l}(${r});
|
|
|
|
void main() {
|
|
${u} coords = getOutputCoords();
|
|
float sum = 0.0;
|
|
bool found = false;
|
|
for (int i = 0; i < ${e}; i++) {
|
|
int flattenedIndex = 0;
|
|
for (int j = 0; j < ${t}; j++) {
|
|
int index = round(${c});
|
|
flattenedIndex += index * ${g};
|
|
}
|
|
if (flattenedIndex == coords[0]) {
|
|
sum += ${h};
|
|
found = true;
|
|
}
|
|
}
|
|
setOutput(mix(${m}, sum, float(found)));
|
|
}
|
|
`}},aQ=class{constructor(e,t,a,n,r,s,i=!0,o=!1){this.variableNames=["updates","indices","defaultValue"],this.packedInputs=!0,this.packedOutput=!0,this.outputShape=s;let l=ft(r.length),u=ft(s.length),p="";a===1?p="i":a===2&&(p="i, j");let c=`getIndices(${p})`,d="";n===1?d="i":n===2&&(d="i, coords[1]");let h=`getUpdates(${d})`,f="";o&&(f="coords[0], coords[1]");let m=`getDefaultValue(${f})`,g=t>1?"strides[j]":"strides",y=t>1?"strides[j + 1]":"strides";this.userCode=`
|
|
${l} strides = ${l}(${r});
|
|
|
|
void main() {
|
|
${u} coords = getOutputCoords();
|
|
vec4 sum = vec4(0.);
|
|
vec4 found = vec4(0.);
|
|
for (int i = 0; i < ${e}; i+=2) {
|
|
ivec2 flattenedIndex = ivec2(0);
|
|
for (int j = 0; j < ${t}; j+=2) {
|
|
ivec4 index = round(${c});
|
|
flattenedIndex += index.xz * ${g};
|
|
if (j + 1 < ${t}) {
|
|
flattenedIndex += index.yw * ${y};
|
|
}
|
|
}
|
|
if (flattenedIndex[0] == coords[0] || flattenedIndex[1] == coords[0] ||
|
|
flattenedIndex[0] == coords[0] + 1 || flattenedIndex[1] == coords[0] + 1) {
|
|
vec4 updVals = ${h};
|
|
if (flattenedIndex[0] == coords[0]) {
|
|
sum.xy += updVals.xy;
|
|
found.xy = vec2(1.);
|
|
} else if (flattenedIndex[0] == coords[0] + 1) {
|
|
sum.zw += updVals.xy;
|
|
found.zw = vec2(1.);
|
|
}
|
|
if (flattenedIndex[1] == coords[0]) {
|
|
sum.xy += updVals.zw;
|
|
found.xy = vec2(1.);
|
|
} else if (flattenedIndex[1] == coords[0] + 1) {
|
|
sum.zw += updVals.zw;
|
|
found.zw = vec2(1.);
|
|
}
|
|
}
|
|
}
|
|
setOutput(mix(${m}, sum, found));
|
|
}
|
|
`}};function nQ(e){let{inputs:t,backend:a,attrs:n}=e,{indices:r,updates:s}=t,{shape:i}=n,{sliceRank:o,numUpdates:l,sliceSize:u,strides:p,outputSize:c}=S.calculateShapes(s,r,i),d=[c/u,u];if(c===0)return a.makeTensorInfo(i,r.dtype);let h=ue({inputs:{x:r},backend:a,attrs:{shape:[l,o]}}),f=ue({inputs:{x:s},backend:a,attrs:{shape:[l,u]}}),m=a.makeTensorInfo([],"float32",new Float32Array([0])),g;W().getBool("WEBGL_PACK")?g=new aQ(l,o,h.shape.length,f.shape.length,p,d):g=new Cg(l,o,h.shape.length,f.shape.length,p,d);let y=a.runWebGLProgram(g,[f,h,m],f.dtype),x=ue({inputs:{x:y},backend:a,attrs:{shape:i}});return a.disposeIntermediateTensorInfo(h),a.disposeIntermediateTensorInfo(f),a.disposeIntermediateTensorInfo(y),a.disposeIntermediateTensorInfo(m),x}var rQ={kernelName:Io,backendName:"webgl",kernelFunc:nQ},sQ=class{constructor(e,t,a,n){this.variableNames=["sortedSequence","values"],this.customUniforms=[{name:"numInputs",type:"int"}],this.outputShape=[e,a];let r="while (left < right) {",s=`for (int i = 0; i < ${Math.ceil(Math.log2(t+1))}; ++i) { if (left >= right) break;`,i=W().getNumber("WEBGL_VERSION")===2?r:s,o=n==="left"?"<":"<=";this.userCode=`
|
|
int findBound(int batch, float value) {
|
|
int left = 0;
|
|
int right = numInputs;
|
|
int mid;
|
|
${i}
|
|
mid = (left + right) / 2;
|
|
if (getSortedSequence(batch, mid) ${o} value) {
|
|
left = mid + 1;
|
|
} else {
|
|
right = mid;
|
|
}
|
|
}
|
|
return right;
|
|
}
|
|
|
|
void main() {
|
|
ivec2 coords = getOutputCoords();
|
|
int batch = coords[0];
|
|
int valueIndex = coords[1];
|
|
|
|
float value = getValues(batch, valueIndex);
|
|
|
|
setOutput(float(findBound(batch, value)));
|
|
}
|
|
`}};function iQ(e){let{inputs:t,backend:a,attrs:n}=e,{sortedSequence:r,values:s}=t,{side:i}=n,o=new sQ(r.shape[0],r.shape[1],s.shape[1],i),l=[[r.shape[1]]];return a.runWebGLProgram(o,[r,s],"int32",l)}var oQ={kernelName:Co,backendName:"webgl",kernelFunc:iQ},lQ=class{constructor(e,t,a){this.variableNames=["c","a","b"],this.outputShape=t;let n,r;if(a>4)throw Error(`Where for rank ${a} is not yet supported`);if(a===1)r="resRC",n="resRC";else{let i=["resRC.x","resRC.y","resRC.z","resRC.w"],o=[],l=[];for(let u=0;u<t.length;u++)l.push(`${i[u]}`),u<e&&o.push(`${i[u]}`);n=o.join(),r=l.join()}let s=ft(a);this.userCode=`
|
|
void main() {
|
|
${s} resRC = getOutputCoords();
|
|
float cVal = getC(${n});
|
|
if (cVal >= 1.0) {
|
|
setOutput(getA(${r}));
|
|
} else {
|
|
setOutput(getB(${r}));
|
|
}
|
|
}
|
|
`}};function uQ(e){let{inputs:t,backend:a}=e,{condition:n,t:r,e:s}=t,i=new lQ(n.shape.length,r.shape,r.shape.length);return a.runWebGLProgram(i,[n,r,s],pa(r.dtype,s.dtype))}var dQ={kernelName:ku,backendName:"webgl",kernelFunc:uQ},pQ=`
|
|
// Stable and Attracting Fixed Point (0, 1) for Normalized Weights.
|
|
// see: https://arxiv.org/abs/1706.02515
|
|
float scaleAlpha = ${S.SELU_SCALEALPHA};
|
|
float scale = ${S.SELU_SCALE};
|
|
return (x >= 0.0) ? scale * x : scaleAlpha * (exp(x) - 1.0);
|
|
`,cQ=et({opSnippet:pQ}),hQ={kernelName:To,backendName:"webgl",kernelFunc:cQ},fQ=Vu+`
|
|
return 1.0 / (1.0 + exp(-1.0 * x));
|
|
`,mQ=`
|
|
vec4 result = 1.0 / (1.0 + exp(-1.0 * x));
|
|
bvec4 isNaN = isnan(x);
|
|
|
|
result.r = isNaN.r ? x.r : result.r;
|
|
result.g = isNaN.g ? x.g : result.g;
|
|
result.b = isNaN.b ? x.b : result.b;
|
|
result.a = isNaN.a ? x.a : result.a;
|
|
|
|
return result;
|
|
`,gQ=et({opSnippet:fQ,packedOpSnippet:mQ,cpuKernelImpl:yH}),yQ={kernelName:Eo,backendName:"webgl",kernelFunc:gQ},xQ=`
|
|
if (isnan(x)) { return 0.0; }
|
|
return sign(x);
|
|
`,AQ=et({opSnippet:xQ}),bQ={kernelName:Ro,backendName:"webgl",kernelFunc:AQ},vQ=Vu+`
|
|
return sin(x);
|
|
`,wQ=`
|
|
vec4 result = sin(x);
|
|
bvec4 isNaN = isnan(x);
|
|
${Xo}
|
|
return result;
|
|
`,kQ=et({opSnippet:vQ,packedOpSnippet:wQ}),IQ={kernelName:No,backendName:"webgl",kernelFunc:kQ},SQ=`
|
|
float e2x = exp(x);
|
|
return (e2x - 1.0 / e2x) / 2.0;
|
|
`,CQ=et({opSnippet:SQ}),TQ={kernelName:Su,backendName:"webgl",kernelFunc:CQ},NQ=`
|
|
float epsilon = 1.1920928955078125e-7;
|
|
float threshold = log(epsilon) + 2.0;
|
|
|
|
bool too_large = x > -threshold;
|
|
bool too_small = x < threshold;
|
|
|
|
float result;
|
|
float exp_x = exp(x);
|
|
|
|
if (too_large){
|
|
result = x;
|
|
}
|
|
else if (too_small){
|
|
result = exp_x;
|
|
}
|
|
else{
|
|
result = log(exp_x + 1.0);
|
|
}
|
|
return result;
|
|
`,RQ=et({opSnippet:NQ}),EQ={kernelName:Mo,backendName:"webgl",kernelFunc:RQ},MQ=e=>{let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{blockShape:s,paddings:i}=n;v.assert(r.shape.length<=4,()=>"spaceToBatchND for rank > 4 with a WebGL backend not implemented yet");let o=s.reduce((y,x)=>y*x),l=[[0,0]];l.push(...i);for(let y=1+s.length;y<r.shape.length;++y)l.push([0,0]);let u=[],p=g8({inputs:{x:r},backend:a,attrs:{paddings:l,constantValue:0}}),c=S.getReshaped(p.shape,s,o,!1),d=S.getPermuted(c.length,s.length,!1),h=S.getReshapedPermuted(p.shape,s,o,!1),f=ue({inputs:{x:p},backend:a,attrs:{shape:c}}),m=Ta({inputs:{x:f},backend:a,attrs:{perm:d}}),g=ue({inputs:{x:m},backend:a,attrs:{shape:h}});return u.push(p),u.push(f),u.push(m),u.forEach(y=>a.disposeIntermediateTensorInfo(y)),g},FQ={kernelName:Cu,backendName:"webgl",kernelFunc:MQ};function _Q(e){let{inputs:t,backend:a}=e,{indices:n,values:r,denseShape:s,defaultValue:i}=t;if(s.shape.length!==1)throw new Error(`Dense shape must be a vector, saw:
|
|
${s.shape}`);if(n.shape.length!==2)throw new Error(`Indices must be a matrix, saw:
|
|
${n.shape}`);if(r.shape.length!==1)throw new Error(`Values must be a vector, saw:
|
|
${r.shape}`);if(i.shape.length!==0)throw new Error(`Default value must be a scalar, saw:
|
|
${i.shape}`);let o=a.readSync(n.dataId),l=a.readSync(r.dataId),u=a.readSync(s.dataId),p=a.readSync(i.dataId)[0],[c,d,h,f,m]=AH(o,n.shape,n.dtype,l,r.dtype,u,p);return[a.makeTensorInfo(d,n.dtype,c),a.makeTensorInfo([d[0]],r.dtype,h),a.makeTensorInfo([f.length],"bool",new Uint8Array(f.map(g=>Number(g)))),a.makeTensorInfo([m.length],n.dtype,new Int32Array(m))]}var $Q={kernelName:hp,backendName:"webgl",kernelFunc:_Q};function DQ(e){let{inputs:t,backend:a}=e,{inputIndices:n,inputShape:r,newShape:s}=t;if(n.shape.length!==2)throw new Error(`Input indices should be a matrix but received shape ${n.shape}`);if(r.shape.length!==1)throw new Error(`Input shape should be a vector but received shape ${r.shape}`);if(s.shape.length!==1)throw new Error(`Target shape should be a vector but received shape ${s.shape}`);let i=Array.from(a.readSync(r.dataId)),o=a.readSync(n.dataId),l=Array.from(a.readSync(s.dataId)),[u,p,c]=bH(o,n.shape,n.dtype,i,l);return[a.makeTensorInfo(p,n.dtype,u),a.makeTensorInfo([c.length],s.dtype,new Int32Array(c))]}var PQ={kernelName:Nu,backendName:"webgl",kernelFunc:DQ};function OQ(e){let{inputs:t,backend:a}=e,{data:n,indices:r,segmentIds:s}=t;if(n.shape.length<1)throw new Error("Data should be at least 1 dimensional but received scalar");if(r.shape.length!==1)throw new Error(`Indices should be a vector but received shape
|
|
${r.shape}`);if(s.shape.length!==1)throw new Error(`Segment ids should be a vector but received shape
|
|
${s.shape}`);let i=a.readSync(n.dataId),o=a.readSync(r.dataId),l=a.readSync(s.dataId),[u,p]=Wv(i,n.shape,n.dtype,o,l,!0);return a.makeTensorInfo(p,n.dtype,u)}var zQ={kernelName:fp,backendName:"webgl",kernelFunc:OQ};function LQ(e){let{inputs:t,backend:a}=e,{data:n,indices:r,segmentIds:s}=t;if(n.shape.length<1)throw new Error("Data should be at least 1 dimensional but received scalar");if(r.shape.length!==1)throw new Error(`Indices should be a vector but received shape
|
|
${r.shape}`);if(s.shape.length!==1)throw new Error(`Segment ids should be a vector but received shape
|
|
${s.shape}`);let i=a.readSync(n.dataId),o=a.readSync(r.dataId),l=a.readSync(s.dataId),[u,p]=Wv(i,n.shape,n.dtype,o,l);return a.makeTensorInfo(p,n.dtype,u)}var WQ={kernelName:mp,backendName:"webgl",kernelFunc:LQ};function BQ(e){let{inputs:t,backend:a,attrs:n}=e,{sparseIndices:r,sparseValues:s,defaultValue:i}=t,{outputShape:o}=n,{sliceRank:l,numUpdates:u,sliceSize:p,strides:c,outputSize:d}=S.calculateShapes(s,r,o),h=!1;if(s.dtype==="string"){let y=a.bufferSync(r),x=a.bufferSync(s),A=v.decodeString(a.readSync(i.dataId)[0]),b=gH(y,x,o,d,p,u,l,c,A,h);return a.makeTensorInfo(o,b.dtype,b.values)}let f=new Cg(u,l,r.shape.length,s.shape.length,c,[d,1],h),m=a.runWebGLProgram(f,[s,r,i],s.dtype),g=ue({inputs:{x:m},backend:a,attrs:{shape:o}});return a.disposeIntermediateTensorInfo(m),g}var VQ={kernelName:Do,backendName:"webgl",kernelFunc:BQ};function UQ(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{numOrSizeSplits:s,axis:i}=n,o=v.parseAxisParam(i,r.shape)[0],l=S.prepareSplitSize(r,s,o),u=r.shape.length,p=new Array(u).fill(0),c=r.shape.slice();return l.map(d=>{let h=[...c];h[o]=d;let f=Uu({inputs:{x:r},backend:a,attrs:{begin:p,size:h}});return p[o]+=d,f})}var GQ={kernelName:Tu,backendName:"webgl",kernelFunc:UQ},Ix="return sqrt(x);",HQ=et({opSnippet:Ix,packedOpSnippet:Ix,cpuKernelImpl:vH}),jQ={kernelName:Fo,backendName:"webgl",kernelFunc:HQ},qQ="return x * x;",XQ=et({opSnippet:qQ}),KQ={kernelName:gp,backendName:"webgl",kernelFunc:XQ},Sx="return (a - b) * (a - b);",YQ=fa({opSnippet:Sx,packedOpSnippet:Sx}),ZQ={kernelName:Po,backendName:"webgl",kernelFunc:YQ};function JQ(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t;if(r.dtype!=="string")throw new Error("Input must be of datatype string");let s=a.readSync(r.dataId),i=S.fromUint8ToStringArray(s),o=wH(i,"string",n);return a.makeTensorInfo(r.shape,"string",o)}var QQ={kernelName:yp,backendName:"webgl",kernelFunc:JQ};function eee({inputs:e,attrs:t,backend:a}){let{x:n}=e,r=Rn+`
|
|
return x > 0.0 ? 1.0 : float(${t.alpha});
|
|
`,s=new Yn(n.shape,r);return a.runWebGLProgram(s,[n],n.dtype)}var tee={kernelName:is,backendName:"webgl",kernelFunc:eee},aee=class{constructor(e,t,a){this.variableNames=["x"],this.outputShape=a;let n=a.length,r=ft(a.length),s=ft(a.length),i="";if(n===1)i="coords * strides + begin";else{let o=0;i=a.map((l,u)=>(o++,a.length===1?`coords * strides[${u}] + begin[${u}]`:`coords[${o-1}] * strides[${u}] + begin[${u}]`)).join(",")}this.userCode=`
|
|
${r} begin = ${r}(${e});
|
|
${r} strides = ${r}(${t});
|
|
|
|
void main() {
|
|
${s} coords = getOutputCoords();
|
|
setOutput(getX(${i}));
|
|
}
|
|
`}};function nee(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{begin:s,end:i,strides:o,beginMask:l,endMask:u,ellipsisMask:p,newAxisMask:c,shrinkAxisMask:d}=n,{finalShapeSparse:h,finalShape:f,isIdentity:m,sliceDim0:g,isSimpleSlice:y,begin:x,end:A,strides:b}=Nt.sliceInfo(r.shape,s,i,o,l,u,p,c,d),w;if(m)w=ue({inputs:{x:r},backend:a,attrs:{shape:f}});else if(g||y){v.assert(r.shape.length>=1,()=>`Input must have rank at least 1, got: ${r.shape.length}`);let T=Nt.computeOutShape(x,A,b),N=Uu({inputs:{x:r},backend:a,attrs:{begin:x,size:T}});w=ue({inputs:{x:N},backend:a,attrs:{shape:f}}),a.disposeIntermediateTensorInfo(N)}else if(a.shouldExecuteOnCPU([r])){let T=a.readSync(r.dataId),N=$e(r.shape,r.dtype,T),M=kH(h,N,b,x);w=a.makeTensorInfo(f,r.dtype,M.values)}else{let T=new aee(x,b,h);w=a.runWebGLProgram(T,[r],r.dtype)}let C=ue({inputs:{x:w},backend:a,attrs:{shape:f}});return a.disposeIntermediateTensorInfo(w),C}var ree={kernelName:Oo,backendName:"webgl",kernelFunc:nee};function see(e){let{inputs:t,backend:a,attrs:n}=e,{separator:r,nGramWidths:s,leftPad:i,rightPad:o,padWidth:l,preserveShortSequences:u}=n,{data:p,dataSplits:c}=t,d=a.readSync(p.dataId),h=a.readSync(c.dataId),[f,m]=IH(d,h,r,s,i,o,l,u);return[a.makeTensorInfo([f.length],"string",f),a.makeTensorInfo(c.shape,"int32",m)]}var iee={kernelName:Ru,backendName:"webgl",kernelFunc:see};function oee(e){let{inputs:t,backend:a,attrs:n}=e,{skipEmpty:r}=n,{input:s,delimiter:i}=t;if(s.dtype!=="string")throw new Error("Input must be of datatype string");if(s.shape.length!==1)throw new Error(`Input must be a vector, got shape: ${s.shape}`);if(i.shape.length!==0)throw new Error(`Delimiter must be a scalar, got shape: ${i.shape}`);let o=a.readSync(s.dataId),l=a.readSync(i.dataId)[0],[u,p,c]=SH(o,l,r),d=p.length;return[a.makeTensorInfo([d,2],"int32",u),a.makeTensorInfo([d],"string",p),a.makeTensorInfo([2],"int32",new Int32Array(c))]}var lee={kernelName:xp,backendName:"webgl",kernelFunc:oee};function uee(e){let{inputs:t,backend:a,attrs:n}=e,{numBuckets:r}=n,{input:s}=t;if(s.dtype!=="string")throw new Error("Input must be of datatype string");if(r<=0)throw new Error("Number of buckets must be at least 1");let i=a.readSync(s.dataId),o=CH(i,r);return a.makeTensorInfo(s.shape,"int32",o)}var dee={kernelName:Ap,backendName:"webgl",kernelFunc:uee},pee="return tan(x);",cee=et({opSnippet:pee}),hee={kernelName:Lo,backendName:"webgl",kernelFunc:cee},fee=`
|
|
float e2x = exp(-2.0 * abs(x));
|
|
return sign(x) * (1.0 - e2x) / (1.0 + e2x);
|
|
`,mee=et({opSnippet:fee}),gee={kernelName:Wo,backendName:"webgl",kernelFunc:mee};function yee(e){let{inputs:t,backend:a,attrs:n}=e,{tensor:r,indices:s,updates:i}=t,{}=n,{sliceRank:o,numUpdates:l,sliceSize:u,strides:p,outputSize:c}=S.calculateShapes(i,s,r.shape),d=[c/u,u];if(c===0)return a.makeTensorInfo(r.shape,s.dtype);let h=ue({inputs:{x:s},backend:a,attrs:{shape:[l,o]}}),f=ue({inputs:{x:i},backend:a,attrs:{shape:[l,u]}}),m=ue({inputs:{x:r},backend:a,attrs:{shape:d}}),g=new Cg(l,o,h.shape.length,f.shape.length,p,d,!1,!0),y=a.runWebGLProgram(g,[f,h,m],m.dtype),x=ue({inputs:{x:y},backend:a,attrs:{shape:r.shape}});return a.disposeIntermediateTensorInfo(h),a.disposeIntermediateTensorInfo(f),a.disposeIntermediateTensorInfo(m),a.disposeIntermediateTensorInfo(y),x}var xee={kernelName:So,backendName:"webgl",kernelFunc:yee},Aee=class{constructor(e,t){this.variableNames=["A"];let a=new Array(e.length);for(let s=0;s<a.length;s++)a[s]=e[s]*t[s];this.outputShape=a,this.rank=a.length;let n=ft(this.rank),r=bee(e);this.userCode=`
|
|
void main() {
|
|
${n} resRC = getOutputCoords();
|
|
setOutput(getA(${r}));
|
|
}
|
|
`}};function bee(e){let t=e.length;if(t>5)throw Error(`Tile for rank ${t} is not yet supported`);if(t===1)return`imod(resRC, ${e[0]})`;let a=["resRC.x","resRC.y","resRC.z","resRC.w","resRC.u"],n=[];for(let r=0;r<e.length;r++)n.push(`imod(${a[r]}, ${e[r]})`);return n.join()}function x8(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{reps:s}=n;if(r.dtype==="string"||r.shape.length>5){let o=a.readSync(r.dataId),l=r.dtype==="string"?o.map(c=>v.decodeString(c)):o,u=$e(r.shape,r.dtype,l),p=NH(u,s);return a.makeTensorInfo(p.shape,p.dtype,p.values)}let i=new Aee(r.shape,s);return a.runWebGLProgram(i,[r],r.dtype)}var vee={kernelName:ss,backendName:"webgl",kernelFunc:x8},wee=class{constructor(e){this.variableNames=["x","indices"],this.customUniforms=[{name:"n",type:"int"},{name:"firstPass",type:"int"},{name:"negativeInf",type:"float"},{name:"dir",type:"int"},{name:"inc",type:"int"}],this.outputShape=e,this.userCode=`
|
|
void main() {
|
|
ivec2 coords = getOutputCoords();
|
|
int batch = coords[0];
|
|
int elemIdx = coords[1];
|
|
|
|
// We compare elements pair-wise within a group of size 2 * inc.
|
|
// The comparing rule for each group alternates between ascending
|
|
// and descending. Within each group, we compare each pair at
|
|
// positions i and i+inc. To decide whether an element at position i
|
|
// is x0 or x1, we mod it by 2 * inc, if the result is smaller than
|
|
// inc, it is in the first half of the group, we denote it as x0,
|
|
// otherwise we denote it as x1.
|
|
// For example, as shown in the Bitonic top K paper referenced above,
|
|
// Figure5(a) shows that element[1] is in the
|
|
// second half of the group when group size is 2, but it is in the
|
|
// first half of the group when group size is 4.
|
|
|
|
bool isFirstInPair = imod(elemIdx, 2 * inc) < inc;
|
|
int i = isFirstInPair ? elemIdx : elemIdx - inc;
|
|
|
|
int i0 = firstPass == 1 ? i : int(getIndices(batch, i));
|
|
int i1 = firstPass == 1 ? i + inc : int(getIndices(batch, i + inc));
|
|
float x0 = i0 < n ? getX(batch, i0) : negativeInf;
|
|
float x1 = i1 < n ? getX(batch, i1) : negativeInf;
|
|
|
|
// Denotes which direction indices are in (ascending or descending).
|
|
bool reverse = imod(elemIdx, 2 * dir) >= dir;
|
|
bool isGreater = x0 > x1 || (x0 == x1 && i1 > i0);
|
|
if (reverse == isGreater) { // Elements in opposite order of direction
|
|
int iTemp = i0;
|
|
i0 = i1;
|
|
i1 = iTemp;
|
|
}
|
|
if (isFirstInPair) {
|
|
setOutput(float(i0));
|
|
} else {
|
|
setOutput(float(i1));
|
|
}
|
|
}
|
|
`}},kee=class{constructor(e){this.variableNames=["x","indices"],this.customUniforms=[{name:"n",type:"int"},{name:"firstPass",type:"int"},{name:"k",type:"int"}],this.outputShape=e,this.userCode=`
|
|
void main() {
|
|
// Takes max of indices (0, k), (1, k + 1), (2, k + 2) ...
|
|
ivec2 coords = getOutputCoords();
|
|
int batch = coords[0];
|
|
int elemIdx = coords[1];
|
|
|
|
// The output size is half of the previous size.
|
|
// If the previous sequence is | | | | _ _ _ _ | | | | _ _ _ _ (k=4),
|
|
// we only need to output the indices at positions |, the indices at
|
|
// positions _ can be thrown away, see Figure5(b) After Phase 2
|
|
// (Merge phase) in the Bitonic Top K paper referenced above.
|
|
// For example, the paper shows we only need to output the orange bars.
|
|
// The output sequence should look like this | | | | | | | |.
|
|
// Because the sequence is halved, to map the output index back
|
|
// to the previous sequence to find the corresponding value,
|
|
// we need to double the index. When we double the index,
|
|
// we basically interpolate a position, so 2i looks like
|
|
// | _ | _ | _ | _ | _ | _ | _. We move the | to the first k position
|
|
// of each 2k positions by - elemIdx % k. E.g. for output at
|
|
// index 4,5,6,7, we want to get the corresponding element at
|
|
// original index 8,9,10,11, for output at index 8,9,10,11,
|
|
// we want to get the corresponding element at original index
|
|
// 16,17,18,19, so on and so forth.
|
|
|
|
int i = elemIdx < k ? elemIdx : (elemIdx * 2 - imod(elemIdx, k));
|
|
int i0 = firstPass == 1 ? i : int(getIndices(batch, i));
|
|
int i1 = firstPass == 1 ? i + k : int(getIndices(batch, i + k));
|
|
|
|
float x0 = getX(batch, i0);
|
|
float x1 = i1 < n ? getX(batch, i1) : x0;
|
|
|
|
setOutput(x0 >= x1 ? float(i0) : float(i1));
|
|
}
|
|
`}};function _s(e,t){t!==null&&e.disposeIntermediateTensorInfo(t)}function Cx(e){let t=1;for(;t<e;)t*=2;return t}function Iee(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{k:s,sorted:i}=n,o=W().getNumber("TOPK_LAST_DIM_CPU_HANDOFF_SIZE_THRESHOLD"),l=W().getNumber("TOPK_K_CPU_HANDOFF_THRESHOLD"),u=r.shape,p=u[u.length-1];if(a.shouldExecuteOnCPU([r])||p<o||s>l){let M=a.readSync(r.dataId),[F,E]=RH(M,u,r.dtype,s,i);return[a.makeTensorInfo(F.shape,F.dtype,F.values),a.makeTensorInfo(E.shape,E.dtype,E.values)]}if(s===0)return u[u.length-1]=0,[a.makeTensorInfo(u,r.dtype,[]),a.makeTensorInfo(u,"int32",[])];if(p===1)return[r,Hp({attrs:{shape:u,dtype:"int32",value:0},backend:a})];let c=a.texData.get(r.dataId),d=c!==null&&c.isPacked,h=d?a.unpackTensor(r):r,f=v.sizeFromShape(u)/p,m=ue({inputs:{x:h},attrs:{shape:[f,p]},backend:a});d&&_s(a,h);let g=Cx(s),y=Cx(p),x=null,A=()=>x===null?[m,m]:[m,x],b=(M,F,E)=>{let I=A(),_=new wee(E),O=[[p],[x===null?1:0],[Number.NEGATIVE_INFINITY],[M],[F]],B=x;x=a.runWebGLProgram(_,I,"int32",O),_s(a,B)};for(let M=1;M<g;M*=2){let F=M*2;for(let E=M;E>=1;E/=2)b(F,E,[f,y])}for(let M=y;M>g;M/=2){let F=A(),E=new kee([f,M/2]),I=[[p],[x===null?1:0],[g]],_=x;x=a.runWebGLProgram(E,F,"int32",I),_s(a,_);let O=g/2,B=O*2;for(let $=O;$>=1;$/=2)b(B,$,x.shape)}let w=x;x=Uu({inputs:{x},backend:a,attrs:{begin:0,size:[f,s]}}),_s(a,w);let C=d8({inputs:{x:m,indices:x},backend:a,attrs:{axis:1,batchDims:1}});_s(a,m);let T=u.slice(0,-1);T.push(s),w=x,x=ue({inputs:{x},attrs:{shape:T},backend:a}),_s(a,w);let N=C;return C=ue({inputs:{x:C},attrs:{shape:T},backend:a}),_s(a,N),[C,x]}var See={kernelName:Bo,backendName:"webgl",kernelFunc:Iee},Cee=class{constructor(e,t,a,n,r,s){this.variableNames=["Image","Transforms"],this.outputShape=s;let i=a==="nearest"?1:2,o;switch(n){case"constant":o=1;break;case"reflect":o=2;break;case"wrap":o=3;break;case"nearest":o=4;break;default:o=1;break}this.userCode=`
|
|
float mapCoord(float outCoord, float len) {
|
|
float inCoord = outCoord;
|
|
if(${o} == 2) {
|
|
if (inCoord < 0.0) {
|
|
if (len <= 1.0) {
|
|
inCoord = 0.0;
|
|
} else {
|
|
float sz2 = 2.0 * len;
|
|
if (inCoord < sz2) {
|
|
inCoord = sz2 * float(int(float(-inCoord / sz2))) +
|
|
inCoord;
|
|
}
|
|
inCoord = inCoord < -len ? inCoord + sz2 : -inCoord - 1.0;
|
|
}
|
|
} else if (inCoord > len - 1.0) {
|
|
if (len <= 1.0) {
|
|
inCoord = 0.0;
|
|
} else {
|
|
float sz2 = 2.0 * len;
|
|
inCoord -= sz2 * float(int(float(inCoord / sz2)));
|
|
if (inCoord >= len) {
|
|
inCoord = sz2 - inCoord - 1.0;
|
|
}
|
|
}
|
|
}
|
|
return clamp(inCoord, 0.0, len - 1.0);
|
|
} else if (${o} == 3) {
|
|
if (inCoord < 0.0) {
|
|
if (len <= 1.0) {
|
|
inCoord = 0.0;
|
|
} else {
|
|
float sz = len - 1.0;
|
|
inCoord += len * (float(int(float(-inCoord / sz))) + 1.0);
|
|
}
|
|
} else if (inCoord > len - 1.0) {
|
|
if (len <= 1.0) {
|
|
inCoord = 0.0;
|
|
} else {
|
|
float sz = len - 1.0;
|
|
inCoord -= len * float(int(float(inCoord / sz)));
|
|
}
|
|
}
|
|
return clamp(inCoord, 0.0, len - 1.0);
|
|
} else if (${o} == 4) {
|
|
return clamp(outCoord, 0.0, len - 1.0);
|
|
} else {
|
|
return outCoord;
|
|
}
|
|
}
|
|
|
|
float readWithFillValue(int batch, int coordY, int coordX,
|
|
int channel) {
|
|
float outputValue;
|
|
if (0 <= coordY && coordY < ${e} && 0 <= coordX && coordX < ${t}) {
|
|
outputValue = getImage(batch, coordY, coordX, channel);
|
|
} else {
|
|
outputValue = float(${r});
|
|
}
|
|
return outputValue;
|
|
}
|
|
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
float outputValue;
|
|
int batch = coords[0];
|
|
int x = coords[2];
|
|
int y = coords[1];
|
|
int channel = coords[3];
|
|
float xf = float(x);
|
|
float yf = float(y);
|
|
float a1 = getTransforms(batch, 0);
|
|
float a2 = getTransforms(batch, 1);
|
|
float a3 = getTransforms(batch, 2);
|
|
float b1 = getTransforms(batch, 3);
|
|
float b2 = getTransforms(batch, 4);
|
|
float b3 = getTransforms(batch, 5);
|
|
float c1 = getTransforms(batch, 6);
|
|
float c2 = getTransforms(batch, 7);
|
|
float projection = c1 * xf + c2 * yf + 1.0;
|
|
if (projection == 0.0) {
|
|
outputValue = float(${r});
|
|
} else {
|
|
float inX = (a1 * xf + a2 * yf + a3) / projection;
|
|
float inY = (b1 * xf + b2 * yf + b3) / projection;
|
|
float mapX = mapCoord(inX, float(${t}));
|
|
float mapY = mapCoord(inY, float(${e}));
|
|
|
|
if (${i} == 1) {
|
|
int coordY = int(round(mapY));
|
|
int coordX = int(round(mapX));
|
|
outputValue = readWithFillValue(batch, coordY, coordX,
|
|
channel);
|
|
} else {
|
|
float yFloor = floor(mapY);
|
|
float xFloor = floor(mapX);
|
|
float yCeil = yFloor + 1.0;
|
|
float xCeil = xFloor + 1.0;
|
|
float valueYFloor = (xCeil - mapX) *
|
|
readWithFillValue(batch, int(yFloor), int(xFloor), channel) +
|
|
(mapX - xFloor) *
|
|
readWithFillValue(batch, int(yFloor), int(xCeil), channel);
|
|
float valueYCeil = (xCeil - mapX) *
|
|
readWithFillValue(batch, int(yCeil), int(xFloor), channel) +
|
|
(mapX - xFloor) *
|
|
readWithFillValue(batch, int(yCeil), int(xCeil), channel);
|
|
outputValue = (yCeil - mapY) * valueYFloor +
|
|
(mapY - yFloor) * valueYCeil;
|
|
}
|
|
}
|
|
setOutput(outputValue);
|
|
}
|
|
`}};function Tee(e){let{inputs:t,backend:a,attrs:n}=e,{image:r,transforms:s}=t,{interpolation:i,fillMode:o,fillValue:l,outputShape:u}=n,[p,c,d,h]=r.shape,[f,m]=u!=null?u:[c,d],g=[p,f,m,h],y=new Cee(c,d,i,o,l,g);return a.runWebGLProgram(y,[r,s],"float32")}var Nee={kernelName:Vo,backendName:"webgl",kernelFunc:Tee};function Ree(e){let{inputs:t,attrs:a,backend:n}=e,{axis:r}=a,{x:s}=t;Pu(s,"unique"),console.warn("WARNING: ","UI might be locked temporarily as data is being downloaded");let i=n.readSync(s.dataId),{outputValues:o,outputShape:l,indices:u}=EH(i,r,s.shape,s.dtype);return[n.makeTensorInfo(l,s.dtype,o),n.makeTensorInfo([u.length],"int32",u)]}var Eee={kernelName:bp,backendName:"webgl",kernelFunc:Ree};function Mee(e){let{inputs:t,backend:a,attrs:n}=e,{value:r}=t,{axis:s}=n;s<0&&(s+=r.shape.length);let i=r,o=i.shape.length,l=r.shape[s],u=new Array(o-1),p=0;for(let m=0;m<o;m++)m!==s&&(u[p++]=i.shape[m]);let c=[],d=new Array(o).fill(0),h=i.shape.slice();h[s]=1;let f=new Array(l);for(let m=0;m<f.length;m++){d[s]=m;let g=Uu({inputs:{x:i},backend:a,attrs:{begin:d,size:h}}),y=ue({inputs:{x:g},backend:a,attrs:{shape:u}});f[m]=y,c.push(g)}return c.forEach(m=>a.disposeIntermediateTensorInfo(m)),f}var Fee={kernelName:Eu,backendName:"webgl",kernelFunc:Mee},_ee=class{constructor(e,t){this.variableNames=["x","segmentIds"];let a=e.windowSize,n=e.batchSize,r=e.inSize,s=e.numSegments,i=s*Math.ceil(r/a);this.outputShape=[n,i];let o="0.0",l="sumValue",u=Math.floor(a/4)*4,p=a%4,c=`
|
|
sumValue += dot(values, segFilter);
|
|
`,d="";r%a>0&&(d=`
|
|
if (inIdx < 0 || inIdx >= ${r}) {
|
|
return initializationValue;
|
|
}
|
|
`);let h="";r%a>0&&(h=`
|
|
if (inIdx < 0 || inIdx >= ${r}) {
|
|
return -1.0;
|
|
}
|
|
`),this.userCode=`
|
|
const float initializationValue = ${o};
|
|
|
|
float getValue(int batch, int inIdx) {
|
|
${d}
|
|
return getX(batch, inIdx);
|
|
}
|
|
|
|
float getSegmentIdAtIndex(int inIdx) {
|
|
${h}
|
|
return getSegmentIds(inIdx);
|
|
}
|
|
|
|
void main() {
|
|
ivec2 coords = getOutputCoords();
|
|
int batch = coords[0];
|
|
int outIdx = coords[1];
|
|
int inOffset = int(floor(float(outIdx) / float(
|
|
${s})) * float(${a}));
|
|
int currentSeg = int(mod(float(outIdx), float(${s})));
|
|
|
|
float sumValue = 0.0;
|
|
|
|
for (int i = 0; i < ${u}; i += 4) {
|
|
int inIdx = inOffset + i;
|
|
vec4 values = vec4(
|
|
getValue(batch, inIdx),
|
|
getValue(batch, inIdx + 1),
|
|
getValue(batch, inIdx + 2),
|
|
getValue(batch, inIdx + 3)
|
|
);
|
|
|
|
vec4 segFilter = vec4(
|
|
int(getSegmentIdAtIndex(inIdx)) == currentSeg ? 1 : 0,
|
|
int(getSegmentIdAtIndex(inIdx + 1)) == currentSeg ? 1 : 0,
|
|
int(getSegmentIdAtIndex(inIdx + 2)) == currentSeg ? 1 : 0,
|
|
int(getSegmentIdAtIndex(inIdx + 3)) == currentSeg ? 1 : 0
|
|
);
|
|
|
|
${c}
|
|
}
|
|
|
|
int inIdx = inOffset + ${u};
|
|
if (${p===1}) {
|
|
vec4 values = vec4(
|
|
getValue(batch, inIdx),
|
|
initializationValue,
|
|
initializationValue,
|
|
initializationValue
|
|
);
|
|
|
|
int inIdxSeg = int(getSegmentIdAtIndex(inIdx));
|
|
|
|
vec4 segFilter = vec4(
|
|
int(getSegmentIdAtIndex(inIdx)) == currentSeg ? 1 : 0,
|
|
0,
|
|
0,
|
|
0
|
|
);
|
|
|
|
${c}
|
|
} else if (${p===2}) {
|
|
vec4 values = vec4(
|
|
getValue(batch, inIdx),
|
|
getValue(batch, inIdx + 1),
|
|
initializationValue,
|
|
initializationValue
|
|
);
|
|
|
|
vec4 segFilter = vec4(
|
|
int(getSegmentIdAtIndex(inIdx)) == currentSeg ? 1 : 0,
|
|
int(getSegmentIdAtIndex(inIdx + 1)) == currentSeg ? 1 : 0,
|
|
0,
|
|
0
|
|
);
|
|
|
|
${c}
|
|
} else if (${p===3}) {
|
|
vec4 values = vec4(
|
|
getValue(batch, inIdx),
|
|
getValue(batch, inIdx + 1),
|
|
getValue(batch, inIdx + 2),
|
|
initializationValue
|
|
);
|
|
|
|
vec4 segFilter = vec4(
|
|
int(getSegmentIdAtIndex(inIdx)) == currentSeg ? 1 : 0,
|
|
int(getSegmentIdAtIndex(inIdx + 1)) == currentSeg ? 1 : 0,
|
|
int(getSegmentIdAtIndex(inIdx + 2)) == currentSeg ? 1 : 0,
|
|
0
|
|
);
|
|
|
|
${c}
|
|
}
|
|
setOutput(${l});
|
|
}
|
|
`}};function $ee(e){let{inputs:t,backend:a,attrs:n}=e,{x:r,segmentIds:s}=t,{numSegments:i}=n,o=r.shape.length,l=[],u=0,p=S.getAxesPermutation([u],o),c=r;p!=null&&(c=Ta({inputs:{x:r},backend:a,attrs:{perm:p}}),l.push(c),u=S.getInnerMostAxes(1,o)[0]);let d=S.segment_util.computeOutShape(c.shape,u,i),h=v.sizeFromShape([c.shape[u]]),f=ue({inputs:{x:c},backend:a,attrs:{shape:[-1,h]}});l.push(f);let m=Ip(r.dtype),g=(b,w,C,T,N)=>{let M=b.shape[0],F=b.shape[1],E=S.segment_util.segOpComputeOptimalWindowSize(F,N),I={windowSize:E,inSize:F,batchSize:M,numSegments:N},_=new _ee(I,w),O=a.compileAndRun(_,[b,C],T);if(l.push(O),O.shape[1]===N)return O;let B=y8({backend:a,attrs:{start:0,stop:N,step:1,dtype:"float32"}}),$=x8({inputs:{x:B},backend:a,attrs:{reps:[F/E]}});return l.push(B),l.push($),g(O,w,$,T,N)},y=g(f,"unsortedSegmentSum",s,m,i),x=ue({inputs:{x:y},backend:a,attrs:{shape:d}}),A=x;if(p!=null){l.push(x);let b=S.getUndoAxesPermutation(p);A=Ta({inputs:{x:A},backend:a,attrs:{perm:b}})}return l.forEach(b=>a.disposeIntermediateTensorInfo(b)),A}var Dee={kernelName:vp,backendName:"webgl",kernelFunc:$ee},Pee=[kj,Sj,Nj,Mj,_j,Pj,zj,Wj,Gj,jj,Kj,Jj,tq,sq,lq,dq,cq,gq,xq,bq,Iq,Mq,_q,Dq,Bq,Uq,qq,ij,Yq,tX,sX,pX,fX,gX,xX,bX,IX,TX,EX,FX,$X,PX,LX,BX,HX,qX,YX,QX,tK,sK,uK,hK,gK,AK,bK,wK,IK,CK,NK,EK,$K,OK,WK,VK,HK,XK,JK,aY,sj,rY,Qq,oY,dY,hY,lj,yY,vY,kY,TY,EY,$Y,OY,BY,HY,XY,YY,eZ,aZ,rZ,lZ,dZ,cZ,fZ,gZ,bZ,IZ,NZ,PZ,pj,WZ,UZ,jZ,KZ,Oq,JZ,eJ,aJ,sJ,uJ,dj,pJ,hJ,mJ,yJ,xJ,zq,FZ,vJ,SJ,RJ,hj,_J,PJ,WJ,UJ,qJ,KJ,JJ,tQ,rQ,oQ,dQ,hQ,yQ,bQ,IQ,TQ,Rq,$Z,EQ,FQ,$Q,PQ,zQ,WQ,VQ,GQ,jQ,KQ,ZQ,QQ,tee,ree,iee,lee,dee,_Z,bj,hee,gee,xee,vee,See,Nee,vj,Eee,Fee,Dee,QZ];for(let e of Pee)yn(e);var nt;(function(e){e[e.float32=0]="float32",e[e.int32=1]="int32",e[e.bool=2]="bool",e[e.string=3]="string",e[e.complex64=4]="complex64"})(nt||(nt={}));var Xd;(function(e){e[e.linear=0]="linear",e[e.relu=1]="relu",e[e.relu6=2]="relu6",e[e.prelu=3]="prelu",e[e.leakyrelu=4]="leakyrelu",e[e.sigmoid=5]="sigmoid",e[e.elu=6]="elu"})(Xd||(Xd={}));var A8;function Oee(e){A8=e.wasm.cwrap(Xr,null,["number","array","number","number","array","number","number","number","number","number","number","number","number"])}function zee(e){let{inputs:t,backend:a,attrs:n}=e,{a:r,b:s,bias:i,preluActivationWeights:o}=t;if(r.dtype!=="float32"||s.dtype!=="float32")throw new Error("_FusedMatMul for non non-float32 tensors not yet supported.");let{transposeA:l,transposeB:u,activation:p,leakyreluAlpha:c}=n,d=a.dataIdMap.get(r.dataId).id,h=a.dataIdMap.get(s.dataId).id,f=0;if(i!=null){let N=a.dataIdMap.get(i.dataId);if(N.shape.length!==1)throw new Error(`_FusedMatMul only supports rank-1 bias but got rank ${N.shape.length}.`);f=N.id}let m=o==null?0:a.dataIdMap.get(o.dataId).id,g=Xd[p];if(g==null)throw new Error(`${p} activation not yet supported for FusedConv2D in the wasm backend.`);let y=l?r.shape[2]:r.shape[1],x=u?s.shape[1]:s.shape[2],A=Ho.assertAndGetBroadcastShape(r.shape.slice(0,-2),s.shape.slice(0,-2)),b=a.makeOutput([...A,y,x],r.dtype),w=a.dataIdMap.get(b.dataId).id,C=new Uint8Array(new Int32Array(r.shape).buffer),T=new Uint8Array(new Int32Array(s.shape).buffer);return A8(d,C,r.shape.length,h,T,s.shape.length,l,u,g,f,m,c||0,w),b}var Lee={kernelName:Xr,backendName:"wasm",setupFunc:Oee,kernelFunc:zee};function at(e,t){let a;function n(s){a=s.wasm.cwrap(e,null,["number","number","number"])}function r(s){let{backend:i,inputs:{x:o}}=s,l=i.dataIdMap.get(o.dataId).id,u=i.makeOutput(o.shape,t||o.dtype),p=i.dataIdMap.get(u.dataId).id;return v.sizeFromShape(u.shape)===0||a(l,nt[o.dtype],p),u}return{kernelName:e,backendName:"wasm",setupFunc:n,kernelFunc:r}}var Wee=at(Zl),Bee=at(Qs),Vee=at(ei);function ea(e,t,a){let n;function r(i){n=i.wasm.cwrap(e,null,["number","array","number","number","array","number","number","number"])}function s(i){let{backend:o,inputs:l}=i,{a:u,b:p}=l,c=o.dataIdMap.get(u.dataId).id,d=o.dataIdMap.get(p.dataId).id,h=a!=null?a:u.dtype,f=S.assertAndGetBroadcastShape(u.shape,p.shape),m=o.makeOutput(f,h);if(v.sizeFromShape(f)===0)return m;let g=new Uint8Array(new Int32Array(u.shape).buffer),y=new Uint8Array(new Int32Array(p.shape).buffer),x=o.dataIdMap.get(m.dataId).id;return n(c,g,u.shape.length,d,y,p.shape.length,nt[u.dtype],x),m}return{kernelName:e,backendName:"wasm",setupFunc:r,kernelFunc:s}}var Uee=!0,Gee=ea(ns,Uee),b8;function Hee(e){b8=e.wasm.cwrap(ti,null,["array","number","number","number"])}function jee(e){let{inputs:t,backend:a}=e,n=a.makeOutput(t[0].shape,t[0].dtype);if(v.sizeFromShape(n.shape)===0)return n;let r=t.map(o=>a.dataIdMap.get(o.dataId).id),s=new Uint8Array(new Int32Array(r).buffer),i=a.dataIdMap.get(n.dataId).id;return b8(s,r.length,nt[n.dtype],i),n}var qee={kernelName:ti,backendName:"wasm",setupFunc:Hee,kernelFunc:jee};function a0(e){let{inputs:{x:t},backend:a}=e;if(t.dtype==="string")return Ve(a.readSync(t.dataId),t.shape,t.dtype);let n=a.makeOutput(t.shape,t.dtype),r=a.typedArrayFromHeap(t);return a.typedArrayFromHeap(n).set(r),n}var Xee={kernelName:zi,backendName:"wasm",kernelFunc:a0},v8;function Kee(e){v8=e.wasm.cwrap(wr,null,["number","array","number","number","number","array","number"])}function as(e){let{inputs:t,backend:a,attrs:n}=e,[r,s]=Zee(t.x.shape,n.perm),i=!0;for(let f=0;f<s.length;f++)s[f]!==f&&(i=!1);let o=Yee(t.x.shape,n.perm),l={dataId:t.x.dataId,shape:r,dtype:t.x.dtype};if(i){let f=a0({inputs:t,backend:a});return f.shape=o,f}let u=a.makeOutput(o,l.dtype),p=a.dataIdMap.get(l.dataId).id,c=a.dataIdMap.get(u.dataId).id,d=new Uint8Array(new Int32Array(s).buffer),h=new Uint8Array(new Int32Array(l.shape).buffer);return v8(p,h,l.shape.length,nt[l.dtype],c,d,s.length),u}function Yee(e,t){let a=new Array(e.length);for(let n=0;n<a.length;n++)a[n]=e[t[n]];return a}function Zee(e,t){let a=[],n=[];for(let r=0;r<e.length;++r)e[r]!==1&&a.push(e[r]),e[t[r]]!==1&&n.push(t[r]);for(let r=0;r<n.length;++r){let s=-1;for(let i=0;i<n.length;++i)n[i]>=r&&(s===-1||n[s]>n[i])&&(s=i);n[s]=r}return[a,n]}var Jee={kernelName:wr,backendName:"wasm",kernelFunc:as,setupFunc:Kee};function ps(e,t,a){let n=e.shape,r=e.shape.length,s=v.parseAxisParam(t,n),i=s,o=S.getAxesPermutation(i,r),l=null,u=!1;if(o!=null){let p=new Array(r);for(let d=0;d<p.length;d++)p[d]=n[o[d]];i=S.getInnerMostAxes(i.length,r),l=as({inputs:{x:e},attrs:{perm:o},backend:a});let c=a.dataIdMap.get(e.dataId).id;a.dataIdMap.get(l.dataId).id!==c&&(u=!0)}return{transposed:l,originalAxes:s,axes:i,inputWasTransposed:u}}var w8;function Qee(e){w8=e.wasm.cwrap(ai,null,["number, number, number"])}function ete(e){let{backend:t,inputs:a,attrs:n}=e,{axis:r,keepDims:s}=n,{x:i}=a,o=t.dataIdMap.get(i.dataId).id,l=i,{transposed:u,axes:p,originalAxes:c,inputWasTransposed:d}=ps(i,r,t);if(d){let x=t.dataIdMap.get(u.dataId).id;l=u,o=x}let h=l.shape.length;S.assertAxesAreInnerMostDims("all",p,h);let[f,m]=S.computeOutAndReduceShapes(l.shape,p),g=v.sizeFromShape(m),y=t.makeOutput(f,i.dtype);if(v.sizeFromShape(l.shape)!==0){let x=t.dataIdMap.get(y.dataId).id;w8(o,g,x)}if(d&&t.disposeData(u.dataId),s){let x=S.expandShapeToKeepDim(y.shape,c);y.shape=x}return y}var tte={kernelName:ai,backendName:"wasm",setupFunc:Qee,kernelFunc:ete},k8;function ate(e){k8=e.wasm.cwrap(ni,null,["number, number, number"])}function nte(e){let{backend:t,inputs:a,attrs:n}=e,{axis:r,keepDims:s}=n,{x:i}=a,o=t.dataIdMap.get(i.dataId).id,l=i,{transposed:u,axes:p,originalAxes:c,inputWasTransposed:d}=ps(i,r,t);if(d){let x=t.dataIdMap.get(u.dataId).id;l=u,o=x}let h=l.shape.length;S.assertAxesAreInnerMostDims("any",p,h);let[f,m]=S.computeOutAndReduceShapes(l.shape,p),g=v.sizeFromShape(m),y=t.makeOutput(f,i.dtype);if(v.sizeFromShape(l.shape)!==0){let x=t.dataIdMap.get(y.dataId).id;k8(o,g,x)}if(d&&t.disposeData(u.dataId),s){let x=S.expandShapeToKeepDim(y.shape,c);y.shape=x}return y}var rte={kernelName:ni,backendName:"wasm",setupFunc:ate,kernelFunc:nte};function I8(e){let t;function a(r){t=r.wasm.cwrap(e,null,["number","number","number","number","number"])}function n(r){let{backend:s,inputs:i,attrs:o}=r,{axis:l}=o,{x:u}=i,p=s.dataIdMap.get(u.dataId).id,c=p,d=u,{transposed:h,axes:f,inputWasTransposed:m}=ps(u,l,s);if(m){let w=s.dataIdMap.get(h.dataId).id;w!==p&&(d=h,c=w)}let g=d.shape.slice(0,-1),y=s.makeOutput(g,"int32"),x=s.dataIdMap.get(y.dataId).id,A=v.sizeFromShape(y.shape),b=d.shape[f[0]];return t(c,nt[d.dtype],A,b,x),m&&s.disposeData(h.dataId),y}return{kernelName:e,backendName:"wasm",setupFunc:a,kernelFunc:n}}var ste=I8(Jl),ite=I8(Ql),ote=at(ri),lte=at(si),ute=at(ii),dte=ea(li,!1),pte=at(oi),S8;function cte(e){S8=e.wasm.cwrap(ui,null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function hte(e){let{inputs:t,attrs:a,backend:n}=e,r=t.x,s=n.dataIdMap.get(r.dataId).id,{filterSize:i,strides:o,pad:l,dimRoundingMode:u}=a,p=S.computePool2DInfo(r.shape,i,o,1,l,u),c=p.filterHeight,d=p.filterWidth,h=p.padInfo.top,f=p.padInfo.right,m=p.padInfo.bottom,g=p.padInfo.left,y=p.strideHeight,x=p.strideWidth,A=p.inChannels;if(p.dataFormat!=="channelsLast")throw new Error(`wasm backend does not support dataFormat:'${p.dataFormat}'. Please use 'channelsLast'.`);if(p.dilationWidth!==1||p.dilationHeight!==1)throw new Error(`was backend only supports average pooling with dilation = [1, 1], got [${p.dilationHeight}, ${p.dilationWidth}].`);let b=n.makeOutput(p.outShape,"float32"),w=n.dataIdMap.get(b.dataId).id;return S8(s,r.shape[0],r.shape[1],r.shape[2],c,d,h,f,m,g,y,x,A,w),b}var fte={kernelName:ui,backendName:"wasm",setupFunc:cte,kernelFunc:hte},C8;function mte(e){C8=e.wasm.cwrap("AvgPool3D",null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function gte(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{filterSize:s,strides:i,pad:o,dimRoundingMode:l,dataFormat:u}=n,p=S.computePool3DInfo(r.shape,s,i,1,o,l,u),c=a.makeOutput(p.outShape,r.dtype);return C8(a.dataIdMap.get(r.dataId).id,a.dataIdMap.get(c.dataId).id,p.batchSize,p.inChannels,p.inDepth,p.inHeight,p.inWidth,p.outDepth,p.outHeight,p.outWidth,p.strideDepth,p.strideHeight,p.strideWidth,p.dilationDepth,p.dilationHeight,p.dilationWidth,p.effectiveFilterDepth,p.effectiveFilterHeight,p.effectiveFilterWidth,p.padInfo.front,p.padInfo.top,p.padInfo.left),c}var yte={kernelName:eu,backendName:"wasm",setupFunc:mte,kernelFunc:gte},T8;function xte(e){T8=e.wasm.cwrap("AvgPool3DGrad",null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function Ate(e){let{inputs:t,backend:a,attrs:n}=e,{dy:r,input:s}=t,{filterSize:i,strides:o,pad:l,dimRoundingMode:u}=n,p=S.computePool3DInfo(s.shape,i,o,1,l,u),c=a.makeOutput(s.shape,s.dtype);return T8(a.dataIdMap.get(r.dataId).id,a.dataIdMap.get(c.dataId).id,p.batchSize,p.inChannels,p.inDepth,p.inHeight,p.inWidth,p.outDepth,p.outHeight,p.outWidth,p.strideDepth,p.strideHeight,p.strideWidth,p.dilationDepth,p.dilationHeight,p.dilationWidth,p.effectiveFilterDepth,p.effectiveFilterHeight,p.effectiveFilterWidth,p.padInfo.front,p.padInfo.top,p.padInfo.left,p.filterDepth,p.filterHeight,p.filterWidth),c}var bte={kernelName:ep,backendName:"wasm",setupFunc:xte,kernelFunc:Ate};function Ba(e){let{inputs:t,attrs:a}=e,{x:n}=t,{shape:r}=a,s=v.sizeFromShape(n.shape),i=v.inferFromImplicitShape(r,s);return v.assert(s===v.sizeFromShape(i),()=>`new shape: ${i}, old shape: ${n.shape}. New shape and old shape must have the same number of elements.`),e.backend.incRef(n.dataId),{dataId:n.dataId,shape:i,dtype:n.dtype}}var vte={kernelName:bu,backendName:"wasm",kernelFunc:Ba},N8;function wte(e){N8=e.wasm.cwrap(di,null,["number","array","number","number","array","number","number","number","number"])}function kte(e){let{inputs:t,backend:a,attrs:n}=e,{a:r,b:s}=t,{transposeA:i,transposeB:o}=n;if(r.dtype!=="float32"||s.dtype!=="float32")throw new Error("BatchMatMul for non non-float32 tensors not yet supported.");let l=r.shape.length,u=s.shape.length,p=i?r.shape[l-2]:r.shape[l-1],c=o?s.shape[u-1]:s.shape[u-2],d=i?r.shape[l-1]:r.shape[l-2],h=o?s.shape[u-2]:s.shape[u-1],f=r.shape.slice(0,-2),m=s.shape.slice(0,-2),g=v.sizeFromShape(f),y=v.sizeFromShape(m),x=Ho.assertAndGetBroadcastShape(r.shape.slice(0,-2),s.shape.slice(0,-2)).concat([d,h]);v.assert(p===c,()=>`Error in matMul: inner shapes (${p}) and (${c}) of Tensors with shapes ${r.shape} and ${s.shape} and transposeA=${i} and transposeB=${o} must match.`);let A=i?[g,p,d]:[g,d,p],b=o?[y,h,c]:[y,c,h],w=Ba({inputs:{x:r},backend:a,attrs:{shape:A}}),C=Ba({inputs:{x:s},backend:a,attrs:{shape:b}}),T=a.dataIdMap.get(w.dataId).id,N=a.dataIdMap.get(C.dataId).id,M=i?w.shape[2]:w.shape[1],F=o?C.shape[1]:C.shape[2],E=Math.max(g,y),I=a.makeOutput([E,M,F],w.dtype),_=a.dataIdMap.get(I.dataId).id,O=new Uint8Array(new Int32Array(w.shape).buffer),B=new Uint8Array(new Int32Array(C.shape).buffer);return N8(T,O,w.shape.length,N,B,C.shape.length,i,o,_),a.disposeData(w.dataId),a.disposeData(C.dataId),I.shape=x,I}var Ite={kernelName:di,backendName:"wasm",setupFunc:wte,kernelFunc:kte};function Zs(e){let{inputs:{x:t},attrs:{begin:a,size:n},backend:r}=e,[s,i]=Nt.parseSliceParams(t,a,n),o=Nt.isSliceContinous(t.shape,s,i),l=r.readSync(t.dataId),u=r.makeOutput(i,t.dtype),p=v.computeStrides(t.shape),c=r.dataIdMap.get(u.dataId);if(o){let f=Nt.computeFlatOffset(s,p);return t.dtype==="string"?c.stringBytes=l.slice(f,f+v.sizeFromShape(i)):r.typedArrayFromHeap(u).set(l.subarray(f,f+v.sizeFromShape(i))),u}if(t.dtype==="string"){let f=dh(l,s,i,t.shape,t.dtype);return c.stringBytes=f,u}let d=r.typedArrayFromHeap(u),h=t.shape.length;if(h===2)Ste(l,p[0],d,s,i);else if(h===3)Cte(l,p[0],p[1],d,s,i);else if(h===4)Tte(l,p[0],p[1],p[2],d,s,i);else{let f=dh(l,s,i,t.shape,t.dtype);d.set(f)}return u}function Ste(e,t,a,n,r){let s=0,i=n[0],o=n[1],l=i+r[0];for(let u=i;u<l;u++){let p=u*t+o;a.set(e.subarray(p,p+r[1]),s),s+=r[1]}}function Cte(e,t,a,n,r,s){let i=0,o=r[0],l=r[1],u=r[2],p=o+s[0],c=l+s[1];for(let d=o;d<p;d++)for(let h=l;h<c;h++){let f=d*t+h*a+u;n.set(e.subarray(f,f+s[2]),i),i+=s[2]}}function Tte(e,t,a,n,r,s,i){let o=0,l=s[0],u=s[1],p=s[2],c=l+i[0],d=u+i[1],h=p+i[2],f=s[3];for(let m=l;m<c;m++)for(let g=u;g<d;g++)for(let y=p;y<h;y++){let x=m*t+g*a+y*n+f;r.set(e.subarray(x,x+i[3]),o),o+=i[3]}}var Nte={kernelName:Iu,backendName:"wasm",kernelFunc:Zs};function Rte(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{blockShape:s,crops:i}=n,o=s.reduce((y,x)=>y*x),l=S.getReshaped(r.shape,s,o),u=S.getPermuted(l.length,s.length),p=S.getReshapedPermuted(r.shape,s,o),c=S.getSliceBeginCoords(i,s.length),d=S.getSliceSize(p,i,s.length),h=Ba({inputs:{x:r},backend:a,attrs:{shape:l}}),f=as({inputs:{x:h},backend:a,attrs:{perm:u}}),m=Ba({inputs:{x:f},backend:a,attrs:{shape:p}}),g=Zs({inputs:{x:m},backend:a,attrs:{begin:c,size:d}});return a.disposeData(h.dataId),a.disposeData(f.dataId),a.disposeData(h.dataId),g}var Ete={kernelName:tu,backendName:"wasm",kernelFunc:Rte},R8;function Mte(e){R8=e.wasm.cwrap(pi,null,["number","number","boolean","number","number","number"])}function Fte(e){let{backend:t,inputs:a,attrs:n}=e,{x:r,weights:s}=a,{size:i}=n,o=s.shape.reduce((c,d)=>c*d,1)!==0,l=r.shape.length===1?[i]:[r.shape[0],i],u=t.makeOutput(l,s.dtype);function p(c){return t.dataIdMap.get(c.dataId).id}return R8(p(r),i,o,p(s),nt[s.dtype],p(u)),u}var _te={kernelName:pi,backendName:"wasm",setupFunc:Mte,kernelFunc:Fte};function $te(e){let{inputs:t,backend:a}=e,{s0:n,s1:r}=t,s=a.typedArrayFromHeap(n),i=a.typedArrayFromHeap(r),o=S.assertAndGetBroadcastShape(Array.from(s),Array.from(i));return a.makeOutput([o.length],"int32",void 0,new Int32Array(o))}var Dte={kernelName:au,backendName:"wasm",kernelFunc:$te};function cs(e){let{inputs:{x:t},attrs:{dtype:a},backend:n}=e,r=n.makeOutput(t.shape,a),s=n.typedArrayFromHeap(t);return n.typedArrayFromHeap(r).set(s),r}var Pte={kernelName:ci,backendName:"wasm",kernelFunc:cs},Ote=at(hi),E8;function zte(e){E8=e.wasm.cwrap(rs,null,["number","number","number","number"])}function Lte(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{clipValueMin:s,clipValueMax:i}=n,o=a.dataIdMap.get(r.dataId).id,l=a.makeOutput(r.shape,r.dtype),u=a.dataIdMap.get(l.dataId).id;return E8(o,s,i,u),l}var Wte={kernelName:rs,backendName:"wasm",setupFunc:zte,kernelFunc:Lte};function M8(e){let{inputs:t,backend:a}=e,n=v.parseAxisParam(e.attrs.axis,t[0].shape)[0],r=t.map(h=>h.shape);S.assertParamsConsistent(r,n);let s=S.computeOutShape(t.map(h=>h.shape),n),i=t.filter(h=>v.sizeFromShape(h.shape)>0);if(i.length===1)return a0({inputs:{x:i[0]},backend:a});let o=a.makeOutput(s,t[0].dtype);if(v.sizeFromShape(s)===0)return o;if(i[0].dtype==="string"){let h=i.map(A=>{let b=[-1,v.sizeFromShape(A.shape.slice(n))];return Ba({inputs:{x:A},backend:a,attrs:{shape:b}})}),f=h.map(A=>({vals:a.readSync(A.dataId),shape:A.shape}));s=S.computeOutShape(h.map(A=>A.shape),1);let m=h[0].shape[0]===1,g=Q3(f,s,t[0].dtype,m),y=S.computeOutShape(i.map(A=>A.shape),n);o.shape=y;let x=a.dataIdMap.get(o.dataId);return x.stringBytes=S.fromStringArrayToUint8(g),h.forEach(A=>a.disposeData(A.dataId)),o}let l=v.sizeFromShape(i[0].shape.slice(0,n)),u=0,p=i.map(h=>{let f=v.sizeFromShape(h.shape.slice(n));return u+=f,f}),c=i.map(h=>a.typedArrayFromHeap(h)),d=a.typedArrayFromHeap(o);for(let h=0;h<l;h++){let f=h*u;for(let m=0;m<c.length;m++){let g=p[m],y=h*g,x=c[m].subarray(y,y+g);d.set(x,f),f+=g}}return o}var Bte={kernelName:nu,backendName:"wasm",kernelFunc:M8},F8;function Vte(e){F8=e.wasm.cwrap(fi,null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function Ute(e){let{inputs:t,attrs:a,backend:n}=e,{x:r,filter:s}=t,i=n.dataIdMap.get(r.dataId).id,o=n.dataIdMap.get(s.dataId).id,{strides:l,dilations:u,pad:p,dimRoundingMode:c,dataFormat:d}=a,h=S.convertConv2DDataFormat(d),f=S.computeConv2DInfo(r.shape,s.shape,l,u,p,c,!1,h),m=f.filterHeight,g=f.filterWidth,y=f.padInfo.top,x=f.padInfo.right,A=f.padInfo.bottom,b=f.padInfo.left,w=f.dilationHeight,C=f.dilationWidth,T=f.strideHeight,N=f.strideWidth,M=f.inChannels,F=f.outChannels,E=f.padInfo.type==="SAME"?1:0;if(f.dataFormat!=="channelsLast")throw new Error(`wasm backend Conv2D does not support dataFormat:'${f.dataFormat}'. Please use 'channelsLast'.`);let I=n.makeOutput(f.outShape,"float32"),_=n.dataIdMap.get(I.dataId).id;return F8(i,r.shape[0],r.shape[1],r.shape[2],o,m,g,y,x,A,b,E,w,C,T,N,M,F,_),I}var Gte={kernelName:fi,backendName:"wasm",setupFunc:Vte,kernelFunc:Ute},_8;function Hte(e){_8=e.wasm.cwrap(mi,null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function jte(e){let{backend:t,inputs:a,attrs:n}=e,{dy:r,filter:s}=a,{strides:i,pad:o,dataFormat:l,dimRoundingMode:u,inputShape:p}=n,c=1,d=S.convertConv2DDataFormat(l),h=S.computeConv2DInfo(p,s.shape,i,c,o,u,!1,d),{batchSize:f,filterHeight:m,filterWidth:g,inChannels:y,inHeight:x,inWidth:A,outChannels:b,outHeight:w,outWidth:C,strideHeight:T,strideWidth:N}=h,M=m-1-h.padInfo.top,F=g-1-h.padInfo.left,E=h.dataFormat==="channelsLast",I=v.computeStrides(h.inShape),_=v.computeStrides(r.shape),[O,B,$]=v.computeStrides(s.shape),U=I[0],G=E?I[1]:I[2],q=E?I[2]:1,H=E?1:I[1],V=_[0],Z=E?_[1]:_[2],X=E?_[2]:1,re=E?1:_[1],ee=t.makeOutput(h.inShape,"float32"),me=t.dataIdMap.get(ee.dataId).id,ie=t.dataIdMap.get(r.dataId).id,Ae=t.dataIdMap.get(s.dataId).id;return _8(ie,Ae,f,m,g,x,A,y,w,C,b,T,N,M,F,O,B,$,U,G,q,H,V,Z,X,re,me),ee}var qte={kernelName:mi,backendName:"wasm",setupFunc:Hte,kernelFunc:jte},$8;function Xte(e){$8=e.wasm.cwrap(gi,null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function Kte(e){let{inputs:t,backend:a,attrs:n}=e,{x:r,filter:s}=t,{strides:i,pad:o,dilations:l}=n;if(r.dtype!=="float32")throw new Error(`Tensor x must have dtype float32, got ${r.dtype}`);if(s.dtype!=="float32")throw new Error(`Tensor filter must have dtype float32, got ${s.dtype}`);let u=S.computeConv3DInfo(r.shape,s.shape,i,l,o),p=a.makeOutput(u.outShape,r.dtype);return $8(a.dataIdMap.get(r.dataId).id,a.dataIdMap.get(s.dataId).id,a.dataIdMap.get(p.dataId).id,u.batchSize,u.inDepth,u.inHeight,u.inWidth,u.inChannels,u.outDepth,u.outHeight,u.outWidth,u.outChannels,u.strideDepth,u.strideHeight,u.strideWidth,u.dilationDepth,u.dilationHeight,u.dilationWidth,u.filterDepth,u.filterHeight,u.filterWidth,u.padInfo.front,u.padInfo.top,u.padInfo.left),p}var Yte={kernelName:gi,backendName:"wasm",setupFunc:Xte,kernelFunc:Kte},D8;function Zte(e){D8=e.wasm.cwrap(ru,null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function Jte(e){let{inputs:t,backend:a,attrs:n}=e,{x:r,dy:s}=t,{strides:i,pad:o,filterShape:l}=n;if(r.dtype!=="float32")throw new Error(`Tensor dy must have dtype float32, got ${r.dtype}`);if(s.dtype!=="float32")throw new Error(`Tensor filter must have dtype float32, got ${s.dtype}`);let u=S.computeConv3DInfo(r.shape,l,i,1,o),p=a.makeOutput(u.filterShape,s.dtype);return D8(a.dataIdMap.get(r.dataId).id,a.dataIdMap.get(s.dataId).id,a.dataIdMap.get(p.dataId).id,u.batchSize,u.inDepth,u.inHeight,u.inWidth,u.inChannels,u.outDepth,u.outHeight,u.outWidth,u.outChannels,u.strideDepth,u.strideHeight,u.strideWidth,u.dilationDepth,u.dilationHeight,u.dilationWidth,u.filterDepth,u.filterHeight,u.filterWidth,u.padInfo.front,u.padInfo.top,u.padInfo.left),p}var Qte={kernelName:ru,backendName:"wasm",setupFunc:Zte,kernelFunc:Jte},P8;function eae(e){P8=e.wasm.cwrap(yi,null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function tae(e){let{inputs:t,backend:a,attrs:n}=e,{dy:r,filter:s}=t,{pad:i,strides:o,inputShape:l}=n;if(r.dtype!=="float32")throw new Error(`Tensor dy must have dtype float32, got ${r.dtype}`);if(s.dtype!=="float32")throw new Error(`Tensor filter must have dtype float32, got ${s.dtype}`);let u=S.computeConv3DInfo(l,s.shape,o,1,i),p=a.makeOutput(u.inShape,r.dtype);return P8(a.dataIdMap.get(s.dataId).id,a.dataIdMap.get(r.dataId).id,a.dataIdMap.get(p.dataId).id,u.batchSize,u.inDepth,u.inHeight,u.inWidth,u.inChannels,u.outDepth,u.outHeight,u.outWidth,u.outChannels,u.strideDepth,u.strideHeight,u.strideWidth,u.dilationDepth,u.dilationHeight,u.dilationWidth,u.filterDepth,u.filterHeight,u.filterWidth,u.padInfo.front,u.padInfo.top,u.padInfo.left),p}var aae={kernelName:yi,backendName:"wasm",setupFunc:eae,kernelFunc:tae},nae=at(xi),rae=at(Ai),N1;(function(e){e[e.bilinear=0]="bilinear",e[e.nearest=1]="nearest"})(N1||(N1={}));var O8;function sae(e){O8=e.wasm.cwrap(wi,null,["number","number","number","number","array","number","number","number","number","number"])}function iae(e){let{backend:t,inputs:a,attrs:n}=e,{method:r,extrapolationValue:s,cropSize:i}=n,{image:o,boxes:l,boxInd:u}=a,p=l.shape[0],[c,d]=i,h=[p,c,d,o.shape[3]],f=t.dataIdMap.get(o.dataId),m;o.dtype!=="float32"&&(m=cs({backend:t,inputs:{x:o},attrs:{dtype:"float32"}}),f=t.dataIdMap.get(m.dataId));let g=f.id,y=t.dataIdMap.get(l.dataId).id,x=t.dataIdMap.get(u.dataId).id,A=t.makeOutput(h,"float32"),b=t.dataIdMap.get(A.dataId).id,w=new Uint8Array(new Int32Array(o.shape).buffer);return O8(g,y,x,p,w,c,d,N1[r],s,b),m!=null&&t.disposeData(m.dataId),A}var oae={kernelName:wi,backendName:"wasm",setupFunc:sae,kernelFunc:iae},z8;function lae(e){z8=e.wasm.cwrap(bi,null,["number","number","number","number","number","number"])}function uae(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{axis:s,exclusive:i,reverse:o}=n,l=r.shape.length;v.assert(r.dtype==="float32"||r.dtype==="int32",()=>`cumprod does not support ${r.dtype} tensors in the WASM backend`);let u=S.getAxesPermutation([s],l),p=r;u!==null&&(p=as({inputs:{x:r},attrs:{perm:u},backend:a}));let c=S.getInnerMostAxes(1,l)[0];S.assertAxesAreInnerMostDims("cumprod",[c],l);let d=a.makeOutput(p.shape,p.dtype),h=p.shape[c],f=a.dataIdMap.get(p.dataId).id,m=a.dataIdMap.get(d.dataId).id;z8(f,i?1:0,o?1:0,h,m,nt[r.dtype]);let g=d;if(u!==null){let y=S.getUndoAxesPermutation(u);g=as({inputs:{x:d},attrs:{perm:y},backend:a}),a.disposeData(p.dataId),a.disposeData(d.dataId)}return g}var dae={kernelName:bi,backendName:"wasm",setupFunc:lae,kernelFunc:uae},L8;function pae(e){L8=e.wasm.cwrap(vi,null,["number","number","number","number","number","number"])}function cae(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{axis:s,exclusive:i,reverse:o}=n,l=r.shape.length;v.assert(r.dtype==="float32"||r.dtype==="int32",()=>`cumsum does not support ${r.dtype} tensors in the WASM backend`);let u=S.getAxesPermutation([s],l),p=r;u!==null&&(p=as({inputs:{x:r},attrs:{perm:u},backend:a}));let c=S.getInnerMostAxes(1,l)[0];S.assertAxesAreInnerMostDims("cumsum",[c],l);let d=a.makeOutput(p.shape,p.dtype),h=p.shape[c],f=a.dataIdMap.get(p.dataId).id,m=a.dataIdMap.get(d.dataId).id;L8(f,i?1:0,o?1:0,h,m,nt[r.dtype]);let g=d;if(u!==null){let y=S.getUndoAxesPermutation(u);g=as({inputs:{x:d},attrs:{perm:y},backend:a}),a.disposeData(p.dataId),a.disposeData(d.dataId)}return g}var hae={kernelName:vi,backendName:"wasm",setupFunc:pae,kernelFunc:cae},W8;function fae(e){W8=e.wasm.cwrap("DenseBincount",null,["number","array","number","number","boolean","number","number","boolean","number"])}function mae(e){let{backend:t,inputs:a,attrs:n}=e,{x:r,weights:s}=a,{size:i,binaryOutput:o}=n,l=s.shape.reduce((d,h)=>d*h,1)!==0,u=r.shape.length===1?[i]:[r.shape[0],i],p=t.makeOutput(u,s.dtype);function c(d){return t.dataIdMap.get(d.dataId).id}return W8(c(r),new Uint8Array(new Int32Array(r.shape).buffer),r.shape.length,i,l,c(s),nt[s.dtype],o,c(p)),p}var gae={kernelName:su,backendName:"wasm",setupFunc:fae,kernelFunc:mae},B8;function yae(e){B8=e.wasm.cwrap(ki,null,["number","number","number","array","number","array","array","number","number"])}function xae(e){let{backend:t,inputs:a,attrs:n}=e,{x:r}=a,{blockSize:s,dataFormat:i}=n,o=r.shape[0],l=i==="NHWC"?r.shape[1]:r.shape[2],u=i==="NHWC"?r.shape[2]:r.shape[3],p=i==="NHWC"?r.shape[3]:r.shape[1],c=l*s,d=u*s,h=p/(s*s),f=i==="NHWC"?[o,c,d,h]:[o,h,c,d],m=t.makeOutput(f,"float32"),g=t.dataIdMap.get(r.dataId).id,y=new Uint8Array(new Int32Array(v.computeStrides(r.shape)).buffer),x=new Uint8Array(new Int32Array(f).buffer),A=new Uint8Array(new Int32Array(v.computeStrides(f)).buffer),b=t.dataIdMap.get(m.dataId).id;return B8(g,s,i==="NHWC"?1:0,y,r.shape.length-1,x,A,f.length,b),m}var Aae={kernelName:ki,backendName:"wasm",setupFunc:yae,kernelFunc:xae},V8;function bae(e){V8=e.wasm.cwrap(Ii,null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function vae(e){let{inputs:t,attrs:a,backend:n}=e,{x:r,filter:s}=t,i=n.dataIdMap.get(r.dataId).id,o=n.dataIdMap.get(s.dataId).id,{strides:l,dilations:u,pad:p,dimRoundingMode:c}=a,d=u==null?[1,1]:u,h=S.computeConv2DInfo(r.shape,s.shape,l,d,p,c,!0),f=h.filterHeight,m=h.filterWidth,g=h.padInfo.top,y=h.padInfo.right,x=h.padInfo.bottom,A=h.padInfo.left,b=h.dilationHeight,w=h.dilationWidth,C=h.strideHeight,T=h.strideWidth,N=h.inChannels,M=h.outChannels,F=h.padInfo.type==="SAME"?1:0;if(h.dataFormat!=="channelsLast")throw new Error(`wasm backend DepthwiseConv2dNative does not support dataFormat:'${h.dataFormat}'. Please use 'channelsLast'.`);let E=n.makeOutput(h.outShape,"float32"),I=n.dataIdMap.get(E.dataId).id;return V8(i,r.shape[0],r.shape[1],r.shape[2],o,f,m,g,y,x,A,F,b,w,C,T,N,M,I),E}var wae={kernelName:Ii,backendName:"wasm",setupFunc:bae,kernelFunc:vae},U8;function kae(e){U8=e.wasm.cwrap("Diag",null,["number","number","number","number"])}function Iae(e){let{inputs:t,backend:a}=e,{x:n}=t,r=v.sizeFromShape(n.shape),s=a.makeOutput([...n.shape,...n.shape],n.dtype);return U8(a.dataIdMap.get(n.dataId).id,nt[n.dtype],r,a.dataIdMap.get(s.dataId).id),s}var Sae={kernelName:iu,backendName:"wasm",setupFunc:kae,kernelFunc:Iae},G8;function Cae(e){G8=e.wasm.cwrap(Si,null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function Tae(e){let{inputs:t,backend:a,attrs:n}=e,{x:r,filter:s}=t,{strides:i,pad:o,dilations:l}=n;if(r.dtype!==s.dtype)throw new Error(`Dilation2D error: x must have the same dtype as filter. Got ${r.dtype} and ${s.dtype}`);let u=S.computeDilation2DInfo(r.shape,s.shape,i,o,"NHWC",l),p=a.makeOutput(u.outShape,r.dtype);return G8(a.dataIdMap.get(r.dataId).id,a.dataIdMap.get(s.dataId).id,a.dataIdMap.get(p.dataId).id,nt[r.dtype],u.batchSize,u.inChannels,u.inHeight,u.inWidth,u.outHeight,u.outWidth,u.strideHeight,u.strideWidth,u.dilationHeight,u.dilationWidth,u.filterHeight,u.filterWidth,u.padInfo.top,u.padInfo.left),p}var Nae={kernelName:Si,backendName:"wasm",setupFunc:Cae,kernelFunc:Tae},H8;function Rae(e){H8=e.wasm.cwrap(Pl,null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function Eae(e){let{inputs:t,backend:a,attrs:n}=e,{x:r,filter:s,dy:i}=t,{strides:o,pad:l,dilations:u}=n;if(r.dtype!==s.dtype||r.dtype!==i.dtype)throw new Error(`Dilation2DBackpropFilter error: x must have the same dtype as filter and dy. Got ${r.dtype}, ${s.dtype}, and ${i.dtype}`);let p=S.computeDilation2DInfo(r.shape,s.shape,o,l,"NHWC",u),c=a.makeOutput(s.shape,s.dtype);return H8(a.dataIdMap.get(r.dataId).id,a.dataIdMap.get(s.dataId).id,a.dataIdMap.get(i.dataId).id,a.dataIdMap.get(c.dataId).id,nt[r.dtype],p.batchSize,p.inChannels,p.inHeight,p.inWidth,p.outHeight,p.outWidth,p.strideHeight,p.strideWidth,p.dilationHeight,p.dilationWidth,p.filterHeight,p.filterWidth,p.padInfo.top,p.padInfo.left),c}var Mae={kernelName:Pl,backendName:"wasm",setupFunc:Rae,kernelFunc:Eae},j8;function Fae(e){j8=e.wasm.cwrap(Dl,null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function _ae(e){let{inputs:t,backend:a,attrs:n}=e,{x:r,filter:s,dy:i}=t,{strides:o,pad:l,dilations:u}=n;if(r.dtype!==s.dtype||r.dtype!==i.dtype)throw new Error(`Dilation2DBackpropInput error: x must have the same dtype as filter and dy. Got ${r.dtype}, ${s.dtype}, and ${i.dtype}`);let p=S.computeDilation2DInfo(r.shape,s.shape,o,l,"NHWC",u),c=a.makeOutput(r.shape,r.dtype);return j8(a.dataIdMap.get(r.dataId).id,a.dataIdMap.get(s.dataId).id,a.dataIdMap.get(i.dataId).id,a.dataIdMap.get(c.dataId).id,nt[r.dtype],p.batchSize,p.inChannels,p.inHeight,p.inWidth,p.outHeight,p.outWidth,p.strideHeight,p.strideWidth,p.dilationHeight,p.dilationWidth,p.filterHeight,p.filterWidth,p.padInfo.top,p.padInfo.left),c}var $ae={kernelName:Dl,backendName:"wasm",setupFunc:Fae,kernelFunc:_ae},Dae=at(Ti),q8;function Pae(e){q8=e.wasm.cwrap(ou,null,["number","number","number"])}function Oae(e){let{inputs:t,backend:a}=e,{dy:n,y:r}=t,s=a.makeOutput(r.shape,"float32"),i=o=>a.dataIdMap.get(o.dataId).id;return q8(i(r),i(n),i(s)),s}var zae={kernelName:ou,backendName:"wasm",setupFunc:Pae,kernelFunc:Oae},Lae=!1,Wae=ea(Ni,Lae,"bool"),Bae=at(Ri,"float32");function R1(e){let{inputs:t,attrs:a,backend:n}=e,{input:r}=t,{dim:s}=a,i=r.shape.length,o=r.shape.slice(),l=s;return s<0&&(v.assert(-(i+1)<=s,()=>`Axis must be in the interval [${-(i+1)}, ${i}]`),l=i+s+1),o.splice(l,0,1),Ba({inputs:{x:r},backend:n,attrs:{shape:o}})}var Vae={kernelName:uu,backendName:"wasm",kernelFunc:R1},Uae=at(Ei,"float32");function X8(e){let{attrs:{shape:t,value:a,dtype:n},backend:r}=e,s=r.makeOutput(t,n);return r.typedArrayFromHeap(s).fill(a),s}var Gae={kernelName:du,backendName:"wasm",kernelFunc:X8},K8;function Hae(e){K8=e.wasm.cwrap(Mi,null,["number","number","number","number","number","number"])}function jae(e){let{inputs:t,backend:a}=e,{image:n}=t,r=a.makeOutput(n.shape,n.dtype),s=a.dataIdMap.get(n.dataId).id,i=a.dataIdMap.get(r.dataId).id,[o,l,u,p]=n.shape;return K8(s,o,l,u,p,i),r}var qae={kernelName:Mi,backendName:"wasm",kernelFunc:jae,setupFunc:Hae},Xae=at(Fi),Kae=!1,Yae=ea(_i,Kae),Y8;function Zae(e){Y8=e.wasm.cwrap($i,null,["number","number","number","number","number","number","number"])}function Jae(e){let{backend:t,inputs:a,attrs:n}=e,{varianceEpsilon:r}=n,{x:s,mean:i,variance:o,offset:l,scale:u}=a,p=t.dataIdMap.get(s.dataId).id,c=t.dataIdMap.get(i.dataId).id,d=t.dataIdMap.get(o.dataId).id,h=l!=null?t.dataIdMap.get(l.dataId).id:0,f=u!=null?t.dataIdMap.get(u.dataId).id:0,m=t.makeOutput(s.shape,s.dtype);if(v.sizeFromShape(s.shape)===0)return m;let g=t.dataIdMap.get(m.dataId).id;return Y8(p,c,d,h,f,r,g),m}var Qae={kernelName:$i,backendName:"wasm",setupFunc:Zae,kernelFunc:Jae},Z8;function ene(e){Z8=e.wasm.cwrap(Kr,null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function tne(e){let{inputs:t,attrs:a,backend:n}=e,{x:r,filter:s,bias:i,preluActivationWeights:o}=t,{strides:l,pad:u,dilations:p,dataFormat:c,dimRoundingMode:d,activation:h,leakyreluAlpha:f}=a,m=S.computeConv2DInfo(r.shape,s.shape,l,p,u,d),g=Xd[h];if(g==null)throw new Error(`${h} activation not yet supported for FusedConv2D in the wasm backend.`);let y=n.dataIdMap.get(r.dataId).id,x=n.dataIdMap.get(s.dataId).id,A=m.outChannels,b=0;if(i!=null){let X=n.dataIdMap.get(i.dataId);if(X.shape.length!==1)throw new Error(`FusedConv2D only supports rank-1 bias but got rank ${X.shape.length}.`);if(X.shape[0]!==A)throw new Error(`FusedConv2D bias shape (${X.shape}) does not match the number of output channels (${A})`);b=X.id}let w=m.filterHeight,C=m.filterWidth,T=m.padInfo.top,N=m.padInfo.right,M=m.padInfo.bottom,F=m.padInfo.left,E=m.dilationHeight,I=m.dilationWidth,_=m.strideHeight,O=m.strideWidth,B=m.inChannels,$=m.padInfo.type==="SAME"?1:0,U=m.batchSize,G=m.inHeight,q=m.inWidth;if(c!=="NHWC")throw new Error(`wasm backend FusedConv2D does not support dataFormat:'${c}'. Please use 'NHWC'.`);let H=n.makeOutput(m.outShape,"float32"),V=n.dataIdMap.get(H.dataId).id,Z=o==null?0:n.dataIdMap.get(o.dataId).id;return Z8(y,U,G,q,x,w,C,b,T,N,M,F,$,E,I,_,O,B,A,g,Z,f||0,V),H}var ane={kernelName:Kr,backendName:"wasm",setupFunc:ene,kernelFunc:tne},J8;function nne(e){J8=e.wasm.cwrap(Yr,null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function rne(e){let{inputs:t,attrs:a,backend:n}=e,{x:r,filter:s,bias:i,preluActivationWeights:o}=t,{strides:l,pad:u,dilations:p,dataFormat:c,dimRoundingMode:d,activation:h,leakyreluAlpha:f}=a,m=S.computeConv2DInfo(r.shape,s.shape,l,p,u,d,!0),g=Xd[h];if(g==null)throw new Error(`${h} activation not yet supported for FusedDepthwiseConv2D in the wasm backend.`);let y=n.dataIdMap.get(r.dataId).id,x=n.dataIdMap.get(s.dataId).id,A=m.outChannels,b=0;if(i!=null){let X=n.dataIdMap.get(i.dataId);if(X.shape.length!==1)throw new Error(`FusedDepthwiseConv2D only supports rank-1 bias but got rank ${X.shape.length}.`);if(X.shape[0]!==A)throw new Error(`FusedDepthwiseConv2D bias shape (${X.shape}) does not match the number of output channels (${A})`);b=X.id}let w=m.filterHeight,C=m.filterWidth,T=m.padInfo.top,N=m.padInfo.right,M=m.padInfo.bottom,F=m.padInfo.left,E=m.dilationHeight,I=m.dilationWidth,_=m.strideHeight,O=m.strideWidth,B=m.inChannels,$=m.padInfo.type==="SAME"?1:0,U=m.batchSize,G=m.inHeight,q=m.inWidth;if(c!=="NHWC")throw new Error(`wasm backend FusedDepthwiseConv2D does not support dataFormat:'${c}'. Please use 'NHWC'.`);let H=n.makeOutput(m.outShape,"float32"),V=n.dataIdMap.get(H.dataId).id,Z=o==null?0:n.dataIdMap.get(o.dataId).id;return J8(y,U,G,q,x,w,C,b,T,N,M,F,$,E,I,_,O,B,A,g,Z,f||0,V),H}var sne={kernelName:Yr,backendName:"wasm",setupFunc:nne,kernelFunc:rne},Q8;function ine(e){Q8=e.wasm.cwrap(Di,null,["number","number","number","number","number","number","array","number"])}function one(e){let{backend:t,inputs:a}=e,{params:n,indices:r}=a,[s,i,o,l]=H3.prepareAndValidate(n,r),u=t.makeOutput(s,n.dtype);if(i===0)return u;let p=r.shape,c=p[p.length-1],d=t.dataIdMap.get(n.dataId).id,h=t.dataIdMap.get(r.dataId).id,f=new Uint8Array(new Int32Array(l).buffer),m=t.dataIdMap.get(u.dataId).id;return Q8(d,nt[n.dtype],h,i,c,o,f,m),u}var lne={kernelName:Di,backendName:"wasm",setupFunc:ine,kernelFunc:one},ew;function une(e){ew=e.wasm.cwrap("Gather",null,["number","number","array","number","number","number","array","number"])}function dne(e){let{backend:t,inputs:a,attrs:n}=e,{x:r,indices:s}=a,{axis:i,batchDims:o}=n,l=v.parseAxisParam(i,r.shape)[0],u=t.readSync(s.dataId),p=r.shape[l];for(let T=0;T<u.length;++T){let N=u[T];v.assert(N<=p-1&&N>=0,()=>`GatherV2: the index value ${N} is not in [0, ${p-1}]`)}let c=S.segment_util.collectGatherOpShapeInfo(r,s,l,o),d=Ba({inputs:{x:r},attrs:{shape:[c.batchSize,c.outerSize,c.dimSize,c.sliceSize]},backend:t}),h=v.sizeFromShape(s.shape),f=Ba({inputs:{x:s},attrs:{shape:[c.batchSize,h/c.batchSize]},backend:t}),m=[c.batchSize,c.outerSize,h/c.batchSize,c.sliceSize],g=t.makeOutput(m,r.dtype);if(v.sizeFromShape(r.shape)===0)return g;let y=d.shape.length-1,x=t.dataIdMap.get(d.dataId).id,A=t.dataIdMap.get(f.dataId).id,b=t.dataIdMap.get(g.dataId).id,w=new Uint8Array(new Int32Array(v.computeStrides(d.shape)).buffer),C=new Uint8Array(new Int32Array(v.computeStrides(m)).buffer);return ew(x,nt[r.dtype],w,y,A,c.batchSize,C,b),t.disposeData(d.dataId),t.disposeData(f.dataId),g.shape=c.outputShape,g}var pne={kernelName:pu,backendName:"wasm",setupFunc:une,kernelFunc:dne},cne=!1,hne=ea(Pi,cne,"bool"),fne=!1,mne=ea(Oi,fne,"bool"),gne=at(Li,"bool"),yne=at(Wi,"bool"),xne=at(Bi,"bool"),tw;function Ane(e){tw=e.wasm.cwrap(Vi,null,["number","number","number","number"])}function bne(e){let{inputs:{x:t},attrs:{alpha:a},backend:n}=e,r=n.dataIdMap.get(t.dataId).id,s=n.makeOutput(t.shape,"float32");if(v.sizeFromShape(t.shape)!==0){let i=n.dataIdMap.get(s.dataId).id;tw(r,nt[t.dtype],a,i)}return s}var vne={kernelName:Vi,backendName:"wasm",setupFunc:Ane,kernelFunc:bne},wne=!1,kne=ea(Ui,wne,"bool"),Ine=!1,Sne=ea(Gi,Ine,"bool"),aw;function Cne(e){aw=e.wasm.cwrap(Hi,null,["number","number","number","number"])}function Tne(e){let{attrs:t,backend:a}=e,{start:n,stop:r,num:s}=t,i=Math.floor(s),o=a.makeOutput([i],"float32");return aw(a.dataIdMap.get(o.dataId).id,n,r,i),o}var Nne={kernelName:Hi,backendName:"wasm",setupFunc:Cne,kernelFunc:Tne},Rne=at(ji),Ene=at(qi),Mne=!1,Fne=ea(Xi,Mne,"bool"),_ne=at(Ki),$ne=!1,Dne=ea(Yi,$ne,"bool"),Pne=!1,One=ea(oA,Pne,"bool"),nw;function zne(e){nw=e.wasm.cwrap(Zi,null,["number","number","number","number","number","number","number"])}function Lne(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{depthRadius:s,bias:i,alpha:o,beta:l}=n;if(r.dtype!=="float32")throw new Error("LRN error: x must have dtype float32");let u=a.makeOutput(r.shape,r.dtype);return nw(a.dataIdMap.get(r.dataId).id,a.dataIdMap.get(u.dataId).id,r.shape[3],s,i,o,l),u}var Wne={kernelName:Zi,backendName:"wasm",setupFunc:zne,kernelFunc:Lne},rw;function Bne(e){rw=e.wasm.cwrap(cu,null,["number","number","number","number","number","number","number","number","number"])}function Vne(e){let{inputs:t,backend:a,attrs:n}=e,{x:r,y:s,dy:i}=t,{depthRadius:o,bias:l,alpha:u,beta:p}=n;if(r.dtype!=="float32"||s.dtype!=="float32"||i.dtype!=="float32")throw new Error("LRNGrad error: x, y, and dy must have dtype float32");let c=a.makeOutput(r.shape,r.dtype);return rw(a.dataIdMap.get(r.dataId).id,a.dataIdMap.get(s.dataId).id,a.dataIdMap.get(i.dataId).id,a.dataIdMap.get(c.dataId).id,i.shape[3],o,l,u,p),c}var Une={kernelName:cu,backendName:"wasm",setupFunc:Bne,kernelFunc:Vne},sw;function Gne(e){sw=e.wasm.cwrap(Ji,null,["number","number","number","number"])}function Hne(e){let{backend:t,inputs:a,attrs:n}=e,{reductionIndices:r,keepDims:s}=n,{x:i}=a,o=t.dataIdMap.get(i.dataId).id,l=i,{transposed:u,axes:p,originalAxes:c,inputWasTransposed:d}=ps(i,r,t);if(d){let x=t.dataIdMap.get(u.dataId).id;l=u,o=x}let h=l.shape.length;S.assertAxesAreInnerMostDims("max",p,h);let[f,m]=S.computeOutAndReduceShapes(l.shape,p),g=v.sizeFromShape(m),y=t.makeOutput(f,i.dtype);if(v.sizeFromShape(l.shape)!==0){let x=t.dataIdMap.get(y.dataId).id;sw(o,nt[i.dtype],g,x)}if(d&&t.disposeData(u.dataId),s){let x=S.expandShapeToKeepDim(y.shape,c);y.shape=x}return y}var jne={kernelName:Ji,backendName:"wasm",setupFunc:Gne,kernelFunc:Hne},qne=!1,Xne=ea(Qi,qne),iw;function Kne(e){iw=e.wasm.cwrap(eo,null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function Yne(e){let{inputs:t,attrs:a,backend:n}=e,r=t.x,s=n.dataIdMap.get(r.dataId).id;v.assert(r.dtype==="float32",()=>`Error in MaxPool: only float32 input is supported. Got ${r.dtype}.`);let{filterSize:i,strides:o,pad:l,dimRoundingMode:u}=a,p=S.computePool2DInfo(r.shape,i,o,1,l,u),c=p.filterHeight,d=p.filterWidth,h=p.padInfo.top,f=p.padInfo.right,m=p.padInfo.bottom,g=p.padInfo.left,y=p.dilationHeight,x=p.dilationWidth,A=p.strideHeight,b=p.strideWidth,w=p.inChannels,C=p.outChannels;if(p.dataFormat!=="channelsLast")throw new Error(`wasm backend does not support dataFormat:'${p.dataFormat}'. Please use 'channelsLast'.`);let T=n.makeOutput(p.outShape,"float32"),N=n.dataIdMap.get(T.dataId).id;return iw(s,r.shape[0],r.shape[1],r.shape[2],c,d,h,f,m,g,y,x,A,b,w,C,N),T}var Zne={kernelName:eo,backendName:"wasm",setupFunc:Kne,kernelFunc:Yne},ow;function Jne(e){ow=e.wasm.cwrap("MaxPool3D",null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function Qne(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{filterSize:s,strides:i,pad:o,dimRoundingMode:l,dataFormat:u}=n,p=S.computePool3DInfo(r.shape,s,i,1,o,l,u),c=a.makeOutput(p.outShape,r.dtype);return ow(a.dataIdMap.get(r.dataId).id,a.dataIdMap.get(c.dataId).id,p.batchSize,p.inChannels,p.inDepth,p.inHeight,p.inWidth,p.outDepth,p.outHeight,p.outWidth,p.strideDepth,p.strideHeight,p.strideWidth,p.dilationDepth,p.dilationHeight,p.dilationWidth,p.effectiveFilterDepth,p.effectiveFilterHeight,p.effectiveFilterWidth,p.padInfo.front,p.padInfo.top,p.padInfo.left),c}var ere={kernelName:hu,backendName:"wasm",setupFunc:Jne,kernelFunc:Qne},lw;function tre(e){lw=e.wasm.cwrap("MaxPool3DGrad",null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function are(e){let{inputs:t,backend:a,attrs:n}=e,{dy:r,input:s}=t,{filterSize:i,strides:o,pad:l,dimRoundingMode:u}=n,p=S.computePool3DInfo(s.shape,i,o,1,l,u),c=a.makeOutput(s.shape,s.dtype);return lw(a.dataIdMap.get(s.dataId).id,a.dataIdMap.get(r.dataId).id,a.dataIdMap.get(c.dataId).id,p.batchSize,p.inChannels,p.inDepth,p.inHeight,p.inWidth,p.outDepth,p.outHeight,p.outWidth,p.strideDepth,p.strideHeight,p.strideWidth,p.dilationDepth,p.dilationHeight,p.dilationWidth,p.effectiveFilterDepth,p.effectiveFilterHeight,p.effectiveFilterWidth,p.padInfo.front,p.padInfo.top,p.padInfo.left),c}var nre={kernelName:dp,backendName:"wasm",setupFunc:tre,kernelFunc:are},uw;function rre(e){uw=e.wasm.cwrap(to,null,["number, number, number"])}function sre(e){let{backend:t,inputs:a,attrs:n}=e,{axis:r,keepDims:s}=n,{x:i}=a,o=t.dataIdMap.get(i.dataId).id,l=o,u=i,{transposed:p,axes:c,originalAxes:d,inputWasTransposed:h}=ps(i,r,t),f=c;if(h){let b=t.dataIdMap.get(p.dataId).id;b!==o&&(u=p,l=b,f=S.getInnerMostAxes(f.length,u.shape.length))}S.assertAxesAreInnerMostDims("mean",f,u.shape.length);let[m,g]=S.computeOutAndReduceShapes(u.shape,f),y=v.sizeFromShape(g),x=u;u.dtype!=="float32"&&(x=cs({backend:t,inputs:{x:u},attrs:{dtype:"float32"}}),l=t.dataIdMap.get(x.dataId).id);let A=t.makeOutput(m,"float32");if(v.sizeFromShape(u.shape)!==0){let b=t.dataIdMap.get(A.dataId).id;uw(l,y,b)}if(h&&t.disposeData(p.dataId),s){let b=S.expandShapeToKeepDim(A.shape,d);A.shape=b}return u.dtype!=="float32"&&t.disposeData(x.dataId),A}var ire={kernelName:to,backendName:"wasm",setupFunc:rre,kernelFunc:sre},dw;function ore(e){dw=e.wasm.cwrap(ao,null,["number","number","number","number"])}function lre(e){let{backend:t,inputs:a,attrs:n}=e,{axis:r,keepDims:s}=n,{x:i}=a,o=t.dataIdMap.get(i.dataId).id,l=o,u=i,{transposed:p,axes:c,originalAxes:d,inputWasTransposed:h}=ps(i,r,t);if(h){let A=t.dataIdMap.get(p.dataId).id;A!==o&&(u=p,l=A)}let f=u.shape.length;S.assertAxesAreInnerMostDims("min",c,f);let[m,g]=S.computeOutAndReduceShapes(u.shape,c),y=v.sizeFromShape(g),x=t.makeOutput(m,u.dtype);if(v.sizeFromShape(u.shape)!==0){let A=t.dataIdMap.get(x.dataId).id;dw(l,nt[i.dtype],y,A)}if(h&&t.disposeData(p.dataId),s){let A=S.expandShapeToKeepDim(x.shape,d);x.shape=A}return x}var ure={kernelName:ao,backendName:"wasm",setupFunc:ore,kernelFunc:lre},dre=!1,pre=ea(no,dre),E1;(function(e){e[e.reflect=0]="reflect",e[e.symmetric=1]="symmetric"})(E1||(E1={}));var pw;function cre(e){pw=e.wasm.cwrap(ro,null,["number","array","number","number","array","array","number","number"])}function hre(e){let{inputs:{x:t},backend:a,attrs:{paddings:n,mode:r}}=e,s=n.map((f,m)=>f[0]+t.shape[m]+f[1]),i=a.dataIdMap.get(t.dataId).id,o=a.makeOutput(s,t.dtype),l=a.dataIdMap.get(o.dataId).id,u=new Uint8Array(new Int32Array(t.shape).buffer),p=n.map(f=>f[0]),c=n.map(f=>f[1]),d=new Uint8Array(new Int32Array(p).buffer),h=new Uint8Array(new Int32Array(c).buffer);return pw(i,u,t.shape.length,nt[t.dtype],d,h,E1[r],l),o}var fre={kernelName:ro,backendName:"wasm",kernelFunc:hre,setupFunc:cre},cw;function mre(e){cw=e.wasm.cwrap($o,null,["number","number","number","number"])}function hw(e){let{backend:t,inputs:{logits:a},attrs:{dim:n}}=e,r=t.dataIdMap.get(a.dataId).id,s=t.makeOutput(a.shape,a.dtype),i=t.dataIdMap.get(s.dataId).id,o=a.shape[n],l=v.sizeFromShape(a.shape)/o;return v.sizeFromShape(s.shape)===0||cw(r,i,o,l),s}var gre={kernelName:$o,backendName:"wasm",setupFunc:mre,kernelFunc:hw},fw;function yre(e){fw=e.wasm.cwrap(so,null,["number","number","number","number","number","number"])}function xre(e){let{inputs:t,backend:a,attrs:n}=e,{logits:r}=t,{numSamples:s,seed:i,normalized:o}=n;if(r.dtype!=="float32")throw new Error(`Tensor logits must have dtype float32, got ${r.dtype}`);let l=o?r:hw({inputs:{logits:r},backend:a,attrs:{dim:r.shape.length-1}}),[u,p]=l.shape,c=a.makeOutput([u,s],"int32");return fw(a.dataIdMap.get(l.dataId).id,u,p,s,i,a.dataIdMap.get(c.dataId).id),o||a.disposeData(l.dataId),c}var Are={kernelName:so,backendName:"wasm",setupFunc:yre,kernelFunc:xre},bre=!0,vre=ea(io,bre),wre=at(mu);function Tg(e,t){let a=new Int32Array(e.wasm.HEAPU8.buffer,t,4),n=a[0],r=a[1],s=a[2],i=a[3];return e.wasm._free(t),{pSelectedIndices:n,selectedSize:r,pSelectedScores:s,pValidOutputs:i}}var mw;function kre(e){mw=e.wasm.cwrap(lo,"number",["number","number","number","number","number"])}function Ire(e){let{backend:t,inputs:a,attrs:n}=e,{iouThreshold:r,maxOutputSize:s,scoreThreshold:i}=n,{boxes:o,scores:l}=a,u=t.dataIdMap.get(o.dataId).id,p=t.dataIdMap.get(l.dataId).id,c=mw(u,p,s,r,i),{pSelectedIndices:d,selectedSize:h,pSelectedScores:f,pValidOutputs:m}=Tg(t,c);return t.wasm._free(f),t.wasm._free(m),t.makeOutput([h],"int32",d)}var Sre={kernelName:lo,backendName:"wasm",setupFunc:kre,kernelFunc:Ire},gw;function Cre(e){gw=e.wasm.cwrap(gu,"number",["number","number","number","number","number","bool"])}function Tre(e){let{backend:t,inputs:a,attrs:n}=e,{iouThreshold:r,maxOutputSize:s,scoreThreshold:i,padToMaxOutputSize:o}=n,{boxes:l,scores:u}=a,p=t.dataIdMap.get(l.dataId).id,c=t.dataIdMap.get(u.dataId).id,d=gw(p,c,s,r,i,o),{pSelectedIndices:h,selectedSize:f,pSelectedScores:m,pValidOutputs:g}=Tg(t,d);t.wasm._free(m);let y=t.makeOutput([f],"int32",h),x=t.makeOutput([],"int32",g);return[y,x]}var Nre={kernelName:gu,backendName:"wasm",setupFunc:Cre,kernelFunc:Tre},yw;function Rre(e){yw=e.wasm.cwrap(uo,"number",["number","number","number","number","number","number"])}function Ere(e){let{backend:t,inputs:a,attrs:n}=e,{iouThreshold:r,maxOutputSize:s,scoreThreshold:i,softNmsSigma:o}=n,{boxes:l,scores:u}=a,p=t.dataIdMap.get(l.dataId).id,c=t.dataIdMap.get(u.dataId).id,d=yw(p,c,s,r,i,o),{pSelectedIndices:h,selectedSize:f,pSelectedScores:m,pValidOutputs:g}=Tg(t,d);t.wasm._free(g);let y=t.makeOutput([f],"int32",h),x=t.makeOutput([f],"float32",m);return[y,x]}var Mre={kernelName:uo,backendName:"wasm",setupFunc:Rre,kernelFunc:Ere},Fre=!1,_re=ea(oo,Fre,"bool"),xw;function $re(e){xw=e.wasm.cwrap(po,null,["number","number","number","number","number"])}function Dre(e){let{inputs:t,backend:a,attrs:n}=e,{indices:r}=t,{dtype:s,depth:i,onValue:o,offValue:l}=n,u=a.makeOutput([...r.shape,i],s),p=a.dataIdMap.get(u.dataId).id,c=a.dataIdMap.get(r.dataId).id;return xw(c,i,o,l,p),u}var Pre={kernelName:po,backendName:"wasm",setupFunc:$re,kernelFunc:Dre};function Ore(e){let{inputs:{x:t},backend:a}=e,n=a.makeOutput(t.shape,t.dtype);return a.typedArrayFromHeap(n).fill(1),n}var zre={kernelName:yu,backendName:"wasm",kernelFunc:Ore};function Lre(e){let{inputs:t,backend:a,attrs:n}=e,{axis:r}=n;if(t.length===1)return R1({inputs:{input:t[0]},backend:a,attrs:{dim:r}});let s=t[0].shape,i=t[0].dtype;t.forEach(p=>{v.assertShapesMatch(s,p.shape,"All tensors passed to stack must have matching shapes"),v.assert(i===p.dtype,()=>"All tensors passed to stack must have matching dtypes")});let o=[],l=t.map(p=>{let c=R1({inputs:{input:p},backend:a,attrs:{dim:r}});return o.push(c),c}),u=M8({inputs:l,backend:a,attrs:{axis:r}});return o.forEach(p=>a.disposeData(p.dataId)),u}var Wre={kernelName:xu,backendName:"wasm",kernelFunc:Lre},Aw;function Bre(e){Aw=e.wasm.cwrap(co,null,["number","array","number","number","array","array","number","number"])}function Vre(e){let{inputs:{x:t},backend:a,attrs:{paddings:n,constantValue:r}}=e,s=n.map((f,m)=>f[0]+t.shape[m]+f[1]);if(v.sizeFromShape(t.shape)===0)return X8({backend:a,attrs:{shape:s,value:r,dtype:t.dtype}});let i=a.dataIdMap.get(t.dataId).id,o=a.makeOutput(s,t.dtype),l=a.dataIdMap.get(o.dataId).id,u=new Uint8Array(new Int32Array(t.shape).buffer),p=n.map(f=>f[0]),c=n.map(f=>f[1]),d=new Uint8Array(new Int32Array(p).buffer),h=new Uint8Array(new Int32Array(c).buffer);return Aw(i,u,t.shape.length,nt[t.dtype],d,h,r,l),o}var bw={kernelName:co,backendName:"wasm",kernelFunc:Vre,setupFunc:Bre},Ure=!1,Gre=ea(ho,Ure),vw;function Hre(e){vw=e.wasm.cwrap(fo,null,["number","number","number"])}function jre(e){let{inputs:t,backend:a}=e,{x:n,alpha:r}=t,s=a.dataIdMap.get(n.dataId).id,i=a.dataIdMap.get(r.dataId).id,o=s,l=n,u=l;l.dtype!=="float32"&&(u=cs({backend:a,inputs:{x:n},attrs:{dtype:"float32"}}),o=a.dataIdMap.get(u.dataId).id);let p=a.makeOutput(n.shape,"float32"),c=a.dataIdMap.get(p.dataId).id;return vw(o,i,c),l.dtype!=="float32"&&a.disposeData(u.dataId),p}var qre={kernelName:fo,backendName:"wasm",setupFunc:Hre,kernelFunc:jre},ww;function Xre(e){ww=e.wasm.cwrap(mo,null,["number","number","number","number"])}function Kre(e){let{backend:t,inputs:a,attrs:n}=e,{axis:r,keepDims:s}=n,{x:i}=a,o=t.dataIdMap.get(i.dataId).id,l=o,u=i,{transposed:p,axes:c,originalAxes:d,inputWasTransposed:h}=ps(i,r,t),f=c;if(h){let A=t.dataIdMap.get(p.dataId).id;A!==o&&(u=p,l=A,f=S.getInnerMostAxes(f.length,u.shape.length))}S.assertAxesAreInnerMostDims("prod",f,u.shape.length);let[m,g]=S.computeOutAndReduceShapes(u.shape,f),y=v.sizeFromShape(g),x=t.makeOutput(m,u.dtype);if(v.sizeFromShape(u.shape)!==0){let A=t.dataIdMap.get(x.dataId).id;ww(l,y,nt[x.dtype],A)}if(h&&t.disposeData(p.dataId),s){let A=S.expandShapeToKeepDim(x.shape,d);x.shape=A}return x}var Yre={kernelName:mo,backendName:"wasm",setupFunc:Xre,kernelFunc:Kre},Zre=e=>{let{backend:t,attrs:a}=e,{start:n,stop:r,step:s,dtype:i}=a,o=ag(n,r,s,i),l=t.makeOutput([o.length],i);return t.typedArrayFromHeap(l).set(o),l},Jre={kernelName:Au,backendName:"wasm",kernelFunc:Zre},Qre=!0,ese=ea(Ci,Qre),tse=at(go),ase=at(yo),nse=at(bo),kw;function rse(e){kw=e.wasm.cwrap(Ao,null,["number","number","number","number","number","number","number","number","number","number"])}function sse(e){let{backend:t,inputs:a,attrs:n}=e,{images:r}=a,{alignCorners:s,halfPixelCenters:i,size:o}=n,[l,u]=o,[p,c,d,h]=r.shape,f=[p,l,u,h],m=t.dataIdMap.get(r.dataId),g;m.dtype!=="float32"&&(g=cs({backend:t,inputs:{x:r},attrs:{dtype:"float32"}}),m=t.dataIdMap.get(g.dataId));let y=m.id,x=t.makeOutput(f,"float32");if(v.sizeFromShape(r.shape)===0)return x;let A=t.dataIdMap.get(x.dataId).id;return kw(y,p,c,d,h,l,u,s?1:0,i?1:0,A),g!=null&&t.disposeData(g.dataId),x}var ise={kernelName:Ao,backendName:"wasm",setupFunc:rse,kernelFunc:sse},Iw;function ose(e){Iw=e.wasm.cwrap(wu,null,["number","number","number","array","array","boolean"])}function lse(e){let{inputs:t,backend:a,attrs:n}=e,{images:r,dy:s}=t,{alignCorners:i}=n,o=a.makeOutput(r.shape,"float32"),l=a.dataIdMap.get(r.dataId),u;return l.dtype!=="float32"&&(u=cs({backend:a,inputs:{x:r},attrs:{dtype:"float32"}}),l=a.dataIdMap.get(u.dataId)),Iw(a.dataIdMap.get(r.dataId).id,a.dataIdMap.get(s.dataId).id,a.dataIdMap.get(o.dataId).id,new Uint8Array(new Int32Array(r.shape).buffer),new Uint8Array(new Int32Array(s.shape).buffer),i),u!=null&&a.disposeData(u.dataId),o}var use={kernelName:wu,backendName:"wasm",setupFunc:ose,kernelFunc:lse},Sw;function dse(e){Sw=e.wasm.cwrap(xo,null,["number","number","number","number","number","number","number","number","number","number"])}function pse(e){let{backend:t,inputs:a,attrs:n}=e,{images:r}=a,{alignCorners:s,halfPixelCenters:i,size:o}=n,[l,u]=o,[p,c,d,h]=r.shape,f=[p,l,u,h],m=t.makeOutput(f,"float32");if(v.sizeFromShape(r.shape)===0)return m;let g=t.dataIdMap.get(r.dataId),y;g.dtype!=="float32"&&(y=cs({backend:t,inputs:{x:r},attrs:{dtype:"float32"}}),g=t.dataIdMap.get(y.dataId));let x=g.id,A=t.dataIdMap.get(m.dataId).id;return Sw(x,p,c,d,h,l,u,s?1:0,i?1:0,A),y!=null&&t.disposeData(y.dataId),m}var cse={kernelName:xo,backendName:"wasm",setupFunc:dse,kernelFunc:pse},Cw;function hse(e){Cw=e.wasm.cwrap(vu,null,["number","number","number","array","array","boolean"])}function fse(e){let{inputs:t,backend:a,attrs:n}=e,{images:r,dy:s}=t,{alignCorners:i}=n,o=a.makeOutput(r.shape,"float32"),l=a.dataIdMap.get(r.dataId),u;return l.dtype!=="float32"&&(u=cs({backend:a,inputs:{x:r},attrs:{dtype:"float32"}}),l=a.dataIdMap.get(u.dataId)),Cw(a.dataIdMap.get(r.dataId).id,a.dataIdMap.get(s.dataId).id,a.dataIdMap.get(o.dataId).id,new Uint8Array(new Int32Array(r.shape).buffer),new Uint8Array(new Int32Array(s.shape).buffer),i),u!=null&&a.disposeData(u.dataId),o}var mse={kernelName:vu,backendName:"wasm",setupFunc:hse,kernelFunc:fse},Tw;function gse(e){Tw=e.wasm.cwrap(vo,null,["number","array","number","array","number","number"])}function yse(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{dims:s}=n,i=v.parseAxisParam(s,r.shape);if(r.shape.length===0)return a0({inputs:{x:r},backend:a});let o=a.makeOutput(r.shape,r.dtype),l=a.dataIdMap.get(r.dataId).id,u=a.dataIdMap.get(o.dataId).id,p=new Uint8Array(new Int32Array(i).buffer),c=new Uint8Array(new Int32Array(r.shape).buffer);Tw(l,p,i.length,c,r.shape.length,u);let d=Ba({inputs:{x:o},attrs:{shape:r.shape},backend:a});return a.disposeData(o.dataId),d}var xse={kernelName:vo,backendName:"wasm",kernelFunc:yse,setupFunc:gse},Nw;function Ase(e){Nw=e.wasm.cwrap(Uo,null,["number","number","number","number","number","number","number","number","array","number","number"])}function bse(e){let{inputs:t,backend:a,attrs:n}=e,{image:r}=t,{radians:s,fillValue:i,center:o}=n,l=a.makeOutput(r.shape,r.dtype),u=a.dataIdMap.get(r.dataId).id,p=a.dataIdMap.get(l.dataId).id,[c,d,h,f]=r.shape,[m,g]=S.getImageCenter(o,d,h),y=i===0,x=255,A=typeof i=="number"?[i,i,i,y?0:x]:[...i,x],b=new Uint8Array(new Int32Array(A).buffer);return Nw(u,c,d,h,f,s,m,g,b,A.length,p),l}var vse={kernelName:Uo,backendName:"wasm",kernelFunc:bse,setupFunc:Ase},wse=at(wo),kse=at(ko),Rw;function Ise(e){Rw=e.wasm.cwrap(Io,null,["number","number","number","number","number","number","array","number","number"])}function Sse(e){let{backend:t,inputs:a,attrs:n}=e,{indices:r,updates:s}=a,{shape:i}=n,o=t.makeOutput(i,s.dtype);if(v.sizeFromShape(i)===0)return o;let{sliceRank:l,numUpdates:u,sliceSize:p,strides:c,outputSize:d}=Wh.calculateShapes(s,r,i),h=t.dataIdMap.get(r.dataId).id,f=t.dataIdMap.get(s.dataId).id,m=new Uint8Array(new Int32Array(c).buffer),g=t.dataIdMap.get(o.dataId).id;return Rw(h,f,nt[s.dtype],l,u,p,m,d,g),o}var Cse={kernelName:Io,backendName:"wasm",setupFunc:Ise,kernelFunc:Sse},Ew;function Tse(e){Ew=e.wasm.cwrap(Co,null,["number","number","number","number","number","number","bool","number"])}function Nse(e){let{inputs:t,backend:a,attrs:n}=e,{sortedSequence:r,values:s}=t,{side:i}=n;if(r.dtype!==s.dtype)throw new Error(`SearchSorted error: sorted_sequence must have the same dtype as values. Got ${r.dtype} and ${s.dtype}`);let o=a.makeOutput(s.shape,"int32");function l(u){return a.dataIdMap.get(u.dataId).id}return Ew(l(r),l(s),r.shape[0],r.shape[1],s.shape[1],nt[r.dtype],i==="left",l(o)),o}var Rse={kernelName:Co,backendName:"wasm",setupFunc:Tse,kernelFunc:Nse},Mw;function Ese(e){Mw=e.wasm.cwrap("SelectV2",null,["number","number","number","number","number"])}function Mse(e){let{inputs:t,backend:a}=e,{condition:n,t:r,e:s}=t,i=a.dataIdMap.get(n.dataId).id,o=a.dataIdMap.get(r.dataId).id,l=a.dataIdMap.get(s.dataId).id,u=a.makeOutput(r.shape,r.dtype),p=a.dataIdMap.get(u.dataId).id,c=n.shape.length,d=r.shape.length,h=c===0||c>1||d===1?1:v.sizeFromShape(r.shape.slice(1));return Mw(i,o,l,h,p),u}var Fse={kernelName:ku,backendName:"wasm",kernelFunc:Mse,setupFunc:Ese},_se=at(To),Fw;function $se(e){Fw=e.wasm.cwrap(Eo,null,["number","number"])}function Dse(e){let{backend:t,inputs:{x:a}}=e,n=t.dataIdMap.get(a.dataId).id,r=t.makeOutput(a.shape,a.dtype),s=t.dataIdMap.get(r.dataId).id;return v.sizeFromShape(r.shape)===0||Fw(n,s),r}var Pse={kernelName:"Sigmoid",backendName:"wasm",setupFunc:$se,kernelFunc:Dse},Ose=at(Ro),zse=at(No),Lse=at(Mo);function Wse(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{blockShape:s,paddings:i}=n,o=v.sizeFromShape(s),l=[[0,0]];l.push(...i);for(let g=1+s.length;g<r.shape.length;++g)l.push([0,0]);let u=bw.kernelFunc({inputs:{x:r},backend:a,attrs:{paddings:l,constantValue:0}}),p=S.getReshaped(u.shape,s,o,!1),c=S.getPermuted(p.length,s.length,!1),d=S.getReshapedPermuted(u.shape,s,o,!1),h=Ba({inputs:{x:u},backend:a,attrs:{shape:p}}),f=as({inputs:{x:h},backend:a,attrs:{perm:c}}),m=Ba({inputs:{x:f},backend:a,attrs:{shape:d}});return a.disposeData(u.dataId),a.disposeData(h.dataId),a.disposeData(f.dataId),m}var Bse={kernelName:Cu,backendName:"wasm",kernelFunc:Wse},_w;function Vse(e){_w=e.wasm.cwrap("SparseFillEmptyRows","number",["number","number","number","number","number","number","number","number","number","number","number","number"])}function Use(e){let{backend:t,inputs:a}=e,{indices:n,values:r,denseShape:s,defaultValue:i}=a,o=n.shape[0],l=n.shape[1],u=t.readSync(s.dataId)[0],p=[o+u,l],c=t.dataIdMap.get(n.dataId).id,d=t.dataIdMap.get(r.dataId).id,h=t.dataIdMap.get(i.dataId).id,f=t.makeOutput(p,n.dtype),m=t.dataIdMap.get(f.dataId).id,g=t.makeOutput(p.slice(0,1),r.dtype),y=t.dataIdMap.get(g.dataId).id,x=t.makeOutput([u],"bool"),A=t.dataIdMap.get(x.dataId).id,b=t.makeOutput([o],n.dtype),w=t.dataIdMap.get(b.dataId).id,C=t.makeOutput([4],"int32"),T=t.dataIdMap.get(C.dataId).id,N=_w(c,d,nt[r.dtype],o,u,l,h,m,y,A,w,T),M=t.readSync(C.dataId),F;switch(M[0]){case 1:{F=S.getSparseFillEmptyRowsIndicesDenseShapeMismatch(M[1]);break}case 2:{F=S.getSparseFillEmptyRowsNegativeIndexErrorMessage(M[1],M[2]);break}case 3:F=S.getSparseFillEmptyRowsOutOfRangeIndexErrorMessage(M[1],M[2],M[3]);break;default:F=""}if(t.disposeData(C.dataId),F)throw t.disposeData(f.dataId),t.disposeData(g.dataId),t.disposeData(x.dataId),t.disposeData(b.dataId),new Error(F);let E=f,I=g;return N!==p[0]&&(E=Zs({inputs:{x:f},attrs:{begin:0,size:[N,l]},backend:t}),I=Zs({inputs:{x:g},attrs:{begin:0,size:N},backend:t}),t.disposeData(f.dataId),t.disposeData(g.dataId)),[E,I,x,b]}var Gse={kernelName:hp,backendName:"wasm",setupFunc:Vse,kernelFunc:Use},$w;function Hse(e){$w=e.wasm.cwrap(Nu,null,["number","number","number","number","number","number","number"])}function jse(e){let{backend:t,inputs:a}=e,{inputIndices:n,inputShape:r,newShape:s}=a;if(n.shape.length!==2)throw new Error(`Input indices should be a matrix but received shape
|
|
${n.shape}`);if(r.shape.length!==1)throw new Error(`Input shape should be a vector but received shape
|
|
${r.shape}`);if(s.shape.length!==1)throw new Error(`Target shape should be a vector but received shape ${s.shape}`);let i=t.dataIdMap.get(n.dataId).id,o=t.dataIdMap.get(r.dataId).id,l=t.dataIdMap.get(s.dataId).id,u=n.shape[0],p=v.sizeFromShape(s.shape),c=t.makeOutput([u,p],n.dtype),d=t.dataIdMap.get(c.dataId).id,h=t.makeOutput([p],s.dtype),f=t.dataIdMap.get(h.dataId).id,m=t.makeOutput([3],"int32"),g=t.dataIdMap.get(m.dataId).id;$w(i,o,l,u,d,f,g);let y=t.readSync(m.dataId),x;switch(y[0]){case 0:{x=S.getSparseReshapeMultipleNegativeOneOutputDimErrorMessage(y[1],y[2]);break}case 1:{x=S.getSparseReshapeNegativeOutputDimErrorMessage(y[1],y[2]);break}case 2:x=S.getSparseReshapeEmptyTensorZeroOutputDimErrorMessage();break;case 3:{let A=Array.from(t.readSync(r.dataId)),b=Array.from(t.readSync(h.dataId));x=S.getSparseReshapeInputOutputMultipleErrorMessage(A,b);break}case 4:{let A=Array.from(t.readSync(r.dataId)),b=Array.from(t.readSync(h.dataId));x=S.getSparseReshapeInputOutputMismatchErrorMessage(A,b);break}default:x=""}if(t.disposeData(m.dataId),x)throw t.disposeData(c.dataId),t.disposeData(h.dataId),new Error(x);return[c,h]}var qse={kernelName:Nu,backendName:"wasm",setupFunc:Hse,kernelFunc:jse},Dw;function Pw(e){Dw=e.wasm.cwrap("SparseSegmentReduction",null,["number","number","number","number","number","number","number","number","number"])}function Ow(e,t){let{backend:a,inputs:n}=e,{data:r,indices:s,segmentIds:i}=n,o=s.shape[0],l=a.readSync(i.dataId,o-1,o)[0],u=o>0?l+1:0;if(u<0)throw new Error(S.getSparseSegmentReductionNegativeSegmentIdsErrorMessage());let p=r.shape.slice();p[0]=u;let c=a.dataIdMap.get(r.dataId).id,d=a.dataIdMap.get(s.dataId).id,h=a.dataIdMap.get(i.dataId).id,f=a.makeOutput(p,r.dtype),m=a.dataIdMap.get(f.dataId).id,g=a.makeOutput([4],"int32"),y=a.dataIdMap.get(g.dataId).id;Dw(c,nt[r.dtype],r.shape[0],d,h,m,y,t,0);let x=a.readSync(g.dataId),A;switch(x[0]){case 0:{A=S.getSparseSegmentReductionNegativeSegmentIdsErrorMessage();break}case 1:{A=S.getSparseSegmentReductionNonIncreasingSegmentIdsErrorMessage();break}case 2:A=S.getSparseSegmentReductionSegmentIdOutOfRangeErrorMessage(x[1],x[2]);break;case 3:A=S.getSparseSegmentReductionIndicesOutOfRangeErrorMessage(x[1],x[2],x[3]);break;default:A=""}if(a.disposeData(g.dataId),A)throw a.disposeData(f.dataId),new Error(A);return f}function Xse(e){return Ow(e,!0)}var Kse={kernelName:fp,backendName:"wasm",setupFunc:Pw,kernelFunc:Xse};function Yse(e){return Ow(e,!1)}var Zse={kernelName:mp,backendName:"wasm",setupFunc:Pw,kernelFunc:Yse},zw;function Jse(e){zw=e.wasm.cwrap(Do,null,["number","number","number","number","number","number","number","number","array","number","number"])}function Qse(e){let{backend:t,inputs:a,attrs:n}=e,{sparseIndices:r,sparseValues:s,defaultValue:i}=a,{outputShape:o}=n,l=t.makeOutput(o,i.dtype);if(v.sizeFromShape(o)===0)return l;let{sliceRank:u,numUpdates:p,sliceSize:c,strides:d,outputSize:h}=S.calculateShapes(s,r,o),f=t.dataIdMap.get(r.dataId).id,m=t.dataIdMap.get(s.dataId).id,g=t.dataIdMap.get(i.dataId).id,y=new Uint8Array(new Int32Array(d).buffer),x=t.dataIdMap.get(l.dataId).id;return zw(f,m,s.shape.length,g,nt[i.dtype],u,p,c,y,h,x),l}var eie={kernelName:Do,backendName:"wasm",setupFunc:Jse,kernelFunc:Qse};function tie(e){let{inputs:t,attrs:a,backend:n}=e,{x:r}=t,{numOrSizeSplits:s,axis:i}=a,o=v.parseAxisParam(i,r.shape)[0],l=S.prepareSplitSize(r,s,o),u=new Array(r.shape.length).fill(0),p=r.shape.slice();return l.map(c=>{let d=[...p];d[o]=c;let h=Zs({inputs:{x:r},attrs:{begin:u,size:d},backend:n});return u[o]+=c,h})}var aie={kernelName:Tu,backendName:"wasm",kernelFunc:tie},nie=at(Fo),rie=at(gp),sie=!0,iie=ea(Po,sie),Lw;function oie(e){Lw=e.wasm.cwrap(is,null,["number","number","number","number"])}function lie(e){let{backend:t,inputs:a,attrs:n}=e,{alpha:r}=n,{x:s}=a,i=t.dataIdMap.get(s.dataId).id,o=t.makeOutput(s.shape,s.dtype),l=t.dataIdMap.get(o.dataId).id;return Lw(i,r,nt[s.dtype],l),o}var uie={kernelName:is,backendName:"wasm",setupFunc:oie,kernelFunc:lie},Ww;function die(e){Ww=e.wasm.cwrap(Oo,null,["number","array","number","array","array","array","array","array","number","number"])}function pie(e){let{backend:t,inputs:a,attrs:n}=e,{x:r}=a,{begin:s,end:i,strides:o,beginMask:l,endMask:u,ellipsisMask:p,newAxisMask:c,shrinkAxisMask:d}=n,{finalShapeSparse:h,finalShape:f,isIdentity:m,sliceDim0:g,isSimpleSlice:y,begin:x,end:A,strides:b}=Nt.sliceInfo(r.shape,s,i,o,l,u,p,c,d),w;if(m)w=Ba({inputs:{x:r},backend:t,attrs:{shape:f}});else if(g||y){v.assert(r.shape.length>=1,()=>`Input must have rank at least 1, got: ${r.shape.length}`);let C=Nt.computeOutShape(x,A,b),T=Zs({inputs:{x:r},backend:t,attrs:{begin:x,size:C}});w=Ba({inputs:{x:T},backend:t,attrs:{shape:f}}),t.disposeData(T.dataId)}else{let C=t.makeOutput(h,"float32"),T=t.dataIdMap.get(r.dataId).id,N=new Uint8Array(new Int32Array(v.computeStrides(r.shape)).buffer),M=new Uint8Array(new Int32Array(x).buffer),F=new Uint8Array(new Int32Array(A).buffer),E=new Uint8Array(new Int32Array(b).buffer),I=new Uint8Array(new Int32Array(h).buffer),_=new Uint8Array(new Int32Array(v.computeStrides(h)).buffer),O=t.dataIdMap.get(C.dataId).id;Ww(T,N,r.shape.length,M,F,E,I,_,h.length,O),w=Ba({inputs:{x:C},backend:t,attrs:{shape:f}}),t.disposeData(C.dataId)}return w}var cie={kernelName:Oo,backendName:"wasm",setupFunc:die,kernelFunc:pie};function hie(e){let{backend:t,inputs:a,attrs:n}=e,{data:r,dataSplits:s}=a,{separator:i,nGramWidths:o,leftPad:l,rightPad:u,padWidth:p,preserveShortSequences:c}=n,d=t.readSync(r.dataId),h=t.readSync(s.dataId),[f,m]=rg(d,h,i,o,l,u,p,c),g=t.makeOutput([f.length],"string"),y=t.dataIdMap.get(g.dataId);y.stringBytes=f;let x=t.makeOutput(s.shape,"int32");return t.typedArrayFromHeap(x).set(m),[g,x]}var fie={kernelName:Ru,backendName:"wasm",kernelFunc:hie};function mie(e){let{backend:t,inputs:a,attrs:n}=e,{input:r,delimiter:s}=a,{skipEmpty:i}=n,o=t.readSync(r.dataId),l=t.readSync(s.dataId),[u,p,c]=sg(o,l[0],i),d=p.length,h=t.makeOutput([d,2],"int32");t.typedArrayFromHeap(h).set(u);let f=t.makeOutput([d],"string"),m=t.dataIdMap.get(f.dataId);m.stringBytes=p;let g=t.makeOutput([2],"int32");return t.typedArrayFromHeap(g).set(c),[h,f,g]}var gie={kernelName:xp,backendName:"wasm",kernelFunc:mie};function yie(e){let{backend:t,inputs:a,attrs:n}=e,{input:r}=a,{numBuckets:s}=n,i=t.readSync(r.dataId),o=ig(i,s),l=t.makeOutput(r.shape,"int32");return t.typedArrayFromHeap(l).set(o),l}var xie={kernelName:Ap,backendName:"wasm",kernelFunc:yie},Aie=!0,bie=ea(zo,Aie),Bw;function vie(e){Bw=e.wasm.cwrap(_o,null,["number","number","number","number"])}function wie(e){let{backend:t,inputs:a,attrs:n}=e,{axis:r,keepDims:s}=n,{x:i}=a,o=t.dataIdMap.get(i.dataId).id,l=o,u=i,{transposed:p,axes:c,originalAxes:d,inputWasTransposed:h}=ps(i,r,t),f=c;if(h){let A=t.dataIdMap.get(p.dataId).id;A!==o&&(u=p,l=A,f=S.getInnerMostAxes(f.length,u.shape.length))}S.assertAxesAreInnerMostDims("sum",f,u.shape.length);let[m,g]=S.computeOutAndReduceShapes(u.shape,f),y=v.sizeFromShape(g),x=t.makeOutput(m,u.dtype);if(v.sizeFromShape(u.shape)!==0){let A=t.dataIdMap.get(x.dataId).id;Bw(l,y,nt[x.dtype],A)}if(h&&t.disposeData(p.dataId),s){let A=S.expandShapeToKeepDim(x.shape,d);x.shape=A}return x}var kie={kernelName:_o,backendName:"wasm",setupFunc:vie,kernelFunc:wie},Iie=at(Lo),Sie=at(Wo),Vw;function Cie(e){Vw=e.wasm.cwrap(So,null,["number","number","number","number","number","number","array","number","number","number"])}function Tie(e){let{backend:t,inputs:a,attrs:n}=e,{tensor:r,indices:s,updates:i}=a,{}=n,o=t.makeOutput(r.shape,r.dtype);if(v.sizeFromShape(r.shape)===0)return o;let{sliceRank:l,numUpdates:u,sliceSize:p,strides:c,outputSize:d}=Wh.calculateShapes(i,s,r.shape),h=t.dataIdMap.get(s.dataId).id,f=t.dataIdMap.get(i.dataId).id,m=t.dataIdMap.get(r.dataId).id,g=new Uint8Array(new Int32Array(c).buffer),y=t.dataIdMap.get(o.dataId).id;return Vw(h,f,nt[i.dtype],l,u,p,g,d,y,m),o}var Nie={kernelName:So,backendName:"wasm",setupFunc:Cie,kernelFunc:Tie},Uw;function Rie(e){Uw=e.wasm.cwrap(ss,null,["number","array","number","array","number","number"])}function Eie(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,s=a.dataIdMap.get(r.dataId).id,{reps:i}=n,o=new Array(r.shape.length);for(let d=0;d<o.length;d++)o[d]=r.shape[d]*i[d];let l=new Uint8Array(new Int32Array(r.shape).buffer),u=new Uint8Array(new Int32Array(o).buffer),p=a.makeOutput(o,r.dtype),c=a.dataIdMap.get(p.dataId).id;return Uw(s,l,r.shape.length,u,o.length,nt[p.dtype],c),p}var Mie={kernelName:ss,backendName:"wasm",setupFunc:Rie,kernelFunc:Eie},Gw;function Fie(e){Gw=e.wasm.cwrap(Bo,null,["number","array","number","number","number","bool","number","number"])}var _ie=({inputs:e,backend:t,attrs:a})=>{let{x:n}=e,{k:r,sorted:s}=a,i=t.dataIdMap.get(n.dataId).id,o=new Uint8Array(new Int32Array(n.shape).buffer),l=n.shape.slice();l[l.length-1]=r;let u=t.makeOutput(l,n.dtype),p=t.dataIdMap.get(u.dataId).id,c=t.makeOutput(l,"int32"),d=t.dataIdMap.get(c.dataId).id;return Gw(i,o,n.shape.length,nt[n.dtype],r,s,p,d),[u,c]},$ie={kernelName:Bo,backendName:"wasm",setupFunc:Fie,kernelFunc:_ie},Hw;function Die(e){Hw=e.wasm.cwrap(Vo,null,["number","number","bool","number","number","number","number","number","number","array","number","array","number","number","number","number","number"])}function Pie(e){let{backend:t,inputs:a,attrs:n}=e,{image:r,transforms:s}=a,{interpolation:i,fillMode:o,fillValue:l,outputShape:u}=n,[p,c,d,h]=r.shape,[f,m]=u!=null?u:[c,d],g=[p,f,m,h],y=new Uint8Array(new Int32Array(v.computeStrides(r.shape)).buffer),x=new Uint8Array(new Int32Array(v.computeStrides(g)).buffer),A=t.makeOutput(g,r.dtype),b=t.dataIdMap.get(A.dataId).id,w=t.dataIdMap.get(r.dataId).id,C=t.dataIdMap.get(s.dataId).id,T=i==="nearest"?1:2,N;switch(o){case"constant":N=1;break;case"reflect":N=2;break;case"wrap":N=3;break;case"nearest":N=4;break;default:N=1;break}return Hw(w,C,s.shape[0]>1,p,f,m,h,d,c,y,r.shape.length-1,x,g.length-1,T,N,l,b),A}var Oie={kernelName:Vo,backendName:"wasm",setupFunc:Die,kernelFunc:Pie};function zie(e){let{inputs:t,attrs:a,backend:n}=e,{axis:r}=a,{x:s}=t,{outputValues:i,outputShape:o,indices:l}=lg(n.readSync(s.dataId),r,s.shape,s.dtype);return[n.makeOutput(o,s.dtype,void 0,i),n.makeOutput([l.length],"int32",void 0,l)]}var Lie={kernelName:bp,backendName:"wasm",kernelFunc:zie};function Wie(e){let{inputs:t,backend:a,attrs:n}=e,{value:r}=t,{axis:s}=n;s<0&&(s+=r.shape.length);let i=r.shape[s],o=r.shape.length,l=new Array(o-1),u=0;for(let h=0;h<o;h++)h!==s&&(l[u++]=r.shape[h]);let p=new Array(i),c=new Array(o).fill(0),d=r.shape.slice();d[s]=1;for(let h=0;h<p.length;h++)c[s]=h,p[h]=Zs({inputs:{x:r},attrs:{begin:c,size:d},backend:a});return p.map(({dataId:h,dtype:f})=>({dataId:h,dtype:f,shape:l}))}var Bie={kernelName:Eu,backendName:"wasm",kernelFunc:Wie};function Vie(e){let{inputs:{x:t},backend:a}=e,n=a.makeOutput(t.shape,t.dtype);return a.typedArrayFromHeap(n).fill(0),n}var Uie={kernelName:Mu,backendName:"wasm",kernelFunc:Vie},Gie=[Lee,Wee,Bee,Vee,Gee,qee,tte,rte,ste,ite,ote,lte,ute,dte,pte,fte,yte,bte,Ite,Ete,_te,Dte,Pte,Ote,Wte,Bte,Gte,qte,Yte,Qte,aae,nae,rae,oae,dae,hae,gae,Aae,wae,Sae,Nae,Mae,$ae,Dae,zae,Wae,Bae,Vae,Uae,Gae,qae,Xae,Yae,Qae,ane,sne,lne,pne,hne,mne,Xee,gne,yne,xne,vne,kne,Sne,Nne,Ene,Rne,Fne,_ne,Dne,One,Wne,Une,jne,Xne,Zne,ere,nre,ire,ure,pre,fre,Are,vre,wre,Sre,Nre,Mre,_re,Pre,zre,Wre,bw,Gre,qre,Yre,Jre,ese,tse,ase,nse,vte,ise,use,cse,mse,xse,vse,wse,kse,Cse,Rse,Fse,_se,Pse,Ose,zse,Nte,gre,Lse,Bse,Gse,qse,Kse,Zse,eie,aie,nie,rie,iie,uie,cie,fie,gie,xie,bie,kie,Iie,Sie,Nie,Mie,$ie,Oie,Jee,Lie,Bie,Uie];for(let e of Gie)yn(e);var M1=W();M1.registerFlag("WASM_HAS_SIMD_SUPPORT",async()=>{try{return WebAssembly.validate(new Uint8Array([0,97,115,109,1,0,0,0,1,4,1,96,0,0,3,2,1,0,10,9,1,7,0,65,0,253,15,26,11]))}catch(e){return!1}});M1.registerFlag("WASM_HAS_MULTITHREAD_SUPPORT",async()=>{if(M1.get("IS_NODE"))return!1;try{return new MessageChannel().port1.postMessage(new SharedArrayBuffer(1)),WebAssembly.validate(new Uint8Array([0,97,115,109,1,0,0,0,1,4,1,96,0,0,3,2,1,0,5,4,1,3,1,1,10,11,1,9,0,65,0,254,16,2,0,26,11]))}catch(e){return!1}});var Tx=Xl(wC()),Hie=Xl(kC()),Nx=Xl(IC()),Rx=Tx.default||Tx,jie=Nx.default||Nx,jw=class extends Kl{constructor(e){super(),this.wasm=e,this.dataIdNextNumber=1,this.wasm.tfjs.initWithThreadsCount(qw),F1=this.wasm.tfjs.getThreadsCount(),this.dataIdMap=new Zd(this,It())}write(e,t,a){let n={id:this.dataIdNextNumber++};return this.move(n,e,t,a,1),n}numDataIds(){return this.dataIdMap.numDataIds()}async time(e){let t=v.now();return e(),{kernelMs:v.now()-t}}move(e,t,a,n,r){let s=this.dataIdNextNumber++;if(n==="string"){let u=t;this.dataIdMap.set(e,{id:s,stringBytes:u,shape:a,dtype:n,memoryOffset:null,refCount:r});return}let i=v.sizeFromShape(a),o=i*v.bytesPerElement(n),l=this.wasm._malloc(o)>>>0;this.dataIdMap.set(e,{id:s,memoryOffset:l,shape:a,dtype:n,refCount:r}),this.wasm.tfjs.registerTensor(s,i,l),t!=null&&this.wasm.HEAPU8.set(new Uint8Array(t.buffer,t.byteOffset,o),l)}async read(e){return this.readSync(e)}readSync(e,t,a){let{memoryOffset:n,dtype:r,shape:s,stringBytes:i}=this.dataIdMap.get(e);if(r==="string")return(t==null||t===0)&&(a==null||a>=i.length)?i:i.slice(t,a);t=t||0,a=a||v.sizeFromShape(s);let o=v.bytesPerElement(r),l=this.wasm.HEAPU8.slice(n+t*o,n+a*o);return Kie(l.buffer,r)}disposeData(e,t=!1){if(this.dataIdMap.has(e)){let a=this.dataIdMap.get(e);if(a.refCount--,!t&&a.refCount>0)return!1;this.wasm._free(a.memoryOffset),this.wasm.tfjs.disposeData(a.id),this.dataIdMap.delete(e)}return!0}refCount(e){return this.dataIdMap.has(e)?this.dataIdMap.get(e).refCount:0}incRef(e){let t=this.dataIdMap.get(e);t!=null&&t.refCount++}floatPrecision(){return 32}getMemoryOffset(e){return this.dataIdMap.get(e).memoryOffset}dispose(){this.wasm.tfjs.dispose(),"PThread"in this.wasm&&this.wasm.PThread.terminateAllThreads(),this.wasm=null}memory(){return{unreliable:!1}}makeOutput(e,t,a,n){let r;if(a==null)r=this.write(n!=null?n:null,e,t);else{let s=this.dataIdNextNumber++;r={id:s},this.dataIdMap.set(r,{id:s,memoryOffset:a,shape:e,dtype:t,refCount:1});let i=v.sizeFromShape(e);this.wasm.tfjs.registerTensor(s,i,a)}return{dataId:r,shape:e,dtype:t}}typedArrayFromHeap({shape:e,dtype:t,dataId:a}){let n=this.wasm.HEAPU8.buffer,{memoryOffset:r}=this.dataIdMap.get(a),s=v.sizeFromShape(e);switch(t){case"float32":return new Float32Array(n,r,s);case"int32":return new Int32Array(n,r,s);case"bool":return new Uint8Array(n,r,s);default:throw new Error(`Unknown dtype ${t}`)}}};function qie(e){return(t,a)=>(v.fetch(e,{credentials:"same-origin"}).then(n=>{n.ok||t.env.a(`failed to load wasm binary file at '${e}'`),n.arrayBuffer().then(r=>{WebAssembly.instantiate(r,t).then(s=>{a(s.instance,s.module)})})}),{})}function Ex(e,t,a){if(yh!=null)return yh;let n="tfjs-backend-wasm.wasm";return e&&t?n="tfjs-backend-wasm-threaded-simd.wasm":e&&(n="tfjs-backend-wasm-simd.wasm"),Cd!=null&&Cd[n]!=null?Cd[n]:a+n}async function Xie(){let[e,t]=await Promise.all([W().getAsync("WASM_HAS_SIMD_SUPPORT"),W().getAsync("WASM_HAS_MULTITHREAD_SUPPORT")]);return new Promise((a,n)=>{let r={};r.locateFile=(o,l)=>{if(o.endsWith(".worker.js")){let u=Hie.wasmWorkerContents.replace(/\n/g,"\\n"),p=new Blob([u],{type:"application/javascript"});return URL.createObjectURL(p)}return o.endsWith(".wasm")?Ex(e,t,kd!=null?kd:l):l+o},Ng&&(r.instantiateWasm=qie(Ex(e,t,kd!=null?kd:"")));let s=!1;r.onAbort=()=>{s||Td||(Td=!0,n({message:"Make sure the server can serve the `.wasm` file relative to the bundled js file. For more details see https://github.com/tensorflow/tfjs/blob/master/tfjs-backend-wasm/README.md#using-bundlers"}))};let i;t&&e&&yh==null?(r.mainScriptUrlOrBlob=new Blob(["var WasmBackendModuleThreadedSimd = "+Rx.toString()],{type:"text/javascript"}),i=Rx(r)):i=jie(r),i.then(o=>{s=!0,Td=!1;let l=null;o.tfjs={init:o.cwrap("init",null,[]),initWithThreadsCount:o.cwrap("init_with_threads_count",null,["number"]),getThreadsCount:o.cwrap("get_threads_count","number",[]),registerTensor:o.cwrap("register_tensor",null,["number","number","number"]),disposeData:o.cwrap("dispose_data",l,["number"]),dispose:o.cwrap("dispose",l,[])},a({wasm:o})}).catch(n)})}function Kie(e,t){switch(t){case"float32":return new Float32Array(e);case"int32":return new Int32Array(e);case"bool":return new Uint8Array(e);default:throw new Error(`Unknown dtype ${t}`)}}var Yie=["tfjs-backend-wasm.wasm","tfjs-backend-wasm-simd.wasm","tfjs-backend-wasm-threaded-simd.wasm"],yh=null,kd=null,Cd={},Td=!1,Ng=!1;function Zie(e,t=!1){if(Y1("setWasmPath has been deprecated in favor of setWasmPaths and will be removed in a future release."),Td)throw new Error("The WASM backend was already initialized. Make sure you call `setWasmPath()` before you call `tf.setBackend()` or `tf.ready()`");yh=e,Ng=t}function n0(e,t=!1){if(Td)throw new Error("The WASM backend was already initialized. Make sure you call `setWasmPaths()` before you call `tf.setBackend()` or `tf.ready()`");if(typeof e=="string")kd=e;else{Cd=e;let a=Yie.filter(n=>Cd[n]==null);if(a.length>0)throw new Error(`There were no entries found for the following binaries: ${a.join(",")}. Please either call setWasmPaths with a map providing a path for each binary, or with a string indicating the directory where all the binaries can be found.`)}Ng=t}var qw=-1,F1=-1;function Jie(e){qw=e}function Qie(){if(F1===-1)throw new Error("WASM backend not initialized.");return F1}var eoe="4.5.0",toe=2;Go("wasm",async()=>{let{wasm:e}=await Xie();return new jw(e)},toe);var En=W();En.registerFlag("WEBGPU_DEFERRED_SUBMIT_BATCH_SIZE",()=>15);En.registerFlag("WEBGPU_CPU_FORWARD",()=>!0);En.registerFlag("WEBGPU_MATMUL_PROGRAM_TYPE",()=>-1);En.registerFlag("WEBGPU_USE_NAIVE_CONV2D_TRANSPOSE",()=>!0);En.registerFlag("WEBGPU_USE_LOW_POWER_GPU",()=>!1);En.registerFlag("WEBGPU_CPU_HANDOFF_SIZE_THRESHOLD",()=>1e3);En.registerFlag("WEBGPU_USE_PROFILE_TOOL",()=>!1);En.registerFlag("WEBGPU_IMPORT_EXTERNAL_TEXTURE",()=>!0);En.registerFlag("WEBGPU_USE_NAIVE_CONV2D_DEBUG",()=>!1);En.registerFlag("WEBGPU_THRESHOLD_TO_INCREASE_WORKGROUPS_FOR_MATMUL",()=>0);En.registerFlag("WEBGPU_CONV_SEPARATE_IM2COL_SHADER",()=>!1);En.registerFlag("WEBGPU_PRINT_SHADER",()=>"");var aoe=class{constructor(e){e&&(this.vendor=e.vendor,this.architecture=e.architecture,this.intelGPUGeneration=this.getIntelGPUGeneration())}getIntelGPUGeneration(){if(this.isIntel()){if(this.architecture.startsWith("gen"))return Number(this.architecture.match(/\d+/));if(this.architecture.startsWith("xe"))return 12}return 0}isIntel(){return this.vendor==="intel"}},noe=class{constructor(e){this.device=e,this.numUsedBuffers=0,this.numFreeBuffers=0,this.freeBuffers=new Map,this.usedBuffers=new Map,this.numBytesUsed=0,this.numBytesAllocated=0}acquireBuffer(e,t,a=!1,n=!0){let r,s=Mx(e,t);return n?(this.freeBuffers.has(s)||this.freeBuffers.set(s,[]),this.freeBuffers.get(s).length>0?(r=this.freeBuffers.get(s).pop(),this.numFreeBuffers--):(r=this.device.createBuffer({size:e,usage:t,mappedAtCreation:a}),this.numBytesAllocated+=e)):(r=this.device.createBuffer({size:e,usage:t,mappedAtCreation:a}),this.numBytesAllocated+=e),this.usedBuffers.has(s)||this.usedBuffers.set(s,[]),this.usedBuffers.get(s).push(r),this.numUsedBuffers++,this.numBytesUsed+=e,r}releaseBuffer(e,t,a,n=!0){if(this.freeBuffers.size===0)return;let r=Mx(t,a),s=this.usedBuffers.get(r),i=s.indexOf(e);if(i<0)throw new Error("Cannot find the buffer in buffer manager");s[i]=s[s.length-1],s.pop(),this.numUsedBuffers--,this.numBytesUsed-=t,n?(this.freeBuffers.get(r).push(e),this.numFreeBuffers++):(e.destroy(),this.numBytesAllocated-=t)}getNumUsedBuffers(){return this.numUsedBuffers}getNumFreeBuffers(){return this.numFreeBuffers}dispose(){this.freeBuffers.forEach((e,t)=>{e.forEach(a=>{a.destroy()})}),this.usedBuffers.forEach((e,t)=>{e.forEach(a=>{a.destroy()})}),this.freeBuffers=new Map,this.usedBuffers=new Map,this.numUsedBuffers=0,this.numFreeBuffers=0,this.numBytesUsed=0,this.numBytesAllocated=0}};function Mx(e,t){return`${e}_${t}`}var roe=class{constructor(e){this.device=e,this.numUsedTextures=0,this.numFreeTextures=0,this.freeTextures=new Map,this.usedTextures=new Map,this.numBytesUsed=0,this.numBytesAllocated=0}acquireTexture(e,t,a,n){let r=_x(a),s=e*t*r,i=Fx(e,t,a,n);if(this.freeTextures.has(i)||this.freeTextures.set(i,[]),this.usedTextures.has(i)||this.usedTextures.set(i,[]),this.numBytesUsed+=s,this.numUsedTextures++,this.freeTextures.get(i).length>0){this.numFreeTextures--;let l=this.freeTextures.get(i).shift();return this.usedTextures.get(i).push(l),l}this.numBytesAllocated+=s;let o=this.device.createTexture({size:[e,t],format:a,usage:n});return this.usedTextures.get(i).push(o),o}releaseTexture(e,t,a,n,r){if(this.freeTextures.size===0)return;let s=Fx(t,a,n,r);this.freeTextures.has(s)||this.freeTextures.set(s,[]),this.freeTextures.get(s).push(e),this.numFreeTextures++,this.numUsedTextures--;let i=this.usedTextures.get(s),o=i.indexOf(e);if(o<0)throw new Error("Cannot release a texture that was never provided by this texture manager");i.splice(o,1);let l=_x(n),u=t*a*l;this.numBytesUsed-=u}getNumUsedTextures(){return this.numUsedTextures}getNumFreeTextures(){return this.numFreeTextures}dispose(){this.freeTextures.forEach((e,t)=>{e.forEach(a=>{a.destroy()})}),this.usedTextures.forEach((e,t)=>{e.forEach(a=>{a.destroy()})}),this.freeTextures=new Map,this.usedTextures=new Map,this.numUsedTextures=0,this.numFreeTextures=0,this.numBytesUsed=0,this.numBytesAllocated=0}};function Fx(e,t,a,n){return`${e}_${t}_${a}_${n}`}function _x(e){if(e==="rgba8unorm")return 16;throw new Error(`${e} is not supported!`)}function soe(e,t){if(Math.max(...e)>5)throw new Error("Cannot symbolically compute strides for rank > 6 tensor.");let a=e.length,n="xyzwuv",r=e.map(i=>`${t}.${n[i]}`),s=new Array(a-1);s[a-2]=r[a-1];for(let i=a-3;i>=0;--i)s[i]=`(${s[i+1]} * ${r[i+1]})`;return s}var Gu=(e,t,a)=>a==="int32"?`atomicAdd(${e}, bitcast<i32>(${t}));`:`
|
|
{
|
|
var oldValue = 0;
|
|
loop {
|
|
let newValueF32 = bitcast<f32>(oldValue) + (${t});
|
|
let newValue = bitcast<i32>(newValueF32);
|
|
let res = atomicCompareExchangeWeak(${e}, oldValue, newValue);
|
|
if res.exchanged {
|
|
break;
|
|
}
|
|
oldValue = res.old_value;
|
|
}
|
|
}`,ioe=(e,t,a,n,r)=>{let s={dtype:n.dtype,shape:n.shape},i=loe(a,s,t),o=e.createShaderModule({code:i,label:t.constructor.name}),l=e.createComputePipeline({compute:{module:o,entryPoint:"_start"},label:t.constructor.name,layout:"auto"}),u=W().get("WEBGPU_PRINT_SHADER");if(u!==""){u=u.toLowerCase();let p=u.split(",");(u==="all"||p.some(c=>r.toLowerCase().includes(c)))&&(console.group(r),console.debug(i),console.groupEnd())}return l},Xe=(e,t="f32")=>{switch(e){case 1:return`${t}`;case 2:return`vec2<${t}>`;case 3:return`vec3<${t}>`;case 4:return`vec4<${t}>`;default:throw new Error(`${e}-component ${t} is not supported.`)}};function Jt(e){if(e<=1)return"i32";if(e===2)return"vec2<i32>";if(e===3)return"vec3<i32>";if(e===4)return"vec4<i32>";if(e===5)return"vec5";if(e===6)return"vec6";throw Error(`GPU for rank ${e} is not yet supported`)}function Ir(e){if(e===0)return"x";if(e===1)return"y";if(e===2)return"z";if(e===3)return"w";if(e===4)return"u";if(e===5)return"v";throw Error(`Index ${e} is not yet supported`)}function de(...e){let t;switch(e.length){case 0:t=`
|
|
fn main()
|
|
`;break;case 1:t=`
|
|
fn main(${e[0]} : i32)
|
|
`;break;default:throw Error("Unreachable")}return t}function $x(e,t){let a;return a=`
|
|
${ooe(t)}
|
|
fn _start(@builtin(local_invocation_id) LocalId : vec3<u32>,
|
|
@builtin(global_invocation_id) GlobalId : vec3<u32>,
|
|
@builtin(local_invocation_index) LocalIndex: u32,
|
|
@builtin(workgroup_id) WorkgroupId : vec3<u32>,
|
|
@builtin(num_workgroups) NumWorkgroups : vec3<u32>) {
|
|
localId = LocalId;
|
|
localIndex = LocalIndex;
|
|
globalId = GlobalId;
|
|
numWorkgroups = NumWorkgroups;
|
|
workgroupId = WorkgroupId;
|
|
${e?"main(getGlobalIndex());":"main();"};
|
|
}
|
|
`,a}function ooe(e){return`
|
|
@compute @workgroup_size(${e.workgroupSize[0]}, ${e.workgroupSize[1]}, ${e.workgroupSize[2]})
|
|
`}function loe(e,t,a){let n=[],r=a.workgroupSize[0]*a.workgroupSize[1]*a.workgroupSize[2];if(a.outputComponent=a.outputComponent?a.outputComponent:1,n.push(`
|
|
|
|
var<private> localId: vec3<u32>;
|
|
var<private> localIndex: u32;
|
|
var<private> globalId: vec3<u32>;
|
|
var<private> numWorkgroups: vec3<u32>;
|
|
var<private> workgroupId: vec3<u32>;
|
|
|
|
// Only used when the y/z dimension of workgroup size is 1.
|
|
fn getGlobalIndex() -> i32 {
|
|
${Xw(a)?" return i32(globalId.x);":` return i32((workgroupId.z * numWorkgroups.x * numWorkgroups.y +
|
|
workgroupId.y * numWorkgroups.x + workgroupId.x) * ${r}u +
|
|
localIndex);
|
|
`}
|
|
}
|
|
`),a.isFromPixels){n.push(`
|
|
struct Uniform {
|
|
size : i32,
|
|
numChannels : i32,
|
|
outShapeStrides : vec2<i32>,
|
|
};
|
|
|
|
@group(0) @binding(0) var<storage, read_write> result: array<${Nl(t.dtype,a.outputComponent)}>;
|
|
@group(0) @binding(2) var<uniform> uniforms: Uniform;
|
|
`);let h=Px(a);return[Dx,n.join(`
|
|
`),V2(t.shape),a.getUserCode(),$x(h,a)].join(`
|
|
`)}let s,i,o="struct Uniforms { NAN : f32, INFINITY : f32, ";a.variableNames.forEach((h,f)=>{let m=Jt(e[f].shape.length);o+=`${h.charAt(0).toLowerCase()+h.slice(1)}Shape : ${m}, `,s=e[f].shape.length-1,i=Jt(s),o+=`${h.charAt(0).toLowerCase()+h.slice(1)}ShapeStrides: ${i}, `});let l=Jt(t.shape.length);o+=`outShape : ${l}, `,s=t.shape.length-1,i=Jt(s),o+=`
|
|
outShapeStrides: ${i}, `,a.size&&(o+="size : i32, "),a.uniforms&&(o+=a.uniforms),o+="};",o=yoe(o),n.push(o),a.atomic?n.push(`
|
|
@group(0) @binding(0) var<storage, read_write> result: array<atomic<i32>>;
|
|
`):n.push(`
|
|
@group(0) @binding(0) var<storage, read_write> result: array<${Nl(t.dtype,a.outputComponent)}>;
|
|
`),a.variableNames.forEach((h,f)=>{n.push(`
|
|
@group(0) @binding(${1+f}) var<storage, read> ${h}: array<${a.variableComponents?Nl(e[f].dtype,a.variableComponents[f]):Nl(e[f].dtype,a.outputComponent)}>;
|
|
`)}),o!==""&&n.push(`
|
|
@group(0) @binding(${1+a.variableNames.length}) var<uniform> uniforms: Uniforms;
|
|
`);let u=foe(t.shape,a.dispatchLayout),p=[Dx,n.join(`
|
|
`)+doe,V2(t.shape),u,moe(t.shape.length)];a.atomic||p.push(goe(t.shape,t.dtype,a.outputComponent)),a.variableNames.forEach((h,f)=>{p.push(`${V2(e[f].shape,h)}`)});let c=e.map((h,f)=>hoe(h,t.shape,a.variableComponents?a.variableComponents[f]:a.outputComponent,a.dispatchLayout.x.length===t.shape.length)).join(`
|
|
`);p.push(c),p.push(a.getUserCode());let d=Px(a);return p.push($x(d,a)),p.join(`
|
|
`)}function uoe(e,t,a,n){let r=e.shaderKey;if(e.isFromPixels)return r;let s=a.map(p=>p.dtype).concat(n.dtype),i=a.map(p=>S.getBroadcastDims(p.shape,n.shape)),o=a.map(p=>v.arraysEqual(p.shape,n.shape)).join("_"),l=i.map(p=>p.join("_")).join(";"),u=Xw(e)?"flatDispatch":"";return r+="_"+(e.workgroupSize?e.workgroupSize.join(","):"")+t.map(p=>p.length).join(",")+s.join(",")+e.variableNames.join(",")+l+o+u,r}var Dx=`
|
|
struct vec5 {x: i32, y: i32, z: i32, w: i32, u: i32};
|
|
struct vec6 {x: i32, y: i32, z: i32, w: i32, u: i32, v: i32};
|
|
|
|
// Checks whether coordinates lie within the bounds of the shape.
|
|
fn coordsInBounds2D(coord : vec2<i32>, shape : vec2<i32>) -> bool {
|
|
return all(coord >= vec2<i32>(0)) && all(coord < shape);
|
|
}
|
|
fn coordsInBounds3D(coord : vec3<i32>, shape : vec3<i32>) -> bool {
|
|
return all(coord >= vec3<i32>(0)) && all(coord < shape);
|
|
}
|
|
fn coordsInBounds4D(coord : vec4<i32>, shape : vec4<i32>) -> bool {
|
|
return all(coord >= vec4<i32>(0)) && all(coord < shape);
|
|
}
|
|
|
|
fn getIndexFromCoords1D(coord : i32, shape : i32) -> i32 {
|
|
return coord;
|
|
}
|
|
fn getIndexFromCoords2D(coords : vec2<i32>, shape : vec2<i32>) -> i32 {
|
|
return dot(coords, vec2<i32>(shape.y, 1));
|
|
}
|
|
fn getIndexFromCoords3D(coords : vec3<i32>, shape : vec3<i32>) -> i32 {
|
|
return dot(coords, vec3<i32>(shape.y * shape.z, shape.z, 1));
|
|
}
|
|
fn getIndexFromCoords4D(coords : vec4<i32>, shape : vec4<i32>) -> i32 {
|
|
return dot(coords, vec4<i32>(
|
|
shape.y * shape.z * shape.w, shape.z * shape.w, shape.w, 1));
|
|
}
|
|
fn getIndexFromCoords5D(coords : vec5, shape : vec5) -> i32 {
|
|
let shapeStrides: vec5 = vec5(shape.y * shape.z * shape.w * shape.u, shape.z * shape.w * shape.u, shape.w * shape.u, shape.u, 1);
|
|
return coords.x*shapeStrides.x + coords.y*shapeStrides.y + coords.z*shapeStrides.z + coords.w*shapeStrides.w + coords.u*shapeStrides.u;
|
|
}
|
|
fn getIndexFromCoords6D(coords : vec6, shape : vec6) -> i32 {
|
|
let shapeStrides: vec6 = vec6(shape.y * shape.z * shape.w * shape.u * shape.v, shape.z * shape.w * shape.u * shape.v, shape.w * shape.u * shape.v, shape.u * shape.v, shape.v, 1);
|
|
return coords.x*shapeStrides.x + coords.y*shapeStrides.y + coords.z*shapeStrides.z + coords.w*shapeStrides.w + coords.u*shapeStrides.u + coords.v*shapeStrides.v;
|
|
}
|
|
|
|
fn idiv(a: i32, b: i32, sign: f32) -> i32 {
|
|
var res: i32 = a / b;
|
|
let modulo: i32 = a % b;
|
|
if (sign < 0. && modulo != 0) {
|
|
res = res - 1;
|
|
}
|
|
return res;
|
|
}
|
|
|
|
// NaN defination in IEEE 754-1985 is :
|
|
// - sign = either 0 or 1.
|
|
// - biased exponent = all 1 bits.
|
|
// - fraction = anything except all 0 bits (since all 0 bits represents infinity).
|
|
// https://en.wikipedia.org/wiki/IEEE_754-1985#Representation_of_non-numbers
|
|
fn isnan(val: f32) -> bool {
|
|
let floatToUint: u32 = bitcast<u32>(val);
|
|
return (floatToUint & 0x7fffffffu) > 0x7f800000u;
|
|
}
|
|
fn isnanVec4(val : vec4<f32>) -> vec4<bool> {
|
|
let floatToUint: vec4<u32> = bitcast<vec4<u32>>(val);
|
|
return (floatToUint & vec4<u32>(0x7fffffffu)) > vec4<u32>(0x7f800000u);
|
|
}
|
|
`,doe=`
|
|
fn isinf(val: f32) -> bool {
|
|
return abs(val) == uniforms.INFINITY;
|
|
}
|
|
`;function V2(e,t=""){let a=e.length,n=t!==""?`get${t.charAt(0).toUpperCase()+t.slice(1)}CoordsFromIndex`:"getCoordsFromIndex",r=t!==""?`${t.charAt(0).toLowerCase()+t.slice(1)}ShapeStrides`:"outShapeStrides";if(a<=1)return`fn ${n}(index : i32) -> i32 { return index; }`;let s=v.computeStrides(e),i=Jt(a),o=[];for(let u=0;u<a;u++)o.push(`d${u}`);if(s.length===1)return` fn ${n}(index : i32) -> vec2<i32> {
|
|
let d0 = index / uniforms.${r}; let d1 = index - d0 * uniforms.${r};
|
|
return vec2<i32>(d0, d1);
|
|
}`;let l;return l="var index2 = index;"+s.map((u,p)=>{let c=`let ${o[p]} = index2 / uniforms.${r}.${Ir(p)}`,d=p===s.length-1?`let ${o[p+1]} = index2 - ${o[p]} * uniforms.${r}.${Ir(p)}`:`index2 = index2 - ${o[p]} * uniforms.${r}.${Ir(p)}`;return`${c}; ${d};`}).join(""),`
|
|
fn ${n}(index : i32) -> ${i} {
|
|
${l}
|
|
return ${i}(${o.join(",")});
|
|
}
|
|
`}function poe(e,t){let a=e.name,n=e.shape.length,r=Jt(n),s="get"+a.charAt(0).toUpperCase()+a.slice(1),i=["d0","d1","d2","d3","d4","d5"].slice(0,n),o=i.map(p=>`${p} : i32`).join(", ");if(n<1)return`
|
|
fn ${s}() -> ${Xe(t)} {
|
|
return ${Xe(t)}(${a}[0]);
|
|
}
|
|
`;let l=`uniforms.${a.charAt(0).toLowerCase()+a.slice(1)}Shape`,u=`${n}D`;return n===0&&(u="1D"),`
|
|
fn ${s}(${o}) -> ${Xe(t)} {
|
|
return ${Xe(t)}(${a}[getIndexFromCoords${u}(${r}(${i.join(",")}),
|
|
${l})${t===1?"":` / ${t}`}]);
|
|
}
|
|
`}function coe(e,t,a,n){let r=e.name,s=r.charAt(0).toUpperCase()+r.slice(1),i="get"+s+"ByOutput",o=e.shape.length,l=t.length,u=Jt(l);if(v.arraysEqual(e.shape,t)&&n)return`
|
|
fn ${i}Index(globalIndex : i32) -> ${Xe(a)} {
|
|
return ${Xe(a)}(${r}[globalIndex]);
|
|
}
|
|
|
|
fn ${i}Coords(coords : ${u}) -> ${Xe(a)} {
|
|
return ${Xe(a)}(${r}[${l>1?"getOutputIndexFromCoords(coords)":"coords"}${a===1?"":` / ${a}`}]);
|
|
}
|
|
`;let p=S.getBroadcastDims(e.shape,t),c=l-o,d="";if(o===0)return`
|
|
fn ${i}Index(globalIndex : i32) -> ${Xe(a)}{
|
|
return get${s}();
|
|
}
|
|
|
|
fn ${i}Coords(coords : ${u}) -> ${Xe(a)}{
|
|
return get${s}();
|
|
}
|
|
`;l<2&&p.length>=1?d="coords = 0;":d=p.map(g=>`coords.${Ir(g+c)} = 0;`).join(`
|
|
`);let h="";if(l<2&&o>0)h="coords";else if(l>1){let g=Jt(o),y=e.shape.map((x,A)=>`coords.${Ir(A+c)}`).join(", ");h=`${g}(${y})`}else h="coords";let f=`uniforms.${r.charAt(0).toLowerCase()+r.slice(1)}Shape`,m=`${o}D`;return`
|
|
fn ${i}Index(globalIndex : i32) -> ${Xe(a)} {
|
|
var coords = getCoordsFromIndex(globalIndex);
|
|
${d}
|
|
return ${Xe(a)}(${r}[getIndexFromCoords${m}(${h}, ${f})${a===1?"":` / ${a}`}]);
|
|
}
|
|
|
|
fn ${i}Coords(coordsIn : ${u}) -> ${Xe(a)} {
|
|
var coords = coordsIn;
|
|
${d}
|
|
return ${Xe(a)}(${r}[getIndexFromCoords${m}(${h}, ${f})${a===1?"":` / ${a}`}]);
|
|
}
|
|
`}function hoe(e,t,a,n){let r=poe(e,a);return e.shape.length<=t.length&&(r+=coe(e,t,a,n)),r}function foe(e,t){let{x:a,y:n=[],z:r=[]}=t,s=e.length,i=a.length+n.length+r.length;if(i!==s)return"";if(a.length===s)return`fn getOutputCoords() -> ${Jt(s)}{
|
|
let globalIndex = getGlobalIndex();
|
|
return getCoordsFromIndex(globalIndex);
|
|
}
|
|
`;let o="",l=[a,n,r];for(let d=0;d<l.length;d++){let h=l[d];if(h.length!==0)if(h.length===1)o+=`let d${h[0]} = i32(globalId[${d}]);`;else{let f=soe(h,"uniforms.outShape");o+=`var index${d} = i32(globalId[${d}]);`;for(let m=0;m<f.length;m++)o+=`let d${h[m]} = index${d} / ${f[m]};`,m===f.length-1?o+=`let d${h[m+1]} = index${d} - d${h[m]} * ${f[m]};`:o+=`index${d} = index${d} - d${h[m]} * ${f[m]};`}}let u=[];for(let d=0;d<i;d++)u.push(`d${d}`);let p=Jt(i),c=`fn getOutputCoords() -> ${p} {
|
|
${o}
|
|
`;return u.length===0?c+=`return ${p}(0); }`:c+=`return ${p}(${u.join(",")}); }`,c}function moe(e){let t="";switch(e){case 0:case 1:t+=`
|
|
fn getOutputIndexFromCoords(coords : i32) -> i32 {
|
|
return coords;
|
|
}
|
|
`;break;case 2:t+=`
|
|
fn getOutputIndexFromCoords(coords : vec2<i32>) -> i32 {
|
|
return dot(coords, vec2<i32>(uniforms.outShapeStrides, 1));
|
|
}
|
|
`;break;case 3:t+=`
|
|
fn getOutputIndexFromCoords(coords : vec3<i32>) -> i32 {
|
|
return dot(coords, vec3<i32>(uniforms.outShapeStrides.x, uniforms.outShapeStrides.y, 1));
|
|
}
|
|
`;break;case 4:t+=`
|
|
fn getOutputIndexFromCoords(coords : vec4<i32>) -> i32 {
|
|
return dot(coords, vec4<i32>(
|
|
uniforms.outShapeStrides.x, uniforms.outShapeStrides.y, uniforms.outShapeStrides.z, 1));
|
|
}
|
|
`;break;case 5:t+=`
|
|
fn getOutputIndexFromCoords(coords : vec5) -> i32 {
|
|
return coords.x * uniforms.outShapeStrides.x +
|
|
coords.y * uniforms.outShapeStrides.y +
|
|
coords.z * uniforms.outShapeStrides.z +
|
|
coords.w * uniforms.outShapeStrides.w +
|
|
coords.u;
|
|
}
|
|
`;break;case 6:t+=`
|
|
fn getOutputIndexFromCoords(coords : vec6) -> i32 {
|
|
return coords.x * uniforms.outShapeStrides.x +
|
|
coords.y * uniforms.outShapeStrides.y +
|
|
coords.z * uniforms.outShapeStrides.z +
|
|
coords.w * uniforms.outShapeStrides.w +
|
|
coords.u * uniforms.outShapeStrides.u +
|
|
coords.v;
|
|
}
|
|
`;break;default:v.assert(!1,()=>`Unsupported ${e}D shape`);break}return t}function Xw(e){return e.dispatch[1]===1&&e.dispatch[2]===1}function Nl(e,t=1){if(e==="float32")return Xe(t,"f32");if(e==="int32"||e==="bool")return Xe(t,"i32");throw new Error(`type ${e} is not supported.`)}function goe(e,t,a){let n=e.length,r=Nl(t,a),s=`fn setOutputAtIndex(flatIndex : i32, value : ${Xe(a)}) {
|
|
result[flatIndex] = ${r}(value);
|
|
}
|
|
|
|
fn setOutputAtIndexI32(flatIndex : i32, value : ${Xe(a,"i32")}) {
|
|
result[flatIndex] = ${r}(value);
|
|
}
|
|
`;if(n>=2){let i=["d0","d1","d2","d3","d4","d5"].slice(0,n),o=Jt(n);s+=`
|
|
fn setOutputAtCoords(${i.map(l=>`${l} : i32`).join(", ")}, value : ${Xe(a)}) {
|
|
let flatIndex = getOutputIndexFromCoords(${o}(${i.join(", ")}));
|
|
setOutputAtIndex(flatIndex${a===1?"":` / ${a}`}, value);
|
|
}
|
|
fn setOutputAtCoordsI32(${i.map(l=>`${l} : i32`).join(", ")}, value : ${Xe(a,"i32")}) {
|
|
let flatIndex = getOutputIndexFromCoords(${o}(${i.join(", ")}));
|
|
setOutputAtIndexI32(flatIndex${a===1?"":` / ${a}`}, value);
|
|
}
|
|
`}return s}function yoe(e){let t=/(\w+)\s*:\s*vec(5|6)/g;e=e.replace(t,n=>"@align(16) "+n);let a=/vec(5|6)\s*,\s*(\w+)/g;return e=e.replace(a,(n,r,s)=>`vec${r}, @align(16) ${s}`),e}function Px(e){return!(e.dispatchLayout.hasOwnProperty("y")&&e.dispatchLayout.y.length!==0||e.dispatchLayout.hasOwnProperty("z")&&e.dispatchLayout.z.length!==0)}var Kw={};Ze(Kw,{GPUBytesPerElement:()=>_1,MatMulProgramType:()=>Pn,assertNotComplex:()=>Fg,computeDispatch:()=>ce,computeWorkPerThreadForConv2d:()=>Eg,computeWorkgroupInfoForMatMul:()=>Yw,computeWorkgroupSizeForConv2d:()=>Rg,flatDispatchLayout:()=>be,isWebGPUSupported:()=>Mg,tilesFitEvenlyIntoShape:()=>xoe});var Vs=e=>{let t=1;for(let a=0;a<e.length;a++)t*=e[a];return t};function xoe(e,t){if(e.length!==t.length)throw new Error(`Cannot compute whether rank ${e.length} tiles fit evenly into rank ${t.length} shape - ranks must match.`);return t.every((a,n)=>a%e[n]===0)}function ce(e,t,a=[1,1,1],n=[1,1,1]){let[r,s,i]=[Math.ceil(Vs(e.x.map(o=>t[o]))/(a[0]*n[0])),e.y?Math.ceil(Vs(e.y.map(o=>t[o]))/(a[1]*n[1])):1,e.z?Math.ceil(Vs(e.z.map(o=>t[o]))/(a[2]*n[2])):1];return[r,s,i]}function Yw(e,t,a,n=!1){let r=[8,8,1],s=[4,4,1];return n||(e<=8&&(s[1]=1),t<=16&&a<=16&&(r[0]=4)),{workgroupSize:r,elementsPerThread:s}}function Rg(e,t,a=!1){if(a)return[8,8,1];let n=Vs(e.x.map(s=>t[s])),r=Vs(e.y.map(s=>t[s]));return n<=4?[4,16,1]:r<=4?[16,4,1]:[16,16,1]}function Eg(e,t,a=!1){if(a)return[4,4,1];let n=Vs(e.x.map(s=>t[s])),r=Vs(e.y.map(s=>t[s]));return n<=4?[1,2,1]:r<=4?[2,1,1]:[2,2,1]}function be(e){return{x:e.map((t,a)=>a)}}function _1(e){if(e==="float32"||e==="int32"||e==="bool"||e==="string")return 4;if(e==="complex64")return 8;throw new Error(`Unknown dtype ${e}`)}function Mg(){return(typeof window!="undefined"||typeof WorkerGlobalScope!="undefined")&&!!navigator.gpu}function Fg(e,t){Array.isArray(e)||(e=[e]),e.forEach(a=>{a!=null&&v.assert(a.dtype!=="complex64",()=>`${t} does not support complex64 tensors in the WebGPU backend.`)})}var Pn;(function(e){e[e.MatMulReduceProgram=0]="MatMulReduceProgram",e[e.MatMulSplitKProgram=1]="MatMulSplitKProgram",e[e.MatMulSmallOutputSizeProgram=2]="MatMulSmallOutputSizeProgram",e[e.MatMulPackedProgram=3]="MatMulPackedProgram",e[e.MatMulMax=4]="MatMulMax"})(Pn||(Pn={}));var Aoe=W().getNumber("WEBGPU_CPU_HANDOFF_SIZE_THRESHOLD"),boe=(e,t)=>{let a=e.limits.maxComputeWorkgroupsPerDimension,n=t.dispatchLayout,r=t.dispatch;if(r.every(i=>i<=a))return r;v.assert(r[0]>a&&n.y===void 0&&n.z===void 0,()=>"Dispatch size exceeds WebGPU limits in Y or Z dimension.");let s=Math.ceil(Math.sqrt(r[0]));return s>a?(s=Math.ceil(Math.cbrt(r[0])),v.assert(s<=a,()=>"Total dispatch size exceeds WebGPU maximum."),[s,s,s]):[s,s,1]},r0=class extends Kl{nextDataId(){return r0.nextDataId++}constructor(e,t){if(super(),this.commandQueueOwnedIds=new WeakSet,this.dispatchNumberInEncoder=0,this.disposed=!1,this.downloadWaitMs=0,this.tensorDataPendingDisposal=[],this.stagingPendingDisposal=[],this.uniformPendingDisposal=[],this.uploadWaitMs=0,!Mg())throw new Error("WebGPU is not supported on this device");this.pipelineCache={},this.device=e,this.queue=e.queue,this.currentCommandEncoder=null,this.currentComputePass=null,this.supportTimeQuery=e.features.has("timestamp-query-inside-passes"),this.adapterInfo=new aoe(t),this.thresholdToIncreaseWorkgroups=this.adapterInfo.intelGPUGeneration>=12?16:8,this.bufferManager=new noe(this.device),this.textureManager=new roe(this.device),this.tensorMap=new Zd(this,It()),this.supportTimeQuery&&(this.querySet=this.device.createQuerySet({type:"timestamp",count:2})),W().getBool("WEBGPU_USE_PROFILE_TOOL")&&(this.dummyCanvas=document.createElement("canvas"),this.dummyCanvas.width=1,this.dummyCanvas.height=1,this.dummyContext=this.dummyCanvas.getContext("webgpu"),this.dummyContext.configure({device:e,format:"bgra8unorm"}),document.body.appendChild(this.dummyCanvas))}floatPrecision(){return 32}defaultGpuBufferUsage(){return GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC|GPUBufferUsage.COPY_DST}disposeData(e,t=!1){if(this.tensorDataPendingDisposal.indexOf(e)>=0)return!1;if(!this.tensorMap.has(e))return!0;let a=this.tensorMap.get(e);if(this.decRef(e),!t&&a.refCount>0)return!1;if(this.commandQueueOwnedIds.has(e))return this.tensorDataPendingDisposal.push(e),!1;let{complexTensorInfos:n}=this.tensorMap.get(e);return n!=null&&(this.disposeData(n.real.dataId,t),this.disposeData(n.imag.dataId,t)),this.releaseResource(e),this.tensorMap.delete(e),!0}memory(){return{numBytesInGPU:this.bufferManager.numBytesUsed,numBytesAllocatedInGPU:this.bufferManager.numBytesAllocated,unreliable:!1}}releaseResource(e){let t=this.tensorMap.get(e);if(!(!t||!t.resourceInfo)){if(t.external){t.resourceInfo=null;return}if("texture"in t.resourceInfo){let a=t.resourceInfo;a.texture instanceof GPUTexture&&this.textureManager.releaseTexture(a.texture,a.width,a.height,a.format,a.usage),a.texture=null}else{let a=t.resourceInfo;this.bufferManager.releaseBuffer(a.buffer,a.size,a.usage),a.buffer=null}t.resourceInfo=null}}refCount(e){return this.tensorMap.has(e)?this.tensorMap.get(e).refCount:0}incRef(e){let t=this.tensorMap.get(e);t.refCount++}decRef(e){if(this.tensorMap.has(e)){let t=this.tensorMap.get(e);t.refCount--}}write(e,t,a){if(a==="complex64"&&e!=null)throw new Error("Cannot write to a complex64 dtype. Please use tf.complex(real, imag).");let n={id:this.nextDataId()};return this.tensorMap.set(n,{dtype:a,shape:t,values:e,refCount:1}),n}move(e,t,a,n,r){if(n==="complex64")throw new Error("Cannot write to a complex64 dtype. Please use tf.complex(real, imag).");this.tensorMap.set(e,{dtype:n,shape:a,values:t,refCount:r})}submitQueue(){this.ensureComputePassEnded(),this.queue.submit([this.currentCommandEncoder.finish()]),this.currentCommandEncoder=null,this.dispatchNumberInEncoder=0,this.commandQueueOwnedIds=new WeakSet,this.tensorDataPendingDisposal.forEach(e=>{this.releaseResource(e),this.tensorMap.delete(e)}),this.uniformPendingDisposal.forEach(e=>this.bufferManager.releaseBuffer(e.buffer,e.size,e.usage)),this.stagingPendingDisposal.forEach(e=>this.bufferManager.releaseBuffer(e.buffer,e.size,e.usage,!1)),this.tensorDataPendingDisposal=[],this.uniformPendingDisposal=[],this.stagingPendingDisposal=[]}ensureCommandEncoderReady(){this.currentCommandEncoder||(this.currentCommandEncoder=this.device.createCommandEncoder())}ensureComputePassEnded(){this.currentComputePass&&(this.currentComputePass.end(),this.currentComputePass=null)}getComputePass(){return this.currentComputePass||(this.currentComputePass=this.currentCommandEncoder.beginComputePass()),this.currentComputePass}async getBufferData(e,t){let a=this.bufferManager.acquireBuffer(t,GPUBufferUsage.COPY_DST|GPUBufferUsage.MAP_READ);this.ensureCommandEncoderReady(),this.ensureComputePassEnded(),this.currentCommandEncoder.copyBufferToBuffer(e,0,a,0,t),this.submitQueue(),await a.mapAsync(GPUMapMode.READ);let n=a.getMappedRange().slice(0);return a.unmap(),a!=null&&this.bufferManager.releaseBuffer(a,t,GPUBufferUsage.COPY_DST|GPUBufferUsage.MAP_READ),W().getBool("WEBGPU_USE_PROFILE_TOOL")&&(v.assert(this.dummyContext!==void 0,()=>"Fail to get context for profiling tool"),this.dummyContext.getCurrentTexture()),n}convertAndCacheOnCPU(e,t){let a=this.tensorMap.get(e);return this.releaseResource(e),a.values=t,a.values}readSync(e){let t=this.tensorMap.get(e),{values:a,complexTensorInfos:n}=t;if(a!=null||t.dtype==="string")return a;if(t.dtype==="complex64"){let f=this.readSync(n.real.dataId),m=this.readSync(n.imag.dataId),g=v.convertBackendValuesAndArrayBuffer(S.mergeRealAndImagArrays(f,m).buffer,"float32");return this.convertAndCacheOnCPU(e,g),g}let r=["opaque","premultiplied"],s=t.resourceInfo,i=s.size;v.assert(i%4===0,()=>"Because there is 4 bytes for one pixel, buffer size must be multiple of 4.");let o=i/4,l=new ArrayBuffer(i),u=256,p=256,c=r.map(f=>new OffscreenCanvas(u,p)),d=new OffscreenCanvas(u,p);this.ensureComputePassEnded(),c.map((f,m)=>{let g=f.getContext("webgpu");return g.configure({device:this.device,format:"bgra8unorm",usage:GPUTextureUsage.COPY_DST,alphaMode:r[m]}),g.getCurrentTexture()}).map((f,m)=>{let g=u*4,y=(T,N,M)=>{this.ensureCommandEncoderReady(),this.currentCommandEncoder.copyBufferToTexture({buffer:s.buffer,bytesPerRow:g,offset:M},{texture:f},{width:T,height:N}),this.submitQueue();let F=d.getContext("2d",{willReadFrequently:!0});F.clearRect(0,0,T,N),F.drawImage(c[m],0,0);let E=F.getImageData(0,0,T,N).data,I=r[m],_=new Uint8ClampedArray(l,M,T*N*4);for(let O=0;O<_.length;O+=4)if(I==="premultiplied")_[O+3]=E[O+3];else{let B=E[O];_[O]=E[O+2],_[O+1]=E[O+1],_[O+2]=B}},x=Math.floor(o/(u*p)),A=u,b=p,w=0;for(let T=0;T<x;T++)y(A,b,w),w+=u*p*4;let C=o%(u*p);b=Math.floor(C/u),b>0&&(y(A,b,w),w+=b*(u*4)),A=C%u,A>0&&y(A,1,w)});let h=v.convertBackendValuesAndArrayBuffer(l,t.dtype);return this.convertAndCacheOnCPU(e,h),h}async read(e){if(!this.tensorMap.has(e))throw new Error(`Tensor ${e} was not registered!`);let t=this.tensorMap.get(e),{values:a}=t;if(a!=null)return a;let n;if(t.dtype==="complex64"){let r=await Promise.all([this.read(t.complexTensorInfos.real.dataId),this.read(t.complexTensorInfos.imag.dataId)]),s=r[0],i=r[1];n=S.mergeRealAndImagArrays(s,i)}else{let r=t.resourceInfo,s=await this.getBufferData(r.buffer,r.size);n=v.convertBackendValuesAndArrayBuffer(s,t.dtype)}return this.convertAndCacheOnCPU(e,n),n}copyBuffer(e,t,a){let n=this.bufferManager.acquireBuffer(t,a);return this.ensureCommandEncoderReady(),this.ensureComputePassEnded(),this.currentCommandEncoder.copyBufferToBuffer(e,0,n,0,t),this.submitQueue(),n}createTensorFromGPUData(e,t,a){let n=e.buffer;if(a==="complex64")throw new Error("Cannot write to a complex64 dtype. ");let r={id:this.nextDataId()};this.tensorMap.set(r,{dtype:a,shape:t,values:null,refCount:1,external:e.zeroCopy});let s=this.tensorMap.get(r),i=_1(s.dtype)*v.sizeFromShape(s.shape);if(e.buffer.size<i)throw new Error(`GPUBuffer size(${e.buffer.size}) is smaller than tensor size(${i})!`);if((e.buffer.usage&(GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC))!==(GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC))throw new Error("GPUBuffer.usage should include GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_SRC!");return e.zeroCopy!==!0&&(n=this.copyBuffer(n,i,n.usage)),s.resourceInfo={size:n.size,usage:n.usage,buffer:n},It().makeTensorFromDataId(r,t,a,this)}readToGPU(e){let t=this.tensorMap.get(e),{values:a,dtype:n,shape:r,resourceInfo:s}=t;if(n==="complex64")throw new Error("Does not support reading buffer for complex64 dtype.");if(s==null)throw a!=null?new Error("Data is not on GPU but on CPU."):new Error("There is no data on GPU or CPU.");let i=s.size,o=this.bufferManager.acquireBuffer(i,s.usage);this.ensureCommandEncoderReady(),this.ensureComputePassEnded(),this.currentCommandEncoder.copyBufferToBuffer(s.buffer,0,o,0,i),this.submitQueue();let l=this.makeTensorInfo(r,n),u=It().makeTensorFromTensorInfo(l),p=this.tensorMap.get(l.dataId);return p.resourceInfo={size:i,usage:this.defaultGpuBufferUsage(),buffer:o},{tensorRef:u,buffer:o}}bufferSync(e){let t=this.readSync(e.dataId);if(e.dtype==="string")try{let a=t.map(n=>v.decodeString(n));return $e(e.shape,e.dtype,a)}catch(a){throw new Error("Failed to decode encoded string bytes into utf-8")}return $e(e.shape,e.dtype,t)}async time(e){this.supportTimeQuery||console.warn("This device doesn't support timestamp-query-inside-passes extension. Start Chrome browser with flag --disable-dawn-features=disallow_unsafe_apis then try again. Otherwise, zero will be shown for the kernel time when profiling mode is enabled. Using performance.now is not workable for webgpu since it doesn't support synchronous data read from GPU.");let t=this.activeTimers,a=[],n=!1;this.programTimersStack==null?(this.programTimersStack=a,n=!0):this.activeTimers.push(a),this.activeTimers=a,e();let r=v.flatten(this.activeTimers.map(l=>l.query)).filter(l=>l!=null),s=v.flatten(this.activeTimers.map(l=>l.name)).filter(l=>l!=null);this.activeTimers=t,n&&(this.programTimersStack=null);let i={uploadWaitMs:this.uploadWaitMs,downloadWaitMs:this.downloadWaitMs,kernelMs:null,wallMs:null},o=await Promise.all(r);return i.kernelMs=v.sum(o),i.getExtraProfileInfo=()=>o.map((l,u)=>({name:s[u],ms:l})).map(l=>`${l.name}: ${l.ms}`).join(", "),this.uploadWaitMs=0,this.downloadWaitMs=0,i}makeTensorInfo(e,t,a){return t==="string"&&a!=null&&a.length>0&&v.isString(a[0])&&(a=a.map(n=>v.encodeString(n))),{dataId:this.write(a,e,t),shape:e,dtype:t}}tensorToBinding(e){if(!e)return null;let t=this.tensorMap.get(e.dataId);if("texture"in t.resourceInfo){let n=t.resourceInfo;return n.texture instanceof GPUExternalTexture?n.texture:n.texture.createView()}let a=t.resourceInfo;return{offset:0,size:a.size,buffer:a.buffer}}async getQueryTime(e){return this.supportTimeQuery?this.getTimeFromQuerySet(e):0}uploadToGPU(e){let t=this.tensorMap.get(e);if(t.resourceInfo)return;let a=_1(t.dtype)*v.sizeFromShape(t.shape),n;if(t.values){if(n=this.bufferManager.acquireBuffer(a,this.defaultGpuBufferUsage(),!0),n.mapState==="unmapped"){let r=this.bufferManager.acquireBuffer(a,GPUBufferUsage.MAP_WRITE|GPUBufferUsage.COPY_SRC,!0,!1),s=r.getMappedRange();t.dtype==="int32"||t.dtype==="bool"?new Int32Array(s).set(t.values):new Float32Array(s).set(t.values),r.unmap(),this.ensureCommandEncoderReady(),this.ensureComputePassEnded(),this.currentCommandEncoder.copyBufferToBuffer(r,0,n,0,a),this.stagingPendingDisposal.push({size:a,usage:GPUBufferUsage.MAP_WRITE|GPUBufferUsage.COPY_SRC,buffer:r})}else{let r=n.getMappedRange();t.dtype==="int32"||t.dtype==="bool"?new Int32Array(r).set(t.values):new Float32Array(r).set(t.values),n.unmap()}t.values=null}else n=this.bufferManager.acquireBuffer(a,this.defaultGpuBufferUsage());t.resourceInfo={size:a,usage:this.defaultGpuBufferUsage(),buffer:n}}makeUniforms(e){let t=0,a=0,n=[],r=1;e.forEach(l=>{l.data.length===0&&(l.data=[1]);let u;switch(l.data.length){case 1:u=4;break;case 2:u=8;break;case 3:u=16;break;case 4:u=16;break;case 5:u=16;break;case 6:u=16;break;default:v.assert(!1,()=>`Unsupported ${l.data.length}D shape`)}(a===5||a===6)&&(u=16),u>r&&(r=u),t=Math.ceil(t/u)*u,a=l.data.length,n.push(t),t+=l.data.length*4}),t=Math.ceil(t/r)*r;let s=new ArrayBuffer(t);e.forEach((l,u)=>{let p=n[u];l.type==="int32"?new Int32Array(s,p,l.data.length).set(l.data):l.type==="uint32"?new Uint32Array(s,p,l.data.length).set(l.data):new Float32Array(s,p,l.data.length).set(l.data)});let i=this.bufferManager.acquireBuffer(t,GPUBufferUsage.COPY_DST|GPUBufferUsage.UNIFORM);this.queue.writeBuffer(i,0,s,0,t);let o={size:t,usage:GPUBufferUsage.COPY_DST|GPUBufferUsage.UNIFORM,buffer:i};return this.uniformPendingDisposal.push(o),{offset:0,size:t,buffer:i}}runWebGPUProgram(e,t,a,n,r){if(r||(r=this.makeTensorInfo(e.outputShape,a)),v.sizeFromShape(r.shape)===0)return this.tensorMap.get(r.dataId).values=v.getTypedArrayFromDType(r.dtype,0),r;this.uploadToGPU(r.dataId),e.dispatch=boe(this.device,e);let s=[],i=[];if(!e.isFromPixels){s.push({type:"float32",data:[NaN]},{type:"float32",data:[1/0]}),i=t.concat(r).map(m=>m.shape);let f="int32";if(i.map(m=>{s.push({type:f,data:m});let g=v.computeStrides(m);s.push({type:f,data:g})}),e.size){let m=v.sizeFromShape(e.outputShape);s.push({type:f,data:[e.outputComponent?m/e.outputComponent:m]})}}let o=t.map((f,m)=>{if(f.dtype==="complex64")throw new Error("GPGPUProgram does not support complex64 input. For complex64 dtypes, please separate the program into real and imaginary parts.");return this.uploadToGPU(f.dataId),{dtype:this.tensorMap.get(f.dataId).dtype,shape:f.shape,name:e.variableNames[m]}}),l=uoe(e,i,o,r),u;l in this.pipelineCache?u=this.pipelineCache[l]:(u=ioe(this.device,e,o,r,l),this.pipelineCache[l]=u),n&&(s=[...s,...n]);let p=[this.tensorToBinding(r),...t.map(f=>this.tensorToBinding(f)),this.makeUniforms(s)],c=this.device.createBindGroup({layout:u.getBindGroupLayout(0),entries:p.map((f,m)=>({binding:m,resource:f}))});this.ensureCommandEncoderReady();let d=this.getComputePass(),h=this.activeTimers!=null;return h&&this.supportTimeQuery&&d.writeTimestamp(this.querySet,0),d.setPipeline(u),d.setBindGroup(0,c),d.dispatchWorkgroups(e.dispatch[0],e.dispatch[1],e.dispatch[2]),h&&this.supportTimeQuery&&d.writeTimestamp(this.querySet,1),this.dispatchNumberInEncoder++,t.forEach(f=>{this.commandQueueOwnedIds.add(f.dataId)}),this.commandQueueOwnedIds.add(r.dataId),W().get("WEBGPU_DEFERRED_SUBMIT_BATCH_SIZE")<=this.dispatchNumberInEncoder&&this.submitQueue(),h&&this.activeTimers.push({name:e.constructor.name,query:this.getQueryTime(this.querySet)}),r}async getTimeFromQuerySet(e){let t=this.bufferManager.acquireBuffer(16,GPUBufferUsage.COPY_SRC|GPUBufferUsage.QUERY_RESOLVE),a=this.bufferManager.acquireBuffer(16,GPUBufferUsage.MAP_READ|GPUBufferUsage.COPY_DST);this.ensureCommandEncoderReady(),this.ensureComputePassEnded(),this.currentCommandEncoder.resolveQuerySet(e,0,2,t,0),this.currentCommandEncoder.copyBufferToBuffer(t,0,a,0,16),this.submitQueue(),await a.mapAsync(GPUMapMode.READ);let n=new BigUint64Array(a.getMappedRange()),r=Number(n[1]-n[0]);return a.unmap(),this.bufferManager.releaseBuffer(a,16,GPUBufferUsage.MAP_READ|GPUBufferUsage.COPY_DST),this.bufferManager.releaseBuffer(t,16,GPUBufferUsage.COPY_SRC|GPUBufferUsage.QUERY_RESOLVE),r/1e6}shouldExecuteOnCPU(e,t=Aoe){return W().getBool("WEBGPU_CPU_FORWARD")&&e.every(a=>this.tensorMap.get(a.dataId).resourceInfo==null&&v.sizeFromShape(a.shape)<t)}numDataIds(){return this.tensorMap.numDataIds()-this.tensorDataPendingDisposal.length}dispose(){this.disposed||(this.bufferManager.dispose(),this.textureManager.dispose(),this.disposed=!0)}};r0.nextDataId=0;Mg()&&Go("webgpu",async()=>{let e={powerPreference:W().get("WEBGPU_USE_LOW_POWER_GPU")?"low-power":"high-performance"},t=await navigator.gpu.requestAdapter(e),a={};t.features.has("timestamp-query-inside-passes")&&(a.requiredFeatures=["timestamp-query-inside-passes"]);let n=t.limits;a.requiredLimits={maxComputeWorkgroupStorageSize:n.maxComputeWorkgroupStorageSize,maxComputeWorkgroupsPerDimension:n.maxComputeWorkgroupsPerDimension,maxStorageBufferBindingSize:n.maxStorageBufferBindingSize,maxBufferSize:n.maxBufferSize,maxComputeWorkgroupSizeX:n.maxComputeWorkgroupSizeX,maxComputeInvocationsPerWorkgroup:n.maxComputeInvocationsPerWorkgroup};let r=await t.requestDevice(a),s=await t.requestAdapterInfo();return new r0(r,s)},3);var _e;(function(e){e[e.ADD=0]="ADD",e[e.ATAN2=1]="ATAN2",e[e.COMPLEX_MULTIPLY_IMAG=2]="COMPLEX_MULTIPLY_IMAG",e[e.COMPLEX_MULTIPLY_REAL=3]="COMPLEX_MULTIPLY_REAL",e[e.DIV=4]="DIV",e[e.ELU_DER=5]="ELU_DER",e[e.EQUAL=6]="EQUAL",e[e.GREATER=7]="GREATER",e[e.GREATER_EQUAL=8]="GREATER_EQUAL",e[e.INT_DIV=9]="INT_DIV",e[e.LESS=10]="LESS",e[e.LESS_EQUAL=11]="LESS_EQUAL",e[e.LOGICAL_AND=12]="LOGICAL_AND",e[e.LOGICAL_OR=13]="LOGICAL_OR",e[e.MAX=14]="MAX",e[e.MIN=15]="MIN",e[e.MOD=16]="MOD",e[e.MUL=17]="MUL",e[e.NOT_EQUAL=18]="NOT_EQUAL",e[e.POW=19]="POW",e[e.PRELU=20]="PRELU",e[e.SQUARED_DIFFERENCE=21]="SQUARED_DIFFERENCE",e[e.SUB=22]="SUB"})(_e||(_e={}));var voe=`
|
|
resultTemp = select(resultTemp, valueForNaN, isNaN | isnan(a) | isnan(b));`,woe=`
|
|
resultTemp = select(
|
|
resultTemp, vec4<f32>(valueForNaN),
|
|
vec4<bool>(isNaN) | isnanVec4(a) | isnanVec4(b));
|
|
`,koe="return a + b;",Ioe="var resultTemp = atan2(a, b);",Soe="return areal * breal - aimag * bimag;",Coe="return areal * bimag + aimag * breal;",Toe="return a / b;",Noe="return select(a * (b + 1.0), a, b >= 0.);",Roe="return select(a * (b + vec4<f32>(1.0)), a, b >= vec4<f32>(0.));",Eoe="return f32(a == b);",Moe="return vec4<f32>(a == b);",Foe="return f32(a > b);",_oe="return vec4<f32>(a > b);",$oe="return f32(a >= b);",Doe="return vec4<f32>(a >= b);",Poe=`
|
|
let s = sign(a) * sign(b);
|
|
let ia = i32(round(a));
|
|
let ib = i32(round(b));
|
|
return f32(idiv(ia, ib, s));
|
|
`,Ooe=`
|
|
let ia = vec4<i32>(round(a));
|
|
let ib = vec4<i32>(round(b));
|
|
let cond = ib != vec4<i32>(0);
|
|
var resultTemp = vec4<i32>(0);
|
|
let s = sign(a) * sign(b);
|
|
|
|
// Windows (D3D) wants guaranteed non-zero int division at compile-time.
|
|
if (cond[0]) {
|
|
resultTemp[0] = idiv(ia[0], ib[0], s[0]);
|
|
}
|
|
if (cond[1]) {
|
|
resultTemp[1] = idiv(ia[1], ib[1], s[1]);
|
|
}
|
|
if (cond[2]) {
|
|
resultTemp[2] = idiv(ia[2], ib[2], s[2]);
|
|
}
|
|
if (cond[3]) {
|
|
resultTemp[3] = idiv(ia[3], ib[3], s[3]);
|
|
}
|
|
return vec4<f32>(resultTemp);
|
|
`,zoe="return f32(a < b);",Loe="return vec4<f32>(a < b);",Woe="return f32(a <= b);",Boe="return vec4<f32>(a <= b);",Voe="return f32(a >= 1.0 && b >= 1.0);",Uoe=`return (vec4<f32>(a >= vec4<f32>(1.0)) *
|
|
vec4<f32>(b >= vec4<f32>(1.0)));`,Goe="return f32(a >= 1.0 || b >= 1.0);",Hoe=`return min(vec4<f32>(a >= vec4<f32>(1.0)) +
|
|
vec4<f32>(b >= vec4<f32>(1.0)), vec4<f32>(1.0));`,joe="var resultTemp = max(a, b);",qoe="var resultTemp = min(a, b);",Xoe=`
|
|
let isNaN = b == 0.;
|
|
var resultTemp = a % b;
|
|
resultTemp = select((resultTemp + b) % b, resultTemp,
|
|
(a < 0. && b < 0.) || (a >= 0. && b > 0.));
|
|
`,Koe=`
|
|
let isNaN = !vec4<bool>(b);
|
|
var resultTemp = vec4<f32>(a % b);
|
|
if (!((a[0] < 0. && b[0] < 0.) || (a[0] >= 0. && b[0] > 0.))) {
|
|
resultTemp[0] = (resultTemp[0] + b[0]) % b[0];
|
|
}
|
|
if (!((a[1] < 0. && b[1] < 0.) || (a[1] >= 0. && b[1] > 0.))) {
|
|
resultTemp[1] = (resultTemp[1] + b[1]) % b[1];
|
|
}
|
|
if (!((a[2] < 0. && b[2] < 0.) || (a[2] >= 0. && b[2] > 0.))) {
|
|
resultTemp[2] = (resultTemp[2] + b[2]) % b[2];
|
|
}
|
|
if (!((a[3] < 0. && b[3] < 0.) || (a[3] >= 0. && b[3] > 0.))) {
|
|
resultTemp[3] = (resultTemp[3] + b[3]) % b[3];
|
|
}
|
|
`,Yoe="return a * b;",Zoe=`
|
|
var resultTemp = f32(a != b);
|
|
let valueForNaN = 1.0;
|
|
`,Joe=`
|
|
var resultTemp = vec4<f32>(a != b);
|
|
let valueForNaN = 1.0;
|
|
`,Qoe=`
|
|
let isNaN = a < 0.0 && floor(b) < b;
|
|
if (b == 0.0) {
|
|
return 1.0;
|
|
}
|
|
var resultTemp = select(sign(a) * pow(abs(a), b), pow(abs(a), b),
|
|
round(abs(b) % 2.0) != 1.0);
|
|
`,ele=`
|
|
let isModRound1Bool = vec4<i32>(round(abs(b) % vec4<f32>(2.0))) == vec4<i32>(1);
|
|
let isModRound1 = vec4<f32>(isModRound1Bool);
|
|
let multiplier = sign(a) * isModRound1 + (vec4<f32>(1.0) - isModRound1);
|
|
var resultTemp = multiplier * pow(abs(a), b);
|
|
|
|
// Ensure that a^0 = 1, including 0^0 = 1 as this correspond to TF and JS
|
|
let isExpZero = b == vec4<f32>(0.0);
|
|
if (isExpZero.r) {
|
|
resultTemp.r = 1.0;
|
|
}
|
|
if (isExpZero.g) {
|
|
resultTemp.g = 1.0;
|
|
}
|
|
if (isExpZero.b) {
|
|
resultTemp.b = 1.0;
|
|
}
|
|
if (isExpZero.a) {
|
|
resultTemp.a = 1.0;
|
|
}
|
|
let isNaN = (a < vec4<f32>(0.0)) & (floor(b) < b);
|
|
`,tle="if (a < 0.0) { return b * a; } return a;",ale=`
|
|
let aLessThanZero = vec4<f32>(a < vec4<f32>(0.0));
|
|
return (aLessThanZero * (b * a)) + ((vec4<f32>(1.0) - aLessThanZero) * a);
|
|
`,nle="return (a - b) * (a - b);",rle="return a - b;";function _g(e,t){do{let a;switch(e){case _e.ATAN2:a=Ioe;break;case _e.MAX:a=joe;break;case _e.MIN:a=qoe;break;case _e.MOD:a=t?Koe:Xoe;break;case _e.NOT_EQUAL:a=t?Joe:Zoe;break;case _e.POW:a=t?ele:Qoe;break;default:continue}return`
|
|
let isNaN = false;
|
|
let valueForNaN = uniforms.NAN;
|
|
{
|
|
${a}
|
|
${t?woe:voe}
|
|
return resultTemp;
|
|
}
|
|
`}while(!1);switch(e){case _e.ADD:return koe;case _e.COMPLEX_MULTIPLY_IMAG:return Coe;case _e.COMPLEX_MULTIPLY_REAL:return Soe;case _e.DIV:return Toe;case _e.ELU_DER:return t?Roe:Noe;case _e.EQUAL:return t?Moe:Eoe;case _e.GREATER:return t?_oe:Foe;case _e.GREATER_EQUAL:return t?Doe:$oe;case _e.INT_DIV:return t?Ooe:Poe;case _e.LESS:return t?Loe:zoe;case _e.LESS_EQUAL:return t?Boe:Woe;case _e.LOGICAL_AND:return t?Uoe:Voe;case _e.LOGICAL_OR:return t?Hoe:Goe;case _e.MUL:return Yoe;case _e.PRELU:return t?ale:tle;case _e.SQUARED_DIFFERENCE:return nle;case _e.SUB:return rle;default:throw new Error(`BinaryType ${e} is not implemented!`)}}var le;(function(e){e[e.ABS=0]="ABS",e[e.ACOS=1]="ACOS",e[e.ACOSH=2]="ACOSH",e[e.ASIN=3]="ASIN",e[e.ASINH=4]="ASINH",e[e.ATAN=5]="ATAN",e[e.ATANH=6]="ATANH",e[e.CEIL=7]="CEIL",e[e.COS=8]="COS",e[e.COSH=9]="COSH",e[e.ELU=10]="ELU",e[e.ERF=11]="ERF",e[e.EXP=12]="EXP",e[e.EXPM1=13]="EXPM1",e[e.FLOOR=14]="FLOOR",e[e.IS_FINITE=15]="IS_FINITE",e[e.IS_INF=16]="IS_INF",e[e.IS_NAN=17]="IS_NAN",e[e.LINEAR=18]="LINEAR",e[e.LOG=19]="LOG",e[e.LOG1P=20]="LOG1P",e[e.LOGICAL_NOT=21]="LOGICAL_NOT",e[e.NEG=22]="NEG",e[e.RELU=23]="RELU",e[e.RELU6=24]="RELU6",e[e.LEAKYRELU=25]="LEAKYRELU",e[e.RECIPROCAL=26]="RECIPROCAL",e[e.ROUND=27]="ROUND",e[e.RSQRT=28]="RSQRT",e[e.SELU=29]="SELU",e[e.SIGMOID=30]="SIGMOID",e[e.SIGN=31]="SIGN",e[e.SIN=32]="SIN",e[e.SINH=33]="SINH",e[e.SOFTPLUS=34]="SOFTPLUS",e[e.SQRT=35]="SQRT",e[e.SQUARE=36]="SQUARE",e[e.STEP=37]="STEP",e[e.TAN=38]="TAN",e[e.TANH=39]="TANH",e[e.TO_INT=40]="TO_INT"})(le||(le={}));var sle="return abs(a);",ile=`
|
|
if (abs(a) > 1.) {
|
|
return uniforms.NAN;
|
|
}
|
|
return acos(a);
|
|
`,ole=`
|
|
if (a < 1.) {
|
|
return uniforms.NAN;
|
|
}
|
|
return acosh(a);
|
|
`,lle=`
|
|
if (abs(a) > 1.) {
|
|
return uniforms.NAN;
|
|
}
|
|
return asin(a);
|
|
`,ule="return asinh(a);",dle=`
|
|
if (isnan(a)) {
|
|
return uniforms.NAN;
|
|
}
|
|
return atan(a);
|
|
`,ple=`
|
|
if (abs(a) > 1.) {
|
|
return uniforms.NAN;
|
|
}
|
|
if (a == 1.) {
|
|
return uniforms.INFINITY;
|
|
}
|
|
if (a == -1.) {
|
|
return -uniforms.INFINITY;
|
|
}
|
|
return atanh(a);
|
|
`,cle="return ceil(a);",hle="return cos(a);",fle=`
|
|
let e2x = exp(-a);
|
|
return (e2x + 1.0 / e2x) / 2.0;
|
|
`,mle="return exp(a) - 1.0;",gle="if (a >= 0.0) { return a; } return (exp(a) - 1.0);",yle=`
|
|
var resFloat = exp(a) - vec4<f32>(1.0);
|
|
if (a.r >= 0.0) {
|
|
resFloat.r = a.r;
|
|
}
|
|
if (a.g >= 0.0) {
|
|
resFloat.g = a.g;
|
|
}
|
|
if (a.b >= 0.0) {
|
|
resFloat.b = a.b;
|
|
}
|
|
if (a.a >= 0.0) {
|
|
resFloat.a = a.a;
|
|
}
|
|
return resFloat;
|
|
`,xle=`
|
|
// Error function is calculated approximately with elementary function.
|
|
// See "Handbook of Mathematical Functions with Formulas,
|
|
// Graphs, and Mathematical Tables", Abramowitz and Stegun.
|
|
let p = ${S.ERF_P};
|
|
let a1 = ${S.ERF_A1};
|
|
let a2 = ${S.ERF_A2};
|
|
let a3 = ${S.ERF_A3};
|
|
let a4 = ${S.ERF_A4};
|
|
let a5 = ${S.ERF_A5};
|
|
|
|
let sign = sign(a);
|
|
let absA = abs(a);
|
|
let t = 1.0 / (1.0 + p * absA);
|
|
return sign * (1.0 - (((((a5 * t + a4) * t) + a3) * t + a2) * t + a1) * t * exp(-absA * absA));
|
|
`,Ale="return exp(a);",ble="return floor(a);",vle="return f32(!isnan(a) && !isinf(a));",wle="return f32(isinf(a));",kle="return f32(isnan(a));",Ile="return a;",Sle=`if (a < 0.0) { return uniforms.NAN; }
|
|
return log(a);`,Cle=`
|
|
if (isnan(a)) { return a; }
|
|
return log(1.0 + a);
|
|
`,Tle="return f32(!(a >= 1.0));",Nle="return -a;",Rle="if (a < 0.0) { return uniforms.alpha * a; } return a;",Ele=`
|
|
let aLessThanZero = vec4<f32>(a < vec4<f32>(0.0));
|
|
return (aLessThanZero * (uniforms.alpha * a)) + ((vec4<f32>(1.0) - aLessThanZero) * a);
|
|
`,Mle="return 1.0 / a;",Fle="return select(a, 0.0, a < 0.0);",_le="return clamp(a, 0.0, 6.0);",$le="return clamp(a, vec4<f32>(0.0, 0.0, 0.0, 0.0), vec4<f32>(6.0, 6.0, 6.0, 6.0));",Dle=`
|
|
return select(a, vec4<f32>(0.0), a < vec4<f32>(0.0));
|
|
`,Ple="return round(a);",Ole="return inverseSqrt(a);",zle=`
|
|
if (a >= 0.0) {
|
|
return ${S.SELU_SCALE} * a;
|
|
} else {
|
|
return ${S.SELU_SCALEALPHA} * (exp(a) - 1.0);
|
|
}
|
|
`,Lle="return 1.0 / (1.0 + exp(-1.0 * a));",Wle="return sign(a);",Ble="return sin(a);",Vle=`
|
|
let e2x = exp(a);
|
|
return (e2x - 1.0 / e2x) / 2.0;
|
|
`,Ule=`
|
|
let epsilon = 1.1920928955078125e-7;
|
|
let threshold = log(epsilon) + 2.0;
|
|
|
|
let too_large = a > -threshold;
|
|
let too_small = a < threshold;
|
|
let exp_a = exp(a);
|
|
|
|
if (too_large) {
|
|
return a;
|
|
} else if (too_small) {
|
|
return exp_a;
|
|
} else {
|
|
return log(exp_a + 1.0);
|
|
}
|
|
`,Gle="return sqrt(a);",Hle="return a * a;",jle=`
|
|
if (isnan(a)) {
|
|
return a;
|
|
}
|
|
|
|
return select(uniforms.stepAlpha, 1.0, a > 0.0);
|
|
`,qle="return tan(a);",Xle=`
|
|
let e2x = exp(-2.0 * abs(a));
|
|
return sign(a) * (1.0 - e2x) / (1.0 + e2x);
|
|
`,Kle="return f32(i32((a)));";function Ds(e,t){switch(e){case le.ABS:return sle;case le.ACOS:return ile;case le.ACOSH:return ole;case le.ASIN:return lle;case le.ASINH:return ule;case le.ATAN:return dle;case le.ATANH:return ple;case le.COS:return hle;case le.COSH:return fle;case le.CEIL:return cle;case le.ELU:return t?yle:gle;case le.ERF:return xle;case le.EXP:return Ale;case le.EXPM1:return mle;case le.FLOOR:return ble;case le.IS_FINITE:return vle;case le.IS_INF:return wle;case le.IS_NAN:return kle;case le.LINEAR:return Ile;case le.LOG:return Sle;case le.LOG1P:return Cle;case le.LOGICAL_NOT:return Tle;case le.NEG:return Nle;case le.LEAKYRELU:return t?Ele:Rle;case le.RECIPROCAL:return Mle;case le.RELU:return t?Dle:Fle;case le.RELU6:return t?$le:_le;case le.ROUND:return Ple;case le.RSQRT:return Ole;case le.SELU:return zle;case le.SIGMOID:return Lle;case le.SIGN:return Wle;case le.SIN:return Ble;case le.SINH:return Vle;case le.SOFTPLUS:return Ule;case le.SQRT:return Gle;case le.SQUARE:return Hle;case le.STEP:return jle;case le.TAN:return qle;case le.TANH:return Xle;case le.TO_INT:return Kle;default:throw new Error(`BinaryType ${e} is not implemented!`)}}function Mr(e,t=!1,a=!1,n=3){if(e===null)return"";let r="";if(e==="linear")r=Ds(le.LINEAR);else if(e==="relu")r=Ds(le.RELU,a);else if(e==="elu")r=Ds(le.ELU,a);else if(e==="relu6")r=Ds(le.RELU6,a);else if(e==="prelu")r=_g(_e.PRELU,a);else if(e==="sigmoid")r=Ds(le.SIGMOID,a);else if(e==="leakyrelu")r=Ds(le.LEAKYRELU,a);else throw new Error(`Activation ${e} has not been implemented for the WebGPU backend.`);let s=Xe(a?4:1),i="";return t?i=`
|
|
fn activation(a : ${s}, coords : vec${n}<i32>) -> ${s} {
|
|
let b = getPreluActivationWeightsByOutputCoords(coords);
|
|
${r}
|
|
}`:i=`
|
|
fn activation(a : ${s}, coords : vec${n}<i32>) -> ${s} {
|
|
${r}
|
|
}`,i}function Yo(e,t){return`
|
|
${e?"value = value + getBiasByOutputCoords(coords);":""}
|
|
${t?"value = activation(value, coords);":""}
|
|
`}function Zw(e,t,a=!1,n=!1,r=!1,s=1){v.assert(e&&s===1||!e,()=>`transposeA ${e} is not compatible with component size ${s}`);let i=`
|
|
${e?"value = getA(batch, col, row);":"value = getA(batch, row, col);"}
|
|
|
|
`,o=t?"value = getB(batch, col, row);":"value = getB(batch, row, col);";return`
|
|
fn mm_readA(batch: i32, row: i32, colIn: i32) -> ${Xe(s)} {
|
|
var value = ${Xe(s)}(0.0);
|
|
let col = colIn * ${s};
|
|
${a&&r?i:`
|
|
${e?"if(row < uniforms.dimAOuter && col < uniforms.dimInner)":"if(row < uniforms.aShape[1] && col < uniforms.aShape[2])"}
|
|
{
|
|
${i}
|
|
}
|
|
`}
|
|
return value;
|
|
}
|
|
|
|
fn mm_readB(batch: i32, row: i32, colIn: i32) -> ${Xe(s)} {
|
|
let col = colIn * ${s};
|
|
var value = ${Xe(s)}(0.0);
|
|
${o}
|
|
return value;
|
|
}
|
|
`}function $g(e,t,a,n,r=!1,s=!1,i=!1,o=1){return`
|
|
${Zw(a,n,r,s,i,o)}
|
|
fn mm_write(batch: i32, row: i32, colIn: i32, valueIn: ${Xe(o)}) {
|
|
let col = colIn * ${o};
|
|
${r&&s?"":"if (row < uniforms.dimAOuter && col < uniforms.dimBOuter)"}
|
|
{
|
|
var value = valueIn;
|
|
let coords = vec3<i32>(batch, row, col);
|
|
${Yo(e,t)}
|
|
setOutputAtCoords(coords[0], coords[1], coords[2], value);
|
|
}
|
|
}
|
|
`}var Yle=(e,t)=>e?`
|
|
mm_Asub[inputRow][inputCol] = mm_readA(batchA,
|
|
kStart + inputRow,
|
|
globalRowStart / ${t} + inputCol);
|
|
`:`
|
|
mm_Asub[inputRow][inputCol] = mm_readA(batchA,
|
|
globalRow + innerRow,
|
|
kStart / ${t} + inputCol);
|
|
`,Zle=(e,t,a)=>e?`
|
|
let ACached0 = mm_Asub[k * ${t}][localRow];
|
|
let ACached1 = mm_Asub[k * ${t} + 1][localRow];
|
|
let ACached2 = mm_Asub[k * ${t} + 2][localRow];
|
|
${t===3?"":`let ACached3 = mm_Asub[k * ${t} + 3][localRow];`}
|
|
for (var i = 0; i < ${a}; i++) {
|
|
acc[i] = fma(BCached0, vec4<f32>(ACached0[i]), acc[i]);
|
|
acc[i] = fma(BCached1, vec4<f32>(ACached1[i]), acc[i]);
|
|
acc[i] = fma(BCached2, vec4<f32>(ACached2[i]), acc[i]);
|
|
${t===3?"":"acc[i] = fma(BCached3, vec4<f32>(ACached3[i]), acc[i]);"}
|
|
}`:`
|
|
for (var i = 0; i < ${a}; i++) {
|
|
let ACached = mm_Asub[tileRow + i][k];
|
|
acc[i] = fma(BCached0, vec4<f32>(ACached.x), acc[i]);
|
|
acc[i] = fma(BCached1, vec4<f32>(ACached.y), acc[i]);
|
|
acc[i] = fma(BCached2, vec4<f32>(ACached.z), acc[i]);
|
|
${t===3?"":"acc[i] = fma(BCached3, vec4<f32>(ACached.w), acc[i]);"}
|
|
}`;function s0(e,t,a=!1,n=32,r=!1,s=32,i=!1){let o=t[1]*e[1],l=t[0]*e[0],u=a?o:n,p=a?n:o,c=u/t[0],d=n/t[1],h=e[1];return v.assert((a&&c===4&&e[1]===4||!a&&(c===3||c===4))&&u%t[0]===0&&n%t[1]===0&&e[0]===4,()=>`If transposeA ${a} is true, innerElementSize ${c} and workPerThread[1] ${e[1]} must be 4.
|
|
Otherwise, innerElementSize ${c} must be 3 or 4.
|
|
tileAWidth ${u} must be divisible by workgroupSize[0]${t[0]}. tileInner ${n} must be divisible by workgroupSize[1] ${t[1]}. colPerThread ${e[0]} must be 4.`),`
|
|
var<workgroup> mm_Asub : array<array<vec${c}<f32>, ${u/c}>, ${p}>;
|
|
var<workgroup> mm_Bsub : array<array<vec4<f32>, ${l/e[0]}>, ${n}>;
|
|
|
|
${de()} {
|
|
let localRow = i32(localId.y);
|
|
let tileRow = localRow * ${h};
|
|
let tileCol = i32(localId.x);
|
|
|
|
let globalRow = i32(globalId.y) * ${h};
|
|
let globalCol = i32(globalId.x);
|
|
let batch = ${r?"0":"i32(globalId.z)"};
|
|
let batchA = ${r||!i?"batch":"batch % uniforms.aShape[0]"};
|
|
let batchB = ${r||!i?"batch":"batch % uniforms.bShape[0]"};
|
|
let globalRowStart = i32(workgroupId.y) * ${o};
|
|
|
|
let numTiles = ${r?`${Math.ceil(s/n)}`:`(uniforms.dimInner - 1) / ${n} + 1`};
|
|
var kStart = ${r?`i32(globalId.z) * ${s}`:"0"};
|
|
|
|
var acc: array<vec4<f32>, ${h}>;
|
|
|
|
// Loop over shared dimension.
|
|
let tileRowB = localRow * ${d};
|
|
for (var t = 0; t < numTiles; t++) {
|
|
// Load one tile of A into local memory.
|
|
for (var innerRow = 0; innerRow < ${h}; innerRow++) {
|
|
let inputRow = tileRow + innerRow;
|
|
let inputCol = tileCol;
|
|
${Yle(a,c)}
|
|
}
|
|
|
|
// Load one tile of B into local memory.
|
|
for (var innerRow = 0; innerRow < ${d}; innerRow++) {
|
|
let inputRow = tileRowB + innerRow;
|
|
let inputCol = tileCol;
|
|
mm_Bsub[inputRow][inputCol] = mm_readB(batchB, kStart + inputRow, globalCol);
|
|
}
|
|
kStart = kStart + ${n};
|
|
workgroupBarrier();
|
|
|
|
// Compute acc values for a single thread.
|
|
for (var k = 0; k < ${n/c}; k++) {
|
|
let BCached0 = mm_Bsub[k * ${c}][tileCol];
|
|
let BCached1 = mm_Bsub[k * ${c} + 1][tileCol];
|
|
let BCached2 = mm_Bsub[k * ${c} + 2][tileCol];
|
|
${c===3?"":`let BCached3 = mm_Bsub[k * ${c} + 3][tileCol];`}
|
|
|
|
${Zle(a,c,h)}
|
|
}
|
|
|
|
workgroupBarrier();
|
|
}
|
|
|
|
for (var innerRow = 0; innerRow < ${h}; innerRow++) {
|
|
mm_write(batch, globalRow + innerRow, globalCol, acc[innerRow]);
|
|
}
|
|
}`}var Ox=e=>e?`
|
|
mm_Asub[inputRow][inputCol] = mm_readA(batchA,
|
|
kStart + inputRow,
|
|
globalRowStart + inputCol);
|
|
`:`
|
|
mm_Asub[inputRow][inputCol] = mm_readA(batchA,
|
|
globalRowStart + inputRow,
|
|
kStart + inputCol);
|
|
`,Jle=e=>e?"let ACached = mm_Asub[k][tileRow + innerRow];":"let ACached = mm_Asub[tileRow + innerRow][k];";function i0(e,t,a=!1,n=32,r=!1,s=32,i=!1,o=!1){let l=e[1]*t[1],u=e[0]*t[0],p=a?l:n,c=a?n:l;v.assert(c%t[1]===0&&p%t[0]===0&&n%t[1]===0,()=>`tileAHight ${c} must be divisible by workgroupSize[1]${t[1]}, tileAWidth ${p} must be divisible by workgroupSize[0]${t[0]}, tileInner ${n} must be divisible by workgroupSize[1]${t[1]}`);let d=c/t[1],h=p/t[0],f=n/t[1],m=e[1],g=e[0],y=i?`
|
|
let localRow = i32(localId.y);
|
|
let localCol = i32(localId.x);
|
|
let globalRowStart = i32(workgroupId.y) * ${l};
|
|
let globalColStart = i32(workgroupId.x) * ${u};
|
|
|
|
// Loop over shared dimension.
|
|
for (var t = 0; t < numTiles; t++) {
|
|
// Load one tile of A into local memory.
|
|
for (var inputRow = localRow; inputRow < ${c}; inputRow = inputRow + ${t[1]}) {
|
|
for (var inputCol = localCol; inputCol < ${p}; inputCol = inputCol + ${t[0]}) {
|
|
${Ox(a)}
|
|
}
|
|
}
|
|
// Load one tile of B into local memory.
|
|
for (var inputRow = localRow; inputRow < ${n}; inputRow = inputRow + ${t[1]}) {
|
|
for (var inputCol = localCol; inputCol < ${u}; inputCol = inputCol + ${t[0]}) {
|
|
mm_Bsub[inputRow][inputCol] = mm_readB(batchB,
|
|
kStart + inputRow,
|
|
globalColStart + inputCol);
|
|
}
|
|
}
|
|
kStart = kStart + ${n};
|
|
workgroupBarrier();
|
|
|
|
// Compute acc values for a single thread.
|
|
var BCached : array<f32, ${g}>;
|
|
for (var k = 0; k < ${n}; k++) {
|
|
for (var inner = 0; inner < ${g}; inner++) {
|
|
BCached[inner] = mm_Bsub[k][localCol + inner * ${t[0]}];
|
|
}
|
|
for (var innerRow = 0; innerRow < ${m}; innerRow++) {
|
|
let ACached = ${a?`mm_Asub[k][localRow + innerRow * ${t[1]}];`:`mm_Asub[localRow + innerRow * ${t[1]}][k];`}
|
|
for (var innerCol = 0; innerCol < ${g}; innerCol++) {
|
|
acc[innerRow][innerCol] =
|
|
fma(ACached, BCached[innerCol], acc[innerRow][innerCol]);
|
|
}
|
|
}
|
|
}
|
|
workgroupBarrier();
|
|
}
|
|
for (var innerRow = 0; innerRow < ${m}; innerRow++) {
|
|
let gRow = globalRowStart + localRow + innerRow * ${t[1]};
|
|
for (var innerCol = 0; innerCol < ${g}; innerCol++) {
|
|
let gCol = globalColStart + localCol + innerCol * ${t[0]};
|
|
mm_write(batch, gRow, gCol, acc[innerRow][innerCol]);
|
|
}
|
|
}
|
|
`:`
|
|
let tileRow = i32(localId.y) * ${m};
|
|
let tileCol = i32(localId.x) * ${g};
|
|
|
|
let globalRow = i32(globalId.y) * ${m};
|
|
let globalCol = i32(globalId.x) * ${g};
|
|
let globalRowStart = i32(workgroupId.y) * ${l};
|
|
|
|
let tileRowA = i32(localId.y) * ${d};
|
|
let tileColA = i32(localId.x) * ${h};
|
|
let tileRowB = i32(localId.y) * ${f};
|
|
// Loop over shared dimension.
|
|
for (var t = 0; t < numTiles; t++) {
|
|
// Load one tile of A into local memory.
|
|
for (var innerRow = 0; innerRow < ${d}; innerRow++) {
|
|
for (var innerCol = 0; innerCol < ${h}; innerCol++) {
|
|
let inputRow = tileRowA + innerRow;
|
|
let inputCol = tileColA + innerCol;
|
|
${Ox(a)}
|
|
}
|
|
}
|
|
|
|
// Load one tile of B into local memory.
|
|
for (var innerRow = 0; innerRow < ${f}; innerRow++) {
|
|
for (var innerCol = 0; innerCol < ${g}; innerCol++) {
|
|
let inputRow = tileRowB + innerRow;
|
|
let inputCol = tileCol + innerCol;
|
|
mm_Bsub[inputRow][inputCol] = mm_readB(batchB,
|
|
kStart + inputRow,
|
|
globalCol + innerCol);
|
|
}
|
|
}
|
|
kStart = kStart + ${n};
|
|
workgroupBarrier();
|
|
|
|
// Compute acc values for a single thread.
|
|
var BCached : array<f32, ${g}>;
|
|
for (var k = 0; k < ${n}; k++) {
|
|
for (var inner = 0; inner < ${g}; inner++) {
|
|
BCached[inner] = mm_Bsub[k][tileCol + inner];
|
|
}
|
|
|
|
for (var innerRow = 0; innerRow < ${m}; innerRow++) {
|
|
${Jle(a)}
|
|
for (var innerCol = 0; innerCol < ${g}; innerCol++) {
|
|
acc[innerRow][innerCol] =
|
|
fma(ACached, BCached[innerCol], acc[innerRow][innerCol]);
|
|
}
|
|
}
|
|
}
|
|
|
|
workgroupBarrier();
|
|
}
|
|
|
|
for (var innerRow = 0; innerRow < ${m}; innerRow++) {
|
|
for (var innerCol = 0; innerCol < ${g}; innerCol++) {
|
|
mm_write(batch, globalRow + innerRow, globalCol + innerCol,
|
|
acc[innerRow][innerCol]);
|
|
}
|
|
}
|
|
`;return`
|
|
var<workgroup> mm_Asub : array<array<f32, ${p}>, ${c}>;
|
|
var<workgroup> mm_Bsub : array<array<f32, ${u}>, ${n}>;
|
|
|
|
${de()} {
|
|
let batch = ${r?"0":"i32(globalId.z)"};
|
|
let batchA = ${r||!o?"batch":"batch % uniforms.aShape[0]"};
|
|
let batchB = ${r||!o?"batch":"batch % uniforms.bShape[0]"};
|
|
let numTiles = ${r?`${Math.ceil(s/n)}`:`(uniforms.dimInner - 1) / ${n} + 1`};
|
|
var kStart = ${r?`i32(globalId.z) * ${s}`:"0"};
|
|
|
|
var acc : array<array<f32, ${g}>, ${m}>;
|
|
|
|
// Without this initialization strange values show up in acc.
|
|
for (var innerRow = 0; innerRow < ${m}; innerRow++) {
|
|
for (var innerCol = 0; innerCol < ${g}; innerCol++) {
|
|
acc[innerRow][innerCol] = 0.0;
|
|
}
|
|
}
|
|
${y}
|
|
}
|
|
`}var Qle=e=>e?`
|
|
mm_readA(batchA, colA, globalRow),
|
|
mm_readA(batchA, colA + 1, globalRow),
|
|
mm_readA(batchA, colA + 2, globalRow),
|
|
mm_readA(batchA, colA + 3, globalRow)
|
|
`:`
|
|
mm_readA(batchA, globalRow, colA),
|
|
mm_readA(batchA, globalRow, colA + 1),
|
|
mm_readA(batchA, globalRow, colA + 2),
|
|
mm_readA(batchA, globalRow, colA + 3)
|
|
`;function eue(e,t=!1){v.assert(e[1]===1&&e[2]===1,()=>`A linear work group size is required. But got ${e}.`);let a=e[0]*4;return`
|
|
var<workgroup> mm_Asub : array<vec4<f32>, ${e[0]}>;
|
|
|
|
${de()} {
|
|
let tileCol = i32(localId.x);
|
|
let globalCol = i32(globalId.x);
|
|
let globalRow = i32(globalId.y);
|
|
|
|
let numTiles = (uniforms.dimInner - 1) / ${a} + 1;
|
|
let batch = i32(globalId.z);
|
|
let batchA = batch % uniforms.aShape[0];
|
|
let batchB = batch % uniforms.bShape[0];
|
|
// Without this initialization strange values show up in acc.
|
|
var acc = 0.0;
|
|
|
|
// Loop over shared dimension.
|
|
for (var t = 0; t < numTiles; t++) {
|
|
// Load one tile of A into local memory.
|
|
let colA = t * ${a} + tileCol * 4;
|
|
mm_Asub[tileCol] = vec4<f32>(${Qle(t)});
|
|
workgroupBarrier();
|
|
|
|
// Compute acc values for a single thread.
|
|
for (var k = 0; k < ${a/4}; k++) {
|
|
let rowB = t * ${a} + k * 4;
|
|
let BCached = vec4<f32>(mm_readB(batchB, rowB, globalCol),
|
|
mm_readB(batchB, rowB + 1, globalCol),
|
|
mm_readB(batchB, rowB + 2, globalCol),
|
|
mm_readB(batchB, rowB + 3, globalCol));
|
|
|
|
let ACached = mm_Asub[k];
|
|
acc = acc + dot(ACached, BCached);
|
|
}
|
|
|
|
workgroupBarrier();
|
|
}
|
|
|
|
mm_write(batch, globalRow, globalCol, acc);
|
|
}
|
|
`}var tue=class{constructor(e,t,a=!1,n=!1,r=null,s=null,i=null,o=!1){this.variableNames=["A","B"],this.uniforms="dimAOuter : i32, dimBOuter : i32, dimInner : i32,",this.outputShape=t,this.dispatchLayout={x:[2],y:[1],z:[0]};let l=a?e[1]:e[2];if(this.isVec4=(l%4===0&&!a||t[1]%4===0&&a)&&t[2]%4===0&&!n,this.outputComponent=this.isVec4?4:1,this.isVectorA=t[1]===1&&!a,!this.isVec4&&this.isVectorA)this.elementsPerThread=[1,1,1],this.workgroupSize=[32,1,1];else{let c=Yw(t[1],l,t[2],a);this.workgroupSize=c.workgroupSize,this.elementsPerThread=c.elementsPerThread}this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize,this.elementsPerThread);let u=r!=null,p=i!=null;u&&this.variableNames.push("bias"),p&&this.variableNames.push("preluActivationWeights"),this.sequentialAccessByThreads=o,this.transposeA=a,this.transposeB=n,this.addBias=u,this.activation=s,this.hasPreluActivationWeights=p,[this.fitAOuter,this.fitBOuter,this.fitInner]=this.getShapeFit(t[1],t[2],l),this.shaderKey=`matMulPacked_${this.elementsPerThread}_${a}_${n}_${this.activation}_${this.fitAOuter}_${this.fitBOuter}_${this.fitInner}_${this.isVec4}_${this.isVectorA}_${this.sequentialAccessByThreads}`}getShapeFit(e,t,a){let n=this.workgroupSize[1]*this.elementsPerThread[1],r=this.workgroupSize[0]*this.elementsPerThread[0];!this.isVec4&&this.isVectorA?this.tileInner=this.workgroupSize[0]*4:this.tileInner=r;let s=e%n===0,i=t%r===0,o=a%this.tileInner===0;return[s,i,o]}getUserCode(){return`
|
|
${Mr(this.activation,this.hasPreluActivationWeights,this.isVec4)}
|
|
${$g(this.addBias,this.activation,!1,this.transposeB,this.fitAOuter,this.fitBOuter,this.fitInner,this.isVec4?4:1)}
|
|
${this.isVec4?s0(this.elementsPerThread,this.workgroupSize,this.transposeA,this.tileInner,!1,null,!0):this.isVectorA?eue(this.workgroupSize,this.transposeA):i0(this.elementsPerThread,this.workgroupSize,this.transposeA,this.tileInner,!1,null,this.sequentialAccessByThreads,!0)}
|
|
`}};function aue(e){return`
|
|
var<workgroup> sumValues : array<f32, ${e}>;
|
|
${de()} {
|
|
let coords = getOutputCoords();
|
|
let batch = coords[0];
|
|
let batchA = batch % uniforms.aShape[0];
|
|
let batchB = batch % uniforms.bShape[0];
|
|
let row = coords[1];
|
|
let col = coords[2];
|
|
var sum = 0.0;
|
|
let Length = uniforms.dimInner;
|
|
for (var k = i32(localId.x); k < Length; k = k + ${e}) {
|
|
let dataA = mm_readA(batchA, row, k);
|
|
let dataB = mm_readB(batchB, k, col);
|
|
sum = sum + dataA * dataB;
|
|
}
|
|
sumValues[localId.x] = sum;
|
|
workgroupBarrier();
|
|
|
|
for(var currentSize = ${e/2}u; currentSize > 1u;
|
|
currentSize = currentSize / 2u) {
|
|
if (localId.x < currentSize)
|
|
{
|
|
sumValues[localId.x] = sumValues[localId.x] + sumValues[localId.x + currentSize];
|
|
}
|
|
workgroupBarrier();
|
|
}
|
|
|
|
if (localId.x == 0u) {
|
|
sum = sumValues[0] + sumValues[1];
|
|
mm_write(batch, row, col, sum);
|
|
}
|
|
}
|
|
`}var nue=class{constructor(e,t=!1,a=!1,n=null,r=null,s=null){this.variableNames=["A","B"],this.uniforms="dimAOuter : i32, dimBOuter : i32, dimInner : i32,",this.workgroupSize=[256,1,1],this.outputShape=e,this.dispatchLayout={x:[],y:[1,2],z:[0]},this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize);let i=n!=null,o=s!=null;i&&this.variableNames.push("bias"),o&&this.variableNames.push("preluActivationWeights"),this.transposeA=t,this.transposeB=a,this.addBias=i,this.activation=r,this.hasPreluActivationWeights=o,this.shaderKey=`matMulReduce_${this.activation}_${t}_${a}`}getUserCode(){return`
|
|
${Mr(this.activation,this.hasPreluActivationWeights)}
|
|
${$g(this.addBias,this.activation,this.transposeA,this.transposeB)}
|
|
${aue(this.workgroupSize[0])}
|
|
`}};function rue(e){let t=e[1],a=e[0],n=t>a?t:a;return`
|
|
var<workgroup> mm_Asub : array<array<f32, ${n}>, ${t}>;
|
|
var<workgroup> mm_Bsub : array<array<f32, ${a}>, ${n}>;
|
|
|
|
// If the output size is small for matrix multiplication, avoid to use vec4
|
|
// and handle some elements per thread to optimally utilize the ALU.
|
|
// Read data from global memory to registers firstly, then store them into
|
|
// shared memory, so it is instruction-Level parallelism for arithmetic
|
|
// operations and others handle IO operations between barrier api, makes ALU
|
|
// and load/store units work simultaneously, could improves the performance.
|
|
${de()} {
|
|
let tileRow = i32(localId.y);
|
|
let tileCol = i32(localId.x);
|
|
let globalRow = i32(globalId.y);
|
|
let globalCol = i32(globalId.x);
|
|
let batch = i32(globalId.z);
|
|
let batchA = batch % uniforms.aShape[0];
|
|
let batchB = batch % uniforms.bShape[0];
|
|
|
|
// uniforms.dimInner should be greater than 0.
|
|
let numTiles = (uniforms.dimInner - 1) / ${n} + 1;
|
|
var acc = 0.0;
|
|
|
|
var globalColA = tileCol;
|
|
var globalRowB = 0;
|
|
var regA = mm_readA(batchA, globalRow, globalColA);
|
|
var regB0 = mm_readB(batchB, globalRowB + 2 * tileRow, globalCol);
|
|
var regB1 = mm_readB(batchB, globalRowB + 2 * tileRow + 1, globalCol);
|
|
globalColA = globalColA + ${n};
|
|
globalRowB = globalRowB + ${n};
|
|
|
|
for (var t = 0; t < numTiles; t = t + 1) {
|
|
mm_Asub[tileRow][tileCol] = regA;
|
|
mm_Bsub[2 * tileRow][tileCol] = regB0;
|
|
mm_Bsub[2 * tileRow + 1][tileCol] = regB1;
|
|
|
|
workgroupBarrier();
|
|
|
|
regA = mm_readA(batchA, globalRow, globalColA);
|
|
regB0 = mm_readB(batchB, globalRowB + 2 * tileRow, globalCol);
|
|
regB1 = mm_readB(batchB, globalRowB + 2 * tileRow + 1, globalCol);
|
|
globalColA = globalColA + ${n};
|
|
globalRowB = globalRowB + ${n};
|
|
|
|
for (var k = 0; k < ${n}; k = k + 1) {
|
|
acc = acc + mm_Asub[tileRow][k] * mm_Bsub[k][tileCol];
|
|
}
|
|
workgroupBarrier();
|
|
}
|
|
|
|
mm_write(batch, globalRow, globalCol, acc);
|
|
}
|
|
`}var sue=class{constructor(e,t,a,n=!1,r=!1,s=null,i=null,o=null){this.variableNames=["A","B"],this.uniforms="dimAOuter : i32, dimBOuter : i32, dimInner : i32,",this.workgroupSize=[16,8,1],this.outputShape=a,this.dispatchLayout={x:[2],y:[1],z:[0]},this.dispatch=[Math.ceil(a[2]/this.workgroupSize[0]),Math.ceil(a[1]/this.workgroupSize[1]),a[0]];let l=s!=null;l&&this.variableNames.push("bias");let u=o!=null;u&&this.variableNames.push("preluActivationWeights"),this.transposeA=n,this.transposeB=r,this.addBias=l,this.activation=i,this.hasPreluActivationWeights=u,this.shaderKey=`matMulSmallOutputSize_${this.activation}_${n}_${r}`}getUserCode(){return`
|
|
${Mr(this.activation,this.hasPreluActivationWeights)}
|
|
${$g(this.addBias,this.activation,this.transposeA,this.transposeB)}
|
|
${rue(this.workgroupSize)}
|
|
`}},iue=class{constructor(e,t,a=!1,n=!1){this.variableNames=["A","B"],this.uniforms="dimAOuter : i32, dimBOuter : i32, dimInner : i32,",this.workgroupSize=[8,8,1],this.atomic=!0,this.splitedDimInner=128,v.assert(e[0]===1,()=>"MatMulSplitKProgram only supports batch = 1."),this.outputShape=e,this.dispatchLayout={x:[2],y:[1],z:[0,3]};let r=(a&&this.outputShape[1]%4===0||!a&&t%4===0)&&this.outputShape[2]%4===0;this.elementsPerThread=[4,4,this.splitedDimInner],this.outputComponent=r?4:1,r||(this.outputShape[1]<16&&(this.elementsPerThread[1]=1),this.outputShape[2]<16&&(this.elementsPerThread[0]=1)),this.dispatch=ce(this.dispatchLayout,[this.outputShape[0],this.outputShape[1],this.outputShape[2],t],this.workgroupSize,this.elementsPerThread),this.transposeA=a,this.transposeB=n,this.shaderKey=`matMulSplitK_${a}_${n}_${this.elementsPerThread}_${this.outputComponent}`}getUserCode(){let e=this.outputComponent;return`
|
|
${Zw(!1,this.transposeB,!1,!1,!1,e)}
|
|
fn mm_write(batch: i32, row : i32, colIn : i32, value : ${Xe(e)}) {
|
|
let col = colIn * ${e};
|
|
if (row < uniforms.dimAOuter && col < uniforms.dimBOuter) {
|
|
let coords = vec3<i32>(batch, row, col);
|
|
let flatIndex = getOutputIndexFromCoords(coords);
|
|
// The problem is that we should initialize output to zero before using.
|
|
// Otherwise, the original value will be added to the result.
|
|
for (var i = 0; i < ${e}; i = i + 1) {
|
|
${Gu("&result[flatIndex + i]",`${e>1?"value[i]":"value"}`,"float32")}
|
|
}
|
|
}
|
|
}
|
|
${e===4?s0(this.elementsPerThread,this.workgroupSize,this.transposeA,32,!0,this.splitedDimInner):i0(this.elementsPerThread,this.workgroupSize,this.transposeA,32,!0,this.splitedDimInner)}
|
|
`}},oue=class{constructor(e,t=null,a=null,n=null){this.uniforms="",this.variableNames=["x"],this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.addBias=t!=null,this.hasPreluActivationWeights=n!=null,this.activation=a,this.addBias&&this.variableNames.push("bias"),this.hasPreluActivationWeights&&this.variableNames.push("preluActivationWeights"),this.shaderKey=`biasActivation_${a}`}getUserCode(){return`
|
|
${Mr(this.activation,this.hasPreluActivationWeights)}
|
|
${de("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
var value = getXByOutputIndex(index);
|
|
${Yo(this.addBias,this.activation)}
|
|
setOutputAtIndex(index, value);
|
|
}
|
|
}
|
|
`}},lue=class{constructor(e){this.variableNames=[],this.outputShape=[],this.uniforms="value : f32,",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="fill"}getUserCode(){return`
|
|
${de("index")} {
|
|
if (index < uniforms.size) {
|
|
setOutputAtIndex(index, uniforms.value);
|
|
}
|
|
}
|
|
`}};function Mn(e){let{backend:t,attrs:a}=e,{shape:n,value:r}=a,{dtype:s}=a;if(s=s||v.inferDtype(r),s==="string"){let i=v.getArrayFromDType(s,v.sizeFromShape(n));return i.fill(r),t.makeTensorInfo(n,s,i)}else{let i=new lue(n),o=[{type:"float32",data:[r]}];return t.runWebGPUProgram(i,[],s,o)}}var uue={kernelName:du,backendName:"webgpu",kernelFunc:Mn};function we(e){let{inputs:t,attrs:a}=e,{x:n}=t,{shape:r}=a,s=v.sizeFromShape(n.shape),i=v.inferFromImplicitShape(r,s),o=v.sizeFromShape(i);return v.assert(s===o,()=>`The new shape (${i}) has ${o} elements and the old shape (${n.shape}) has ${s} elements. The new shape and old shape must have the same number of elements.`),e.backend.incRef(n.dataId),{dataId:n.dataId,shape:i,dtype:n.dtype}}var due={kernelName:bu,backendName:"webgpu",kernelFunc:we};function o0({a:e,b:t,transposeA:a,transposeB:n,backend:r,bias:s=null,preluActivationWeights:i=null,leakyreluAlpha:o=0,activation:l=null}){let u=e.shape.length,p=t.shape.length,c=a?e.shape[u-2]:e.shape[u-1],d=n?t.shape[p-1]:t.shape[p-2],h=a?e.shape[u-1]:e.shape[u-2],f=n?t.shape[p-2]:t.shape[p-1],m=e.shape.slice(0,-2),g=t.shape.slice(0,-2),y=v.sizeFromShape(m),x=v.sizeFromShape(g),A=Ho.assertAndGetBroadcastShape(e.shape.slice(0,-2),t.shape.slice(0,-2)).concat([h,f]);v.assert(c===d,()=>`Error in matMul: inner shapes (${c}) and (${d}) of Tensors with shapes ${e.shape} and ${t.shape} and transposeA=${a} and transposeB=${n} must match.`);let b=a?[y,c,h]:[y,h,c],w=n?[x,f,d]:[x,d,f],C=we({inputs:{x:e},backend:r,attrs:{shape:b}}),T=we({inputs:{x:t},backend:r,attrs:{shape:w}}),N=[C,T],M=Math.max(y,x),F=[C,T],E=[{type:"int32",data:[h]},{type:"int32",data:[f]},{type:"int32",data:[c]}],I,_,O=[M,h,f],B=W().get("WEBGPU_MATMUL_PROGRAM_TYPE");if(B<0){let U=W().getNumber("WEBGPU_THRESHOLD_TO_INCREASE_WORKGROUPS_FOR_MATMUL"),G=U>0?U:r.thresholdToIncreaseWorkgroups,q=M*Math.ceil(h/32)*Math.ceil(f/32);q<=G||h<=8&&q<=G*2?M*h*f<=128?B=Pn.MatMulReduceProgram:M===1&&d>=2e3?B=Pn.MatMulSplitKProgram:B=Pn.MatMulSmallOutputSizeProgram:B=Pn.MatMulPackedProgram}switch(B){case Pn.MatMulReduceProgram:I=new nue(O,a,n,s,l,i);break;case Pn.MatMulSplitKProgram:{if(_=Mn({backend:r,attrs:{shape:O,value:0,dtype:e.dtype}}),I=new iue(O,d,a,n),s||l){_=r.runWebGPUProgram(I,F,e.dtype,E,_);let G=new oue(_.shape,s,l,i),q=null,H=[_];s&&H.push(s),i&&H.push(i),l==="leakyrelu"&&(q=[{type:"float32",data:[o]}],G.uniforms+=" alpha : f32,");let V=r.runWebGPUProgram(G,H,_.dtype,q);N.push(_);let Z=we({inputs:{x:V},backend:r,attrs:{shape:A}});N.push(V);for(let X of N)r.disposeData(X.dataId);return Z}break}case Pn.MatMulSmallOutputSizeProgram:I=new sue(b,w,O,a,n,s,l,i);break;case Pn.MatMulPackedProgram:let U=r.adapterInfo.isIntel();I=new tue(b,O,a,n,s,l,i,U);break;default:throw new Error(`Unsupported MatMulProgramType ${B}.`)}s&&F.push(s),i&&F.push(i),l==="leakyrelu"&&(E.push({type:"float32",data:[o]}),I.uniforms+=" alpha : f32,"),_=r.runWebGPUProgram(I,F,e.dtype,E,_);let $=we({inputs:{x:_},backend:r,attrs:{shape:A}});N.push(_);for(let U of N)r.disposeData(U.dataId);return $}function pue(e){let{inputs:t,backend:a,attrs:n}=e,{a:r,b:s,bias:i,preluActivationWeights:o}=t,{transposeA:l,transposeB:u,activation:p,leakyreluAlpha:c}=n;return o0({a:r,b:s,transposeA:l,transposeB:u,backend:a,bias:i,preluActivationWeights:o,leakyreluAlpha:c,activation:p})}var cue={kernelName:Xr,backendName:"webgpu",kernelFunc:pue},zx=class{constructor(e,t,a){this.variableNames=["AReal","AImag","BReal","BImag"],this.workgroupSize=[128,1,1],this.size=!0,this.outputShape=S.assertAndGetBroadcastShape(t,a),this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey=`binaryOpComplex_${e}`,this.op=e}getUserCode(){return`
|
|
fn binaryOpComplex(
|
|
areal : f32, aimag : f32, breal : f32, bimag : f32) -> f32 {
|
|
${_g(this.op,!1)}
|
|
}
|
|
|
|
${de("index")} {
|
|
if(index < uniforms.size) {
|
|
let areal = getARealByOutputIndex(index);
|
|
let aimag = getAImagByOutputIndex(index);
|
|
let breal = getBRealByOutputIndex(index);
|
|
let bimag = getBImagByOutputIndex(index);
|
|
setOutputAtIndex(index, binaryOpComplex(areal, aimag, breal, bimag));
|
|
}
|
|
}
|
|
`}},xh=class{constructor(e,t,a){if(this.size=!0,this.variableNames=["A","B"],this.outputShape=S.assertAndGetBroadcastShape(t,a),this.dispatchLayout=be(this.outputShape),this.op=e,this.useSharedMemoryWithA=t.length<=1&&a.length>1&&t[0]<128,this.useSharedMemoryWithB=a.length<=1&&t.length>1&&a[0]<128,this.useSharedMemoryWithA||this.useSharedMemoryWithB)this.outputComponent=1,this.variableComponents=[1,1],this.lastDimensionSize=this.useSharedMemoryWithB?a[0]:t[0],this.shaderKey=`binary_${e}_${this.lastDimensionSize}`,this.type="shared",this.workgroupSize=[256,1,1];else{let n=t.length>0&&t[t.length-1]%4===0,r=a.length>0&&a[a.length-1]%4===0;n&&r?(this.outputComponent=4,this.variableComponents=[4,4]):n&&(v.isScalarShape(a)||a[a.length-1]===1)||r&&(v.isScalarShape(t)||t[t.length-1]===1)?(this.outputComponent=4,this.variableComponents=n?[4,1]:[1,4]):(this.outputComponent=1,this.variableComponents=[1,1]),this.type="nonshared",this.shaderKey=`binary_${e}_${this.variableComponents}`,this.workgroupSize=[128,1,1]}this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize,[this.outputComponent,1,1])}getUserCode(){let e,t=this.outputComponent===4?"vec4<f32>":"f32",a=`
|
|
fn binaryOperation(a : ${t}, b : ${t}) -> ${t} {
|
|
${_g(this.op,this.outputComponent===4)}
|
|
};
|
|
`;if(this.type==="shared"){let n=this.lastDimensionSize>1?`coords[${this.outputShape.length-1}]`:"0",r=this.useSharedMemoryWithB?`let a = getAByOutputIndex(index);
|
|
let b = sharedBuf[${n}];`:`let a = sharedBuf[${n}];
|
|
let b = getBByOutputIndex(index);`;e=`
|
|
${a}
|
|
var<workgroup> sharedBuf : array<f32, ${this.lastDimensionSize}>;
|
|
${de("index")} {
|
|
// Fill in the shared memory buffer.
|
|
let localIndex = i32(localId.x);
|
|
if(localIndex < ${this.lastDimensionSize}) {
|
|
sharedBuf[localIndex] = f32(${this.useSharedMemoryWithB?"B":"A"}[localIndex]);
|
|
}
|
|
workgroupBarrier();
|
|
|
|
if(index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
${r}
|
|
setOutputAtIndex(index, binaryOperation(a, b));
|
|
}
|
|
}
|
|
`}else e=`
|
|
${a}
|
|
${de("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index * ${this.outputComponent});
|
|
let a = ${t}(getAByOutputCoords(coords));
|
|
let b = ${t}(getBByOutputCoords(coords));
|
|
setOutputAtIndex(index, binaryOperation(a, b));
|
|
}
|
|
}
|
|
`;return e}};function tn(e){let{inputs:t}=e,{x:a}=t;return e.backend.incRef(a.dataId),{dataId:a.dataId,shape:a.shape,dtype:a.dtype}}var hue={kernelName:zi,backendName:"webgpu",kernelFunc:tn};function Zo(e){let{inputs:t,backend:a}=e,{real:n,imag:r}=t,s=a.makeTensorInfo(n.shape,"complex64"),i=a.tensorMap.get(s.dataId),o=tn({inputs:{x:n},backend:a}),l=tn({inputs:{x:r},backend:a});return i.complexTensorInfos={real:o,imag:l},s}var fue={kernelName:tp,backendName:"webgpu",kernelFunc:Zo},Hu=class{constructor(e,t,a=""){this.variableNames=["A"],this.size=!0;let n=128;this.workgroupSize=[n,1,1],this.outputShape=e,this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.op=t,a!==""&&(this.uniforms=a),this.shaderKey=`unary_${t}`}getUserCode(){return`
|
|
fn unaryOperation(a : f32) -> f32 {
|
|
${Ds(this.op,!1)}
|
|
}
|
|
${de("index")} {
|
|
if (index < uniforms.size) {
|
|
let a = getAByOutputIndex(index);
|
|
setOutputAtIndex(index, unaryOperation(a));
|
|
}
|
|
}
|
|
`}};function tt({opType:e,cpuKernelImpl:t,dtype:a}){return({inputs:n,backend:r})=>{let{x:s}=n,i=r,o=a||s.dtype;if(i.shouldExecuteOnCPU([s])&&t!=null){let u=i.tensorMap.get(s.dataId),p=t(u.values,o);return i.makeTensorInfo(s.shape,o,p)}let l=new Hu(s.shape,e);return i.runWebGPUProgram(l,[s],o)}}function ta({opType:e,cpuKernelImpl:t,supportsComplex:a=!1,dtype:n}){return({inputs:r,backend:s})=>{let{a:i,b:o}=r,l=s;if(a&&i.dtype==="complex64"){let c=l.tensorMap.get(i.dataId),d=l.tensorMap.get(o.dataId),h,f;if(e!==_e.MUL)[h,f]=[[c.complexTensorInfos.real,d.complexTensorInfos.real],[c.complexTensorInfos.imag,d.complexTensorInfos.imag]].map(g=>{let[y,x]=g,A={dataId:y.dataId,dtype:y.dtype,shape:i.shape},b={dataId:x.dataId,dtype:x.dtype,shape:o.shape},w=new xh(e,i.shape,o.shape);return l.runWebGPUProgram(w,[A,b],pa(y.dtype,x.dtype))});else{let g=new zx(_e.COMPLEX_MULTIPLY_REAL,i.shape,o.shape),y=new zx(_e.COMPLEX_MULTIPLY_IMAG,i.shape,o.shape),x=[{dataId:c.complexTensorInfos.real.dataId,dtype:c.complexTensorInfos.real.dtype,shape:i.shape},{dataId:c.complexTensorInfos.imag.dataId,dtype:c.complexTensorInfos.imag.dtype,shape:i.shape},{dataId:d.complexTensorInfos.real.dataId,dtype:d.complexTensorInfos.real.dtype,shape:o.shape},{dataId:d.complexTensorInfos.imag.dataId,dtype:d.complexTensorInfos.imag.dtype,shape:o.shape}];h=l.runWebGPUProgram(g,x,"float32"),f=l.runWebGPUProgram(y,x,"float32")}let m=Zo({inputs:{real:h,imag:f},backend:l});return l.disposeData(h.dataId),l.disposeData(f.dataId),m}let u=n||pa(i.dtype,o.dtype);if((i.dtype==="string"||o.dtype==="string"||l.shouldExecuteOnCPU([i,o]))&&t!=null){let c=l.tensorMap.get(i.dataId).values,d=l.tensorMap.get(o.dataId).values,h=i.dtype==="string"?S.fromUint8ToStringArray(c):c,f=i.dtype==="string"?S.fromUint8ToStringArray(d):d,[m,g]=t(i.shape,o.shape,h,f,u);return l.makeTensorInfo(g,u,m)}let p=new xh(e,i.shape,o.shape);return l.runWebGPUProgram(p,[i,o],u)}}var{addImpl:mue,castImpl:gue,ceilImpl:yue,concatImpl:xue,equalImpl:Aue,expImpl:bue,expm1Impl:vue,floorImpl:wue,floorDivImpl:kue,gatherNdImpl:Iue,gatherV2Impl:Sue,greaterEqualImpl:Cue,greaterImpl:Tue,lessEqualImpl:Nue,lessImpl:Rue,logImpl:Eue,maxImpl:Mue,maximumImpl:Fue,minimumImpl:_ue,multiplyImpl:$ue,negImpl:Due,notEqualImpl:Pue,prodImpl:Oue,rangeImpl:zue,rsqrtImpl:Lue,scatterImpl:Wue,simpleAbsImpl:Bue,sliceImpl:Vue,stridedSliceImpl:Uue,stringNGramsImpl:Gue,subImpl:Hue,tileImpl:jue,topKImpl:que,transposeImpl:Xue,uniqueImpl:J3e}=Kh,Kue=tt({opType:le.ABS,cpuKernelImpl:Bue}),Yue={kernelName:Zl,backendName:"webgpu",kernelFunc:Kue},Zue=tt({opType:le.ACOS}),Jue={kernelName:Qs,backendName:"webgpu",kernelFunc:Zue},Que=tt({opType:le.ACOSH}),ede={kernelName:ei,backendName:"webgpu",kernelFunc:Que},tde=ta({opType:_e.ADD,cpuKernelImpl:mue,supportsComplex:!0}),ade={kernelName:ns,backendName:"webgpu",kernelFunc:tde},nde=class{constructor(e){this.workPerThread=1,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e[0],this.variableNames=e.map((t,a)=>`T${a}`),this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize,[this.workPerThread,1,1]),this.shaderKey="addN"}getUserCode(){let e=[];this.variableNames.forEach(a=>{e.push(`let v${a} = get${a}ByOutputCoords(coords);`)});let t=this.variableNames.map(a=>`v${a}`).join(" + ");return`
|
|
${de("index")} {
|
|
for (var i = 0; i < ${this.workPerThread}; i = i + 1) {
|
|
let flatIndex = index * ${this.workPerThread} + i;
|
|
if (flatIndex < uniforms.size) {
|
|
let coords = getCoordsFromIndex(flatIndex);
|
|
${e.join(`
|
|
`)}
|
|
setOutputAtIndex(flatIndex, ${t});
|
|
}
|
|
}
|
|
}
|
|
`}};function rde(e){let{inputs:t,backend:a}=e,n=t;if(n.length===1)return tn({inputs:{x:n[0]},backend:a});let r=n.map(o=>o.dtype).reduce((o,l)=>pa(o,l)),s=n.map(o=>o.shape),i=new nde(s);return a.runWebGPUProgram(i,n,r)}var sde={kernelName:ti,backendName:"webgpu",kernelFunc:rde},ide=class{constructor(e,t){this.variableNames=["A"],this.workgroupSize=[16,16,1];let a=new Array(e.length);for(let n=0;n<a.length;n++)a[n]=e[t[n]];this.outputShape=a,this.dispatchLayout={x:[0],y:[1]},this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize,[1,1,1]),this.shaderKey="transposeShared"}getUserCode(){v.assert(this.workgroupSize[0]===this.workgroupSize[1],()=>`Must be a square tile, current tile shape is ${this.workgroupSize[0]} x ${this.workgroupSize[1]}`);let e=this.workgroupSize[0];return`
|
|
var<workgroup> tile : array<array<f32, ${this.workgroupSize[0]+1}>, ${this.workgroupSize[0]}>;
|
|
${de()} {
|
|
var x = i32(workgroupId.x) * ${e} + i32(localId.x);
|
|
var y = i32(workgroupId.y) * ${e} + i32(localId.y);
|
|
let width = uniforms.outShape[0];
|
|
let height = uniforms.outShape[1];
|
|
if (x < width && y < height) {
|
|
tile[localId.y][localId.x] = f32(A[y * width + x]);
|
|
}
|
|
workgroupBarrier();
|
|
|
|
x = i32(workgroupId.y) * ${e} + i32(localId.x);
|
|
y = i32(workgroupId.x) * ${e} + i32(localId.y);
|
|
if (x < height && y < width) {
|
|
setOutputAtIndex((y * height + x), tile[localId.x]
|
|
[localId.y]);
|
|
}
|
|
}
|
|
`}},ode=class{constructor(e,t){this.variableNames=["A"],this.workPerThread=1,this.workgroupSize=[64,1,1],this.size=!0;let a=new Array(e.length);for(let n=0;n<a.length;n++)a[n]=e[t[n]];this.outputShape=a,this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize,[this.workPerThread,1,1]),this.newDim=t,this.shaderKey=`transpose_${t}`}getUserCode(){let e=Jt(this.outputShape.length),t=lde(this.newDim);return`
|
|
${de("index")} {
|
|
for(var i = 0; i < ${this.workPerThread}; i = i + 1) {
|
|
let flatIndex = index * ${this.workPerThread} + i;
|
|
if(flatIndex < uniforms.size) {
|
|
let resRC = getCoordsFromIndex(flatIndex);
|
|
setOutputAtIndex(flatIndex, A[getIndexFromCoords${this.outputShape.length}D(
|
|
${e}(${t}), uniforms.aShape)]);
|
|
}
|
|
}
|
|
}
|
|
`}};function lde(e){let t=e.length;if(t>6)throw Error(`Transpose for rank ${t} is not yet supported`);let a=new Array(t);for(let n=0;n<e.length;n++)a[e[n]]=`resRC.${Ir(n)}`;return a.join()}function Bn(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{perm:s}=n,i=a,o=r.shape.length,l=new Array(o);for(let p=0;p<l.length;p++)l[p]=r.shape[s[p]];if(a.shouldExecuteOnCPU([r])){let p=i.tensorMap.get(r.dataId).values,c=Xue(p,r.shape,r.dtype,s,l);return a.makeTensorInfo(l,r.dtype,c)}if(r.shape.length===2&&v.arraysEqual(s,[1,0])){let p=new ide(r.shape,s);return i.runWebGPUProgram(p,[r],r.dtype)}let u=new ode(r.shape,s);return i.runWebGPUProgram(u,[r],r.dtype)}var ude={kernelName:wr,backendName:"webgpu",kernelFunc:Bn},dde=class{constructor(e,t,a){this.variableNames=["x"],this.uniforms="reduceSize : i32,",this.size=!0,this.inputShape=[e.batchSize,e.inSize];let[n]=S.computeOutAndReduceShapes(this.inputShape,[1]);this.outputShape=n.length===0?[1]:n,e.inSize>=32768&&a>=512?this.workgroupSize=[512,1,1]:e.inSize>=4096?this.workgroupSize=[256,1,1]:this.workgroupSize=[64,1,1],this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,[1,1,1]),this.reduceType=t,this.shaderKey=`reduce_${t}`}getUserCode(){let e="",t="0.0",a=this.workgroupSize[0];this.reduceType==="min"||this.reduceType==="max"?(e=`
|
|
if (isnan(candidate)) {
|
|
bestValue = uniforms.NAN;
|
|
} else if (!isnan(bestValue) && candidate ${this.reduceType==="min"?"<":">"} bestValue)
|
|
{ bestValue = candidate; }`,t="f32(x[offset])"):this.reduceType==="sum"||this.reduceType==="mean"?e=" bestValue = bestValue + candidate; ":this.reduceType==="prod"?(e=" bestValue = bestValue * candidate; ",t="1.0"):this.reduceType==="all"?(e=" bestValue = f32(bestValue >= 1.0 && candidate >= 1.0); ",t="1.0"):this.reduceType==="any"&&(e=" bestValue = f32(bestValue >= 1.0 || candidate >= 1.0); ",t="0.0");let n=this.reduceType==="mean"?"setOutputAtIndex(outputIndex, bestValue / f32(uniforms.reduceSize));":"setOutputAtIndex(outputIndex, bestValue);";return`
|
|
fn DIV_CEIL(a : u32, b : u32) -> u32 {
|
|
return ((a - 1u) / b + 1u);
|
|
}
|
|
|
|
${`
|
|
var<workgroup> xBestValues : array<f32, ${a}>;
|
|
`}
|
|
fn getOffset(outputIndex : i32) -> i32 {
|
|
let outputCoords = getCoordsFromIndex(outputIndex);
|
|
let offset = ${this.outputShape.length===1?"outputCoords":"outputCoords[0]"} * uniforms.reduceSize;
|
|
return offset;
|
|
}
|
|
${de("index")} {
|
|
let outputIndex = index / ${a};
|
|
let offset = getOffset(outputIndex);
|
|
var bestValue = ${t};
|
|
let Length = uniforms.reduceSize;
|
|
let WorkPerThread = DIV_CEIL(u32(Length), ${a}u);
|
|
for (var k = i32(localId.x); k < Length && outputIndex < uniforms.size;
|
|
k = k + ${a}) {
|
|
let candidate = f32(x[offset + k]);
|
|
${e}
|
|
}
|
|
xBestValues[localId.x] = bestValue;
|
|
workgroupBarrier();
|
|
|
|
var reduceSize = min(u32(Length), ${a}u);
|
|
for (var currentSize = reduceSize / 2u; reduceSize > 1u;
|
|
currentSize = reduceSize / 2u) {
|
|
let interval = DIV_CEIL(reduceSize, 2u);
|
|
if (localId.x < currentSize) {
|
|
let candidate = xBestValues[localId.x + interval];
|
|
${e}
|
|
xBestValues[localId.x] = bestValue;
|
|
}
|
|
reduceSize = interval;
|
|
workgroupBarrier();
|
|
}
|
|
|
|
if (localId.x == 0u && outputIndex < uniforms.size) {
|
|
${n}
|
|
}
|
|
}
|
|
`}};function Jo(e,t,a,n,r){let s=e.shape.length,i=[],o=v.parseAxisParam(t,e.shape),l=o,u=S.getAxesPermutation(l,s),p=e;u!=null&&(p=Bn({inputs:{x:e},attrs:{perm:u},backend:r}),l=S.getInnerMostAxes(l.length,s),i.push(p)),S.assertAxesAreInnerMostDims(n,l,s);let[c,d]=S.computeOutAndReduceShapes(p.shape,l),h=c;a&&(h=S.expandShapeToKeepDim(c,o));let f;if((n==="max"||n==="prod")&&r.shouldExecuteOnCPU([p])){let m=r.tensorMap.get(p.dataId).values;switch(n){case"max":let g=Mue(m,v.sizeFromShape(d),h,e.dtype);f=r.makeTensorInfo(h,e.dtype,g);break;case"prod":let{outVals:y,outShape:x,outDtype:A}=Oue(p.shape,p.dtype,m,l);f=r.makeTensorInfo(x,A,y);break;default:throw new Error(`${n} CPU implementation is not yet supported.`)}}else{let m=v.sizeFromShape(d),g=v.sizeFromShape(p.shape)/m,y={windowSize:m,inSize:m,batchSize:g,outSize:1},x=n==="mean"?"float32":Ip(e.dtype),A=[{type:"int32",data:[m]}],b=new dde(y,n,r.device.limits.maxComputeWorkgroupSizeX),w=r.runWebGPUProgram(b,[p],x,A);i.push(w),f=we({inputs:{x:w},attrs:{shape:h},backend:r})}return i.forEach(m=>r.disposeData(m.dataId)),f}function pde(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{keepDims:s,axis:i}=n;return Jo(r,i,s,"all",a)}var cde={kernelName:ai,backendName:"webgpu",kernelFunc:pde};function hde(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{keepDims:s,axis:i}=n;return Jo(r,i,s,"any",a)}var fde={kernelName:ni,backendName:"webgpu",kernelFunc:hde},Jw=class{constructor(e,t,a){this.workgroupSize=[64,1,1],this.variableNames=["x"],this.uniforms="infinityValue : f32,",this.size=!0;let n=[t];this.op=a==="min"?"<":">";let[r,s]=S.computeOutAndReduceShapes(e,n);this.outputShape=r.length===0?[1]:r,this.dispatchLayout=be(this.outputShape),v.sizeFromShape(s)<32||v.sizeFromShape(r)>1e3?(this.type="plain",this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize)):(this.type="shared",this.dispatch=ce(this.dispatchLayout,this.outputShape,[1,1,1])),this.inputShape=e,this.shaderKey=`argMinMax_${this.op}_${this.type}`}getUserCode(){let e=this.workgroupSize[0],t=()=>this.inputShape.length===1?"uniforms.xShape":`uniforms.xShape.${Ir(this.inputShape.length-1)}`,a=()=>{let n="";if(this.outputShape.length===1)this.inputShape.length!==1&&(n+="outputCoords,");else for(let r=0;r<this.outputShape.length;r++)n+=`outputCoords.${Ir(r)},`;return n};return this.type==="shared"?`
|
|
fn DIV_CEIL(a : u32, b : u32) -> u32 {
|
|
return ((a - 1u) / b + 1u);
|
|
}
|
|
|
|
${`
|
|
var<workgroup> xBestIndices : array<i32, ${e}>;
|
|
var<workgroup> xBestValues : array<f32, ${e}>;
|
|
`}
|
|
|
|
${de("index")} {
|
|
let outputIndex = index / ${e};
|
|
let reduceLength = ${t()};
|
|
|
|
var bestIndex = i32(localId.x);
|
|
var bestValue = uniforms.infinityValue;
|
|
let outputCoords = getCoordsFromIndex(outputIndex);
|
|
for (var k = i32(localId.x); k < reduceLength && outputIndex < uniforms.size;
|
|
k = k + ${e}) {
|
|
let candidate = getX(${a()} k);
|
|
if (!isnan(candidate) && candidate ${this.op} bestValue) {
|
|
bestValue = candidate;
|
|
bestIndex = k;
|
|
}
|
|
}
|
|
xBestValues[localId.x] = bestValue;
|
|
xBestIndices[localId.x] = bestIndex;
|
|
workgroupBarrier();
|
|
|
|
var reduceSize = min(u32(reduceLength), ${e}u);
|
|
for (var currentSize = reduceSize / 2u; reduceSize > 1u;
|
|
currentSize = reduceSize / 2u) {
|
|
let interval = DIV_CEIL(reduceSize, 2u);
|
|
if (localId.x < currentSize) {
|
|
let candidate = xBestValues[localId.x + interval];
|
|
if (candidate ${this.op} bestValue) {
|
|
bestValue = candidate;
|
|
xBestValues[localId.x] = bestValue;
|
|
xBestIndices[localId.x] = xBestIndices[localId.x + interval];
|
|
}
|
|
}
|
|
reduceSize = interval;
|
|
workgroupBarrier();
|
|
}
|
|
|
|
if (localId.x == 0u && outputIndex < uniforms.size) {
|
|
setOutputAtIndexI32(outputIndex, xBestIndices[localId.x]);
|
|
}
|
|
}
|
|
`:`
|
|
${de("index")} {
|
|
if (index < uniforms.size) {
|
|
let outputCoords = getCoordsFromIndex(index);
|
|
var bestIndex = 0;
|
|
var bestValue = getX(${a()} 0);
|
|
let reduceLength = ${t()};
|
|
for (var i = 1; i < reduceLength; i++) {
|
|
let candidate = getX(${a()} i);
|
|
if (candidate ${this.op} bestValue) {
|
|
bestValue = candidate;
|
|
bestIndex = i;
|
|
}
|
|
}
|
|
setOutputAtIndexI32(index, bestIndex);
|
|
}
|
|
}
|
|
`}};function mde(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{axis:s}=n,i=v.parseAxisParam(s,r.shape),o=S.getAxesPermutation(i,r.shape.length),l=r,u=[];o!=null&&(l=Bn({inputs:{x:r},backend:a,attrs:{perm:o}}),u.push(l),i=S.getInnerMostAxes(i.length,l.shape.length)),S.assertAxesAreInnerMostDims("argMax",[i[0]],l.shape.length);let p=new Jw(l.shape,i[0],"max"),c=[{type:"float32",data:[Number.NEGATIVE_INFINITY]}],d=a.runWebGPUProgram(p,[l],"int32",c);return u.forEach(h=>a.disposeData(h.dataId)),d}var gde={kernelName:Jl,backendName:"webgpu",kernelFunc:mde};function yde(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{axis:s}=n,i=v.parseAxisParam(s,r.shape),o=S.getAxesPermutation(i,r.shape.length),l=r,u=[];o!=null&&(l=Bn({inputs:{x:r},backend:a,attrs:{perm:o}}),u.push(l),i=S.getInnerMostAxes(i.length,l.shape.length)),S.assertAxesAreInnerMostDims("argMin",[i[0]],l.shape.length);let p=new Jw(l.shape,i[0],"min"),c=[{type:"float32",data:[Number.POSITIVE_INFINITY]}],d=a.runWebGPUProgram(p,[l],"int32",c);return u.forEach(h=>a.disposeData(h.dataId)),d}var xde={kernelName:Ql,backendName:"webgpu",kernelFunc:yde},Ade=tt({opType:le.ASIN}),bde={kernelName:ri,backendName:"webgpu",kernelFunc:Ade},vde=tt({opType:le.ASINH}),wde={kernelName:si,backendName:"webgpu",kernelFunc:vde},kde=tt({opType:le.ATAN}),Ide={kernelName:ii,backendName:"webgpu",kernelFunc:kde},Sde=ta({opType:_e.ATAN2}),Cde={kernelName:li,backendName:"webgpu",kernelFunc:Sde},Tde=tt({opType:le.ATANH}),Nde={kernelName:oi,backendName:"webgpu",kernelFunc:Tde},Rde=class{constructor(e){this.variableNames=["x"],this.uniforms="strides : vec2<i32>,",this.workgroupSize=[256,1,1],this.size=!0,this.outputShape=e.outShape,this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="poolWithFilterSizeEqualsOne"}getUserCode(){return`
|
|
${de("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
let batch = coords[0];
|
|
let d = coords[3];
|
|
|
|
let xRCCorner = coords.yz * uniforms.strides;
|
|
let xRCorner = xRCCorner.x;
|
|
let xCCorner = xRCCorner.y;
|
|
|
|
let value = getX(batch, xRCorner, xCCorner, d);
|
|
setOutputAtIndex(index, value);
|
|
}
|
|
}
|
|
`}},Kd=class{constructor(e,t,a=!1,n=!1,r=!1){if(this.variableNames=["x"],this.uniforms="strides : vec2<i32>, pads : vec2<i32>, dilations : vec2<i32>, convDims : vec2<i32>, filterDims : vec2<i32>,",this.workgroupSize=[128,1,1],this.size=!0,t==="avg"&&a)throw new Error("Cannot compute positions for average pool.");this.outputShape=e.outShape,this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.poolType=t,this.computePositions=a,this.flattenPositions=n,this.includeBatchIndex=r,this.shaderKey=`pool2D_${t}_${a}_${n}_${r}`}getUserCode(){let e;this.poolType==="avg"?e="resultValue = resultValue + value; count = count + 1.0;":this.computePositions?e=`let currMaxValue = mix(value, maxValue, maxValueFound);
|
|
if (value >= currMaxValue) {
|
|
maxValue = value;
|
|
maxValueFound = 1.0;
|
|
maxPosition = ${this.flattenPositions?this.includeBatchIndex?"((batch * uniforms.xShape[1] + xR) * uniforms.xShape[2] + xC) * uniforms.xShape[3] + d":"(xR * uniforms.xShape[2] + xC) * uniforms.xShape[3] + d":"wR * uniforms.filterDims.y + wC"};
|
|
}`:e="resultValue = max(value, resultValue);";let t="resultValue";return this.poolType==="avg"&&(t="resultValue / max(count, 1.0)"),`
|
|
${de("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
let batch = coords[0];
|
|
let d = coords[3];
|
|
let xRCCorner = vec2<i32>(coords.yz) * uniforms.strides - uniforms.pads;
|
|
let xRCorner = xRCCorner.x;
|
|
let xCCorner = xRCCorner.y;
|
|
|
|
${this.computePositions?`var maxValue = 0.0;
|
|
var maxValueFound = 0.0;
|
|
var maxPosition = 0;`:`var resultValue = ${this.poolType==="avg"?"0.0":"-1.0 / pow(10.0, -20.0)"};`}
|
|
|
|
var count = 0.0;
|
|
for (var wR = 0; wR < uniforms.filterDims.x; wR = wR + uniforms.dilations.x) {
|
|
let xR = xRCorner + wR;
|
|
|
|
if (xR < 0 || xR >= uniforms.convDims.x) {
|
|
continue;
|
|
}
|
|
|
|
for (var wC = 0; wC < uniforms.filterDims.y; wC = wC + uniforms.dilations.y) {
|
|
let xC = xCCorner + wC;
|
|
if (xC < 0 || xC >= uniforms.convDims.y) {
|
|
continue;
|
|
}
|
|
|
|
let value = getX(batch, xR, xC, d);
|
|
${e}
|
|
}
|
|
}
|
|
|
|
${this.computePositions?"setOutputAtIndexI32(index, maxPosition);":`setOutputAtIndex(index, ${t});`}
|
|
}
|
|
}
|
|
`}},Dg=class{constructor(e,t,a=!1,n=!1,r=!1){if(this.variableNames=["x"],this.uniforms="strides : vec3<i32>, pads : vec3<i32>, convDims : vec3<i32>, filterDims : vec3<i32>,",this.workgroupSize=[128,1,1],this.size=!0,t==="avg"&&a)throw new Error("Cannot compute positions for average pool.");this.outputShape=e.outShape,this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.poolType=t,this.computePositions=a,this.flattenPositions=n,this.includeBatchIndex=r,this.shaderKey=`pool3D_${t}_${a}_${n}_${r}`}getUserCode(){let e;this.poolType==="avg"?e="resultValue += value; count += 1.0;":this.computePositions?e=`let currMaxValue = mix(value, maxValue, maxValueFound);
|
|
if (value >= currMaxValue) {
|
|
maxValue = value;
|
|
maxValueFound = 1.0;
|
|
maxPosition = ${this.flattenPositions?this.includeBatchIndex?"(((batch * uniforms.xShape.y + xD) * uniforms.xShape.z + xR) * uniforms.xShape.w + xC) * uniforms.xShape.u + ch":"((xD * uniforms.xShape.z + xR) * uniforms.xShape.w + xC) * uniforms.xShape.u + ch":"wD * uniforms.filterDims.y * uniforms.filterDims.y + wR * uniforms.filterDims.z + wC"};
|
|
}`:e="resultValue = max(value, resultValue);";let t="resultValue";return this.poolType==="avg"&&(t="resultValue / max(count, 1.0)"),`
|
|
${de("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
let batch = coords.x;
|
|
let ch = coords.u;
|
|
|
|
let xCorner = vec3<i32>(coords.y, coords.z, coords.w) * uniforms.strides - uniforms.pads;
|
|
let xDCorner = xCorner.x;
|
|
let xRCorner = xCorner.y;
|
|
let xCCorner = xCorner.z;
|
|
|
|
${this.computePositions?`var maxValue = 0.0;
|
|
var maxValueFound = 0.0;
|
|
var maxPosition = 0;`:`var resultValue = ${this.poolType==="avg"?"0.0":"-1.0 / pow(10.0, -20.0)"};`}
|
|
|
|
var count = 0.0;
|
|
for (var wD = 0; wD < uniforms.filterDims.x; wD++) {
|
|
let xD = xDCorner + wD;
|
|
if (xD < 0 || xD >= uniforms.convDims.x) {
|
|
continue;
|
|
}
|
|
|
|
for (var wR = 0; wR < uniforms.filterDims.y; wR++) {
|
|
let xR = xRCorner + wR;
|
|
if (xR < 0 || xR >= uniforms.convDims.y) {
|
|
continue;
|
|
}
|
|
|
|
for (var wC = 0; wC < uniforms.filterDims.z; wC++) {
|
|
let xC = xCCorner + wC;
|
|
if (xC < 0 || xC >= uniforms.convDims.z) {
|
|
continue;
|
|
}
|
|
|
|
let value = getX(batch, xD, xR, xC, ch);
|
|
${e}
|
|
}
|
|
}
|
|
}
|
|
|
|
${this.computePositions?"setOutputAtIndexI32(index, maxPosition);":`setOutputAtIndex(index, ${t});`}
|
|
}
|
|
}
|
|
`}};function Qw(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{reductionIndices:s,keepDims:i}=n;return Jo(r,s,i,"max",a)}var Ede={kernelName:Ji,backendName:"webgpu",kernelFunc:Qw};function ek(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{keepDims:s,axis:i}=n;return Jo(r,i,s,"mean",a)}var Mde={kernelName:to,backendName:"webgpu",kernelFunc:ek};function tk(e,t,a,n){if(t.filterWidth===1&&t.filterHeight===1&&v.arraysEqual(t.inShape,t.outShape))return tn({inputs:{x:e},backend:n});if(t.filterWidth===t.inWidth&&t.filterHeight===t.inHeight&&t.batchSize===1&&t.padInfo.type==="VALID"){let i=e.shape.length,o=we({inputs:{x:e},backend:n,attrs:{shape:[e.shape[i-3]*e.shape[i-2],e.shape[i-1]]}}),l;a==="avg"?l=ek({inputs:{x:o},backend:n,attrs:{axis:0,keepDims:!1}}):(v.assert(a==="max",()=>`Invalid pool type ${a}`),l=Qw({inputs:{x:o},backend:n,attrs:{reductionIndices:0,keepDims:!1}}));let u=we({inputs:{x:l},backend:n,attrs:{shape:t.outShape}});return n.disposeData(o.dataId),n.disposeData(l.dataId),u}let r,s=[{type:"int32",data:[t.strideHeight,t.strideWidth]}];return t.filterHeight===1&&t.filterWidth===1?r=new Rde(t):(a==="avg"?r=new Kd(t,"avg"):(v.assert(a==="max",()=>`Invalid pool type ${a}`),r=new Kd(t,"max")),s.push({type:"int32",data:[t.padInfo.top,t.padInfo.left]},{type:"int32",data:[t.dilationHeight,t.dilationWidth]},{type:"int32",data:[t.inHeight,t.inWidth]},{type:"int32",data:[t.effectiveFilterHeight,t.effectiveFilterWidth]})),n.runWebGPUProgram(r,[e],e.dtype,s)}function Fde(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{filterSize:s,strides:i,pad:o,dimRoundingMode:l}=n,u=1,p=S.computePool2DInfo(r.shape,s,i,u,o,l);return tk(r,p,"avg",a)}var _de={kernelName:ui,backendName:"webgpu",kernelFunc:Fde};function $de(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{filterSize:s,strides:i,pad:o,dataFormat:l,dimRoundingMode:u}=n,p=[1,1,1],c=S.computePool3DInfo(r.shape,s,i,p,o,u,l),d=new Dg(c,"avg"),h=[{type:"int32",data:[c.strideDepth,c.strideHeight,c.strideWidth]},{type:"int32",data:[c.padInfo.front,c.padInfo.top,c.padInfo.left]},{type:"int32",data:[c.inDepth,c.inHeight,c.inWidth]},{type:"int32",data:[c.effectiveFilterDepth,c.effectiveFilterHeight,c.effectiveFilterWidth]}];return a.runWebGPUProgram(d,[r],r.dtype,h)}var Dde={kernelName:eu,backendName:"webgpu",kernelFunc:$de},Pde=class{constructor(e){this.variableNames=["dy"],this.uniforms=`strides : vec2<i32>, pads : vec2<i32>, dilations : vec2<i32>, filterDims : vec2<i32>,
|
|
outHeight : i32, outWidth : i32, avgMultiplier : f32,`,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e.inShape,this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="avgPool2DBackprop"}getUserCode(){return`
|
|
${de("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
let batch = coords[0];
|
|
let d = coords[3];
|
|
|
|
let dyRCCorner = vec2<i32>(coords.yz) - uniforms.pads;
|
|
let dyRCorner = dyRCCorner.x;
|
|
let dyCCorner = dyRCCorner.y;
|
|
|
|
// Convolve dy(?, ?, d) with pos mask(:, :, d) to get dx(xR, xC, d).
|
|
// ? = to be determined. : = across all values in that axis.
|
|
var dotProd = 0.0;
|
|
for (var wR = 0; wR < uniforms.filterDims[0]; wR = wR + uniforms.dilations[0]) {
|
|
let dyR = f32(dyRCorner + wR) / f32(uniforms.strides[0]);
|
|
|
|
if (dyR < 0.0 || dyR >= f32(uniforms.outHeight) || fract(dyR) > 0.0) {
|
|
continue;
|
|
}
|
|
let idyR = i32(dyR);
|
|
|
|
for (var wC = 0; wC < uniforms.filterDims[1]; wC = wC + uniforms.dilations[1]) {
|
|
let dyC = f32(dyCCorner + wC) / f32(uniforms.strides[1]);
|
|
|
|
if (dyC < 0.0 || dyC >= f32(uniforms.outWidth) || fract(dyC) > 0.0) {
|
|
continue;
|
|
}
|
|
let idyC = i32(dyC);
|
|
|
|
let dyValue = getDy(batch, idyR, idyC, d);
|
|
|
|
dotProd = dotProd + dyValue * uniforms.avgMultiplier;
|
|
}
|
|
}
|
|
setOutputAtIndex(index, dotProd);
|
|
}
|
|
}
|
|
`}},Ode=class{constructor(e){this.variableNames=["dy"],this.uniforms=`strides : vec3<i32>, pads : vec3<i32>, filterDims : vec3<i32>,
|
|
outDepth : i32, outHeight : i32, outWidth : i32, avgMultiplier : f32,`,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e.inShape,this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="avgPool3DBackprop"}getUserCode(){return`
|
|
${de("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
let batch = coords.x;
|
|
let ch = coords.u;
|
|
|
|
let dyCorner = vec3<i32>(coords.y, coords.z, coords.w) - uniforms.pads;
|
|
let dyDCorner = dyCorner.x;
|
|
let dyRCorner = dyCorner.y;
|
|
let dyCCorner = dyCorner.z;
|
|
|
|
// Convolve dy(?, ?, ?, d) with pos mask(:, :, :, ch) to get
|
|
// dx(xD, xR, xC, ch).
|
|
// ? = to be determined. : = across all values in that axis.
|
|
var dotProd = 0.0;
|
|
for (var wD = 0; wD < uniforms.filterDims[0]; wD++) {
|
|
let dyD = f32(dyDCorner + wD) / f32(uniforms.strides[0]);
|
|
|
|
if (dyD < 0.0 || dyD >= f32(uniforms.outDepth) || fract(dyD) > 0.0) {
|
|
continue;
|
|
}
|
|
let idyD = i32(dyD);
|
|
|
|
for (var wR = 0; wR < uniforms.filterDims[1]; wR++) {
|
|
let dyR = f32(dyRCorner + wR) / f32(uniforms.strides[1]);
|
|
|
|
if (dyR < 0.0 || dyR >= f32(uniforms.outHeight) || fract(dyR) > 0.0) {
|
|
continue;
|
|
}
|
|
let idyR = i32(dyR);
|
|
|
|
for (var wC = 0; wC < uniforms.filterDims[2]; wC++) {
|
|
let dyC = f32(dyCCorner + wC) / f32(uniforms.strides[2]);
|
|
|
|
if (dyC < 0.0 || dyC >= f32(uniforms.outWidth) || fract(dyC) > 0.0) {
|
|
continue;
|
|
}
|
|
let idyC = i32(dyC);
|
|
|
|
let dyValue = getDy(batch, idyD, idyR, idyC, ch);
|
|
dotProd += dyValue * uniforms.avgMultiplier;
|
|
}
|
|
}
|
|
}
|
|
setOutputAtIndex(index, dotProd);
|
|
}
|
|
}
|
|
`}};function zde(e){let{inputs:t,backend:a,attrs:n}=e,{dy:r,input:s}=t,i=s,{filterSize:o,strides:l,pad:u,dimRoundingMode:p}=n,c=S.computePool3DInfo(i.shape,o,l,1,u,p),d=new Ode(c),h=1/(c.filterDepth*c.filterHeight*c.filterWidth),f=[{type:"int32",data:[c.strideDepth,c.strideHeight,c.strideWidth]},{type:"int32",data:[c.effectiveFilterDepth-1-c.padInfo.front,c.effectiveFilterHeight-1-c.padInfo.top,c.effectiveFilterWidth-1-c.padInfo.left]},{type:"int32",data:[c.effectiveFilterDepth,c.effectiveFilterHeight,c.effectiveFilterWidth]},{type:"int32",data:[c.outDepth]},{type:"int32",data:[c.outHeight]},{type:"int32",data:[c.outWidth]},{type:"float32",data:[h]}];return a.runWebGPUProgram(d,[r],i.dtype,f)}var Lde={kernelName:ep,backendName:"webgpu",kernelFunc:zde};function Wde(e){let{inputs:t,backend:a,attrs:n}=e,{dy:r,input:s}=t,i=s;Fg([r,s],"avgPoolGrad");let{filterSize:o,strides:l,pad:u}=n,p=S.computePool2DInfo(i.shape,o,l,1,u),c=new Pde(p),d=1/(p.filterHeight*p.filterWidth),h=[{type:"int32",data:[p.strideHeight,p.strideWidth]},{type:"int32",data:[p.effectiveFilterHeight-1-p.padInfo.top,p.effectiveFilterWidth-1-p.padInfo.left]},{type:"int32",data:[p.dilationHeight,p.dilationWidth]},{type:"int32",data:[p.effectiveFilterHeight,p.effectiveFilterWidth]},{type:"int32",data:[p.outHeight]},{type:"int32",data:[p.outWidth]},{type:"float32",data:[d]}];return a.runWebGPUProgram(c,[r],i.dtype,h)}var Bde={kernelName:kh,backendName:"webgpu",kernelFunc:Wde};function Vde(e){let{inputs:t,backend:a,attrs:n}=e,{a:r,b:s}=t,{transposeA:i,transposeB:o}=n;return o0({a:r,b:s,transposeA:i,transposeB:o,backend:a})}var Ude={kernelName:di,backendName:"webgpu",kernelFunc:Vde},Gde=class{constructor(e,t){this.variableNames=["source"],this.workPerThread=1,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=t,this.rank=t.length,this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize,[this.workPerThread,1,1]),this.start=e,this.uniforms=`start : ${Jt(e.length)}, `,this.shaderKey="slice"}getUserCode(){let e=Jt(this.rank),t=Hde(this.rank),a;return this.start.length===1?a=this.outputShape.map((n,r)=>"sourceLoc = uniforms.start + coords;"):a=this.outputShape.map((n,r)=>`sourceLoc.${$1[r]} = uniforms.start.${Ir(r)} + coords.${$1[r]};`),`
|
|
${de("index")} {
|
|
if (index < uniforms.size) {
|
|
var sourceLoc : ${e};
|
|
let coords = getCoordsFromIndex(index);
|
|
${a.join(`
|
|
`)}
|
|
setOutputAtIndex(index, getSource(${t}));
|
|
}
|
|
}
|
|
`}},$1=["x","y","z","w","u","v"];function Hde(e){if(e===1)return"sourceLoc";if(e<=6)return $1.slice(0,e).map(t=>`sourceLoc.${t}`).join(",");throw Error(`Slicing for rank ${e} is not yet supported`)}function ju(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{begin:s,size:i}=n,[o,l]=Nt.parseSliceParams(r,s,i);if(Nt.assertParamsValid(r,o,l),a.shouldExecuteOnCPU([r])||r.dtype==="string"){let c=a.tensorMap.get(r.dataId),d=Vue(c.values,o,l,r.shape,r.dtype);return a.makeTensorInfo(l,r.dtype,d)}if(v.sizeFromShape(l)===0)return a.makeTensorInfo(l,r.dtype,[]);let u=new Gde(o,l),p=[{type:"int32",data:o}];return a.runWebGPUProgram(u,[r],r.dtype,p)}var jde={kernelName:Iu,backendName:"webgpu",kernelFunc:ju},qde=e=>{let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{blockShape:s,crops:i}=n;v.assert(r.shape.length<=4,()=>"batchToSpaceND for rank > 4 with a WebGPU backend not implemented yet");let o=s.reduce((x,A)=>x*A),l=S.getReshaped(r.shape,s,o),u=S.getPermuted(l.length,s.length),p=S.getReshapedPermuted(r.shape,s,o),c=S.getSliceBeginCoords(i,s.length),d=S.getSliceSize(p,i,s.length),h=[],f=we({inputs:{x:r},backend:a,attrs:{shape:l}}),m=Bn({inputs:{x:f},backend:a,attrs:{perm:u}}),g=we({inputs:{x:m},backend:a,attrs:{shape:p}}),y=ju({inputs:{x:g},backend:a,attrs:{begin:c,size:d}});return h.push(f),h.push(m),h.push(g),h.forEach(x=>a.disposeData(x.dataId)),y},Xde={kernelName:tu,backendName:"webgpu",kernelFunc:qde},Kde=`
|
|
fn bincount_write(index: i32, value: f32) {
|
|
${Gu("&result[index]","value","float32")}
|
|
}
|
|
`,Yde=`
|
|
fn bincount_write(index: i32, value: f32) {
|
|
atomicStore(&result[index], bitcast<i32>(value));
|
|
}
|
|
`,ak=class{constructor(e,t,a=!1){this.outputShape=[],this.variableNames=["x"],this.uniforms="binCountSize : i32,",this.workgroupSize=[64,1,1],this.atomic=!0,this.hasWeights=!0,this.binaryOutput=!1,this.outputShape=e,this.rank=e.length,this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.binaryOutput=a,a&&(this.atomic=!1),this.hasWeights=t,this.hasWeights&&this.variableNames.push("w"),this.shaderKey=`bincount_${this.hasWeights}_${this.binaryOutput}_${this.rank}`}getUserCode(){return`
|
|
${this.binaryOutput?Yde:Kde}
|
|
${de("index")} {
|
|
${this.rank===1?`if (index < uniforms.xShape) {
|
|
let indexVal = i32(getX(index));
|
|
if (indexVal < uniforms.binCountSize) {
|
|
let value = ${this.binaryOutput?1:this.hasWeights?"getW(index)":"1."};
|
|
bincount_write(indexVal, value);
|
|
}
|
|
}`:`let coord = getCoordsFromIndex(index);
|
|
if (coordsInBounds2D(coord, uniforms.xShape)) {
|
|
let indexVal = i32(getX(coord[0], coord[1]));
|
|
if (indexVal < uniforms.binCountSize) {
|
|
let value = ${this.binaryOutput?1:this.hasWeights?"getW(coord[0], coord[1])":"1."};
|
|
bincount_write(coord.x * uniforms.binCountSize + indexVal, value);
|
|
}
|
|
}`}
|
|
}
|
|
`}};function Zde(e){let{inputs:t,backend:a,attrs:n}=e,{x:r,weights:s}=t,{size:i}=n,o=v.sizeFromShape(r.shape),l=v.sizeFromShape(s.shape)>0,u=[i],p=s.dtype,c=Mn({backend:a,attrs:{shape:u,value:0,dtype:p}}),d=new ak([o],l),h=[{type:"int32",data:[i]}],f=l?[r,s]:[r];return a.runWebGPUProgram(d,f,p,h,c)}var Jde={kernelName:pi,backendName:"webgpu",kernelFunc:Zde},Qde=class{constructor(e){this.outputShape=[],this.variableNames=["s0","s1"],this.uniforms="s0Size : i32, s1Size : i32, ",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=[e],this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="broadcastArgs"}getUserCode(){return`
|
|
${de("index")} {
|
|
if (index < uniforms.size) {
|
|
var s0 = 1.0;
|
|
var s1 = 1.0;
|
|
let indexS0 = index - uniforms.size + uniforms.s0Size;
|
|
let indexS1 = index - uniforms.size + uniforms.s1Size;
|
|
if (indexS0 >= 0) {
|
|
s0 = getS0(indexS0);
|
|
}
|
|
if (indexS1 >= 0) {
|
|
s1 = getS1(indexS1);
|
|
}
|
|
|
|
if (s0 == 1.0) {
|
|
setOutputAtIndex(index, s1);
|
|
} else if (s1 == 1.0) {
|
|
setOutputAtIndex(index, s0);
|
|
} else if (s0 != s1) {
|
|
setOutputAtIndex(index, uniforms.NAN);
|
|
} else {
|
|
setOutputAtIndex(index, s0);
|
|
}
|
|
}
|
|
}
|
|
`}};function epe(e){let{inputs:t,backend:a}=e,{s0:n,s1:r}=t;if(a.shouldExecuteOnCPU([n,r])){let p=a.tensorMap.get(n.dataId),c=a.tensorMap.get(r.dataId),d=p.values,h=c.values,f=S.assertAndGetBroadcastShape(Array.from(d),Array.from(h));return a.makeTensorInfo([f.length],"int32",Int32Array.from(f))}let s=v.sizeFromShape(n.shape),i=v.sizeFromShape(r.shape),o=Math.max(s,i),l=new Qde(o),u=[{type:"int32",data:[s]},{type:"int32",data:[i]}];return a.runWebGPUProgram(l,[n,r],"int32",u)}var tpe={kernelName:au,backendName:"webgpu",kernelFunc:epe},nk=ta({opType:_e.NOT_EQUAL,dtype:"bool",cpuKernelImpl:Pue}),ape={kernelName:oo,backendName:"webgpu",kernelFunc:nk};function jp(e){let{inputs:t,backend:a}=e,{input:n}=t,r=a.tensorMap.get(n.dataId);return tn({inputs:{x:r.complexTensorInfos.real},backend:a})}var npe={kernelName:cp,backendName:"webgpu",kernelFunc:jp};function rpe(e,t){let a=new Hu(e.shape,le.TO_INT),n=t.runWebGPUProgram(a,[e],"int32");return{dataId:n.dataId,shape:n.shape,dtype:n.dtype}}function D1(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{dtype:s}=n;if(s==="complex64"){if(r.dtype==="complex64")return tn({inputs:{x:r},backend:a});let i=gn(r.shape),o=D1({inputs:{x:r},backend:a,attrs:{dtype:"float32"}}),l=Zo({inputs:{real:o,imag:i},backend:a});return i.dispose(),a.disposeData(o.dataId),l}if(r.dtype==="complex64"){let i=jp({inputs:{input:r},backend:a}),o=D1({inputs:{x:i},backend:a,attrs:{dtype:s}});return a.disposeData(i.dataId),o}if(!v.hasEncodingLoss(r.dtype,s)){let i=tn({inputs:{x:r},backend:a});return{dataId:i.dataId,shape:i.shape,dtype:s}}if(a.shouldExecuteOnCPU([r])){let i=a.tensorMap.get(r.dataId).values,[o,l,u]=gue(i,r.shape,r.dtype,s);return a.makeTensorInfo(o,l,u)}if(s==="int32")return rpe(r,a);if(s==="bool"){let i=a.makeTensorInfo([],"bool",v.getTypedArrayFromDType("bool",1)),o=nk({inputs:{a:r,b:i},backend:a});return a.disposeData(i.dataId),o}throw new Error(`Error in Cast: failed to cast ${r.dtype} to ${s}`)}var spe={kernelName:ci,backendName:"webgpu",kernelFunc:D1},ipe=tt({opType:le.CEIL,cpuKernelImpl:yue}),ope={kernelName:hi,backendName:"webgpu",kernelFunc:ipe},lpe=class{constructor(e){this.variableNames=["A"],this.uniforms="minVal : f32, maxVal : f32,",this.workPerThread=4,this.workgroupSize=[64,1,1],this.outputComponent=4,this.size=!0,this.outputShape=e,this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize,[this.workPerThread,1,1]),this.shaderKey="clipVec4"}getUserCode(){return`
|
|
${de("index")} {
|
|
if(index < uniforms.size) {
|
|
let value = getAByOutputIndex(index);
|
|
var clampedValue = clamp(
|
|
value, vec4<f32>(uniforms.minVal), vec4<f32>(uniforms.maxVal));
|
|
clampedValue = select(clampedValue, value, isnanVec4(value));
|
|
setOutputAtIndex(index, clampedValue);
|
|
}
|
|
}
|
|
`}},upe=class{constructor(e){this.variableNames=["A"],this.uniforms="minVal : f32, maxVal : f32,",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="clip"}getUserCode(){return`
|
|
${de("index")} {
|
|
if(index < uniforms.size) {
|
|
let value = getAByOutputIndex(index);
|
|
if (isnan(value)) {
|
|
setOutputAtIndex(index, value);
|
|
return;
|
|
}
|
|
setOutputAtIndex(index, clamp(value, uniforms.minVal, uniforms.maxVal));
|
|
}
|
|
}
|
|
`}};function dpe(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{clipValueMin:s,clipValueMax:i}=n,o,l=[{type:"float32",data:[s]},{type:"float32",data:[i]}];return v.sizeFromShape(r.shape)%4===0?o=new lpe(r.shape):o=new upe(r.shape),a.runWebGPUProgram(o,[r],r.dtype,l)}var ppe={kernelName:rs,backendName:"webgpu",kernelFunc:dpe},cpe=class{constructor(e){this.outputShape=[],this.variableNames=["real","imag"],this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="complexAbs"}getUserCode(){return`
|
|
${de("index")} {
|
|
if (index < uniforms.size) {
|
|
let re = abs(getRealByOutputIndex(index));
|
|
let im = abs(getImagByOutputIndex(index));
|
|
let mx = max(re, im);
|
|
|
|
// The length function in wgsl may be not underflow-safe on some GPUs.
|
|
// So the safe solution is to ensure underflow-safety in all cases.
|
|
setOutputAtIndex(index, select(mx * length(vec2<f32>(1, min(re, im)/mx)), 0.0, mx == 0.0));
|
|
}
|
|
}
|
|
`}};function Lx(e,t){return{dataId:t.dataId,dtype:t.dtype,shape:e.shape}}function hpe(e){let{inputs:t,backend:a}=e,{x:n}=t,r=a.tensorMap.get(n.dataId),s=new cpe(n.shape),i=[Lx(n,r.complexTensorInfos.real),Lx(n,r.complexTensorInfos.imag)];return a.runWebGPUProgram(s,i,i[0].dtype)}var fpe={kernelName:ap,backendName:"webgpu",kernelFunc:hpe},mpe=class{constructor(e){this.uniforms="",this.workPerThread=1,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=S.computeOutShape(e,1),this.variableNames=e.map((t,a)=>`T${a}`),this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize,[this.workPerThread,1,1]),this.offsetLength=e.length-1;for(let t=0;t<this.offsetLength;t++)this.uniforms+=`offset${t} : i32,`;this.shaderKey="concat"}getUserCode(){let e=[];if(this.offsetLength>0){e.push("if (yC < uniforms.offset0){ setOutputAtCoords(coords.x, coords.y, getT0(yR, yC)); }");for(let n=1;n<this.offsetLength;n++)e.push(`else if (yC < uniforms.offset${[n]}){ setOutputAtCoords(coords.x, coords.y, getT${n}(yR, yC - uniforms.offset${n-1})); }`);let t=this.offsetLength,a=this.offsetLength-1;e.push(`else { setOutputAtCoords(coords.x, coords.y, getT${t}(yR, yC - uniforms.offset${a})); }`)}else e.push("setOutputAtCoords(coords.x, coords.y, getT0(yR, yC));");return`
|
|
${de("index")} {
|
|
for(var i = 0; i < ${this.workPerThread}; i = i + 1) {
|
|
let flatIndex = index * ${this.workPerThread} + i;
|
|
if(flatIndex < uniforms.size) {
|
|
let coords = getCoordsFromIndex(flatIndex);
|
|
let yR = coords.x;
|
|
let yC = coords.y;
|
|
|
|
${e.join(`
|
|
`)}
|
|
}
|
|
}
|
|
}
|
|
`}};function l0(e){let{inputs:t,backend:a}=e,{input:n}=t,r=a.tensorMap.get(n.dataId);return tn({inputs:{x:r.complexTensorInfos.imag},backend:a})}var gpe={kernelName:up,backendName:"webgpu",kernelFunc:l0};function Id(e,t,a){let n=e[0].dtype;if(n==="complex64"){let f=e.map(A=>jp({inputs:{input:A},backend:a})),m=e.map(A=>l0({inputs:{input:A},backend:a})),g=Id(f,t,a),y=Id(m,t,a),x=Zo({inputs:{real:g,imag:y},backend:a});return f.forEach(A=>a.disposeData(A.dataId)),m.forEach(A=>a.disposeData(A.dataId)),a.disposeData(g.dataId),a.disposeData(y.dataId),x}let r=a.shouldExecuteOnCPU(e);if(n==="string"&&(r=!0),r){let f=e.map(w=>{let C=[-1,v.sizeFromShape(w.shape.slice(t))];return we({inputs:{x:w},backend:a,attrs:{shape:C}})}),m=f.map(w=>({vals:a.readSync(w.dataId),shape:w.shape})),g=S.computeOutShape(f.map(w=>w.shape),1),y=f[0].shape[0]===1,x=xue(m,g,n,y),A=S.computeOutShape(e.map(w=>w.shape),t),b=a.makeTensorInfo(A,n,x);return f.forEach(w=>a.disposeData(w.dataId)),b}let s=a.device.limits.maxStorageBuffersPerShaderStage-1;if(e.length>s){let f=[];for(let g=0;g<e.length;g+=s){let y=e.slice(g,g+s);f.push(Id(y,t,a))}let m=Id(f,t,a);for(let g of f)a.disposeData(g.dataId);return m}let{tensors2D:i,outShape:o}=ype(e,t,a),l=i.map(f=>f.shape),u=new mpe(l),p=[],c=new Array(l.length-1);if(c.length>0){c[0]=l[0][1],p.push({type:"int32",data:[c[0]]});for(let f=1;f<c.length;f++)c[f]=c[f-1]+l[f][1],p.push({type:"int32",data:[c[f]]})}let d=a.runWebGPUProgram(u,i,i[0].dtype,p);i.forEach(f=>a.disposeData(f.dataId));let h=we({inputs:{x:d},backend:a,attrs:{shape:o}});return a.disposeData(d.dataId),h}function ype(e,t,a){let n=S.computeOutShape(e.map(r=>r.shape),t);return{tensors2D:e.map(r=>we({inputs:{x:r},backend:a,attrs:{shape:[v.sizeFromShape(r.shape.slice(0,t)),v.sizeFromShape(r.shape.slice(t))]}})),outShape:n}}function rk(e){let{inputs:t,backend:a,attrs:n}=e,{axis:r}=n,s=v.parseAxisParam(r,t[0].shape)[0],i=t.map(u=>u.shape);S.assertParamsConsistent(i,s);let o=S.computeOutShape(t.map(u=>u.shape),s);if(v.sizeFromShape(o)===0)return a.makeTensorInfo(o,t[0].dtype,[]);let l=t.filter(u=>v.sizeFromShape(u.shape)>0);return l.length===1?tn({inputs:{x:l[0]},backend:a}):Id(l,s,a)}var xpe={kernelName:nu,backendName:"webgpu",kernelFunc:rk};function Ape(e,t,a,n,r=!1,s=null,i=!1,o=4,l=4,u=4){let p=N=>{switch(N){case 1:return"resData = x[xIndex];";case 3:return"resData = vec3<f32>(x[xIndex], x[xIndex + 1], x[xIndex + 2]);";case 4:return"resData = x[xIndex / 4];";default:throw new Error(`innerElementSize ${N} is not supported.`)}},c=N=>{switch(N){case 1:return"return W[row * uniforms.wShape[3] + colIn];";case 4:return"return W[row * uniforms.wShape[3] / 4 + colIn];";default:throw new Error(`innerElementSize ${N} is not supported.`)}},d=e?`
|
|
let coord = vec4<i32>(batch, xRow, xCol, xCh);
|
|
`:`
|
|
let coord = vec4<i32>(batch, xCh, xRow, xCol);
|
|
`,h=e?`
|
|
let coords = vec4<i32>(
|
|
batch,
|
|
row / outWidth,
|
|
row % outWidth,
|
|
col);
|
|
`:`
|
|
let coords = vec4<i32>(
|
|
batch,
|
|
row,
|
|
col / outWidth,
|
|
col % outWidth);
|
|
`,f=e?"uniforms.xShape[1]":"uniforms.xShape[2]",m=e?"uniforms.xShape[2]":"uniforms.xShape[3]",g=e?"row":"col",y=e?"col":"row",x=`
|
|
let inChannels = uniforms.wShape[2];
|
|
let outWidth = ${e?"uniforms.outShape[2]":"uniforms.outShape[3]"};
|
|
let outRow = ${g} / outWidth;
|
|
let outCol = ${g} % outWidth;
|
|
|
|
let WRow = ${y} / (uniforms.filterDims[1] * inChannels);
|
|
let WCol = ${y} / inChannels % uniforms.filterDims[1];
|
|
let xRow = outRow * uniforms.strides[0] + uniforms.dilations[0] * WRow - uniforms.pads[0];
|
|
let xCol = outCol * uniforms.strides[1] + uniforms.dilations[1] * WCol - uniforms.pads[1];
|
|
let xCh = ${y} % inChannels;
|
|
var resData = ${Xe(o)}(0.0);
|
|
// The bounds checking is always needed since we use it to pad zero for
|
|
// the 'same' padding type.
|
|
if (xRow >= 0 && xRow < ${f} && xCol >= 0 && xCol < ${m}) {
|
|
${d}
|
|
let xIndex = getIndexFromCoords4D(coord, uniforms.xShape);
|
|
${p(o)}
|
|
}
|
|
return resData;`,A=e?t&&n?`
|
|
let col = colIn * ${o};
|
|
${x}`:`
|
|
let col = colIn * ${o};
|
|
if (row < uniforms.dimAOuter && col < uniforms.dimInner) {
|
|
${x}
|
|
}
|
|
return ${Xe(o)}(0.0);`:n&&a?`
|
|
let col = colIn * ${o};
|
|
${x}`:`
|
|
let col = colIn * ${o};
|
|
if (row < uniforms.dimInner && col < uniforms.dimBOuter) {
|
|
${x}
|
|
}
|
|
return ${Xe(o)}(0.0);`,b=`${c(l)}`,w=Xe(u),C=Xe(e?o:l),T=Xe(e?l:o);return`
|
|
${Mr(s,i,u===4,4)}
|
|
fn mm_readA(batch: i32, row : i32, colIn : i32) -> ${C} {
|
|
${e?A:b}
|
|
}
|
|
|
|
fn mm_readB(batch: i32, row : i32, colIn : i32) -> ${T} {
|
|
${e?b:A}
|
|
}
|
|
|
|
fn mm_write(batch: i32, row : i32, colIn : i32, valueIn : ${w}) {
|
|
let col = colIn * ${u};
|
|
if (row < uniforms.dimAOuter && col < uniforms.dimBOuter)
|
|
{
|
|
var value = valueIn;
|
|
let outWidth = ${e?"uniforms.outShape[2]":"uniforms.outShape[3]"};
|
|
${h}
|
|
${Yo(r,s)}
|
|
setOutputAtCoords(coords[0], coords[1], coords[2], coords[3], value);
|
|
}
|
|
}`}var bpe=class{constructor(e,t,a,n,r=!1,s=null,i=!1,o=!1){this.variableNames=["x","W"],this.uniforms="filterDims : vec2<i32>, pads : vec2<i32>, strides : vec2<i32>, dilations : vec2<i32>, dimAOuter : i32, dimBOuter : i32, dimInner : i32,",this.outputShape=e.outShape,this.isChannelsLast=e.dataFormat==="channelsLast",this.isVec4=((e.inChannels%4===0||e.inChannels%3===0)&&this.isChannelsLast||e.outWidth%4===0&&!this.isChannelsLast)&&e.outChannels%4===0,this.dispatchLayout=this.isChannelsLast?{x:[3],y:[1,2],z:[0]}:{x:[2,3],y:[1],z:[0]},this.workgroupSize=Rg(this.dispatchLayout,this.outputShape,this.isVec4),this.elementsPerThread=Eg(this.dispatchLayout,this.outputShape,this.isVec4),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize,this.elementsPerThread),this.isVec4?(this.outputComponent=4,this.isChannelsLast&&e.inChannels%4!==0?(this.innerElementSize=3,this.variableComponents=[1,4]):(this.innerElementSize=4,this.variableComponents=[4,4]),r&&(this.variableNames.push("bias"),this.variableComponents.push(4)),i&&(this.variableNames.push("preluActivationWeights"),this.variableComponents.push(4))):(this.innerElementSize=this.elementsPerThread[0],r&&this.variableNames.push("bias"),i&&this.variableNames.push("preluActivationWeights")),this.sequentialAccessByThreads=o,this.addBias=r,this.activation=s,this.hasPreluActivationWeights=i,this.tileAOuter=this.workgroupSize[1]*this.elementsPerThread[1],this.tileBOuter=this.workgroupSize[0]*this.elementsPerThread[0],this.tileInner=Math.max(this.workgroupSize[0]*this.innerElementSize,this.workgroupSize[1]),this.fitAOuter=t%this.tileAOuter===0,this.fitBOuter=a%this.tileBOuter===0,this.fitInner=n%this.tileInner===0,this.shaderKey=`conv2DMM_${this.elementsPerThread}_${this.activation}}_${this.fitAOuter}_${this.fitBOuter}_${this.fitInner}_${this.isVec4}_${this.innerElementSize}_${this.isChannelsLast}_${this.sequentialAccessByThreads}`}getUserCode(){let e=this.isVec4?s0(this.elementsPerThread,this.workgroupSize,!this.isChannelsLast,this.tileInner):i0(this.elementsPerThread,this.workgroupSize,!this.isChannelsLast,this.tileInner,!1,null,this.sequentialAccessByThreads),t=this.isVec4?[this.innerElementSize,4,4]:[1,1,1];return`
|
|
${Ape(this.isChannelsLast,this.fitAOuter,this.fitBOuter,this.fitInner,this.addBias,this.activation,this.hasPreluActivationWeights,t[0],t[1],t[2])}
|
|
${e}
|
|
`}},vpe=class{constructor(e,t=!1,a=null,n=!1){this.variableNames=["x","W"],this.uniforms="filterDims: vec2<i32>, pads: vec2<i32>, strides: vec2<i32>, dilations: vec2<i32>,",this.workgroupSize=[4,4,8],this.outputShape=e.outShape,this.isChannelsLast=e.dataFormat==="channelsLast",this.dispatchLayout=this.isChannelsLast?{x:[2],y:[1],z:[0,3]}:{x:[3],y:[2],z:[0,1]},this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.addBias=t,this.activation=a,this.hasPreluActivationWeights=n,t&&this.variableNames.push("bias"),n&&this.variableNames.push("preluActivationWeights"),this.shaderKey=`conv2dnaive_${this.activation}_${this.isChannelsLast}`}getUserCode(){return`
|
|
${Mr(this.activation,this.hasPreluActivationWeights,!1,4)}
|
|
fn readInp(batch : i32, row : i32, col : i32, chan : i32) -> f32{
|
|
let coords = vec4<i32>(batch, row, col, chan);
|
|
if (coordsInBounds4D(coords, uniforms.xShape)) {
|
|
return getX(batch, row, col, chan);
|
|
} else {
|
|
return 0.0;
|
|
}
|
|
}
|
|
fn readFilt(row : i32, col : i32, xChannel : i32, outChannel : i32) -> f32{
|
|
let coords = vec4<i32>(row, col, xChannel, outChannel);
|
|
if(coordsInBounds4D(coords, uniforms.wShape)) {
|
|
return getW(row, col, xChannel, outChannel);
|
|
} else {
|
|
return 0.0;
|
|
}
|
|
}
|
|
fn writeResult(batch : i32, row : i32, col : i32, chan : i32, valueIn : f32) {
|
|
let coords = ${this.isChannelsLast?"vec4<i32>(batch, row, col, chan);":"vec4<i32>(batch, chan, row, col);"}
|
|
if (coordsInBounds4D(coords, uniforms.outShape)) {
|
|
var value = valueIn;
|
|
${Yo(this.addBias,this.activation)}
|
|
setOutputAtCoords(coords.x, coords.y, coords.z, coords.w, value);
|
|
}
|
|
}
|
|
${de("index")} {
|
|
let coords = getOutputCoords();
|
|
let batch = coords[0];
|
|
let outChannel = ${this.isChannelsLast?"coords[3];":"coords[1];"}
|
|
let outRow = ${this.isChannelsLast?"coords[1];":"coords[2];"}
|
|
let outCol = ${this.isChannelsLast?"coords[2];":"coords[3];"}
|
|
var acc : f32 = 0.0;
|
|
for (var row = 0; row < uniforms.filterDims[0]; row = row + 1) {
|
|
for (var col = 0; col < uniforms.filterDims[1]; col = col + 1) {
|
|
let xRow = outRow * uniforms.strides[0] + uniforms.dilations[0] * row - uniforms.pads[0];
|
|
let xCol = outCol * uniforms.strides[1] + uniforms.dilations[1] * col - uniforms.pads[1];
|
|
for (var xChannel = 0; xChannel < ${this.isChannelsLast?"uniforms.xShape[3];":"uniforms.xShape[1];"} xChannel = xChannel + 1) {
|
|
${this.isChannelsLast?"let v = readInp(batch, xRow, xCol, xChannel);":"let v = readInp(batch, xChannel, xRow, xCol);"}
|
|
let f = readFilt(row, col, xChannel, outChannel);
|
|
acc = acc + v * f;
|
|
}
|
|
}
|
|
}
|
|
writeResult(batch, outRow, outCol, outChannel, acc);
|
|
}
|
|
`}},wpe=class{constructor(e,t){this.variableNames=["x"],this.uniforms=`pads : vec2<i32>, strides : vec2<i32>, dilations : vec2<i32>, outWidth : i32, itemsPerBlockRow : i32,
|
|
inChannels : i32,`,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.isChannelsLast=t,this.shaderKey=`im2col_${this.isChannelsLast}`}getUserCode(){let e=this.isChannelsLast?1:2,t=this.isChannelsLast?2:3,a=this.isChannelsLast?"coords[1]":"coords[2]",n=this.isChannelsLast?"coords[2]":"coords[1]",r=this.isChannelsLast?"getX(batch, xRow, xCol, ch)":"getX(batch, ch, xRow, xCol)";return`
|
|
${de("index")} {
|
|
let coords = getCoordsFromIndex(index);
|
|
if(index < uniforms.size) {
|
|
let batch = coords[0];
|
|
let row = ${a};
|
|
let col = ${n};
|
|
let offsetY = (row / uniforms.outWidth) * uniforms.strides[0] - uniforms.pads[0];
|
|
let xRow = offsetY + uniforms.dilations[0] * (col / uniforms.itemsPerBlockRow);
|
|
var value = 0.0;
|
|
if(xRow < uniforms.xShape[${e}] && xRow >= 0) {
|
|
let offsetX = (row % uniforms.outWidth) * uniforms.strides[1] -
|
|
uniforms.pads[1];
|
|
let xCol = offsetX + uniforms.dilations[1] * ((col %
|
|
uniforms.itemsPerBlockRow) / uniforms.inChannels);
|
|
let ch = col % uniforms.inChannels;
|
|
if(xCol < uniforms.xShape[${t}] && xCol >= 0) {
|
|
value = ${r};
|
|
}
|
|
}
|
|
setOutputAtIndex(index, value);
|
|
}
|
|
}
|
|
`}};function Ah(e,t){let a=e.length;return a>=3?t?[...e.slice(0,-3),e[a-3]*e[a-2],e[a-1]]:[...e.slice(0,-3),e[a-3],e[a-2]*e[a-1]]:!t&&a===1&&e[0]>1?[e[0],1]:null}function kpe({x:e,filter:t,convInfo:a,backend:n,bias:r=null,preluActivationWeights:s=null,leakyreluAlpha:i=0,activation:o=null}){let l=a.dataFormat==="channelsLast",u=!l,p=!1,c=l&&a.filterHeight===a.inHeight&&a.filterWidth===a.inWidth&&a.padInfo.type==="VALID",d=[],h,f;if(c){let y=a.inHeight*a.inWidth*a.inChannels;h=we({inputs:{x:e},backend:n,attrs:{shape:[1,a.batchSize,y]}}),f=we({inputs:{x:t},backend:n,attrs:{shape:[1,y,a.outChannels]}})}else h=we({inputs:{x:e},backend:n,attrs:{shape:l?[a.batchSize,a.inHeight*a.inWidth,a.inChannels]:[a.batchSize,a.inChannels,a.inHeight*a.inWidth]}}),f=we({inputs:{x:t},backend:n,attrs:{shape:[1,a.inChannels,a.outChannels]}});if(d.push(h),d.push(f),s!=null){let y=Ah(s.shape,l);y!=null&&(s=we({inputs:{x:s},backend:n,attrs:{shape:y}}),d.push(s))}if(r!=null){let y=Ah(r.shape,l);y!=null&&(r=we({inputs:{x:r},backend:n,attrs:{shape:y}}),d.push(r))}let m=o0({a:l?h:f,b:l?f:h,transposeA:u,transposeB:p,backend:n,bias:r,activation:o,preluActivationWeights:s,leakyreluAlpha:i}),g=we({inputs:{x:m},backend:n,attrs:{shape:a.outShape}});d.push(m);for(let y of d)n.disposeData(y.dataId);return g}function Ipe({x:e,filter:t,convInfo:a,backend:n,bias:r=null,preluActivationWeights:s=null,leakyreluAlpha:i=0,activation:o=null}){let{filterWidth:l,filterHeight:u,inChannels:p,strideWidth:c,strideHeight:d,padInfo:h,outWidth:f,outHeight:m,dilationWidth:g,dilationHeight:y,dataFormat:x}=a,A=x==="channelsLast",b=l*u*p,w=m*f,C=A?[a.batchSize,w,b]:[a.batchSize,b,w],T=new wpe(C,A),N=[{type:"int32",data:[h.top,h.left]},{type:"int32",data:[d,c]},{type:"int32",data:[y,g]},{type:"int32",data:[f]},{type:"int32",data:[p*l]},{type:"int32",data:[p]}],M=n.runWebGPUProgram(T,[e],e.dtype,N),F=[];F.push(M);let E=we({inputs:{x:t},backend:n,attrs:{shape:[1,b,-1]}});if(F.push(E),s!=null){let O=Ah(s.shape,A);O!=null&&(s=we({inputs:{x:s},backend:n,attrs:{shape:O}}),F.push(s))}if(r!=null){let O=Ah(r.shape,A);O!=null&&(r=we({inputs:{x:r},backend:n,attrs:{shape:O}}),F.push(r))}let I=o0({a:A?M:E,b:A?E:M,transposeA:!A,transposeB:!1,backend:n,bias:r,activation:o,preluActivationWeights:s,leakyreluAlpha:i}),_=we({inputs:{x:I},backend:n,attrs:{shape:a.outShape}});F.push(I);for(let O of F)n.disposeData(O.dataId);return _}function sk({x:e,filter:t,convInfo:a,backend:n,bias:r=null,preluActivationWeights:s=null,leakyreluAlpha:i=0,activation:o=null}){let l=r!=null,u=s!=null,p=a.dataFormat==="channelsLast",c=p&&a.filterHeight===a.inHeight&&a.filterWidth===a.inWidth&&a.padInfo.type==="VALID",d=W().getBool("WEBGPU_USE_NAIVE_CONV2D_DEBUG");if(!d&&(c||a.filterHeight===1&&a.filterWidth===1&&a.dilationHeight===1&&a.dilationWidth===1&&a.strideHeight===1&&a.strideWidth===1&&(a.padInfo.type==="SAME"||a.padInfo.type==="VALID")))return kpe({x:e,filter:t,convInfo:a,backend:n,bias:r,activation:o,preluActivationWeights:s,leakyreluAlpha:i});let h=W().getNumber("WEBGPU_THRESHOLD_TO_INCREASE_WORKGROUPS_FOR_MATMUL"),f=h>0?h:n.thresholdToIncreaseWorkgroups,m=a.batchSize*Math.ceil(a.outHeight*a.outWidth/32)*Math.ceil(a.outChannels/32);if(W().getBool("WEBGPU_CONV_SEPARATE_IM2COL_SHADER")||m<=f)return Ipe({x:e,filter:t,convInfo:a,backend:n,bias:r,preluActivationWeights:s,leakyreluAlpha:i,activation:o});let g,y=[a.padInfo.top,a.padInfo.left],x=[{type:"int32",data:[a.filterHeight,a.filterWidth]},{type:"int32",data:[...y]},{type:"int32",data:[a.strideHeight,a.strideWidth]},{type:"int32",data:[a.dilationHeight,a.dilationWidth]}];if(d)g=new vpe(a,l,o,u);else{let C=p?a.outHeight*a.outWidth:a.outChannels,T=p?a.outChannels:a.outHeight*a.outWidth,N=a.filterHeight*a.filterWidth*a.inChannels;x.push({type:"int32",data:[C]},{type:"int32",data:[T]},{type:"int32",data:[N]});let M=n.adapterInfo.isIntel();g=new bpe(a,C,T,N,l,o,u,M)}let A=[],b=[e,t];l&&(!p&&r.shape.length===1&&(r=we({inputs:{x:r},backend:n,attrs:{shape:[r.shape[0],1,1]}}),A.push(r)),b.push(r)),u&&(!p&&s.shape.length===1&&(s=we({inputs:{x:s},backend:n,attrs:{shape:[s.shape[0],1,1]}}),A.push(s)),b.push(s)),o==="leakyrelu"&&(x.push({type:"float32",data:[i]}),g.uniforms+=" alpha : f32,");let w=n.runWebGPUProgram(g,b,e.dtype,x);for(let C of A)n.disposeData(C.dataId);return w}function Spe(e){let{inputs:t,attrs:a,backend:n}=e,{x:r,filter:s}=t,{strides:i,pad:o,dataFormat:l,dilations:u,dimRoundingMode:p}=a,c=S.convertConv2DDataFormat(l),d=S.computeConv2DInfo(r.shape,s.shape,i,u,o,p,!1,c);return sk({x:r,filter:s,convInfo:d,backend:n})}var Cpe={kernelName:fi,backendName:"webgpu",kernelFunc:Spe},Tpe=class{constructor(e){this.variableNames=["dy","W"],this.uniforms="filterDims : vec2<i32>, pads : vec2<i32>, strides : vec2<i32>, outBackprop : vec4<i32>,",this.workgroupSize=[64,1,1],this.size=!1,this.isVec4=!1,this.workPerThread=1,this.outputShape=e.inShape,this.isChannelsLast=e.dataFormat==="channelsLast",this.isVec4=this.isChannelsLast&&e.outChannels%4===0&&e.inChannels%4===0,this.isVec4?(this.workPerThread=2,this.outputComponent=4,this.workgroupSize=[4,4,4],this.dispatchLayout={x:[3],y:[2],z:[0,1]},this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize,[4,this.workPerThread,1])):(this.size=!0,this.workPerThread=1,this.workgroupSize=[64,1,1],this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize)),this.shaderKey=`conv2DDerInput_${this.isChannelsLast}_${this.isVec4}_${this.workPerThread}`}getUserCode(){let e=this.isChannelsLast?1:2,t=this.isChannelsLast?2:3,a=this.isChannelsLast?3:1,n=`
|
|
${de()} {
|
|
let batch = i32(globalId.z) / uniforms.outShape[1];
|
|
let r = i32(globalId.z) % uniforms.outShape[1];
|
|
let c = i32(globalId.y) * ${this.workPerThread};
|
|
let d1 = i32(globalId.x) * 4;
|
|
|
|
let dyCorner = vec2<i32>(r, c) - uniforms.pads;
|
|
|
|
// Convolve dy(?, ?, d2) with w(:, :, d1, d2) to compute dx(xR, xC, d1).
|
|
// ? = to be determined. : = across all values in that axis.
|
|
var dotProd: array<vec4<f32>, ${this.workPerThread}>;
|
|
for (var i = 0; i < ${this.workPerThread}; i++) {
|
|
dotProd[i] = vec4<f32>(0.0);
|
|
}
|
|
for (var wR = 0; wR < uniforms.filterDims.x; wR = wR + 1) {
|
|
let dyR = f32(dyCorner.x + wR) / f32(uniforms.strides.x);
|
|
let wRPerm = uniforms.filterDims.x - 1 - wR;
|
|
if (dyR < 0.0 || dyR >= f32(uniforms.outBackprop[1]) ||
|
|
fract(dyR) > 0.0) {
|
|
continue;
|
|
}
|
|
let idyR = i32(dyR);
|
|
|
|
for (var wC = 0; wC < uniforms.filterDims.y; wC = wC + 1) {
|
|
let dyC = f32(dyCorner.y + wC) / f32(uniforms.strides.y);
|
|
let dyC2 = f32(dyCorner.y + 1 + wC) / f32(uniforms.strides.y);
|
|
let wCPerm = uniforms.filterDims.y - 1 - wC;
|
|
var bDyCVal = true;
|
|
var bDyCVal2 = true;
|
|
if (dyC < 0.0 || dyC >= f32(uniforms.outBackprop[2]) ||
|
|
fract(dyC) > 0.0) {
|
|
bDyCVal = false;
|
|
}
|
|
if (dyC2 < 0.0 || dyC2 >= f32(uniforms.outBackprop[2]) ||
|
|
fract(dyC2) > 0.0) {
|
|
bDyCVal2 = false;
|
|
}
|
|
|
|
let idyC = i32(dyC);
|
|
let idyC2 = i32(dyC2);
|
|
if (bDyCVal && bDyCVal2) {
|
|
let d2Length = uniforms.outBackprop[3];
|
|
for (var d2 = 0; d2 < d2Length; d2 = d2 + 4) {
|
|
let wValue0 = getW(wRPerm, wCPerm, d1, d2);
|
|
let wValue1 = getW(wRPerm, wCPerm, d1 + 1, d2);
|
|
let wValue2 = getW(wRPerm, wCPerm, d1 + 2, d2);
|
|
let wValue3 = getW(wRPerm, wCPerm, d1 + 3, d2);
|
|
var xValue = getDy(batch, idyR, idyC, d2);
|
|
let tmpval = vec4<f32>(dot(xValue, wValue0),
|
|
dot(xValue, wValue1),
|
|
dot(xValue, wValue2),
|
|
dot(xValue, wValue3));
|
|
dotProd[0] = dotProd[0] + tmpval;
|
|
xValue = getDy(batch, idyR, idyC2, d2);
|
|
dotProd[1] = dotProd[1] + vec4<f32>(dot(xValue, wValue0),
|
|
dot(xValue, wValue1),
|
|
dot(xValue, wValue2),
|
|
dot(xValue, wValue3));
|
|
}
|
|
} else if (bDyCVal) {
|
|
let d2Length = uniforms.outBackprop[3];
|
|
for (var d2 = 0; d2 < d2Length; d2 = d2 + 4) {
|
|
let wValue0 = getW(wRPerm, wCPerm, d1, d2);
|
|
let wValue1 = getW(wRPerm, wCPerm, d1 + 1, d2);
|
|
let wValue2 = getW(wRPerm, wCPerm, d1 + 2, d2);
|
|
let wValue3 = getW(wRPerm, wCPerm, d1 + 3, d2);
|
|
var xValue = getDy(batch, idyR, idyC, d2);
|
|
let tmpval = vec4<f32>(dot(xValue, wValue0),
|
|
dot(xValue, wValue1),
|
|
dot(xValue, wValue2),
|
|
dot(xValue, wValue3));
|
|
dotProd[0] = dotProd[0] + tmpval;
|
|
}
|
|
} else if (bDyCVal2) {
|
|
let d2Length = uniforms.outBackprop[3];
|
|
for (var d2 = 0; d2 < d2Length; d2 = d2 + 4) {
|
|
let wValue0 = getW(wRPerm, wCPerm, d1, d2);
|
|
let wValue1 = getW(wRPerm, wCPerm, d1 + 1, d2);
|
|
let wValue2 = getW(wRPerm, wCPerm, d1 + 2, d2);
|
|
let wValue3 = getW(wRPerm, wCPerm, d1 + 3, d2);
|
|
var xValue = getDy(batch, idyR, idyC2, d2);
|
|
let tmpval = vec4<f32>(dot(xValue, wValue0),
|
|
dot(xValue, wValue1),
|
|
dot(xValue, wValue2),
|
|
dot(xValue, wValue3));
|
|
dotProd[1] = dotProd[1] + tmpval;
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
for (var i = 0; i < ${this.workPerThread}; i = i + 1) {
|
|
let coords = vec4<i32>(batch, r, c + i, d1);
|
|
if (coordsInBounds4D(coords, uniforms.outShape)) {
|
|
setOutputAtCoords(coords[0], coords[1], coords[2], coords[3], dotProd[i]);
|
|
}
|
|
}
|
|
}
|
|
`;return this.isVec4?`
|
|
${n}
|
|
`:`
|
|
${de("index")} {
|
|
if(index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
let batch = coords[0];
|
|
let d1 = coords[${a}];
|
|
|
|
let dyCorner = vec2<i32>(coords[${e}], coords[${t}]) - uniforms.pads;
|
|
let dyRCorner = dyCorner.x;
|
|
let dyCCorner = dyCorner.y;
|
|
|
|
// Convolve dy(?, ?, d2) with w(:, :, d1, d2) to compute dx(xR, xC, d1).
|
|
// ? = to be determined. : = across all values in that axis.
|
|
var dotProd = 0.0;
|
|
for (var wR = 0; wR < uniforms.filterDims.x; wR = wR + 1) {
|
|
let dyR = (f32(dyRCorner) + f32(wR)) / f32(uniforms.strides.x);
|
|
let wRPerm = uniforms.filterDims.x - 1 - wR;
|
|
if (dyR < 0.0 || dyR >= f32(uniforms.outBackprop[1]) || fract(dyR) > 0.0 ||
|
|
wRPerm < 0) {
|
|
continue;
|
|
}
|
|
let idyR = i32(dyR);
|
|
|
|
for (var wC = 0; wC < uniforms.filterDims.y; wC = wC + 1) {
|
|
let dyC = (f32(dyCCorner) + f32(wC)) / f32(uniforms.strides.y);
|
|
let wCPerm = uniforms.filterDims.y - 1 - wC;
|
|
if (dyC < 0.0 || dyC >= f32(uniforms.outBackprop[2]) ||
|
|
fract(dyC) > 0.0 || wCPerm < 0) {
|
|
continue;
|
|
}
|
|
let idyC = i32(dyC);
|
|
|
|
for (var d2 = 0; d2 < uniforms.outBackprop[3]; d2 = d2 + 1) {
|
|
let xValue = ${this.isChannelsLast?"getDy(batch, idyR, idyC, d2)":"getDy(batch, d2, idyR, idyC)"};
|
|
let wValue = getW(wRPerm, wCPerm, d1, d2);
|
|
dotProd = dotProd + xValue * wValue;
|
|
}
|
|
}
|
|
}
|
|
setOutputAtIndex(index, dotProd);
|
|
}
|
|
}
|
|
`}},Npe=class{constructor(e){this.variableNames=["x","dy"],this.uniforms="pads : vec2<i32>, strides : vec2<i32>, batchSize : i32, outHeight : i32, outWidth : i32, inHeight : i32, inWidth : i32,",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e.filterShape,this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.isChannelsLast=e.dataFormat==="channelsLast",this.shaderKey=`conv2DDerFilter_${this.isChannelsLast}`}getUserCode(){return`
|
|
${de("index")} {
|
|
if(index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
let wR = coords[0];
|
|
let wC = coords[1];
|
|
let d1 = coords[2];
|
|
let d2 = coords[3];
|
|
|
|
// Convolve x(?, ?, d1) with dy(:, :, d2) to get dw(wR, wC, d1, d2).
|
|
// ? = to be determined. : = across all values in that axis.
|
|
var dotProd = 0.0;
|
|
for (var b = 0; b < uniforms.batchSize; b = b + 1) {
|
|
for (var yR = 0; yR < uniforms.outHeight; yR = yR + 1) {
|
|
let xR = wR + yR * uniforms.strides[0] - uniforms.pads[0];
|
|
if (xR < 0 || xR >= uniforms.inHeight) {
|
|
continue;
|
|
}
|
|
|
|
for (var yC = 0; yC < uniforms.outWidth; yC = yC + 1) {
|
|
let xC = wC + yC * uniforms.strides[1] - uniforms.pads[1];
|
|
|
|
if (xC < 0 || xC >= uniforms.inWidth) {
|
|
continue;
|
|
}
|
|
|
|
if (${this.isChannelsLast}) {
|
|
let dyValue = getDy(b, yR, yC, d2);
|
|
let xValue = getX(b, xR, xC, d1);
|
|
dotProd = dotProd + xValue * dyValue;
|
|
} else {
|
|
let dyValue = getDy(b, d2, yR, yC);
|
|
let xValue = getX(b, d1, xR, xC);
|
|
dotProd = dotProd + xValue * dyValue;
|
|
}
|
|
}
|
|
}
|
|
}
|
|
setOutputAtIndex(index, dotProd);
|
|
}
|
|
}
|
|
`}},Rpe=class{constructor(e){this.variableNames=["x","dy"],this.uniforms=`pads : vec3<i32>, strides : vec3<i32>, batchSize : i32, outDepth : i32,
|
|
outHeight : i32, outWidth : i32, inDepth : i32, inHeight : i32, inWidth : i32,`,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e.filterShape,this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="conv3DDerFilter"}getUserCode(){return`
|
|
${de("index")} {
|
|
if(index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
let wF = coords.x;
|
|
let wR = coords.y;
|
|
let wC = coords.z;
|
|
let d1 = coords.w;
|
|
let d2 = coords.u;
|
|
|
|
var dotProd = 0.0;
|
|
for (var b = 0; b < uniforms.batchSize; b++) {
|
|
for (var yF = 0; yF < uniforms.outDepth; yF++) {
|
|
let xF = wF + yF * uniforms.strides[0] - uniforms.pads[0];
|
|
if (xF < 0 || xF >= uniforms.inDepth) {
|
|
continue;
|
|
}
|
|
|
|
for (var yR = 0; yR < uniforms.outHeight; yR++) {
|
|
let xR = wR + yR * uniforms.strides[1] - uniforms.pads[1];
|
|
if (xR < 0 || xR >= uniforms.inHeight) {
|
|
continue;
|
|
}
|
|
|
|
for (var yC = 0; yC < uniforms.outWidth; yC++) {
|
|
let xC = wC + yC * uniforms.strides[2] - uniforms.pads[2];
|
|
if (xC < 0 || xC >= uniforms.inWidth) {
|
|
continue;
|
|
}
|
|
|
|
let dyValue = getDy(b, yF, yR, yC, d2);
|
|
let xValue = getX(b, xF, xR, xC, d1);
|
|
dotProd += xValue * dyValue;
|
|
}
|
|
}
|
|
}
|
|
}
|
|
setOutputAtIndex(index, dotProd);
|
|
}
|
|
}
|
|
`}},Epe=class{constructor(e){this.variableNames=["dy","W"],this.uniforms=`filterDims : vec3<i32>, pads : vec3<i32>, strides : vec3<i32>,
|
|
outDepth : i32, outHeight : i32, outWidth : i32, outChannels : i32,`,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e.inShape,this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="conv3DDerInput"}getUserCode(){return`
|
|
${de("index")} {
|
|
if(index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
let batch = coords.x;
|
|
let d1 = coords.u;
|
|
|
|
let dyCorner = vec3<i32>(coords.y, coords.z, coords.w) - uniforms.pads;
|
|
let dyFCorner = dyCorner.x;
|
|
let dyRCorner = dyCorner.y;
|
|
let dyCCorner = dyCorner.z;
|
|
|
|
var dotProd = 0.0;
|
|
for (var wF = 0; wF < uniforms.filterDims[0]; wF++) {
|
|
let dyF = f32(dyFCorner + wF) / f32(uniforms.strides[0]);
|
|
if (dyF < 0.0 || dyF >= f32(uniforms.outDepth) || fract(dyF) > 0.0) {
|
|
continue;
|
|
}
|
|
let idyF = i32(dyF);
|
|
|
|
let wFPerm = uniforms.filterDims[0] - 1 - wF;
|
|
|
|
for (var wR = 0; wR < uniforms.filterDims[1]; wR++) {
|
|
let dyR = f32(dyRCorner + wR) / f32(uniforms.strides[1]);
|
|
|
|
if (dyR < 0.0 || dyR >= f32(uniforms.outHeight) || fract(dyR) > 0.0) {
|
|
continue;
|
|
}
|
|
let idyR = i32(dyR);
|
|
|
|
let wRPerm = uniforms.filterDims[1] - 1 - wR;
|
|
|
|
for (var wC = 0; wC < uniforms.filterDims[2]; wC++) {
|
|
let dyC = f32(dyCCorner + wC) / f32(uniforms.strides[2]);
|
|
|
|
if (dyC < 0.0 || dyC >= f32(uniforms.outWidth) || fract(dyC) > 0.0) {
|
|
continue;
|
|
}
|
|
let idyC = i32(dyC);
|
|
|
|
let wCPerm = uniforms.filterDims[2] - 1 - wC;
|
|
|
|
for (var d2 = 0; d2 < uniforms.outChannels; d2++) {
|
|
let xValue = getDy(batch, idyF, idyR, idyC, d2);
|
|
let wValue = getW(wFPerm, wRPerm, wCPerm, d1, d2);
|
|
dotProd += xValue * wValue;
|
|
}
|
|
}
|
|
}
|
|
}
|
|
setOutputAtIndex(index, dotProd);
|
|
}
|
|
}
|
|
`}};function Mpe(e){let{inputs:t,backend:a,attrs:n}=e,{x:r,dy:s}=t,{strides:i,pad:o,dataFormat:l,dimRoundingMode:u,filterShape:p}=n,c=S.convertConv2DDataFormat(l),d=S.computeConv2DInfo(r.shape,p,i,1,o,u,!1,c),h=new Npe(d),f=[{type:"int32",data:[d.padInfo.top,d.padInfo.left]},{type:"int32",data:[d.strideHeight,d.strideWidth]},{type:"int32",data:[d.batchSize]},{type:"int32",data:[d.outHeight]},{type:"int32",data:[d.outWidth]},{type:"int32",data:[d.inHeight]},{type:"int32",data:[d.inWidth]}];return a.runWebGPUProgram(h,[r,s],r.dtype,f)}var Fpe={kernelName:np,backendName:"webgpu",kernelFunc:Mpe};function _pe(e=4){let t=n=>{switch(n){case 1:return"return W[getIndexFromCoords4D(coord, uniforms.wShape)];";case 4:return`
|
|
let coord1 = vec4<i32>(coordX, coordY, col + 1, rowInner);
|
|
let coord2 = vec4<i32>(coordX, coordY, col + 2, rowInner);
|
|
let coord3 = vec4<i32>(coordX, coordY, col + 3, rowInner);
|
|
let v0 = W[getIndexFromCoords4D(coord, uniforms.wShape)];
|
|
let v1 = W[getIndexFromCoords4D(coord1, uniforms.wShape)];
|
|
let v2 = W[getIndexFromCoords4D(coord2, uniforms.wShape)];
|
|
let v3 = W[getIndexFromCoords4D(coord3, uniforms.wShape)];
|
|
return vec4<f32>(v0, v1, v2, v3);
|
|
`;default:throw new Error(`innerElementSize ${n} is not supported.`)}},a=`if (row < uniforms.dimAOuter && col < uniforms.dimInner) {
|
|
${`
|
|
let outRow = row / uniforms.outShape[2];
|
|
let outCol = row % uniforms.outShape[2];
|
|
|
|
let WRow = col / (uniforms.filterDims[1] * uniforms.outBackprop[3]);
|
|
let WCol = col / uniforms.outBackprop[3] % uniforms.filterDims[1];
|
|
let xR = f32(outRow - uniforms.pads[0] + WRow) / f32(uniforms.strides[0]);
|
|
let xC = f32(outCol - uniforms.pads[1] + WCol) / f32(uniforms.strides[1]);
|
|
if (xR < 0.0 || xR >= f32(uniforms.outBackprop[1]) || fract(xR) > 0.0) {
|
|
return ${Xe(e)}(0.0);
|
|
}
|
|
if (xC < 0.0 || xC >= f32(uniforms.outBackprop[2]) || fract(xC) > 0.0) {
|
|
return ${Xe(e)}(0.0);
|
|
}
|
|
let coord = vec4<i32>(
|
|
batch,
|
|
i32(xR),
|
|
i32(xC),
|
|
col % uniforms.outBackprop[3]);
|
|
return x[getIndexFromCoords4D(coord, uniforms.xShape)/${e}];`}
|
|
}
|
|
return ${Xe(e)}(0.0);`;return`
|
|
fn mm_readA(batch: i32, row : i32, colIn : i32) -> ${Xe(e)} {
|
|
let col = colIn * ${e};
|
|
${a}
|
|
}
|
|
|
|
fn mm_readB(batch: i32, row : i32, colIn : i32) -> ${Xe(e)} {
|
|
let col = colIn * ${e};
|
|
let coordX = uniforms.filterDims.x - 1 -
|
|
row / (uniforms.filterDims[1] * uniforms.outBackprop[3]);
|
|
let coordY = uniforms.filterDims.y - 1 -
|
|
(row / uniforms.outBackprop[3]) % uniforms.filterDims[1];
|
|
if (row < uniforms.dimInner && col < uniforms.dimBOuter &&
|
|
coordX >= 0 && coordY >= 0) {
|
|
let rowInner = row % uniforms.outBackprop[3];
|
|
let coord = vec4<i32>(coordX, coordY, col, rowInner);
|
|
${t(e)}
|
|
}
|
|
return ${Xe(e)}(0.0);
|
|
}
|
|
|
|
fn mm_write(batch: i32, row : i32, colIn : i32, valueInput : ${Xe(e)}) {
|
|
let col = colIn * ${e};
|
|
if (row < uniforms.dimAOuter && (col + ${e-1}) < uniforms.dimBOuter) {
|
|
var value = valueInput;
|
|
let outCoord = vec4<i32>(
|
|
batch,
|
|
row / uniforms.outShape[2],
|
|
row % uniforms.outShape[2],
|
|
col);
|
|
result[getIndexFromCoords4D(outCoord, uniforms.outShape)/${e}] = value;
|
|
}
|
|
}`}var $pe=class{constructor(e){this.variableNames=["x","W"],this.uniforms="filterDims : vec2<i32>, pads : vec2<i32>, strides : vec2<i32>, outBackprop : vec4<i32>, dimAOuter : i32, dimBOuter : i32, dimInner : i32,",this.outputShape=e.inShape,v.assert(e.dataFormat==="channelsLast",()=>"TODO: NCHW is unimplemented"),this.isVec4=e.inChannels%4===0&&e.outChannels%4===0,this.dispatchLayout={x:[3],y:[1,2],z:[0]},this.workgroupSize=Rg(this.dispatchLayout,this.outputShape,this.isVec4),this.elementsPerThread=Eg(this.dispatchLayout,this.outputShape,this.isVec4),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize,this.elementsPerThread),this.isVec4&&(this.outputComponent=4,this.variableComponents=[4,1]),this.shaderKey=`conv2DDerInputMM_${this.isVec4}_${this.elementsPerThread}`}getUserCode(){let e=this.isVec4?s0(this.elementsPerThread,this.workgroupSize):i0(this.elementsPerThread,this.workgroupSize);return`
|
|
${_pe(this.isVec4?4:1)}
|
|
${e}
|
|
`}};function Dpe(e){let{inputs:t,backend:a,attrs:n}=e,{dy:r,filter:s}=t,{inputShape:i,strides:o,pad:l,dataFormat:u,dimRoundingMode:p}=n,c=S.convertConv2DDataFormat(u),d=S.computeConv2DInfo(i,s.shape,o,1,l,p,!1,c),h=[{type:"int32",data:[d.filterHeight,d.filterWidth]},{type:"int32",data:[d.filterHeight-1-d.padInfo.top,d.filterWidth-1-d.padInfo.left]},{type:"int32",data:[d.strideHeight,d.strideWidth]},{type:"int32",data:[d.batchSize,d.outHeight,d.outWidth,d.outChannels]}],f;if(W().getBool("WEBGPU_USE_NAIVE_CONV2D_TRANSPOSE")||d.dataFormat!=="channelsLast")f=new Tpe(d);else{f=new $pe(d);let m=d.inHeight*d.inWidth,g=d.inChannels,y=d.filterHeight*d.filterWidth*d.outChannels;h.push({type:"uint32",data:[m]},{type:"uint32",data:[g]},{type:"uint32",data:[y]})}return a.runWebGPUProgram(f,[r,s],"float32",h)}var Ppe={kernelName:mi,backendName:"webgpu",kernelFunc:Dpe},Ope=class{constructor(e){this.variableNames=["x","W"],this.uniforms="filterDims: vec3<i32>, pads: vec3<i32>, strides: vec3<i32>, dilations: vec3<i32>,",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e.outShape,this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="conv3dnaive"}getUserCode(){return`
|
|
${de("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getOutputCoords();
|
|
let batch = coords.x;
|
|
let d2 = coords.u;
|
|
|
|
let xFRCCorner = vec3<i32>(coords.y, coords.z, coords.w) * uniforms.strides - uniforms.pads;
|
|
let xFCorner = xFRCCorner.x;
|
|
let xRCorner = xFRCCorner.y;
|
|
let xCCorner = xFRCCorner.z;
|
|
|
|
let inputDepthNearestVec4 = (uniforms.xShape.u / 4) * 4;
|
|
let inputDepthVec4Remainder = uniforms.xShape.u % 4;
|
|
|
|
var dotProd = 0.0;
|
|
for (var wF = 0; wF < uniforms.filterDims[0]; wF++) {
|
|
let xF = xFCorner + wF * uniforms.dilations[0];
|
|
if (xF < 0 || xF >= uniforms.xShape.y) {
|
|
continue;
|
|
}
|
|
|
|
for (var wR = 0; wR < uniforms.filterDims[1]; wR++) {
|
|
let xR = xRCorner + wR * uniforms.dilations[1];
|
|
if (xR < 0 || xR >= uniforms.xShape.z) {
|
|
continue;
|
|
}
|
|
|
|
for (var wC = 0; wC < uniforms.filterDims[2]; wC++) {
|
|
let xC = xCCorner + wC * uniforms.dilations[2];
|
|
if (xC < 0 || xC >= uniforms.xShape.w) {
|
|
continue;
|
|
}
|
|
|
|
for (var d1 = 0; d1 < inputDepthNearestVec4; d1 += 4) {
|
|
let xValues = vec4<f32>(
|
|
getX(batch, xF, xR, xC, d1),
|
|
getX(batch, xF, xR, xC, d1 + 1),
|
|
getX(batch, xF, xR, xC, d1 + 2),
|
|
getX(batch, xF, xR, xC, d1 + 3)
|
|
);
|
|
let wValues = vec4<f32>(
|
|
getW(wF, wR, wC, d1, d2),
|
|
getW(wF, wR, wC, d1 + 1, d2),
|
|
getW(wF, wR, wC, d1 + 2, d2),
|
|
getW(wF, wR, wC, d1 + 3, d2)
|
|
);
|
|
|
|
dotProd += dot(xValues, wValues);
|
|
}
|
|
|
|
if (inputDepthVec4Remainder == 1) {
|
|
dotProd += getX(batch, xF, xR, xC, inputDepthNearestVec4) *
|
|
getW(wF, wR, wC, inputDepthNearestVec4, d2);
|
|
} else if (inputDepthVec4Remainder == 2) {
|
|
let xValues = vec2<f32>(
|
|
getX(batch, xF, xR, xC, inputDepthNearestVec4),
|
|
getX(batch, xF, xR, xC, inputDepthNearestVec4 + 1)
|
|
);
|
|
let wValues = vec2<f32>(
|
|
getW(wF, wR, wC, inputDepthNearestVec4, d2),
|
|
getW(wF, wR, wC, inputDepthNearestVec4 + 1, d2)
|
|
);
|
|
dotProd += dot(xValues, wValues);
|
|
} else if (inputDepthVec4Remainder == 3) {
|
|
let xValues = vec3<f32>(
|
|
getX(batch, xF, xR, xC, inputDepthNearestVec4),
|
|
getX(batch, xF, xR, xC, inputDepthNearestVec4 + 1),
|
|
getX(batch, xF, xR, xC, inputDepthNearestVec4 + 2)
|
|
);
|
|
let wValues = vec3<f32>(
|
|
getW(wF, wR, wC, inputDepthNearestVec4, d2),
|
|
getW(wF, wR, wC, inputDepthNearestVec4 + 1, d2),
|
|
getW(wF, wR, wC, inputDepthNearestVec4 + 2, d2)
|
|
);
|
|
dotProd += dot(xValues, wValues);
|
|
}
|
|
}
|
|
}
|
|
}
|
|
setOutputAtIndex(index, dotProd);
|
|
}
|
|
}`}};function zpe(e){let{inputs:t,backend:a,attrs:n}=e,{x:r,filter:s}=t,{strides:i,pad:o,dilations:l}=n,u=S.computeConv3DInfo(r.shape,s.shape,i,l,o),p=[u.padInfo.front,u.padInfo.top,u.padInfo.left],c=[{type:"int32",data:[u.filterDepth,u.filterHeight,u.filterWidth]},{type:"int32",data:[...p]},{type:"int32",data:[u.strideDepth,u.strideHeight,u.strideWidth]},{type:"int32",data:[u.dilationDepth,u.dilationHeight,u.dilationWidth]}],d=new Ope(u),h=pa(r.dtype,s.dtype);return a.runWebGPUProgram(d,[r,s],h,c)}var Lpe={kernelName:gi,backendName:"webgpu",kernelFunc:zpe};function Wpe(e){let{inputs:t,backend:a,attrs:n}=e,{x:r,dy:s}=t,{strides:i,pad:o,filterShape:l}=n,u=S.computeConv3DInfo(r.shape,l,i,1,o),p=new Rpe(u),c=[{type:"int32",data:[u.padInfo.front,u.padInfo.top,u.padInfo.left]},{type:"int32",data:[u.strideDepth,u.strideHeight,u.strideWidth]},{type:"int32",data:[u.batchSize]},{type:"int32",data:[u.outDepth]},{type:"int32",data:[u.outHeight]},{type:"int32",data:[u.outWidth]},{type:"int32",data:[u.inDepth]},{type:"int32",data:[u.inHeight]},{type:"int32",data:[u.inWidth]}];return a.runWebGPUProgram(p,[r,s],s.dtype,c)}var Bpe={kernelName:ru,backendName:"webgpu",kernelFunc:Wpe};function Vpe(e){let{inputs:t,backend:a,attrs:n}=e,{dy:r,filter:s}=t,{strides:i,pad:o,inputShape:l}=n,u=S.computeConv3DInfo(l,s.shape,i,1,o),p=new Epe(u),c=[{type:"int32",data:[u.filterDepth,u.filterHeight,u.filterWidth]},{type:"int32",data:[u.filterDepth-1-u.padInfo.front,u.filterHeight-1-u.padInfo.top,u.filterWidth-1-u.padInfo.left]},{type:"int32",data:[u.strideDepth,u.strideHeight,u.strideWidth]},{type:"int32",data:[u.outDepth]},{type:"int32",data:[u.outHeight]},{type:"int32",data:[u.outWidth]},{type:"int32",data:[u.outChannels]}];return a.runWebGPUProgram(p,[r,s],r.dtype,c)}var Upe={kernelName:yi,backendName:"webgpu",kernelFunc:Vpe},Gpe=tt({opType:le.COS}),Hpe={kernelName:xi,backendName:"webgpu",kernelFunc:Gpe},jpe=tt({opType:le.COSH}),qpe={kernelName:Ai,backendName:"webgpu",kernelFunc:jpe},Xpe=class{constructor(e,t,a,n){this.variableNames=["Image","Boxes","BoxInd"],this.uniforms="extrapolationValue : f32,",this.workgroupSize=[64,1,1],this.size=!0;let[r]=t;this.outputShape=[r,a[0],a[1],e],this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.methodId=n==="bilinear"?1:0,this.cropHeightBiggerThan1=this.outputShape[1]>1,this.cropWidthBiggerThan1=this.outputShape[2]>1,this.shaderKey=`cropAndResize_${this.methodId}_${this.cropHeightBiggerThan1}_${this.cropWidthBiggerThan1}`}getUserCode(){let[e,t]=["f32(uniforms.imageShape[1] - 1)","f32(uniforms.imageShape[2] - 1)"],[a,n,r]=this.cropHeightBiggerThan1?[`(${e} / f32(uniforms.outShape[1] - 1))`,"(y2-y1) * height_ratio",`y1*${e} + f32(y)*(height_scale)`]:["0.0","0.0",`0.5 * (y1+y2) * ${e}`],[s,i,o]=this.cropWidthBiggerThan1?[`(${t} / f32(uniforms.outShape[2] - 1))`,"(x2-x1) * width_ratio",`x1*${t} + f32(x)*(width_scale)`]:["0.0","0.0",`0.5 * (x1+x2) * ${t}`];return`
|
|
${de("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
let height_ratio = f32(${a});
|
|
let width_ratio = f32(${s});
|
|
let b = coords[0];
|
|
let y = coords[1];
|
|
let x = coords[2];
|
|
let d = coords[3];
|
|
// get box vals
|
|
let y1 = getBoxes(b, 0);
|
|
let x1 = getBoxes(b, 1);
|
|
let y2 = getBoxes(b, 2);
|
|
let x2 = getBoxes(b, 3);
|
|
// get image in batch index
|
|
let bInd = i32(round(getBoxInd(b)));
|
|
if(bInd < 0 || bInd >= uniforms.outShape[0]) {
|
|
return;
|
|
}
|
|
let height_scale = ${n};
|
|
let width_scale = ${i};
|
|
let in_y = ${r};
|
|
if( in_y < 0.0 || in_y > ${e} ) {
|
|
setOutputAtIndex(index, uniforms.extrapolationValue);
|
|
return;
|
|
}
|
|
let in_x = ${o};
|
|
if( in_x < 0.0 || in_x > ${t} ) {
|
|
setOutputAtIndex(index, uniforms.extrapolationValue);
|
|
return;
|
|
}
|
|
let sourceFracIndexCR = vec2<f32>(in_x,in_y);
|
|
if(${this.methodId} == 1) {
|
|
// Compute the four integer indices.
|
|
let sourceFloorCR = vec2<i32>(sourceFracIndexCR);
|
|
let sourceCeilCR = vec2<i32>(ceil(sourceFracIndexCR));
|
|
let topLeft = getImage(bInd, sourceFloorCR.y, sourceFloorCR.x, d);
|
|
let bottomLeft = getImage(bInd, sourceCeilCR.y, sourceFloorCR.x, d);
|
|
let topRight = getImage(bInd, sourceFloorCR.y, sourceCeilCR.x, d);
|
|
let bottomRight = getImage(bInd, sourceCeilCR.y, sourceCeilCR.x, d);
|
|
let fracCR = sourceFracIndexCR - vec2<f32>(sourceFloorCR);
|
|
let top = topLeft + (topRight - topLeft) * fracCR.x;
|
|
let bottom = bottomLeft + (bottomRight - bottomLeft) * fracCR.x;
|
|
let newValue = top + (bottom - top) * fracCR.y;
|
|
setOutputAtIndex(index, newValue);
|
|
} else {
|
|
// Compute the coordinators of nearest neighbor point.
|
|
let sourceNearestCR = vec2<i32>(floor(
|
|
sourceFracIndexCR + vec2<f32>(0.5,0.5)));
|
|
let newValue = getImage(
|
|
bInd, sourceNearestCR.y, sourceNearestCR.x, d);
|
|
setOutputAtIndex(index, newValue);
|
|
}
|
|
}
|
|
}
|
|
`}},Kpe=e=>{let{inputs:t,backend:a,attrs:n}=e,{image:r,boxes:s,boxInd:i}=t,{cropSize:o,method:l,extrapolationValue:u}=n,p=new Xpe(r.shape[3],s.shape,o,l),c=[{type:"float32",data:[u]}];return a.runWebGPUProgram(p,[r,s,i],"float32",c)},Ype={kernelName:wi,backendName:"webgpu",kernelFunc:Kpe},Yd;(function(e){e.Prod="*",e.Sum="+"})(Yd||(Yd={}));var Wx=class{constructor(e,t,a,n){this.variableNames=["x"],this.uniforms="index : f32,",this.size=!0,this.workgroupSize=[128,1,1],this.outputShape=t,this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.exclusive=a,this.reverse=n,this.op=e,this.shaderKey=`cum_${this.op}_${this.exclusive}_${this.reverse}`}getUserCode(){let e=this.outputShape.length,t=this.op===Yd.Prod?"1.0":"0.0",a=this.exclusive?t:`getX(${Bx(e,"coords",this.op)})`,n=this.outputShape[this.outputShape.length-1],r="",s="";return this.exclusive?(r=this.reverse?`end != ${n-1}`:"end != 0",s=this.reverse?"end + 1":"end - 1"):(r=this.reverse?`end + pow2 < ${n}`:"end >= pow2",s=this.reverse?"end + pow2":"end - pow2"),`
|
|
${de("index")} {
|
|
if (index < uniforms.size) {
|
|
var coords = getCoordsFromIndex(index);
|
|
|
|
let end = ${Vx(e,"coords",this.op)};
|
|
var val = ${a};
|
|
let pow2 = i32(pow(2.0, uniforms.index));
|
|
if (${r}) {
|
|
let idx = ${s};
|
|
${Vx(e,"coords",this.op)} = idx;
|
|
val ${this.op}= getX(${Bx(e,"coords",this.op)});
|
|
}
|
|
setOutputAtIndex(index, val);
|
|
}
|
|
}
|
|
`}};function Bx(e,t,a){if(e===1)return`${t}`;if(e===2)return`${t}.x, ${t}.y`;if(e===3)return`${t}.x, ${t}.y, ${t}.z`;if(e===4)return`${t}.x, ${t}.y, ${t}.z, ${t}.w`;throw Error(`Cumulative ${a} for rank ${e} is not yet supported`)}function Vx(e,t,a){if(e===1)return`${t}`;if(e===2)return`${t}.y`;if(e===3)return`${t}.z`;if(e===4)return`${t}.w`;throw Error(`Cumulative ${a} for rank ${e} is not yet supported`)}function ik(e,t,a,n,r,s){let i=t.shape.length,o=S.getAxesPermutation([n],i),l=t;o!=null&&(l=Bn({inputs:{x:t},backend:a,attrs:{perm:o}}));let u=S.getInnerMostAxes(1,i)[0];if(u!==i-1)throw new Error(`WebGPU cumprod shader expects an inner-most axis=${t.shape.length-1} but got axis=${n}`);let p=l.shape[u],c=tn({inputs:{x:l},backend:a});for(let d=0;d<=Math.ceil(Math.log2(p))-1;d++){let h=new Wx(e,l.shape,!1,s),f=c,m=[{type:"float32",data:[d]}];c=a.runWebGPUProgram(h,[c],c.dtype,m),a.disposeData(f.dataId)}if(r){let d=new Wx(e,l.shape,r,s),h=c,f=[{type:"float32",data:[0]}];c=a.runWebGPUProgram(d,[c],c.dtype,f),a.disposeData(h.dataId)}if(o!=null){let d=S.getUndoAxesPermutation(o),h=Bn({inputs:{x:c},backend:a,attrs:{perm:d}});return a.disposeData(c.dataId),a.disposeData(l.dataId),h}return c}function Zpe(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{axis:s,exclusive:i,reverse:o}=n;return ik(Yd.Prod,r,a,s,i,o)}var Jpe={kernelName:bi,backendName:"webgpu",kernelFunc:Zpe};function Qpe(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{axis:s,exclusive:i,reverse:o}=n;return ik(Yd.Sum,r,a,s,i,o)}var ece={kernelName:vi,backendName:"webgpu",kernelFunc:Qpe};function tce(e){let{inputs:t,backend:a,attrs:n}=e,{x:r,weights:s}=t,{size:i,binaryOutput:o}=n,l=r.shape.length===1,u=v.sizeFromShape(s.shape)>0,p=s.dtype,c=l?[r.shape[0]]:[r.shape[0],r.shape[1]],d=l?[i]:[r.shape[0],i],h=Mn({backend:a,attrs:{shape:d,value:0,dtype:p}}),f=new ak(c,u,o),m=[{type:"int32",data:[i]}],g=u?[r,s]:[r];return a.runWebGPUProgram(f,g,p,m,h)}var ace={kernelName:su,backendName:"webgpu",kernelFunc:tce},nce=class{constructor(e,t){this.variableNames=["x"],this.workgroupSize=[64,1,1],this.size=!0,this.uniforms="blockSize : i32,",this.outputShape=e,this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey=`depthToSpace_${t}`,this.dataFormat=t}getUserCode(){return`
|
|
${de("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
let b = coords[0];
|
|
let h = ${this.getHeightCoordString()};
|
|
let w = ${this.getWidthCoordString()};
|
|
let d = ${this.getDepthCoordString()};
|
|
|
|
let in_h = h / uniforms.blockSize;
|
|
let offset_h = h % uniforms.blockSize;
|
|
let in_w = w / uniforms.blockSize;
|
|
let offset_w = w % uniforms.blockSize;
|
|
let offset_d = (offset_h * uniforms.blockSize + offset_w) *
|
|
${this.getOutputDepthSize()};
|
|
let in_d = d + offset_d;
|
|
|
|
let rlt = ${this.getInputSamplingString()};
|
|
setOutputAtIndex(index, rlt);
|
|
}
|
|
}`}getHeightCoordString(){return this.dataFormat==="NHWC"?"coords[1]":"coords[2]"}getWidthCoordString(){return this.dataFormat==="NHWC"?"coords[2]":"coords[3]"}getDepthCoordString(){return this.dataFormat==="NHWC"?"coords[3]":"coords[1]"}getOutputDepthSize(){return this.dataFormat==="NHWC"?"uniforms.outShape[3]":"uniforms.outShape[1]"}getInputSamplingString(){return this.dataFormat==="NHWC"?"getX(b, in_h, in_w, in_d)":"getX(b, in_d, in_h, in_w)"}};function rce(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{blockSize:s,dataFormat:i}=n,o=r.shape[0],l=i==="NHWC"?r.shape[1]:r.shape[2],u=i==="NHWC"?r.shape[2]:r.shape[3],p=i==="NHWC"?r.shape[3]:r.shape[1],c=l*s,d=u*s,h=p/(s*s),f=i==="NHWC"?[o,c,d,h]:[o,h,c,d],m=[{type:"int32",data:[s]}],g=new nce(f,i);return a.runWebGPUProgram(g,[r],r.dtype,m)}var sce={kernelName:ki,backendName:"webgpu",kernelFunc:rce},ice=class{constructor(e,t,a,n=!1,r=null,s=!1){this.variableNames=["x","W"],this.uniforms="pads : vec2<i32>, inDims : vec2<i32>,",this.workgroupSize=[16,16,1],this.outputShape=e,this.dispatchLayout={x:[3],y:[2],z:[0,1]},this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),n&&this.variableNames.push("bias"),s&&this.variableNames.push("preluActivationWeights"),this.addBias=n,this.activation=r,this.hasPreluActivation=s,this.filterHeight=t,this.filterWidth=a,this.shaderKey=`depthwiseNCHW_${this.activation}_${this.filterHeight}_${this.filterWidth}`}getUserCode(){let e=this.filterWidth*this.filterHeight,t=this.workgroupSize[0]*this.workgroupSize[1]*this.workgroupSize[2],a=this.workgroupSize[1]+this.filterHeight-1,n=this.workgroupSize[0]+this.filterWidth-1;return`
|
|
${Mr(this.activation,this.hasPreluActivation,!1,4)}
|
|
|
|
var<workgroup> mm_Asub : array<array<f32, ${n}>, ${a}>;
|
|
var<workgroup> mm_Bsub : array<array<f32, ${this.filterWidth}>, ${this.filterHeight}>;
|
|
fn readX(batch : i32, channel : i32, row : i32, col : i32) -> f32 {
|
|
var value = 0.0;
|
|
if (row >=0 && row < uniforms.inDims[0] && col >=0 && col < uniforms.inDims[1])
|
|
{
|
|
value = getX(batch, channel, row, col);
|
|
}
|
|
return value;
|
|
}
|
|
|
|
${de()} {
|
|
let coords = getOutputCoords();
|
|
let batch = coords[0];
|
|
let xRCCorner = vec2<i32>(coords.zw) - uniforms.pads;
|
|
let channelMul = uniforms.wShape[3];
|
|
let d1 = coords[1] / channelMul;
|
|
let q = coords[1] % channelMul;
|
|
|
|
let inputRowStart = xRCCorner.x;
|
|
let inputColStart = xRCCorner.y;
|
|
|
|
let localRow = i32(localId.y);
|
|
let localCol = i32(localId.x);
|
|
|
|
// Load one tile of X into local memory.
|
|
for (var inputRow = localRow; inputRow < ${a}; inputRow = inputRow + ${this.workgroupSize[1]}) {
|
|
for (var inputCol = localCol; inputCol < ${n}; inputCol = inputCol + ${this.workgroupSize[0]}) {
|
|
let rowOffset = inputRow - localRow;
|
|
let colOffset = inputCol - localCol;
|
|
mm_Asub[inputRow][inputCol] = readX(batch, d1, inputRowStart + rowOffset, inputColStart + colOffset);
|
|
}
|
|
}
|
|
|
|
// Load one tile of W into local memory.
|
|
var wIndex = i32(localIndex);
|
|
${e<t?`if (wIndex < ${e})`:`for(; wIndex < ${e}; wIndex = wIndex + ${t})`}
|
|
|
|
{
|
|
let wRow = wIndex / ${this.filterWidth};
|
|
let wCol = wIndex % ${this.filterWidth};
|
|
mm_Bsub[wRow][wCol] = getW(wRow, wCol, d1, q);
|
|
}
|
|
|
|
workgroupBarrier();
|
|
|
|
var value = 0.0;
|
|
for (var wR = 0; wR < ${this.filterHeight}; wR = wR + 1) {
|
|
for (var wC = 0; wC < ${this.filterWidth}; wC = wC + 1) {
|
|
let xVal = mm_Asub[localRow + wR][localCol + wC];
|
|
let wVal = mm_Bsub[wR][wC];
|
|
value = fma(xVal, wVal, value);
|
|
}
|
|
}
|
|
${Yo(this.addBias,this.activation)}
|
|
if (coordsInBounds4D(coords, uniforms.outShape)) {
|
|
setOutputAtCoords(coords[0], coords[1], coords[2], coords[3], value);
|
|
}
|
|
}
|
|
`}},ok=class{constructor(e,t=!1,a=null,n=!1){this.variableNames=["x","W"],this.uniforms="pads : vec2<i32>, inDims : vec2<i32>,",this.workgroupSize=[4,4,4],this.workPerThread=4,this.outputComponent=4,this.outputShape=e.outShape,this.dispatchLayout={x:[3],y:[2],z:[0,1]},this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize,[4,this.workPerThread,1]),v.assert(e.dataFormat==="channelsLast",()=>"TODO: NCHW is unimplemented"),t&&this.variableNames.push("bias"),n&&this.variableNames.push("preluActivationWeights"),this.convInfo=e,this.addBias=t,this.activation=a,this.hasPreluActivation=n,this.shaderKey=`depthwiseVec4_${a}_${this.convInfo.filterHeight}_${this.convInfo.filterWidth}_${this.convInfo.strideHeight}_${this.convInfo.strideWidth}_${this.workPerThread}`}getUserCode(){let e=(this.workPerThread-1)*this.convInfo.strideWidth+this.convInfo.filterWidth,t=this.convInfo.strideHeight,a=this.convInfo.strideWidth;return`
|
|
${Mr(this.activation,this.hasPreluActivation,!0,4)}
|
|
fn readX(batch : i32, row : i32, col : i32, channel : i32) -> vec4<f32> {
|
|
var value = vec4<f32>(0.0);
|
|
if (col >=0 && col < uniforms.inDims[1]) {
|
|
value = getX(batch, row, col, channel);
|
|
}
|
|
return value;
|
|
}
|
|
|
|
${de()} {
|
|
let batch = i32(globalId.z) / uniforms.outShape[1];
|
|
let r = i32(globalId.z) % uniforms.outShape[1];
|
|
let c = i32(globalId.y) * ${this.workPerThread};
|
|
let d1 = i32(globalId.x) * 4;
|
|
let xRCCorner = vec2<i32>(r, c) * vec2<i32>(${t}, ${a}) - uniforms.pads;
|
|
|
|
let xRCorner = xRCCorner.x;
|
|
let xCCorner = xRCCorner.y;
|
|
var xVals : array<vec4<f32>, ${e}>;
|
|
var dotProd : array<vec4<f32>, ${this.workPerThread}>;
|
|
for (var i = 0; i < ${this.workPerThread}; i++) {
|
|
dotProd[i] = vec4<f32>(0.0);
|
|
}
|
|
|
|
// Use constant instead of uniform can give better performance.
|
|
for (var wR = 0; wR < ${this.convInfo.filterHeight}; wR = wR + 1) {
|
|
let xR = xRCorner + wR;
|
|
if (xR >=0 && xR < uniforms.inDims[0]) {
|
|
for (var i = 0; i < ${e}; i++) {
|
|
xVals[i] = readX(batch, xR, xCCorner + i, d1);
|
|
}
|
|
for (var wC = 0; wC < ${this.convInfo.filterWidth}; wC = wC + 1) {
|
|
let wValue = getW(wR, wC, d1, 0);
|
|
for (var i = 0; i < ${this.workPerThread}; i++) {
|
|
dotProd[i] = fma(xVals[i * ${a} + wC], wValue, dotProd[i]);
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
for (var i = 0; i < ${this.workPerThread}; i = i + 1) {
|
|
let coords = vec4<i32>(batch, r, c + i, d1);
|
|
if (coordsInBounds4D(coords, uniforms.outShape)) {
|
|
var value = dotProd[i];
|
|
${Yo(this.addBias,this.activation)}
|
|
setOutputAtCoords(coords[0], coords[1], coords[2], coords[3], value);
|
|
}
|
|
}
|
|
}
|
|
`}},lk=class{constructor(e,t=!1,a=null,n=!1){this.variableNames=["x","W"],this.uniforms=`pads : vec2<i32>, inDims : vec2<i32>, filterHeight : i32,
|
|
filterWidth : i32, strides : vec2<i32>, dilations : vec2<i32>,`,this.workgroupSize=[256,1,1],this.size=!0,this.outputShape=e.outShape,this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.isChannelsLast=e.dataFormat==="channelsLast",t&&this.variableNames.push("bias"),n&&this.variableNames.push("preluActivationWeights"),this.convInfo=e,this.addBias=t,this.activation=a,this.hasPreluActivation=n,this.shaderKey=`depthwise_${this.activation}_${this.isChannelsLast}`}getUserCode(){let e=this.isChannelsLast?"getX(batch, xR, xC, d1);":"getX(batch, d1, xR, xC);";return`
|
|
${Mr(this.activation,this.hasPreluActivation,!1,4)}
|
|
|
|
${de("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getOutputCoords();
|
|
let batch = coords[0];
|
|
let xRCCorner = vec2<i32>(coords.${this.isChannelsLast?"yz":"zw"}) * uniforms.strides - uniforms.pads;
|
|
let d2 = coords[${this.isChannelsLast?3:1}];
|
|
let channelMul = uniforms.wShape[3];
|
|
let d1 = d2 / channelMul;
|
|
let q = d2 % channelMul;
|
|
|
|
let inputRowStart = xRCCorner.x;
|
|
let inputColStart = xRCCorner.y;
|
|
let inputRowEnd = inputRowStart + uniforms.filterHeight *
|
|
uniforms.dilations[0];
|
|
let inputColEnd = inputColStart + uniforms.filterWidth *
|
|
uniforms.dilations[1];
|
|
|
|
// Convolve x(?, ?, d1)|x(d1, ?, ?) with w(:, :, d1, q) to get
|
|
// y(yR, yC, d2)|y(d2, yR, yC). ? = to be determined. : = across all
|
|
// values in that axis. x(?, ?, d1) and y(yR, yC, d2) is for NHWC.
|
|
// x(d1, ?, ?) and y(d2, yR, yC) is for NCHW.
|
|
var value = 0.0;
|
|
|
|
// Extract if checking out of for loop for performance.
|
|
if (inputRowStart >= 0 && inputColStart >= 0 &&
|
|
inputRowEnd < uniforms.inDims[0] &&
|
|
inputColEnd < uniforms.inDims[1]) {
|
|
for (var wR = 0; wR < uniforms.filterHeight; wR = wR + 1) {
|
|
let xR = inputRowStart + wR * uniforms.dilations[0];
|
|
|
|
for (var wC = 0; wC < uniforms.filterWidth; wC = wC + 1) {
|
|
let xC = inputColStart + wC * uniforms.dilations[1];
|
|
|
|
let xVal = ${e};
|
|
let wVal = getW(wR, wC, d1, q);
|
|
value = value + xVal * wVal;
|
|
}
|
|
}
|
|
} else {
|
|
for (var wR = 0; wR < uniforms.filterHeight; wR = wR + 1) {
|
|
let xR = inputRowStart + wR * uniforms.dilations[0];
|
|
|
|
if (xR < 0 || xR >= uniforms.inDims[0]) {
|
|
continue;
|
|
}
|
|
|
|
for (var wC = 0; wC < uniforms.filterWidth; wC = wC + 1) {
|
|
let xC = inputColStart + wC * uniforms.dilations[1];
|
|
|
|
if (xC < 0 || xC >= uniforms.inDims[1]) {
|
|
continue;
|
|
}
|
|
|
|
let xVal = ${e};
|
|
let wVal = getW(wR, wC, d1, q);
|
|
value = value + xVal * wVal;
|
|
}
|
|
}
|
|
}
|
|
${Yo(this.addBias,this.activation)}
|
|
setOutputAtCoords(coords[0], coords[1], coords[2], coords[3], value);
|
|
}
|
|
}
|
|
`}};function oce(e){let{inputs:t,backend:a,attrs:n}=e,{x:r,filter:s}=t,{strides:i,pad:o,dataFormat:l,dilations:u,dimRoundingMode:p}=n,c=S.convertConv2DDataFormat(l),d=u;d==null&&(d=[1,1]);let h=S.computeConv2DInfo(r.shape,s.shape,i,d,o,p,!0,c),f=[{type:"int32",data:[h.padInfo.top,h.padInfo.left]},{type:"int32",data:[h.inHeight,h.inWidth]}],m=h.dataFormat==="channelsLast",g;return!m&&h.inHeight>16&&h.inWidth>16&&h.strideHeight===1&&h.strideWidth===1&&h.dilationWidth===1&&h.dilationHeight===1&&h.inChannels===h.outChannels?g=new ice(h.outShape,h.filterHeight,h.filterWidth):m&&h.outHeight>4&&h.outWidth>4&&h.strideWidth<=2&&h.inChannels===h.outChannels&&h.dilationHeight===1&&h.dilationWidth===1&&h.inChannels%4===0?g=new ok(h):(g=new lk(h),f.push({type:"int32",data:[h.filterHeight]},{type:"int32",data:[h.filterWidth]},{type:"int32",data:[h.strideHeight,h.strideWidth]},{type:"int32",data:[h.dilationHeight,h.dilationWidth]})),a.runWebGPUProgram(g,[r,s],r.dtype,f)}var lce={kernelName:Ii,backendName:"webgpu",kernelFunc:oce},uce=class{constructor(e){this.variableNames=["x","dy"],this.uniforms=`strides : vec2<i32>, pads : vec2<i32>, filterDims : vec2<i32>, outHeight : i32,
|
|
outWidth : i32, inHeight : i32, inWidth : i32, batchSize : i32, channelMul : i32,`,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e.filterShape,this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="depthwise_conv2d_backprop_filter"}getUserCode(){return`
|
|
${de("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
let wR = coords[0];
|
|
let wC = coords[1];
|
|
let d1 = coords[2];
|
|
let dm = coords[3];
|
|
let d2 = d1 * uniforms.channelMul + dm;
|
|
|
|
var dotProd = 0.0;
|
|
for (var b = 0; b < uniforms.batchSize; b++) {
|
|
for (var yR = 0; yR < uniforms.outHeight; yR++) {
|
|
let xR = wR + yR * uniforms.strides[0] - uniforms.pads[0];
|
|
|
|
if (xR < 0 || xR >= uniforms.inHeight) {
|
|
continue;
|
|
}
|
|
|
|
for (var yC = 0; yC < uniforms.outWidth; yC++) {
|
|
let xC = wC + yC * uniforms.strides[1] - uniforms.pads[1];
|
|
|
|
if (xC < 0 || xC >= uniforms.inWidth) {
|
|
continue;
|
|
}
|
|
|
|
let dyValue = getDy(b, yR, yC, d2);
|
|
let xValue = getX(b, xR, xC, d1);
|
|
dotProd += xValue * dyValue;
|
|
}
|
|
}
|
|
}
|
|
setOutputAtIndex(index, dotProd);
|
|
}
|
|
}
|
|
`}},dce=class{constructor(e){this.variableNames=["dy","W"],this.uniforms=`strides : vec2<i32>, pads : vec2<i32>, filterDims : vec2<i32>,
|
|
outHeight : i32, outWidth : i32, channelMul : i32,`,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e.inShape,this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="depthwise_conv2d_backprop_input"}getUserCode(){return`
|
|
${de("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
let batch = coords[0];
|
|
let d1 = coords[3];
|
|
let dyCorner = coords.yz - uniforms.pads;
|
|
let dyRCorner = dyCorner.x;
|
|
let dyCCorner = dyCorner.y;
|
|
|
|
var dotProd = 0.0;
|
|
for (var wR = 0; wR < uniforms.filterDims[0]; wR++) {
|
|
let dyR = f32(dyRCorner + wR) / f32(uniforms.strides[0]);
|
|
|
|
if (dyR < 0.0 || dyR >= f32(uniforms.outHeight) || fract(dyR) > 0.0) {
|
|
continue;
|
|
}
|
|
|
|
let idyR = i32(dyR);
|
|
let wRPerm = uniforms.filterDims[0] - 1 - wR;
|
|
|
|
for (var wC = 0; wC < uniforms.filterDims[1]; wC++) {
|
|
let dyC = f32(dyCCorner + wC) / f32(uniforms.strides[1]);
|
|
|
|
if (dyC < 0.0 || dyC >= f32(uniforms.outWidth) || fract(dyC) > 0.0) {
|
|
continue;
|
|
}
|
|
|
|
let idyC = i32(dyC);
|
|
let wCPerm = uniforms.filterDims[1] - 1 - wC;
|
|
|
|
for (var dm = 0; dm < uniforms.channelMul; dm++) {
|
|
let d2 = d1 * uniforms.channelMul + dm;
|
|
let xValue = getDy(batch, idyR, idyC, d2);
|
|
let wValue = getW(wRPerm, wCPerm, d1, dm);
|
|
dotProd += xValue * wValue;
|
|
}
|
|
}
|
|
}
|
|
setOutputAtIndex(index, dotProd);
|
|
}
|
|
}
|
|
`}};function pce(e){let{inputs:t,backend:a,attrs:n}=e,{x:r,dy:s}=t,{strides:i,dilations:o,pad:l,dimRoundingMode:u,filterShape:p}=n,c=S.computeConv2DInfo(r.shape,p,i,o,l,u,!0),d=new uce(c),h=[{type:"int32",data:[c.strideHeight,c.strideWidth]},{type:"int32",data:[c.padInfo.top,c.padInfo.left]},{type:"int32",data:[c.filterHeight,c.filterWidth]},{type:"int32",data:[c.outHeight]},{type:"int32",data:[c.outWidth]},{type:"int32",data:[c.inHeight]},{type:"int32",data:[c.inWidth]},{type:"int32",data:[c.batchSize]},{type:"int32",data:[c.outChannels/c.inChannels]}];return a.runWebGPUProgram(d,[r,s],"float32",h)}var cce={kernelName:rp,backendName:"webgpu",kernelFunc:pce};function hce(e){let{inputs:t,backend:a,attrs:n}=e,{dy:r,filter:s}=t,{strides:i,dilations:o,pad:l,dimRoundingMode:u,inputShape:p}=n,c=S.computeConv2DInfo(p,s.shape,i,o,l,u,!0),d=new dce(c),h=[{type:"int32",data:[c.strideHeight,c.strideWidth]},{type:"int32",data:[c.filterHeight-1-c.padInfo.top,c.filterWidth-1-c.padInfo.left]},{type:"int32",data:[c.filterHeight,c.filterWidth]},{type:"int32",data:[c.outHeight]},{type:"int32",data:[c.outWidth]},{type:"int32",data:[c.outChannels/c.inChannels]}];return a.runWebGPUProgram(d,[r,s],r.dtype,h)}var fce={kernelName:sp,backendName:"webgpu",kernelFunc:hce},mce=class{constructor(e){this.variableNames=["x"],this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=[e,e],this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="diag"}getUserCode(){return`
|
|
${de("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getOutputCoords();
|
|
let value = select(0.0, getX(coords[0]), coords[0] == coords[1]);
|
|
setOutputAtIndex(index, value);
|
|
}
|
|
}
|
|
`}};function gce(e){let{inputs:t,backend:a}=e,{x:n}=t,r=[...n.shape,...n.shape],s=v.sizeFromShape(n.shape),i=we({inputs:{x:n},backend:a,attrs:{shape:[s]}}),o=new mce(s),l=a.runWebGPUProgram(o,[i],i.dtype),u=we({inputs:{x:l},backend:a,attrs:{shape:r}});return a.disposeData(i.dataId),a.disposeData(l.dataId),u}var yce={kernelName:iu,backendName:"webgpu",kernelFunc:gce},xce=class{constructor(e){this.variableNames=["x","w"],this.uniforms="filterDims: vec2<i32>, pads: vec2<i32>, strides: vec2<i32>, dilations: vec2<i32>",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e.outShape,this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="dilation2d"}getUserCode(){return`
|
|
${de("index")} {
|
|
if (index < uniforms.size) {
|
|
let neg_infinity = -3.4e38;
|
|
let coords = getOutputCoords();
|
|
let batch = coords.x;
|
|
let d1 = coords.w;
|
|
let outTopLeftCorner = coords.yz * uniforms.strides - uniforms.pads;
|
|
let hBeg = outTopLeftCorner.x;
|
|
let wBeg = outTopLeftCorner.y;
|
|
|
|
var curVal = neg_infinity;
|
|
for (var h = 0; h < uniforms.filterDims[0]; h = h + 1) {
|
|
let hIn = hBeg + h * uniforms.dilations[0];
|
|
|
|
if (hIn >= 0 && hIn < uniforms.xShape[1]) {
|
|
for (var w = 0; w < uniforms.filterDims[1]; w = w + 1) {
|
|
let wIn = wBeg + w * uniforms.dilations[1];
|
|
|
|
if (wIn >= 0 && wIn < uniforms.xShape[2]) {
|
|
let val = getX(batch, hIn, wIn, d1) + getW(h, w, d1);
|
|
if (val > curVal) {
|
|
curVal = val;
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
setOutputAtIndex(index, curVal);
|
|
}
|
|
}
|
|
`}};function Ace(e){let{inputs:t,backend:a,attrs:n}=e,{x:r,filter:s}=t,{strides:i,pad:o,dilations:l}=n,u=S.computeDilation2DInfo(r.shape,s.shape,i,o,"NHWC",l),p=[u.padInfo.top,u.padInfo.left],c=[{type:"int32",data:[u.filterHeight,u.filterWidth]},{type:"int32",data:[...p]},{type:"int32",data:[u.strideHeight,u.strideWidth]},{type:"int32",data:[u.dilationHeight,u.dilationWidth]}],d=new xce(u);return a.runWebGPUProgram(d,[r,s],r.dtype,c)}var bce={kernelName:Si,backendName:"webgpu",kernelFunc:Ace},vce=class{constructor(e,t){if(this.variableNames=["x","w","dy"],this.uniforms="filterDims: vec2<i32>, pads: vec2<i32>, strides: vec2<i32>, dilations: vec2<i32>, dySize: i32,",this.workgroupSize=[64,1,1],this.atomic=!0,this.outputShape=e.inShape,this.dispatchLayout=be(e.outShape),this.dispatch=ce(this.dispatchLayout,e.outShape,this.workgroupSize),t!=="float32"&&t!=="int32")throw new Error(`Dilation2DBackpropInput only supports float32 and int32
|
|
types, does not support ${t} type.`);this.type=t,this.shaderKey="dilation2DBackpropInput"}getUserCode(){return`
|
|
${de("index")} {
|
|
if (index < uniforms.dySize) {
|
|
let coords = getDyCoordsFromIndex(index);
|
|
let b = coords[0];
|
|
let r = coords[1];
|
|
let c = coords[2];
|
|
let d = coords[3];
|
|
|
|
let dyCorner = vec2<i32>(r, c) * uniforms.strides - uniforms.pads;
|
|
var curVal = -3.4e38; // neg_infinity
|
|
var xRMax = 0;
|
|
var xCMax = 0;
|
|
|
|
// In the case of multiple argmax branches, we only back-propagate
|
|
// along the last branch, i.e., the one with largest value of
|
|
// 'wR * uniforms.filterDims[1] + wC', similarly to the max-pooling
|
|
// backward routines.
|
|
for (var wR = 0; wR < uniforms.filterDims[0]; wR++) {
|
|
let xR = dyCorner.x + wR * uniforms.dilations[0];
|
|
|
|
if (xR >= 0 && xR < uniforms.xShape[1]) {
|
|
for (var wC = 0; wC < uniforms.filterDims[1]; wC++) {
|
|
let xC = dyCorner.y + wC * uniforms.dilations[1];
|
|
|
|
if (xC >= 0 && xC < uniforms.xShape[2]) {
|
|
let val = getX(b, xR, xC, d) + getW(wR, wC, d);
|
|
if (val > curVal) {
|
|
curVal = val;
|
|
xRMax = xR;
|
|
xCMax = xC;
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
let flatIndexIn = d + uniforms.xShape[3] *
|
|
(xCMax + uniforms.xShape[2] * (xRMax + uniforms.xShape[1] * b));
|
|
let value = getDy(b, r, c, d);
|
|
${Gu("&result[flatIndexIn]","value",this.type)}
|
|
}
|
|
}
|
|
`}},wce=class{constructor(e,t,a){if(this.variableNames=["x","w","dy"],this.uniforms="filterDims: vec2<i32>, pads: vec2<i32>, strides: vec2<i32>, dilations: vec2<i32>, dySize: i32,",this.workgroupSize=[64,1,1],this.atomic=!0,this.outputShape=e.filterShape,this.dispatchLayout=be(e.outShape),this.dispatch=ce(this.dispatchLayout,e.outShape,this.workgroupSize),a!=="float32"&&a!=="int32")throw new Error(`Dilation2DBackpropFilter only supports float32 and int32
|
|
types, does not support ${a} type.`);this.type=a,this.shaderKey="dilation2DBackpropFilter"}getUserCode(){return`
|
|
${de("index")} {
|
|
if (index < uniforms.dySize) {
|
|
let coords = getDyCoordsFromIndex(index);
|
|
let b = coords[0];
|
|
let r = coords[1];
|
|
let c = coords[2];
|
|
let d = coords[3];
|
|
|
|
let dyCorner = vec2<i32>(r, c) * uniforms.strides - uniforms.pads;
|
|
var curVal = -3.4e38; // neg_infinity
|
|
var wRMax = 0;
|
|
var wCMax = 0;
|
|
|
|
// In the case of multiple argmax branches, we only back-propagate
|
|
// along the last branch, i.e., the one with largest value of
|
|
// 'wR * uniforms.filterDims[1] + wC', similarly to the max-pooling
|
|
// backward routines.
|
|
for (var wR = 0; wR < uniforms.filterDims[0]; wR++) {
|
|
let xR = dyCorner.x + wR * uniforms.dilations[0];
|
|
|
|
if (xR >= 0 && xR < uniforms.xShape[1]) {
|
|
for (var wC = 0; wC < uniforms.filterDims[1]; wC++) {
|
|
let xC = dyCorner.y + wC * uniforms.dilations[1];
|
|
|
|
if (xC >= 0 && xC < uniforms.xShape[2]) {
|
|
let val = getX(b, xR, xC, d) + getW(wR, wC, d);
|
|
if (val > curVal) {
|
|
curVal = val;
|
|
wRMax = wR;
|
|
wCMax = wC;
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
let flatIndexIn = d + uniforms.wShape[2] * (wCMax + wRMax * uniforms.wShape[1]);
|
|
let value = getDy(b, r, c, d);
|
|
${Gu("&result[flatIndexIn]","value",this.type)}
|
|
}
|
|
}
|
|
`}};function kce(e){let{inputs:t,backend:a,attrs:n}=e,{x:r,filter:s,dy:i}=t,{strides:o,pad:l,dilations:u}=n,p=S.computeDilation2DInfo(r.shape,s.shape,o,l,"NHWC",u),c=s.dtype,d=new wce(p,s.shape,c),h=[{type:"int32",data:[p.filterHeight,p.filterWidth]},{type:"int32",data:[p.padInfo.top,p.padInfo.left]},{type:"int32",data:[p.strideHeight,p.strideWidth]},{type:"int32",data:[p.dilationHeight,p.dilationWidth]},{type:"int32",data:[v.sizeFromShape(p.outShape)]}],f=Mn({backend:a,attrs:{shape:s.shape,value:0,dtype:c}});return a.runWebGPUProgram(d,[r,s,i],c,h,f)}var Ice={kernelName:Pl,backendName:"webgpu",kernelFunc:kce};function Sce(e){let{inputs:t,backend:a,attrs:n}=e,{x:r,filter:s,dy:i}=t,{strides:o,pad:l,dilations:u}=n,p=S.computeDilation2DInfo(r.shape,s.shape,o,l,"NHWC",u),c=r.dtype,d=new vce(p,c),h=[{type:"int32",data:[p.filterHeight,p.filterWidth]},{type:"int32",data:[p.padInfo.top,p.padInfo.left]},{type:"int32",data:[p.strideHeight,p.strideWidth]},{type:"int32",data:[p.dilationHeight,p.dilationWidth]},{type:"int32",data:[v.sizeFromShape(p.outShape)]}],f=Mn({backend:a,attrs:{shape:p.inShape,value:0,dtype:c}});return a.runWebGPUProgram(d,[r,s,i],c,h,f)}var Cce={kernelName:Dl,backendName:"webgpu",kernelFunc:Sce},uk=ta({opType:_e.MUL,cpuKernelImpl:$ue,supportsComplex:!0}),Tce={kernelName:io,backendName:"webgpu",kernelFunc:uk};function dk(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{axis:s,keepDims:i}=n;return Jo(r,s,i,"sum",a)}var Nce={kernelName:_o,backendName:"webgpu",kernelFunc:dk};function Rce(e){let{inputs:t,backend:a,attrs:n}=e,{equation:r}=n,s=t,{allDims:i,summedDims:o,idDims:l}=S.decodeEinsumEquation(r,s.length);S.checkEinsumDimSizes(i.length,l,s);let{path:u,steps:p}=S.getEinsumComputePath(o,l),c=p.length,d=null,h=i.length,f=[];for(let m=0;m<c;++m){for(let g of p[m]){let{permutationIndices:y,expandDims:x}=S.getEinsumPermutation(h,l[g]),A;S.isIdentityPermutation(y)?A=s[g]:(A=Bn({inputs:{x:s[g]},backend:a,attrs:{perm:y}}),f.push(A));let b=A.shape.slice();for(let w=0;w<x.length;++w)b.splice(x[w],0,1);v.arraysEqual(A.shape,b)||(A=we({inputs:{x:A},backend:a,attrs:{shape:b}}),f.push(A)),d===null?d=A:(d=uk({inputs:{a:A,b:d},backend:a}),f.push(d))}m<c-1&&(u[m]>=0&&(d=dk({inputs:{x:d},backend:a,attrs:{axis:u[m]-(i.length-h),keepDims:!1}}),f.push(d)),h--)}for(let m of f)m!==d&&a.disposeData(m.dataId);return d}var Ece={kernelName:ip,backendName:"webgpu",kernelFunc:Rce},Mce=tt({opType:le.ELU}),Fce={kernelName:Ti,backendName:"webgpu",kernelFunc:Mce},_ce=e=>{let{inputs:t,backend:a}=e,{dy:n,y:r}=t,s=new xh(_e.ELU_DER,n.shape,r.shape);return a.runWebGPUProgram(s,[n,r],n.dtype)},$ce={kernelName:ou,backendName:"webgpu",kernelFunc:_ce},Dce=ta({opType:_e.EQUAL,dtype:"bool",cpuKernelImpl:Aue}),Pce={kernelName:Ni,backendName:"webgpu",kernelFunc:Dce},Oce=tt({opType:le.ERF}),zce={kernelName:lu,backendName:"webgpu",kernelFunc:Oce},Lce=tt({opType:le.EXP,cpuKernelImpl:bue,dtype:"float32"}),Wce={kernelName:Ri,backendName:"webgpu",kernelFunc:Lce};function P1(e){let{inputs:t,attrs:a,backend:n}=e,{dim:r}=a,{input:s}=t,i=s.shape.length,o=s.shape.slice(),l=r;return r<0&&(v.assert(-(i+1)<=r,()=>`Axis must be in the interval [${-(i+1)}, ${i}]`),l=i+r+1),o.splice(l,0,1),we({inputs:{x:s},backend:n,attrs:{shape:o}})}var Bce={kernelName:uu,backendName:"webgpu",kernelFunc:P1},Vce=tt({opType:le.EXPM1,cpuKernelImpl:vue}),Uce={kernelName:Ei,backendName:"webgpu",kernelFunc:Vce},Ux=class{constructor(e,t){this.variableNames=["real","imag"],this.outputShape=[],this.uniforms="exponentMultiplier : f32, denominator: f32,",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=t,this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.component=e,this.shaderKey=`fft_${e}`}getUserCode(){return`
|
|
fn unaryOpComplex(real: f32, expR: f32, imag: f32, expI: f32) -> f32 {
|
|
${this.component==="real"?"return real * expR - imag * expI;":"return real * expI + imag * expR;"}
|
|
}
|
|
|
|
fn mulMatDFT(batch: i32, index: i32) -> f32 {
|
|
let indexRatio = f32(index) / f32(uniforms.realShape[1]);
|
|
let exponentMultiplierTimesIndexRatio =
|
|
uniforms.exponentMultiplier * indexRatio;
|
|
|
|
var result = 0.0;
|
|
|
|
for (var i = 0; i < uniforms.realShape[1]; i = i + 1) {
|
|
// x = (-2|2 * PI / N) * index * i;
|
|
let x = exponentMultiplierTimesIndexRatio * f32(i);
|
|
let expR = cos(x);
|
|
let expI = sin(x);
|
|
let real = getReal(batch, i);
|
|
let imag = getImag(batch, i);
|
|
|
|
result = result +
|
|
unaryOpComplex(real, expR, imag, expI) / uniforms.denominator;
|
|
}
|
|
|
|
return result;
|
|
}
|
|
|
|
${de("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getOutputCoords();
|
|
setOutputAtIndex(index, mulMatDFT(coords[0], coords[1]));
|
|
}
|
|
}
|
|
`}};function pk(e,t,a){let n=a.tensorMap.get(e.dataId),r=v.sizeFromShape(e.shape),s=e.shape[e.shape.length-1],i=r/s,o=[],l=we({inputs:{x:e},backend:a,attrs:{shape:[i,s]}});o.push(l);let u=l.shape,p=new Ux("real",u),c=new Ux("imag",u),d=[{dataId:n.complexTensorInfos.real.dataId,dtype:n.complexTensorInfos.real.dtype,shape:u},{dataId:n.complexTensorInfos.imag.dataId,dtype:n.complexTensorInfos.imag.dtype,shape:u}],h=t?2*Math.PI:-2*Math.PI,f=t?u[1]:1,m=[{type:"float32",data:[h]},{type:"float32",data:[f]}],g=a.runWebGPUProgram(p,d,"float32",m);o.push(g);let y=a.runWebGPUProgram(c,d,"float32",m);o.push(y);let x=Zo({inputs:{real:g,imag:y},backend:a});o.push(x);let A=we({inputs:{x},backend:a,attrs:{shape:e.shape}});return o.forEach(b=>a.disposeData(b.dataId)),A}function Gce(e){let{inputs:t,backend:a}=e,{input:n}=t;return pk(n,!1,a)}var Hce={kernelName:op,backendName:"webgpu",kernelFunc:Gce},jce=class{constructor(e){this.outputShape=[],this.variableNames=["x"],this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="flipLeftRight"}getUserCode(){return`
|
|
${de("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
let coordX = uniforms.xShape[2] - coords[2] - 1;
|
|
let outputValue = getX(coords[0], coords[1], coordX, coords[3]);
|
|
setOutputAtIndex(index, outputValue);
|
|
}
|
|
}
|
|
`}},qce={kernelName:Mi,backendName:"webgpu",kernelFunc:({inputs:e,backend:t})=>{let{image:a}=e,n=t,r=new jce(a.shape);return n.runWebGPUProgram(r,[a],a.dtype)}},Xce=tt({opType:le.FLOOR,cpuKernelImpl:wue}),Kce={kernelName:Fi,backendName:"webgpu",kernelFunc:Xce},Yce=ta({opType:_e.INT_DIV,cpuKernelImpl:kue,dtype:"int32"}),Zce={kernelName:_i,backendName:"webgpu",kernelFunc:Yce},Jce=class{constructor(e,t,a=!1){this.isFromPixels=!0,this.outputShape=[0],this.variableNames=[],this.workgroupSize=[256,1,1],this.outputShape=e,this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize,[t,1,1]),this.importVideo=a,this.shaderKey=`fromPixels_${this.importVideo}`}getUserCode(){let e=this.importVideo?"textureLoad(src, vec2<i32>(coords.yx));":"textureLoad(src, vec2<i32>(coords.yx), 0)";return`
|
|
@binding(1) @group(0) var src: ${this.importVideo?"texture_external":"texture_2d<f32>"};
|
|
${de("index")} {
|
|
let flatIndex = index * uniforms.numChannels;
|
|
if (flatIndex < uniforms.size) {
|
|
let coords = getCoordsFromIndex(flatIndex);
|
|
let values = ${e};
|
|
for (var i = 0; i < uniforms.numChannels; i = i + 1) {
|
|
result[flatIndex + i] = i32(floor(255.0 * values[i]));
|
|
}
|
|
}
|
|
}
|
|
`}},Qce={kernelName:Ed,backendName:"webgpu",kernelFunc:ehe},Il,U2=W().getBool("CANVAS2D_WILL_READ_FREQUENTLY_FOR_GPU");function ehe(e){let{inputs:t,backend:a,attrs:n}=e,{pixels:r}=t,{numChannels:s}=n;if(r==null)throw new Error("pixels passed to tf.browser.fromPixels() can not be null");let i=typeof HTMLVideoElement!="undefined"&&r instanceof HTMLVideoElement,o=typeof HTMLImageElement!="undefined"&&r instanceof HTMLImageElement,l=typeof HTMLCanvasElement!="undefined"&&r instanceof HTMLCanvasElement||typeof OffscreenCanvas!="undefined"&&r instanceof OffscreenCanvas,u=typeof ImageBitmap!="undefined"&&r instanceof ImageBitmap,[p,c]=i?[r.videoWidth,r.videoHeight]:[r.width,r.height],d=[c,p,s],h=!1,f=i||o;if(u||l||f){let x;if(h)x={width:p,height:c,format:null,usage:null,texture:a.device.importExternalTexture({source:r})};else{if(f){let _=W().getBool("CANVAS2D_WILL_READ_FREQUENTLY_FOR_GPU");(Il==null||_!==U2)&&(U2=_,Il=document.createElement("canvas").getContext("2d",{willReadFrequently:U2})),Il.canvas.width=p,Il.canvas.height=c,Il.drawImage(r,0,0,p,c),r=Il.canvas}let F=GPUTextureUsage.COPY_DST|GPUTextureUsage.RENDER_ATTACHMENT|GPUTextureUsage.TEXTURE_BINDING,E="rgba8unorm",I=a.textureManager.acquireTexture(d[1],d[0],E,F);a.queue.copyExternalImageToTexture({source:r},{texture:I},[d[1],d[0]]),x={width:p,height:c,format:E,usage:F,texture:I}}let A=v.sizeFromShape(d),b=v.computeStrides(d),w=new Jce(d,s,h),C=[{type:"uint32",data:[A]},{type:"uint32",data:[s]},{type:"uint32",data:[...b]}],T=a.makeTensorInfo([c,p],"int32"),N=a.tensorMap.get(T.dataId);N.resourceInfo=x;let M=a.runWebGPUProgram(w,[T],"int32",C);return a.disposeData(T.dataId),M}let m=r.data,g=m;if(s!=null&&s!==4){g=new Uint8Array(r.width*r.height*s);let x=m.length,A=0;for(let b=0;b<x;b++)b%4<s&&(g[A++]=m[b])}let y=a.makeTensorInfo(d,"int32",new Int32Array(g));return a.uploadToGPU(y.dataId),y}var the=class{constructor(e,t,a,n,r){this.uniforms="varianceEpsilon : f32,",this.workgroupSize=[128,1,1],this.size=!0,this.variableNames=["x","mean","variance"],S.assertAndGetBroadcastShape(e,t),S.assertAndGetBroadcastShape(e,a),this.outputShape=e,this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),n!=null&&(S.assertAndGetBroadcastShape(e,n),this.variableNames.push("offset")),r!=null&&(S.assertAndGetBroadcastShape(e,r),this.variableNames.push("scale")),this.offsetShape=n,this.scaleShape=r,this.shaderKey="batchNorm"}getUserCode(){let e="0.0";this.offsetShape!=null&&(e="getOffsetByOutputIndex(index)");let t="1.0";return this.scaleShape!=null&&(t="getScaleByOutputIndex(index)"),`
|
|
${de("index")} {
|
|
if (index < uniforms.size)
|
|
{
|
|
let xValue = getXByOutputIndex(index);
|
|
let meanValue = getMeanByOutputIndex(index);
|
|
let varianValue = getVarianceByOutputIndex(index);
|
|
let offsetValue = ${e};
|
|
let scaleValue = ${t};
|
|
let inv = scaleValue * inverseSqrt(varianValue + f32(uniforms.varianceEpsilon));
|
|
setOutputAtIndex(index,dot(vec3<f32>(xValue, -meanValue, offsetValue), vec3<f32>(inv, inv, 1.0)));
|
|
}
|
|
}
|
|
`}},ahe={kernelName:$i,backendName:"webgpu",kernelFunc:({inputs:e,attrs:t,backend:a})=>{let{x:n,scale:r,offset:s,mean:i,variance:o}=e,{varianceEpsilon:l}=t,u=a,p=[n,i,o],c=null;s!=null&&(c=s.shape,p.push(s));let d=null;r!=null&&(d=r.shape,p.push(r));let h=new the(n.shape,i.shape,o.shape,c,d),f=[{type:"float32",data:[l]}];return u.runWebGPUProgram(h,p,n.dtype,f)}};function nhe(e){let{inputs:t,backend:a,attrs:n}=e,{x:r,filter:s,bias:i,preluActivationWeights:o}=t,{strides:l,pad:u,dataFormat:p,dilations:c,dimRoundingMode:d,activation:h,leakyreluAlpha:f}=n,m=S.convertConv2DDataFormat(p),g=S.computeConv2DInfo(r.shape,s.shape,l,c,u,d,!1,m);return sk({x:r,filter:s,convInfo:g,backend:a,bias:i,preluActivationWeights:o,leakyreluAlpha:f,activation:h})}var rhe={kernelName:Kr,backendName:"webgpu",kernelFunc:nhe};function she(e){let{inputs:t,backend:a,attrs:n}=e,{x:r,filter:s,bias:i,preluActivationWeights:o}=t,{strides:l,pad:u,dilations:p,dimRoundingMode:c,activation:d,leakyreluAlpha:h}=n,f=p;f==null&&(f=[1,1]),v.assert(S.eitherStridesOrDilationsAreOne(l,f),()=>`Error in depthwiseConv2d: Either strides or dilations must be 1. Got strides ${l} and dilations '${f}'`);let m=S.computeConv2DInfo(r.shape,s.shape,l,f,u,c,!0),g=[r,s],y=i!=null,x=o!=null;y&&g.push(i),x&&g.push(o);let A=[{type:"int32",data:[m.padInfo.top,m.padInfo.left]},{type:"int32",data:[m.inHeight,m.inWidth]}],b;return m.outHeight>4&&m.outWidth>4&&m.strideWidth<=2&&m.inChannels===m.outChannels&&m.dilationHeight===1&&m.dilationWidth===1&&m.inChannels%4===0?b=new ok(m,y,d,x):(b=new lk(m,y,d,x),A.push({type:"int32",data:[m.filterHeight]},{type:"int32",data:[m.filterWidth]},{type:"int32",data:[m.strideHeight,m.strideWidth]},{type:"int32",data:[m.dilationHeight,m.dilationWidth]})),d==="leakyrelu"&&(A.push({type:"float32",data:[h]}),b.uniforms+=" alpha : f32,"),a.runWebGPUProgram(b,g,"float32",A)}var ihe={kernelName:Yr,backendName:"webgpu",kernelFunc:she},ohe=class{constructor(e,t){this.variableNames=["A","indices"],this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=t,this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey=`gathernd_${e}`,this.sliceDim=e,this.uniforms=`sliceDim : i32, strides : ${Jt(e)},`}getUserCode(){let e;return this.sliceDim>1?e="uniforms.strides[j]":e="uniforms.strides",`
|
|
${de("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
var flattenIndex = 0;
|
|
for (var j = 0; j < uniforms.sliceDim; j = j + 1) {
|
|
let indexTemp = i32(round(getIndices(coords[0], j)));
|
|
let strideNum = ${e};
|
|
flattenIndex = flattenIndex + indexTemp * strideNum;
|
|
}
|
|
|
|
setOutputAtIndex(index, getA(flattenIndex, coords[1]));
|
|
}
|
|
}
|
|
`}};function lhe(e){let{inputs:t,backend:a}=e,{params:n,indices:r}=t,s=r.shape,i=s[s.length-1],o=v.sizeFromShape(n.shape),[l,u,p,c]=S.prepareAndValidate(n,r),d=we({inputs:{x:r},backend:a,attrs:{shape:[u,i]}}),h=we({inputs:{x:n},backend:a,attrs:{shape:[v.sizeFromShape(n.shape)/p,p]}});if(a.shouldExecuteOnCPU([n,r])||n.dtype==="string"){let x=a.readSync(r.dataId),A=a.bufferSync(n),b=Iue(x,A,n.dtype,u,i,p,c,n.shape,o);return a.makeTensorInfo(l,n.dtype,b.values)}let f=new ohe(i,[u,p]),m=[{type:"int32",data:[i]},{type:"int32",data:c}],g=a.runWebGPUProgram(f,[h,d],h.dtype,m),y=we({inputs:{x:g},backend:a,attrs:{shape:l}});return a.disposeData(d.dataId),a.disposeData(h.dataId),a.disposeData(g.dataId),y}var uhe={kernelName:Di,backendName:"webgpu",kernelFunc:lhe},dhe=class{constructor(e,t){this.variableNames=["A","indices"],this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e.slice(),this.aShape=e,this.outputShape=t,this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="gather"}getUserCode(){let e=phe(this.aShape);return`
|
|
${de("index")} {
|
|
if (index < uniforms.size) {
|
|
let resRC = getCoordsFromIndex(index);
|
|
let indexZ = i32(getIndices(resRC.x, resRC.z));
|
|
let inBounds = select(0.0, 1.0, indexZ >= 0 && indexZ < uniforms.aShape[2]);
|
|
setOutputAtIndex(index, inBounds * getA(${e}));
|
|
}
|
|
}
|
|
`}};function phe(e){let t=["resRC.x","resRC.y","resRC.z","resRC.w"],a=[];for(let n=0;n<e.length;n++)n===2?a.push("indexZ"):a.push(`${t[n]}`);return a.join()}function ck(e){let{inputs:t,backend:a,attrs:n}=e,{x:r,indices:s}=t,{axis:i,batchDims:o}=n,l=v.parseAxisParam(i,r.shape)[0],u=S.segment_util.collectGatherOpShapeInfo(r,s,l,o),p=v.sizeFromShape(s.shape),c=[],d=we({inputs:{x:r},backend:a,attrs:{shape:[u.batchSize,u.outerSize,u.dimSize,u.sliceSize]}}),h=we({inputs:{x:s},backend:a,attrs:{shape:[u.batchSize,p/u.batchSize]}});c.push(d),c.push(h);let f=[u.batchSize,u.outerSize,p/u.batchSize,u.sliceSize];if(a.shouldExecuteOnCPU([r,s])){let x=a.tensorMap.get(h.dataId).values,A=$e(h.shape,h.dtype,x),b=a.tensorMap.get(d.dataId).values,w=$e(d.shape,d.dtype,b),C=Sue(w,A,f);return c.forEach(T=>a.disposeData(T.dataId)),a.makeTensorInfo(u.outputShape,C.dtype,C.values)}let m=new dhe(d.shape,f),g=a.runWebGPUProgram(m,[d,h],d.dtype);c.push(g);let y=we({inputs:{x:g},backend:a,attrs:{shape:u.outputShape}});return c.forEach(x=>a.disposeData(x.dataId)),y}var che={kernelName:pu,backendName:"webgpu",kernelFunc:ck},hhe=ta({opType:_e.GREATER,cpuKernelImpl:Tue,dtype:"bool"}),fhe={kernelName:Pi,backendName:"webgpu",kernelFunc:hhe},mhe=ta({opType:_e.GREATER_EQUAL,dtype:"bool",cpuKernelImpl:Cue}),ghe={kernelName:Oi,backendName:"webgpu",kernelFunc:mhe};function yhe(e){let{inputs:t,backend:a}=e,{input:n}=t;return pk(n,!0,a)}var xhe={kernelName:lp,backendName:"webgpu",kernelFunc:yhe},Ahe=tt({opType:le.IS_FINITE,dtype:"bool"}),bhe={kernelName:Li,backendName:"webgpu",kernelFunc:Ahe},vhe=tt({opType:le.IS_INF,dtype:"bool"}),whe={kernelName:Wi,backendName:"webgpu",kernelFunc:vhe},khe=tt({opType:le.IS_NAN,dtype:"bool"}),Ihe={kernelName:Bi,backendName:"webgpu",kernelFunc:khe};function She(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{alpha:s}=n,i=[{type:"float32",data:[s]}],o=new Hu(r.shape,le.LEAKYRELU,"alpha : f32,");return a.runWebGPUProgram(o,[r],"float32",i)}var Che={kernelName:Vi,backendName:"webgpu",kernelFunc:She},The=ta({opType:_e.LESS,dtype:"bool",cpuKernelImpl:Rue}),Nhe={kernelName:Ui,backendName:"webgpu",kernelFunc:The},Rhe=ta({opType:_e.LESS_EQUAL,dtype:"bool",cpuKernelImpl:Nue}),Ehe={kernelName:Gi,backendName:"webgpu",kernelFunc:Rhe},Mhe=class{constructor(e){this.variableNames=[],this.outputShape=[],this.uniforms="start : f32, step : f32,",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=[e],this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="linSpace"}getUserCode(){return`
|
|
${de("index")} {
|
|
if (index < uniforms.size) {
|
|
setOutputAtIndex(index, uniforms.start + f32(index) * uniforms.step);
|
|
}
|
|
}
|
|
`}};function Fhe(e){let{backend:t,attrs:a}=e,{start:n,stop:r,num:s}=a,i=(r-n)/(s-1),o=new Mhe(s),l=[{type:"float32",data:[n]},{type:"float32",data:[i]}];return t.runWebGPUProgram(o,[],"float32",l)}var _he={kernelName:Hi,backendName:"webgpu",kernelFunc:Fhe},$he=tt({opType:le.LOG,cpuKernelImpl:Eue}),Dhe={kernelName:ji,backendName:"webgpu",kernelFunc:$he},Phe=tt({opType:le.LOG1P}),Ohe={kernelName:qi,backendName:"webgpu",kernelFunc:Phe},zhe=ta({opType:_e.LOGICAL_AND,dtype:"bool"}),Lhe={kernelName:Xi,backendName:"webgpu",kernelFunc:zhe},Whe=tt({opType:le.LOGICAL_NOT}),Bhe={kernelName:Ki,backendName:"webgpu",kernelFunc:Whe},Vhe=ta({opType:_e.LOGICAL_OR}),Uhe={kernelName:Yi,backendName:"webgpu",kernelFunc:Vhe},hk=`
|
|
var powValue = 0.0;
|
|
let basis = uniforms.bias + uniforms.alpha * sum;
|
|
if (uniforms.beta == 0.5) {
|
|
powValue = inverseSqrt(basis);
|
|
} else if (uniforms.beta == 1.0) {
|
|
powValue = 1.0 / basis;
|
|
} else {
|
|
powValue = exp(log(basis) * (-uniforms.beta));
|
|
}
|
|
`,Ghe=class{constructor(e){this.outputShape=[],this.variableNames=["x"],this.uniforms="radius : i32, bias : f32, alpha : f32, beta : f32,",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="lrn"}getUserCode(){return`
|
|
${de("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getOutputCoords();
|
|
let b = coords[0];
|
|
let r = coords[1];
|
|
let c = coords[2];
|
|
let d = coords[3];
|
|
|
|
let x = getX(b, r, c, d);
|
|
var sum = 0.0;
|
|
for (var i = -uniforms.radius; i <= uniforms.radius; i = i + 1) {
|
|
let idx = d + i;
|
|
if (idx >= 0 && idx < uniforms.xShape[3]) {
|
|
let z = getX(b, r, c, idx);
|
|
sum = sum + z * z;
|
|
}
|
|
}
|
|
${hk}
|
|
|
|
setOutputAtIndex(index, x * powValue);
|
|
}
|
|
}
|
|
`}},Hhe=class{constructor(e,t){this.outputShape=[],this.variableNames=["x"],this.uniforms="radius : i32, bias : f32, alpha : f32, beta : f32,",this.workgroupSize=[256,1,1],this.maxAllowRadius=16,v.assert(t<=this.maxAllowRadius,()=>`Radius must be less than or equal to ${this.maxAllowRadius}, current radius is ${t}`),this.outputShape=e,this.elementsPerWorkgroup=this.workgroupSize[0]-2*this.maxAllowRadius,this.dispatchLayout={x:[3],y:[2],z:[0,1]},this.dispatch=ce(this.dispatchLayout,this.outputShape,[this.elementsPerWorkgroup,this.workgroupSize[1],this.workgroupSize[2]]),this.shaderKey="lrn_shared"}getUserCode(){return`
|
|
var <workgroup>lrnSub: array<f32, ${this.workgroupSize[0]}>;
|
|
const elementsPerWorkgroup = ${this.elementsPerWorkgroup};
|
|
const maxAllowRadius = ${this.maxAllowRadius};
|
|
|
|
${de()} {
|
|
let localDepth = i32(localId.x);
|
|
let workgroupDepth = i32(workgroupId.x) * elementsPerWorkgroup;
|
|
let xDepth = workgroupDepth + localDepth - maxAllowRadius;
|
|
let b = i32(globalId.z) / uniforms.xShape[1];
|
|
let r = i32(globalId.z) - b * uniforms.xShape[1];
|
|
let c = i32(globalId.y);
|
|
let d = workgroupDepth + localDepth;
|
|
|
|
var x = 0.0;
|
|
if (xDepth >= 0 && xDepth < uniforms.xShape[3]) {
|
|
x = getX(b, r, c, xDepth);
|
|
}
|
|
lrnSub[localDepth] = x;
|
|
workgroupBarrier();
|
|
|
|
if (localDepth < elementsPerWorkgroup && d < uniforms.outShape[3]) {
|
|
var sum = 0.0;
|
|
let index = localDepth + maxAllowRadius;
|
|
for (var i = -uniforms.radius; i <= uniforms.radius; i = i + 1) {
|
|
let z = lrnSub[index + i];
|
|
sum = sum + z * z;
|
|
}
|
|
${hk}
|
|
|
|
setOutputAtCoords(b, r, c, d, lrnSub[index] * powValue);
|
|
}
|
|
} `}};function jhe(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{depthRadius:s,bias:i,alpha:o,beta:l}=n,u;s>16?u=new Ghe(r.shape):u=new Hhe(r.shape,s);let p=[{type:"int32",data:[s]},{type:"float32",data:[i]},{type:"float32",data:[o]},{type:"float32",data:[l]}];return a.runWebGPUProgram(u,[r],r.dtype,p)}var qhe={kernelName:Zi,backendName:"webgpu",kernelFunc:jhe},Xhe=class{constructor(e){this.outputShape=[],this.variableNames=["inputImage","outputImage","dy"],this.uniforms="depthRadius : i32, bias : f32, alpha : f32, beta : f32,",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="lrn_grad"}getUserCode(){return`
|
|
${de("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getOutputCoords();
|
|
let b = coords[0];
|
|
let r = coords[1];
|
|
let c = coords[2];
|
|
|
|
let MIN_DEPTH_BEGIN = 0;
|
|
let MAX_DEPTH_END = uniforms.outShape[3];
|
|
var result = 0.0;
|
|
for (var d = MIN_DEPTH_BEGIN; d < MAX_DEPTH_END; d++) {
|
|
let depthBegin = max(MIN_DEPTH_BEGIN, d - uniforms.depthRadius);
|
|
let depthEnd = min(MAX_DEPTH_END, d + uniforms.depthRadius + 1);
|
|
|
|
var norm = 0.0;
|
|
for (var k = MIN_DEPTH_BEGIN; k < MAX_DEPTH_END; k++) {
|
|
if (k < depthBegin) {
|
|
continue;
|
|
} else if (k >= depthBegin && k < depthEnd) {
|
|
norm += getInputImage(b, r, c, k) * getInputImage(b, r, c, k);
|
|
} else {
|
|
break;
|
|
}
|
|
}
|
|
|
|
norm = uniforms.alpha * norm + uniforms.bias;
|
|
|
|
for (var k = MIN_DEPTH_BEGIN; k < MAX_DEPTH_END; k++) {
|
|
if (k < depthBegin) {
|
|
continue;
|
|
} else if (k >= depthBegin && k < depthEnd) {
|
|
var dyi = -2.0 * uniforms.alpha * uniforms.beta
|
|
* getInputImage(b, r, c, k) * getOutputImage(b, r, c, d) / norm;
|
|
if (k == d) {
|
|
dyi += pow(norm, -1.0 * uniforms.beta);
|
|
}
|
|
if (k == coords[3]) {
|
|
dyi *= getDy(b, r, c, d);
|
|
result += dyi;
|
|
}
|
|
} else {
|
|
break;
|
|
}
|
|
}
|
|
}
|
|
|
|
setOutputAtIndex(index, result);
|
|
}
|
|
}
|
|
`}};function Khe(e){let{inputs:t,backend:a,attrs:n}=e,{x:r,y:s,dy:i}=t,{depthRadius:o,bias:l,alpha:u,beta:p}=n,c=new Xhe(r.shape),d=[{type:"int32",data:[o]},{type:"float32",data:[l]},{type:"float32",data:[u]},{type:"float32",data:[p]}];return a.runWebGPUProgram(c,[r,s,i],r.dtype,d)}var Yhe={kernelName:cu,backendName:"webgpu",kernelFunc:Khe},Zhe=ta({opType:_e.MAX,cpuKernelImpl:Fue}),Jhe={kernelName:Qi,backendName:"webgpu",kernelFunc:Zhe};function Qhe(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{filterSize:s,strides:i,pad:o,dimRoundingMode:l}=n,u=1,p=S.computePool2DInfo(r.shape,s,i,u,o,l);return tk(r,p,"max",a)}var e0e={kernelName:eo,backendName:"webgpu",kernelFunc:Qhe};function t0e(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{filterSize:s,strides:i,pad:o,dataFormat:l,dimRoundingMode:u}=n,p=[1,1,1],c=S.computePool3DInfo(r.shape,s,i,p,o,u,l),d=new Dg(c,"max"),h=[{type:"int32",data:[c.strideDepth,c.strideHeight,c.strideWidth]},{type:"int32",data:[c.padInfo.front,c.padInfo.top,c.padInfo.left]},{type:"int32",data:[c.inDepth,c.inHeight,c.inWidth]},{type:"int32",data:[c.effectiveFilterDepth,c.effectiveFilterHeight,c.effectiveFilterWidth]}];return a.runWebGPUProgram(d,[r],r.dtype,h)}var a0e={kernelName:hu,backendName:"webgpu",kernelFunc:t0e},n0e=class{constructor(e){this.variableNames=["dy","maxPos"],this.uniforms=`strides : vec2<i32>, pads : vec2<i32>, dilations : vec2<i32>, filterDims : vec2<i32>,
|
|
outHeight : i32, outWidth : i32`,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e.inShape,this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="maxPool2DBackprop"}getUserCode(){return`
|
|
${de("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
let batch = coords[0];
|
|
let d = coords[3];
|
|
|
|
let dyRCCorner = vec2<i32>(coords.yz) - uniforms.pads;
|
|
let dyRCorner = dyRCCorner.x;
|
|
let dyCCorner = dyRCCorner.y;
|
|
|
|
// Convolve dy(?, ?, d) with pos mask(:, :, d) to get dx(xR, xC, d).
|
|
// ? = to be determined. : = across all values in that axis.
|
|
var dotProd = 0.0;
|
|
let lastIndex = uniforms.filterDims[0] * uniforms.filterDims[1] - 1;
|
|
for (var wR = 0; wR < uniforms.filterDims[0]; wR += uniforms.dilations[0]) {
|
|
let dyR = f32(dyRCorner + wR) / f32(uniforms.strides[0]);
|
|
|
|
if (dyR < 0.0 || dyR >= f32(uniforms.outHeight) || fract(dyR) > 0.0) {
|
|
continue;
|
|
}
|
|
let idyR = i32(dyR);
|
|
|
|
for (var wC = 0; wC < uniforms.filterDims[1]; wC += uniforms.dilations[1]) {
|
|
let dyC = f32(dyCCorner + wC) / f32(uniforms.strides[1]);
|
|
|
|
if (dyC < 0.0 || dyC >= f32(uniforms.outWidth) || fract(dyC) > 0.0) {
|
|
continue;
|
|
}
|
|
let idyC = i32(dyC);
|
|
|
|
let dyValue = getDy(batch, idyR, idyC, d);
|
|
let maxPosValue = lastIndex - i32(getMaxPos(batch, idyR, idyC, d));
|
|
|
|
// Get the current value, check it against the value from the
|
|
// position matrix.
|
|
let curPosValue = wR * uniforms.filterDims[1] + wC;
|
|
let mask = select(0.0, 1.0, maxPosValue == curPosValue);
|
|
dotProd += dyValue * mask;
|
|
}
|
|
}
|
|
setOutputAtIndex(index, dotProd);
|
|
}
|
|
}
|
|
`}},r0e=class{constructor(e){this.variableNames=["dy","maxPos"],this.uniforms=`strides : vec3<i32>, pads : vec3<i32>, filterDims : vec3<i32>,
|
|
outDepth : i32, outHeight : i32, outWidth : i32`,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e.inShape,this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="maxPool3DBackprop"}getUserCode(){return`
|
|
${de("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
let batch = coords.x;
|
|
let ch = coords.u;
|
|
|
|
let dyCorner = vec3<i32>(coords.y, coords.z, coords.w) - uniforms.pads;
|
|
let dyDCorner = dyCorner.x;
|
|
let dyRCorner = dyCorner.y;
|
|
let dyCCorner = dyCorner.z;
|
|
|
|
// Convolve dy(?, ?, ?, ch) with pos mask(:, :, :, d) to get
|
|
// dx(xD, xR, xC, ch).
|
|
// ? = to be determined. : = across all values in that axis.
|
|
var dotProd = 0.0;
|
|
let lastIndex = uniforms.filterDims[0] * uniforms.filterDims[1] * uniforms.filterDims[2] - 1;
|
|
|
|
for (var wD = 0; wD < uniforms.filterDims[0]; wD++) {
|
|
let dyD = f32(dyDCorner + wD) / f32(uniforms.strides[0]);
|
|
|
|
if (dyD < 0.0 || dyD >= f32(uniforms.outDepth) || fract(dyD) > 0.0) {
|
|
continue;
|
|
}
|
|
let idyD = i32(dyD);
|
|
|
|
for (var wR = 0; wR < uniforms.filterDims[1]; wR++) {
|
|
let dyR = f32(dyRCorner + wR) / f32(uniforms.strides[1]);
|
|
|
|
if (dyR < 0.0 || dyR >= f32(uniforms.outHeight) || fract(dyR) > 0.0) {
|
|
continue;
|
|
}
|
|
let idyR = i32(dyR);
|
|
|
|
for (var wC = 0; wC < uniforms.filterDims[2]; wC++) {
|
|
let dyC = f32(dyCCorner + wC) / f32(uniforms.strides[2]);
|
|
|
|
if (dyC < 0.0 || dyC >= f32(uniforms.outWidth) || fract(dyC) > 0.0) {
|
|
continue;
|
|
}
|
|
let idyC = i32(dyC);
|
|
|
|
let dyValue = getDy(batch, idyD, idyR, idyC, ch);
|
|
let maxPosValue = lastIndex - i32(getMaxPos(batch, idyD, idyR, idyC, ch));
|
|
|
|
// Get the current value, check it against the value from the
|
|
// position matrix.
|
|
let curPosValue = wD * uniforms.filterDims[1] * uniforms.filterDims[2] + wR * uniforms.filterDims[2] + wC;
|
|
let mask = select(0.0, 1.0, maxPosValue == curPosValue);
|
|
dotProd += dyValue * mask;
|
|
}
|
|
}
|
|
}
|
|
|
|
setOutputAtIndex(index, dotProd);
|
|
}
|
|
}
|
|
`}};function s0e(e){let{inputs:t,backend:a,attrs:n}=e,{dy:r,input:s}=t,i=s,{filterSize:o,strides:l,pad:u,dimRoundingMode:p}=n,c=[1,1,1],d=S.computePool3DInfo(i.shape,o,l,c,u,p),h=new Dg(d,"max",!0),f=[{type:"int32",data:[d.strideDepth,d.strideHeight,d.strideWidth]},{type:"int32",data:[d.padInfo.front,d.padInfo.top,d.padInfo.left]},{type:"int32",data:[d.inDepth,d.inHeight,d.inWidth]},{type:"int32",data:[d.effectiveFilterDepth,d.effectiveFilterHeight,d.effectiveFilterWidth]}],m=a.runWebGPUProgram(h,[i],"int32",f),g=new r0e(d);f=[{type:"int32",data:[d.strideDepth,d.strideHeight,d.strideWidth]},{type:"int32",data:[d.effectiveFilterDepth-1-d.padInfo.front,d.effectiveFilterHeight-1-d.padInfo.top,d.effectiveFilterWidth-1-d.padInfo.left]},{type:"int32",data:[d.effectiveFilterDepth,d.effectiveFilterHeight,d.effectiveFilterWidth]},{type:"int32",data:[d.outDepth]},{type:"int32",data:[d.outHeight]},{type:"int32",data:[d.outWidth]}];let y=a.runWebGPUProgram(g,[r,m],i.dtype,f);return a.disposeData(m.dataId),y}var i0e={kernelName:dp,backendName:"webgpu",kernelFunc:s0e};function o0e(e){let{inputs:t,backend:a,attrs:n}=e,{dy:r,input:s,output:i}=t,o=s;Fg([s,i],"maxPoolGrad");let{filterSize:l,strides:u,pad:p,dimRoundingMode:c}=n,d=S.computePool2DInfo(o.shape,l,u,1,p,c),h=new Kd(d,"max",!0),f=[{type:"int32",data:[d.strideHeight,d.strideWidth]},{type:"int32",data:[d.padInfo.top,d.padInfo.left]},{type:"int32",data:[d.dilationHeight,d.dilationWidth]},{type:"int32",data:[d.inHeight,d.inWidth]},{type:"int32",data:[d.effectiveFilterHeight,d.effectiveFilterWidth]}],m=a.runWebGPUProgram(h,[o],"int32",f),g=new n0e(d);f=[{type:"int32",data:[d.strideHeight,d.strideWidth]},{type:"int32",data:[d.effectiveFilterHeight-1-d.padInfo.top,d.effectiveFilterWidth-1-d.padInfo.left]},{type:"int32",data:[d.dilationHeight,d.dilationWidth]},{type:"int32",data:[d.effectiveFilterHeight,d.effectiveFilterWidth]},{type:"int32",data:[d.outHeight]},{type:"int32",data:[d.outWidth]}];let y=a.runWebGPUProgram(g,[r,m],o.dtype,f);return a.disposeData(m.dataId),y}var l0e={kernelName:Sh,backendName:"webgpu",kernelFunc:o0e};function u0e(e){let{inputs:t,backend:a,attrs:n}=e,{filterSize:r,strides:s,pad:i,includeBatchInIndex:o}=n,{x:l}=t;v.assert(l.shape.length===4,()=>`Error in maxPool: input must be rank 4 but got rank ${l.shape.length}.`);let u=[1,1];v.assert(S.eitherStridesOrDilationsAreOne(s,u),()=>`Error in maxPool: Either strides or dilations must be 1. Got strides ${s} and dilations '${u}'`);let p=S.computePool2DInfo(l.shape,r,s,u,i),c=[{type:"int32",data:[p.strideHeight,p.strideWidth]},{type:"int32",data:[p.padInfo.top,p.padInfo.left]},{type:"int32",data:[p.dilationHeight,p.dilationWidth]},{type:"int32",data:[p.inHeight,p.inWidth]},{type:"int32",data:[p.effectiveFilterHeight,p.effectiveFilterWidth]}],d=new Kd(p,"max",!1),h=a.runWebGPUProgram(d,[l],l.dtype,c);d=new Kd(p,"max",!0,!0,o);let f=a.runWebGPUProgram(d,[l],"int32",c);return[h,f]}var d0e={kernelName:pp,backendName:"webgpu",kernelFunc:u0e};function p0e(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{axis:s,keepDims:i}=n;return Jo(r,s,i,"min",a)}var c0e={kernelName:ao,backendName:"webgpu",kernelFunc:p0e},h0e=ta({opType:_e.MIN,cpuKernelImpl:_ue}),f0e={kernelName:no,backendName:"webgpu",kernelFunc:h0e},m0e=class{constructor(e,t,a){this.uniforms="",this.variableNames=["x"],this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=t.map((n,r)=>n[0]+e[r]+n[1]),this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.xShape=e,t.map((n,r)=>{this.uniforms+=` pad${r} : vec2<i32>,`}),this.offset=a==="reflect"?0:1,this.shaderKey=`mirrorPad_${a}`}getUserCode(){let e=this.xShape.length,t=this.xShape.map((l,u)=>`uniforms.pad${u}[0]`).join(","),a=this.xShape.map((l,u)=>`uniforms.pad${u}[0] + uniforms.xShape${e>1?`[${u}]`:""}`).join(","),n=e===1?"start":"start[i]",r=e===1?"end":"end[i]",s=e===1?"outC":"outC[i]",i=Jt(e),o=e>1?["coords[0]","coords[1]","coords[2]","coords[3]"].slice(0,e):"coords";return`
|
|
${de("index")} {
|
|
if (index < uniforms.size) {
|
|
let start = ${i}(${t});
|
|
let end = ${i}(${a});
|
|
var outC = getCoordsFromIndex(index);
|
|
for (var i = 0; i < ${e}; i = i + 1) {
|
|
if (${s} < ${n}) {
|
|
${s} = ${n} * 2 - ${s} - ${this.offset};
|
|
} else if(${s} >= ${r}) {
|
|
${s} = (${r} - 1) * 2 - ${s} + ${this.offset};
|
|
}
|
|
}
|
|
let coords = outC - start;
|
|
setOutputAtIndex(index, getX(${o}));
|
|
}
|
|
}
|
|
`}},g0e={kernelName:ro,backendName:"webgpu",kernelFunc:({inputs:e,attrs:t,backend:a})=>{let{x:n}=e,{paddings:r,mode:s}=t,i=a,o=r.map(u=>({type:"int32",data:[u[0],u[1]]})),l=new m0e(n.shape,r,s);return i.runWebGPUProgram(l,[n],n.dtype,o)}},y0e=ta({opType:_e.MOD}),x0e={kernelName:fu,backendName:"webgpu",kernelFunc:y0e},A0e=class{constructor(e,t){this.variableNames=["probs"],this.outputShape=[],this.uniforms="seed : f32, numOutcomes: i32,",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=[e,t],this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="multinomial"}getUserCode(){return`
|
|
//Based on the work of Dave Hoskins
|
|
//https://www.shadertoy.com/view/4djSRW
|
|
fn random (seed : f32, resultUV : vec2<f32>) -> f32 {
|
|
let HASHSCALE1 = 443.8975;
|
|
let p = resultUV * seed;
|
|
var p3 = fract(vec3<f32>(p.xyx) * HASHSCALE1);
|
|
p3 = p3 + dot(p3, p3.yzx + 19.19);
|
|
return fract((p3.x + p3.y) * p3.z);
|
|
}
|
|
|
|
${de("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getOutputCoords();
|
|
let batch = coords[0];
|
|
|
|
let resUV = vec2<f32>(f32(coords[1]) / f32(uniforms.outShape[1]),
|
|
f32(coords[0]) / f32(uniforms.outShape[0]));
|
|
let r = random(uniforms.seed, resUV);
|
|
var cdf = 0.0;
|
|
for (var i = 0; i < uniforms.numOutcomes - 1; i = i + 1) {
|
|
cdf = cdf + getProbs(batch, i);
|
|
|
|
if (r < cdf) {
|
|
setOutputAtIndexI32(index, i);
|
|
return;
|
|
}
|
|
}
|
|
|
|
// If no other event happened, last event happened.
|
|
setOutputAtIndexI32(index, uniforms.numOutcomes - 1);
|
|
}
|
|
}
|
|
`}},b0e=class{constructor(e){this.variableNames=["logits"],this.outputShape=e,this.dispatchLayout=be(this.outputShape),this.dispatch=[this.outputShape[0],1,1],this.outputShape[1]>=4096?this.workgroupSize=[256,1,1]:this.workgroupSize=[64,1,1],this.shaderKey="softmax"}getUserCode(){return`
|
|
var<workgroup> buf : array<f32, ${this.workgroupSize[0]}>;
|
|
var<workgroup> rowMaxShared : f32;
|
|
var<workgroup> rowSumShared : f32;
|
|
const blockSize = ${this.workgroupSize[0]};
|
|
${de("index")} {
|
|
let row = index / blockSize;
|
|
let tid = i32(localId.x);
|
|
let cols = uniforms.outShape[1];
|
|
|
|
var threadMax = -3.402823e+38f;
|
|
for (var col = tid; col < cols; col += blockSize) {
|
|
let value = getLogits(row, col);
|
|
threadMax = max(threadMax, value);
|
|
}
|
|
if (tid < cols) {
|
|
buf[tid] = threadMax;
|
|
}
|
|
workgroupBarrier();
|
|
|
|
var reduceSize = min(cols, blockSize);
|
|
for (var currSize = reduceSize >> 1; currSize > 0; currSize = reduceSize >> 1) {
|
|
reduceSize = currSize + (reduceSize & 1);
|
|
if (tid < currSize) {
|
|
buf[tid] = max(buf[tid], buf[tid + reduceSize]);
|
|
}
|
|
workgroupBarrier();
|
|
}
|
|
|
|
if (tid == 0) {
|
|
rowMaxShared = buf[0];
|
|
}
|
|
workgroupBarrier();
|
|
|
|
var threadSum = 0.0;
|
|
for (var col = tid; col < cols; col += blockSize) {
|
|
let subExp = exp(getLogits(row, col) - rowMaxShared);
|
|
threadSum += subExp;
|
|
}
|
|
buf[tid] = threadSum;
|
|
workgroupBarrier();
|
|
|
|
for (var currSize = blockSize >> 1; currSize > 0; currSize = currSize >> 1) {
|
|
if (tid < currSize) {
|
|
buf[tid] = buf[tid] + buf[tid + currSize];
|
|
}
|
|
workgroupBarrier();
|
|
}
|
|
|
|
if (tid == 0) {
|
|
rowSumShared = buf[0];
|
|
}
|
|
workgroupBarrier();
|
|
|
|
for (var col = tid; col < cols; col += blockSize) {
|
|
let value = exp(getLogits(row, col) - rowMaxShared) / rowSumShared;
|
|
setOutputAtCoords(row, col, value);
|
|
}
|
|
}
|
|
`}};function fk(e){let{inputs:t,backend:a,attrs:n}=e,{logits:r}=t,{dim:s}=n,i=we({inputs:{x:r},backend:a,attrs:{shape:[v.sizeFromShape(r.shape)/r.shape[s],r.shape[s]]}}),o=new b0e(i.shape),l=a.runWebGPUProgram(o,[i],r.dtype),u=we({inputs:{x:l},backend:a,attrs:{shape:r.shape}});return a.disposeData(i.dataId),a.disposeData(l.dataId),u}var v0e={kernelName:$o,backendName:"webgpu",kernelFunc:fk};function w0e(e){let{inputs:t,backend:a,attrs:n}=e,{logits:r}=t,{numSamples:s,seed:i,normalized:o}=n,l=o?r:fk({inputs:{logits:r},backend:a,attrs:{dim:r.shape.length-1}}),u=l.shape[0],p=l.shape[1],c=new A0e(u,s),d=[{type:"float32",data:[i]},{type:"int32",data:[p]}],h=a.runWebGPUProgram(c,[l],"int32",d);return o||a.disposeData(l.dataId),h}var k0e={kernelName:so,backendName:"webgpu",kernelFunc:w0e};function I0e(e){let{inputs:t,backend:a}=e,{x:n}=t;if(a.shouldExecuteOnCPU([n])){let s=a.tensorMap.get(n.dataId),[i,o]=Due(s.values,n.shape,n.dtype);return a.makeTensorInfo(o,n.dtype,i)}let r=new Hu(n.shape,le.NEG);return a.runWebGPUProgram(r,[n],n.dtype)}var S0e={kernelName:mu,backendName:"webgpu",kernelFunc:I0e};function C0e(e){console.warn("tf.nonMaxSuppression() in webgpu locks the UI thread. Call tf.nonMaxSuppressionAsync() instead");let{inputs:t,backend:a,attrs:n}=e,{boxes:r,scores:s}=t,{maxOutputSize:i,iouThreshold:o,scoreThreshold:l}=n,u=a.readSync(r.dataId),p=a.readSync(s.dataId),{selectedIndices:c}=Nn.nonMaxSuppressionV3Impl(u,p,i,o,l);return a.makeTensorInfo([c.length],"int32",new Int32Array(c))}var T0e={kernelName:lo,backendName:"webgpu",kernelFunc:C0e};function N0e(e){console.warn("tf.nonMaxSuppression() in webgpu locks the UI thread. Call tf.nonMaxSuppressionAsync() instead");let{inputs:t,backend:a,attrs:n}=e,{boxes:r,scores:s}=t,{maxOutputSize:i,iouThreshold:o,scoreThreshold:l,softNmsSigma:u}=n,p=a.readSync(r.dataId),c=a.readSync(s.dataId),d=i,h=o,f=l,m=u,{selectedIndices:g,selectedScores:y}=Nn.nonMaxSuppressionV5Impl(p,c,d,h,f,m);return[a.makeTensorInfo([g.length],"int32",new Int32Array(g)),a.makeTensorInfo([y.length],"float32",new Float32Array(y))]}var R0e={kernelName:uo,backendName:"webgpu",kernelFunc:N0e},E0e=class{constructor(e,t){this.variableNames=["x"],this.uniforms="onValue : f32, offValue : f32,",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=[e,t],this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="onehot"}getUserCode(){return`
|
|
${de("index")} {
|
|
if(index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
setOutputAtIndex(index, mix(uniforms.offValue, uniforms.onValue,
|
|
f32(i32(round(getX(coords.x))) == coords.y)));
|
|
}
|
|
}
|
|
`}};function M0e(e){let{inputs:t,backend:a,attrs:n}=e,{indices:r}=t,{dtype:s,depth:i,onValue:o,offValue:l}=n,u=v.sizeFromShape(r.shape),p=new E0e(u,i),c=we({inputs:{x:r},backend:a,attrs:{shape:[u]}}),d=[{type:"float32",data:[o]},{type:"float32",data:[l]}],h=a.runWebGPUProgram(p,[c],s,d);a.disposeData(c.dataId);let f=[...r.shape,i],m=we({inputs:{x:h},backend:a,attrs:{shape:f}});return a.disposeData(h.dataId),m}var F0e={kernelName:po,backendName:"webgpu",kernelFunc:M0e};function bh(e){let{inputs:t,backend:a}=e,{x:n}=t;if(n.dtype==="complex64"){let r=jp({inputs:{input:n},backend:a}),s=bh({inputs:{x:r},backend:a}),i=l0({inputs:{input:n},backend:a}),o=bh({inputs:{x:i},backend:a}),l=Zo({inputs:{real:s,imag:o},backend:a});return a.disposeData(r.dataId),a.disposeData(s.dataId),a.disposeData(i.dataId),a.disposeData(o.dataId),l}else return Mn({attrs:{shape:n.shape,dtype:n.dtype,value:n.dtype==="string"?"":0},backend:a})}var _0e={kernelName:Mu,backendName:"webgpu",kernelFunc:bh};function mk(e){let{inputs:t,backend:a}=e,{x:n}=t;if(n.dtype==="string")throw new Error("onesLike is not supported under string dtype");if(n.dtype==="complex64"){let r=jp({inputs:{input:n},backend:a}),s=mk({inputs:{x:r},backend:a}),i=l0({inputs:{input:n},backend:a}),o=bh({inputs:{x:i},backend:a}),l=Zo({inputs:{real:s,imag:o},backend:a});return a.disposeData(r.dataId),a.disposeData(s.dataId),a.disposeData(i.dataId),a.disposeData(o.dataId),l}else return Mn({attrs:{shape:n.shape,dtype:n.dtype,value:1},backend:a})}var $0e={kernelName:yu,backendName:"webgpu",kernelFunc:mk};function D0e(e){let{inputs:t,backend:a,attrs:n}=e,{axis:r}=n;if(t.length===1)return P1({inputs:{input:t[0]},backend:a,attrs:{dim:r}});let s=t[0].shape,i=t[0].dtype;t.forEach(p=>{v.assertShapesMatch(s,p.shape,"All tensors passed to stack must have matching shapes"),v.assert(i===p.dtype,()=>"All tensors passed to stack must have matching dtypes")});let o=[],l=t.map(p=>{let c=P1({inputs:{input:p},backend:a,attrs:{dim:r}});return o.push(c),c}),u=rk({inputs:l,backend:a,attrs:{axis:r}});return o.forEach(p=>a.disposeData(p.dataId)),u}var P0e={kernelName:xu,backendName:"webgpu",kernelFunc:D0e},O0e=class{constructor(e,t){this.variableNames=["x"],this.uniforms="constantValue : f32,",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=t.map((a,n)=>a[0]+e[n]+a[1]),this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),t.map((a,n)=>{this.uniforms+=` pad${n} : vec2<i32>,`}),this.xShape=e,this.shaderKey="pad"}getUserCode(){let e=this.xShape.length,t=Jt(e),a=this.xShape.map((u,p)=>`uniforms.pad${p}[0]`).join(","),n=this.xShape.map((u,p)=>`uniforms.pad${p}[0] + uniforms.xShape${e>1?`[${p}]`:""}`).join(","),r=e>1?`${t}(${a})`:`${a}`,s=e>1?`${t}(${n})`:`${n}`,i=e>1?"any(outC < start)":"outC < start",o=e>1?"any(outC >= end)":"outC >= end",l=e>1?["coords[0]","coords[1]","coords[2]","coords[3]"].slice(0,e):"coords";return`
|
|
${de("index")} {
|
|
if (index < uniforms.size) {
|
|
let start = ${r};
|
|
let end = ${s};
|
|
let outC = getCoordsFromIndex(index);
|
|
|
|
if (${i} || ${o}) {
|
|
setOutputAtIndex(index, uniforms.constantValue);
|
|
} else {
|
|
let coords = outC - start;
|
|
setOutputAtIndex(index, getX(${l}));
|
|
}
|
|
}
|
|
}
|
|
`}},gk=e=>{let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{paddings:s,constantValue:i}=n;if(s.every(u=>v.arraysEqual(u,[0,0])))return tn({inputs:{x:r},backend:a});if(v.sizeFromShape(r.shape)===0){let u=s.map((p,c)=>p[0]+r.shape[c]+p[1]);return Mn({backend:a,attrs:{shape:u,value:i,dtype:r.dtype}})}let o=[{type:"float32",data:[i]}];s.map(u=>o.push({type:"int32",data:[u[0],u[1]]}));let l=new O0e(r.shape,s);return a.runWebGPUProgram(l,[r],r.dtype,o)},z0e={kernelName:co,backendName:"webgpu",kernelFunc:gk},L0e=ta({opType:_e.POW}),W0e={kernelName:ho,backendName:"webgpu",kernelFunc:L0e};function B0e(e){let{inputs:t,backend:a}=e,{x:n,alpha:r}=t,s=new xh(_e.PRELU,n.shape,r.shape);return a.runWebGPUProgram(s,[n,r],"float32")}var V0e={kernelName:fo,backendName:"webgpu",kernelFunc:B0e};function U0e(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{axis:s,keepDims:i}=n;return Jo(r,s,i,"prod",a)}var G0e={kernelName:mo,backendName:"webgpu",kernelFunc:U0e},H0e=e=>{let{backend:t,attrs:a}=e,{start:n,stop:r,step:s,dtype:i}=a,o=zue(n,r,s,i);return t.makeTensorInfo([o.length],i,o)},j0e={kernelName:Au,backendName:"webgpu",kernelFunc:H0e},q0e=ta({opType:_e.DIV}),X0e={kernelName:Ci,backendName:"webgpu",kernelFunc:q0e},K0e=tt({opType:le.RECIPROCAL}),Y0e={kernelName:go,backendName:"webgpu",kernelFunc:K0e},Z0e=tt({opType:le.RELU}),J0e={kernelName:yo,backendName:"webgpu",kernelFunc:Z0e},Q0e=tt({opType:le.RELU6}),efe={kernelName:bo,backendName:"webgpu",kernelFunc:Q0e},tfe=class{constructor(e,t,a){this.variableNames=["x"],this.uniforms="adjustHeightWidth : vec2<f32>, halfPixelCenters : f32,",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=[e[0],t,a,e[3]],this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="resizeBilinear"}getUserCode(){return`
|
|
${de("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
let b = coords[0];
|
|
let d = coords[3];
|
|
let rc = coords.yz;
|
|
|
|
let effectiveInSize = vec2<f32>(
|
|
f32(uniforms.xShape.y) - uniforms.adjustHeightWidth[0],
|
|
f32(uniforms.xShape.z) - uniforms.adjustHeightWidth[1]);
|
|
|
|
let effectiveOutSize = vec2<f32>(
|
|
f32(uniforms.outShape.y) - uniforms.adjustHeightWidth[0],
|
|
f32(uniforms.outShape.z) - uniforms.adjustHeightWidth[1]);
|
|
|
|
let effectiveInputOverOutputRatioRC =
|
|
effectiveInSize / effectiveOutSize;
|
|
|
|
// Fractional source index
|
|
let sourceFracIndexRC =
|
|
(vec2<f32>(rc) + vec2<f32>(uniforms.halfPixelCenters)) *
|
|
effectiveInputOverOutputRatioRC - vec2<f32>(uniforms.halfPixelCenters);
|
|
|
|
// Compute the four integer indices.
|
|
let sourceFloorRC = vec2<i32>(sourceFracIndexRC);
|
|
let sourceCeilRC = vec2<i32>(
|
|
min(vec2<f32>(uniforms.xShape.yz) - vec2<f32>(1.0), ceil(sourceFracIndexRC)));
|
|
|
|
let topLeft = getX(b, sourceFloorRC.x, sourceFloorRC.y, d);
|
|
let bottomLeft = getX(b, sourceCeilRC.x, sourceFloorRC.y, d);
|
|
let topRight = getX(b, sourceFloorRC.x, sourceCeilRC.y, d);
|
|
let bottomRight = getX(b, sourceCeilRC.x, sourceCeilRC.y, d);
|
|
|
|
let fracRC = sourceFracIndexRC - vec2<f32>(sourceFloorRC);
|
|
|
|
let top = topLeft + (topRight - topLeft) * fracRC.y;
|
|
let bottom = bottomLeft + (bottomRight - bottomLeft) * fracRC.y;
|
|
let newValue = top + (bottom - top) * fracRC.x;
|
|
|
|
setOutputAtIndex(index, newValue);
|
|
}
|
|
}
|
|
`}};function afe(e){let{inputs:t,backend:a,attrs:n}=e,{images:r}=t,{alignCorners:s,size:i,halfPixelCenters:o}=n,[l,u]=i,p=s&&l>1?1:0,c=s&&u>1?1:0,d=[{type:"float32",data:[p,c]},{type:"float32",data:[o?.5:0]}],h=new tfe(r.shape,l,u);return a.runWebGPUProgram(h,[r],"float32",d)}var nfe={kernelName:Ao,backendName:"webgpu",kernelFunc:afe},rfe=class{constructor(e,t){this.variableNames=["dy"],this.uniforms=`effectiveXSize : vec2<i32>, effectiveYSize : vec2<i32>, heightScale : f32, widthScale : f32,
|
|
invHeightScale : f32, invWidthScale : f32, winHeight : i32, winWidth : i32,`,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.alignCorners=t,this.shaderKey=`resizeBilinearBackprop_${t}`}getUserCode(){return`
|
|
${de("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getOutputCoords();
|
|
let b = coords[0];
|
|
let d = coords[3];
|
|
let r = coords[1];
|
|
let c = coords[2];
|
|
|
|
var accumulator = 0.0;
|
|
|
|
// Compute bounds for where in dy we will look
|
|
let startRLerp = floor(f32(r) * uniforms.invHeightScale);
|
|
let startDyR = i32(startRLerp - f32(uniforms.winHeight / 2));
|
|
|
|
let startCLerp = floor(f32(c) * uniforms.invWidthScale);
|
|
let startDyC = i32(startCLerp - f32(uniforms.winWidth / 2));
|
|
|
|
// Loop over dy
|
|
for (var dyROffset = 0; dyROffset < uniforms.winHeight; dyROffset++) {
|
|
let dyR = startDyR + dyROffset;
|
|
|
|
// Guard against the window exceeding the bounds of dy
|
|
if (dyR < 0 || dyR >= uniforms.dyShape[1]) {
|
|
continue;
|
|
}
|
|
|
|
for (var dyCOffset = 0; dyCOffset < uniforms.winWidth; dyCOffset++) {
|
|
let dyC = startDyC + dyCOffset;
|
|
|
|
// Guard against the window exceeding the bounds of dy
|
|
if (dyC < 0 || dyC >= uniforms.dyShape[2]) {
|
|
continue;
|
|
}
|
|
|
|
let dxR = f32(dyR) * uniforms.heightScale;
|
|
let topDxRIndex = i32(floor(dxR));
|
|
let bottomDxRIndex = i32(min(ceil(dxR), f32(uniforms.outShape[1] - 1)));
|
|
let dxRLerp = dxR - f32(topDxRIndex);
|
|
let inverseDxRLerp = 1.0 - dxRLerp;
|
|
|
|
let dxC = f32(dyC) * uniforms.widthScale;
|
|
let leftDxCIndex = i32(floor(dxC));
|
|
let rightDxCIndex = i32(min(ceil(dxC), f32(uniforms.outShape[2] - 1)));
|
|
let dxCLerp = dxC - f32(leftDxCIndex);
|
|
let inverseDxCLerp = 1.0 - dxCLerp;
|
|
|
|
if (r == topDxRIndex && c == leftDxCIndex) {
|
|
// topLeft
|
|
accumulator +=
|
|
getDy(b, dyR, dyC, d) * inverseDxRLerp * inverseDxCLerp;
|
|
}
|
|
|
|
if (r == topDxRIndex && c == rightDxCIndex) {
|
|
// topRight
|
|
accumulator += getDy(b, dyR, dyC, d) * inverseDxRLerp * dxCLerp;
|
|
}
|
|
|
|
if (r == bottomDxRIndex && c == leftDxCIndex) {
|
|
// bottomLeft
|
|
accumulator += getDy(b, dyR, dyC, d) * dxRLerp * inverseDxCLerp;
|
|
}
|
|
|
|
if (r == bottomDxRIndex && c == rightDxCIndex) {
|
|
// bottomRight
|
|
accumulator += getDy(b, dyR, dyC, d) * dxRLerp * dxCLerp;
|
|
}
|
|
}
|
|
}
|
|
// End loop over dy
|
|
|
|
setOutputAtIndex(index, accumulator);
|
|
}
|
|
}
|
|
`}};function sfe(e){let{inputs:t,backend:a,attrs:n}=e,{images:r,dy:s}=t,{alignCorners:i}=n,[,o,l]=r.shape,[,u,p]=s.shape,c=[i&&u>1?o-1:o,i&&p>1?l-1:l],d=[i&&u>1?u-1:u,i&&p>1?p-1:p],h=c[0]/d[0],f=c[1]/d[1],m=1/h,g=1/f,y=Math.ceil(m)*2+2,x=Math.ceil(g)*2+2,A=new rfe(r.shape,i),b=[{type:"int32",data:c},{type:"int32",data:d},{type:"float32",data:[h]},{type:"float32",data:[f]},{type:"float32",data:[m]},{type:"float32",data:[g]},{type:"int32",data:[y]},{type:"int32",data:[x]}];return a.runWebGPUProgram(A,[s],s.dtype,b)}var ife={kernelName:wu,backendName:"webgpu",kernelFunc:sfe},ofe=class{constructor(e,t,a,n){this.variableNames=["x"],this.uniforms="adjustHeightWidth : vec2<f32>, roundBase : f32,",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=[e[0],t,a,e[3]],this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.halfPixelCenters=n,this.shaderKey=`resizeNearest_${n}`}getUserCode(){let e;return this.halfPixelCenters?e="max((vec2<f32>(rc) + vec2<f32>(0.5)) * effectiveInputOverOutputRatioRC, vec2<f32>(0.0))":e="vec2<f32>(rc) * effectiveInputOverOutputRatioRC",`
|
|
${de("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
let b = coords[0];
|
|
let d = coords[3];
|
|
let rc = coords.yz;
|
|
|
|
let effectiveInSize = vec2<f32>(
|
|
f32(uniforms.xShape.y) - uniforms.adjustHeightWidth[0],
|
|
f32(uniforms.xShape.z) - uniforms.adjustHeightWidth[1]);
|
|
|
|
let effectiveOutSize = vec2<f32>(
|
|
f32(uniforms.outShape.y) - uniforms.adjustHeightWidth[0],
|
|
f32(uniforms.outShape.z) - uniforms.adjustHeightWidth[1]);
|
|
|
|
let effectiveInputOverOutputRatioRC =
|
|
effectiveInSize / effectiveOutSize;
|
|
|
|
// Fractional source index
|
|
let sourceFracIndexRC = ${e};
|
|
|
|
// Compute the coordinators of nearest neighbor point.
|
|
let inputShapeRC = vec2<f32>(f32(uniforms.xShape.y), f32(uniforms.xShape.z));
|
|
let sourceNearestRC = vec2<i32>(
|
|
min(inputShapeRC - 1.0, floor(sourceFracIndexRC + uniforms.roundBase)));
|
|
let newValue = getX(b, sourceNearestRC.x, sourceNearestRC.y, d);
|
|
|
|
setOutputAtIndex(index, newValue);
|
|
}
|
|
}
|
|
`}};function lfe(e){let{inputs:t,backend:a,attrs:n}=e,{images:r}=t,{alignCorners:s,halfPixelCenters:i,size:o}=n,[l,u]=o,p=s&&l>1?1:0,c=s&&u>1?1:0,d=[{type:"float32",data:[p,c]},{type:"float32",data:[s?.5:0]}],h=new ofe(r.shape,l,u,i);return a.runWebGPUProgram(h,[r],r.dtype,d)}var ufe={kernelName:xo,backendName:"webgpu",kernelFunc:lfe},dfe=class{constructor(e,t){this.variableNames=["dy"],this.uniforms=`effectiveXSize : vec2<i32>, effectiveYSize : vec2<i32>, invHeightScale : f32, invWidthScale : f32,
|
|
winHeight : i32, winWidth : i32,`,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.alignCorners=t,this.shaderKey=`resizeNearestNeigborBackprop_${t}`}getUserCode(){return`
|
|
${de("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getOutputCoords();
|
|
let b = coords[0];
|
|
let d = coords[3];
|
|
let r = coords[1];
|
|
let c = coords[2];
|
|
|
|
var accumulator = 0.0;
|
|
|
|
// Compute bounds for where in dy we will look
|
|
let startRLerp = floor(f32(r) * uniforms.invHeightScale);
|
|
let startDyR = i32(floor(startRLerp - f32(uniforms.winHeight / 2)));
|
|
|
|
let startCLerp = floor(f32(c) * uniforms.invWidthScale);
|
|
let startDyC = i32(floor(startCLerp - f32(uniforms.winWidth / 2)));
|
|
|
|
// Loop over dy
|
|
for (var dyROffset = 0; dyROffset < uniforms.winHeight; dyROffset++) {
|
|
let dyR = startDyR + dyROffset;
|
|
|
|
// Guard against the window exceeding the bounds of dy
|
|
if (dyR < 0 || dyR >= uniforms.dyShape[1]) {
|
|
continue;
|
|
}
|
|
|
|
for (var dyCOffset = 0; dyCOffset < uniforms.winWidth; dyCOffset++) {
|
|
let dyC = startDyC + dyCOffset;
|
|
|
|
// Guard against the window exceeding the bounds of dy
|
|
if (dyC < 0 || dyC >= uniforms.dyShape[2]) {
|
|
continue;
|
|
}
|
|
|
|
let sourceFracRow = f32(uniforms.effectiveXSize[0]) *
|
|
(f32(dyR) / f32(uniforms.effectiveYSize[0]));
|
|
|
|
let sourceFracCol = f32(uniforms.effectiveXSize[1]) *
|
|
(f32(dyC) / f32(uniforms.effectiveYSize[1]));
|
|
|
|
let sourceNearestRow =
|
|
i32(min(f32(uniforms.outShape[1] - 1),
|
|
${this.alignCorners?"floor(sourceFracRow + 0.5)":"floor(sourceFracRow)"}));
|
|
|
|
let sourceNearestCol =
|
|
i32(min(f32(uniforms.outShape[2] - 1),
|
|
${this.alignCorners?"floor(sourceFracCol + 0.5)":"floor(sourceFracCol)"}));
|
|
|
|
if (r == sourceNearestRow && c == sourceNearestCol) {
|
|
accumulator += getDy(b, dyR, dyC, d);
|
|
}
|
|
}
|
|
}
|
|
// End loop over dy
|
|
|
|
setOutputAtIndex(index, accumulator);
|
|
}
|
|
}
|
|
`}};function pfe(e){let{inputs:t,backend:a,attrs:n}=e,{images:r,dy:s}=t,{alignCorners:i}=n,[,o,l]=r.shape,[,u,p]=s.shape,c=[i&&u>1?o-1:o,i&&p>1?l-1:l],d=[i&&u>1?u-1:u,i&&p>1?p-1:p],h=c[0]/d[0],f=c[1]/d[1],m=1/h,g=1/f,y=Math.ceil(m)*2+2,x=Math.ceil(g)*2+2,A=new dfe(r.shape,i),b=[{type:"int32",data:c},{type:"int32",data:d},{type:"float32",data:[m]},{type:"float32",data:[g]},{type:"int32",data:[y]},{type:"int32",data:[x]}];return a.runWebGPUProgram(A,[s],s.dtype,b)}var cfe={kernelName:vu,backendName:"webgpu",kernelFunc:pfe},hfe=class{constructor(e){this.variableNames=["x"],this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.uniforms=" axis : vec4<i32>,",this.shaderKey="reverse"}getUserCode(){return`
|
|
|
|
// Using uniform variables as judging conditions, so the function has
|
|
// coherent execution within all threads.
|
|
fn getReverseCoords(coords : vec4<i32>) -> vec4<i32> {
|
|
var reverseCoords = coords;
|
|
if (uniforms.axis[0] == 1) {
|
|
reverseCoords[0] = uniforms.xShape[0] - coords[0] - 1;
|
|
}
|
|
if (uniforms.axis[1] == 1) {
|
|
reverseCoords[1] = uniforms.xShape[1] - coords[1] - 1;
|
|
}
|
|
if (uniforms.axis[2] == 1) {
|
|
reverseCoords[2] = uniforms.xShape[2] - coords[2] - 1;
|
|
}
|
|
if (uniforms.axis[3] == 1) {
|
|
reverseCoords[3] = uniforms.xShape[3] - coords[3] - 1;
|
|
}
|
|
|
|
return reverseCoords;
|
|
}
|
|
|
|
${de("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
let reverseCoords = getReverseCoords(coords);
|
|
setOutputAtIndex(index, getX(reverseCoords[0],
|
|
reverseCoords[1], reverseCoords[2], reverseCoords[3]));
|
|
}
|
|
}
|
|
`}};function ffe(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{dims:s}=n,i=r.shape.length;if(i===0)return tn({inputs:{x:r},backend:a});let o=r.shape,l=[1,1,1,1];o.forEach((g,y)=>{let x=y+4-i;l[x]=g});let u=v.parseAxisParam(s,r.shape),p=[0,0,0,0];u.forEach(g=>{let y=g+4-i;p[y]=1});let c=[{type:"int32",data:p}],d=we({inputs:{x:r},backend:a,attrs:{shape:l}}),h=new hfe(l),f=a.runWebGPUProgram(h,[d],d.dtype,c);a.disposeData(d.dataId);let m=we({inputs:{x:f},backend:a,attrs:{shape:o}});return a.disposeData(f.dataId),m}var mfe={kernelName:vo,backendName:"webgpu",kernelFunc:ffe},gfe=class{constructor(e,t){this.outputShape=[],this.variableNames=["x"],this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.uniforms=`centerX : f32, centerY : f32, sinRadians : f32,
|
|
cosRadians : f32,`,this.shaderKey="rotate",this.outputShape=e,typeof t=="number"?(this.uniforms+=" fillValue : f32,",this.fillSnippet="var outputValue = uniforms.fillValue;",this.shaderKey+="_float"):(this.uniforms+=" fillValue : vec3<f32>,",this.fillSnippet="var outputValue = uniforms.fillValue[coords[3]];",this.shaderKey+="_vec3")}getUserCode(){return`
|
|
${de("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
let coordXFloat = (f32(coords[2]) - uniforms.centerX) *
|
|
uniforms.cosRadians - (f32(coords[1]) - uniforms.centerY) *
|
|
uniforms.sinRadians;
|
|
let coordYFloat = (f32(coords[2]) - uniforms.centerX) *
|
|
uniforms.sinRadians + (f32(coords[1]) - uniforms.centerY) *
|
|
uniforms.cosRadians;
|
|
let coordX = i32(round(coordXFloat + uniforms.centerX));
|
|
let coordY = i32(round(coordYFloat + uniforms.centerY));
|
|
${this.fillSnippet}
|
|
if(coordX >= 0 && coordX < uniforms.xShape[2] && coordY >= 0 &&
|
|
coordY < uniforms.xShape[1]) {
|
|
outputValue = getX(coords[0], coordY, coordX, coords[3]);
|
|
}
|
|
setOutputAtIndex(index, outputValue);
|
|
}
|
|
}
|
|
`}},yfe={kernelName:Uo,backendName:"webgpu",kernelFunc:({inputs:e,attrs:t,backend:a})=>{let{image:n}=e,{radians:r,fillValue:s,center:i}=t,o=a,l=new gfe(n.shape,s),[u,p]=S.getImageCenter(i,n.shape[1],n.shape[2]),c=[{type:"float32",data:[u]},{type:"float32",data:[p]},{type:"float32",data:[Math.sin(r)]},{type:"float32",data:[Math.cos(r)]}];return typeof s=="number"?c.push({type:"float32",data:[Number.parseFloat(s.toFixed(2))]}):c.push({type:"float32",data:s}),o.runWebGPUProgram(l,[n],n.dtype,c)}},xfe=tt({opType:le.ROUND}),Afe={kernelName:wo,backendName:"webgpu",kernelFunc:xfe},bfe=tt({opType:le.RSQRT,cpuKernelImpl:Lue}),vfe={kernelName:ko,backendName:"webgpu",kernelFunc:bfe},Nd=class{constructor(e,t,a,n,r,s,i,o=!0){this.variableNames=["updates","indices"],this.workgroupSize=[64,1,1],this.atomic=!0,this.outputShape=s,this.type=i,this.sumDupeIndices=o,this.dispatchLayout=be(e),this.dispatch=ce(this.dispatchLayout,e,this.workgroupSize),this.sliceDimGreaterThanOne=t>1,this.shaderKey=`scatter_${a}_${n}_${this.sliceDimGreaterThanOne}_${i}_${o}`;let l=Jt(r.length);this.uniforms=`sliceDim : i32, strides: ${l}, updatesSize: i32,`,this.updatesRank=n,this.indicesRank=a}getUserCode(){let e="";this.indicesRank===1?e="coords[0]":this.indicesRank===2&&(e="coords[0], j");let t=`getIndices(${e})`,a=this.sliceDimGreaterThanOne?"uniforms.strides[j]":"uniforms.strides",n="",r="";this.dispatchLayout.x.length===1?(n="flattenedIndex",r=`
|
|
fn getUpdatesCoordsFromFlatIndex(index : i32) -> i32 {
|
|
return index;
|
|
}
|
|
`):this.dispatchLayout.x.length===2&&(n="vec2<i32>(flattenedIndex, coords[1])",r=`
|
|
fn getUpdatesCoordsFromFlatIndex(index : i32) -> vec2<i32> {
|
|
// N.B. |updates| could be a scalar tensor, conceptually representing a
|
|
// 2D tensor with all values equal to that. By design, its size must be
|
|
// the same as |outShape[1]| in one dimension, and |indicesShape[0]|
|
|
// gives the other.
|
|
let sliceSize = uniforms.outShape[1];
|
|
let d0 = index / sliceSize;
|
|
let d1 = index - d0 * sliceSize;
|
|
return vec2<i32>(d0, d1);
|
|
}
|
|
`);let s=`getUpdates(${Array.from({length:this.updatesRank},(i,o)=>`coords[${o}]`).join(", ")})`;return`
|
|
${r}
|
|
${de("index")} {
|
|
if (index < uniforms.updatesSize) {
|
|
let coords = getUpdatesCoordsFromFlatIndex(index);
|
|
var flattenedIndex = 0;
|
|
for (var j = 0; j < uniforms.sliceDim; j = j + 1) {
|
|
let indexInside = i32(round(${t}));
|
|
flattenedIndex = flattenedIndex + indexInside * ${a};
|
|
}
|
|
let updateValue =
|
|
${Nl(this.type)}(${s});
|
|
let flatIndex = getOutputIndexFromCoords(${n});
|
|
|
|
${this.sumDupeIndices?Gu("&result[flatIndex]","updateValue",this.type):"atomicStore(&result[flatIndex], bitcast<i32>(updateValue));"}
|
|
}
|
|
}`}};function wfe(e){let{inputs:t,backend:a,attrs:n}=e,{indices:r,updates:s}=t,{shape:i}=n,{sliceRank:o,numUpdates:l,sliceSize:u,strides:p,outputSize:c}=S.calculateShapes(s,r,i),d=[c/u,u];if(c===0)return a.makeTensorInfo(i,r.dtype);let h=we({inputs:{x:r},backend:a,attrs:{shape:[l,o]}}),f=we({inputs:{x:s},backend:a,attrs:{shape:[l,u]}}),m=f.dtype,g=Mn({backend:a,attrs:{shape:d,value:0,dtype:m}}),y=v.sizeFromShape(f.shape),x=[{type:"int32",data:[o]},{type:"int32",data:p},{type:"int32",data:[y]}],A=new Nd(f.shape,o,h.shape.length,f.shape.length,p,d,m),b=a.runWebGPUProgram(A,[f,h],m,x,g),w=we({inputs:{x:b},backend:a,attrs:{shape:i}});return a.disposeData(h.dataId),a.disposeData(f.dataId),a.disposeData(b.dataId),w}var kfe={kernelName:Io,backendName:"webgpu",kernelFunc:wfe},Ife=class{constructor(e,t){this.outputShape=[],this.variableNames=["sortedSequence","values"],this.uniforms="numInputs : i32,",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.side=t,this.shaderKey=`search_sorted_${t}`}getUserCode(){return`
|
|
fn findBound(batch: i32, value: f32) -> i32 {
|
|
var left = i32(0);
|
|
var right = uniforms.numInputs;
|
|
while (left < right) {
|
|
var mid = (left + right) / 2;
|
|
if (getSortedSequence(batch, mid) ${this.side==="left"?"<":"<="} value) {
|
|
left = mid + 1;
|
|
} else {
|
|
right = mid;
|
|
}
|
|
}
|
|
return right;
|
|
}
|
|
|
|
${de("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
let value = getValuesByOutputIndex(index);
|
|
setOutputAtIndexI32(index, findBound(coords[0], value));
|
|
}
|
|
}
|
|
`}};function Sfe(e){let{inputs:t,backend:a,attrs:n}=e,{sortedSequence:r,values:s}=t,{side:i}=n,o=new Ife([s.shape[0],s.shape[1]],i),l=[{type:"int32",data:[r.shape[1]]}];return a.runWebGPUProgram(o,[r,s],"int32",l)}var Cfe={kernelName:Co,backendName:"webgpu",kernelFunc:Sfe},Tfe=class{constructor(e,t,a){this.variableNames=["c","a","b"],this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=t,this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.cRank=e,this.rank=a,this.shaderKey="select"}getUserCode(){let e,t;if(this.rank>4)throw Error(`Where for rank ${this.rank} is not yet supported`);if(this.rank===1)t="resRC",e="resRC";else{let a=["resRC.x","resRC.y","resRC.z","resRC.w"],n=[],r=[];for(let s=0;s<this.outputShape.length;s++)r.push(`${a[s]}`),s<this.cRank&&n.push(`${a[s]}`);e=n.join(),t=r.join()}return`
|
|
${de("index")} {
|
|
if (index < uniforms.size) {
|
|
let resRC = getCoordsFromIndex(index);
|
|
let cVal = getC(${e});
|
|
if (cVal >= 1.0) {
|
|
setOutputAtIndex(index, getA(${t}));
|
|
} else {
|
|
setOutputAtIndex(index, getB(${t}));
|
|
}
|
|
}
|
|
}
|
|
`}};function Nfe(e){let{inputs:t,backend:a}=e,{condition:n,t:r,e:s}=t,i=new Tfe(n.shape.length,r.shape,r.shape.length);return a.runWebGPUProgram(i,[n,r,s],pa(r.dtype,s.dtype))}var Rfe={kernelName:ku,backendName:"webgpu",kernelFunc:Nfe},Efe=tt({opType:le.SELU}),Mfe={kernelName:To,backendName:"webgpu",kernelFunc:Efe},Ffe=tt({opType:le.SIGMOID}),_fe={kernelName:Eo,backendName:"webgpu",kernelFunc:Ffe},$fe=tt({opType:le.SIGN}),Dfe={kernelName:Ro,backendName:"webgpu",kernelFunc:$fe},Pfe=tt({opType:le.SIN}),Ofe={kernelName:No,backendName:"webgpu",kernelFunc:Pfe},zfe=tt({opType:le.SINH}),Lfe={kernelName:Su,backendName:"webgpu",kernelFunc:zfe},Wfe=tt({opType:le.SOFTPLUS}),Bfe={kernelName:Mo,backendName:"webgpu",kernelFunc:Wfe},Vfe=e=>{let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{blockShape:s,paddings:i}=n;v.assert(r.shape.length<=4,()=>"spaceToBatchND for rank > 4 with a WebGPU backend not implemented yet");let o=s.reduce((y,x)=>y*x),l=[[0,0]];l.push(...i);for(let y=1+s.length;y<r.shape.length;++y)l.push([0,0]);let u=[],p=gk({inputs:{x:r},backend:a,attrs:{paddings:l,constantValue:0}}),c=S.getReshaped(p.shape,s,o,!1),d=S.getPermuted(c.length,s.length,!1),h=S.getReshapedPermuted(p.shape,s,o,!1),f=we({inputs:{x:p},backend:a,attrs:{shape:c}}),m=Bn({inputs:{x:f},backend:a,attrs:{perm:d}}),g=we({inputs:{x:m},backend:a,attrs:{shape:h}});return u.push(p),u.push(f),u.push(m),u.forEach(y=>a.disposeData(y.dataId)),g},Ufe={kernelName:Cu,backendName:"webgpu",kernelFunc:Vfe},Gfe=class{constructor(e,t){this.variableNames=["A"],this.workgroupSize=[64,1,1],this.size=!0;let a=new Array(e.length);for(let n=0;n<a.length;n++)a[n]=e[n]*t[n];this.outputShape=a,this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.rank=this.outputShape.length,this.shaderKey="tile"}getUserCode(){let e=Hfe(this.rank,"uniforms.");return`
|
|
${de("index")} {
|
|
if (index < uniforms.size) {
|
|
let resRC = getCoordsFromIndex(index);
|
|
setOutputAtIndex(index, getA(${e}));
|
|
}
|
|
}
|
|
`}};function Hfe(e,t=""){if(e>=5)throw Error(`Tile for rank ${e} is not yet supported`);if(e===1)return`(resRC % ${t}aShape)`;let a=["resRC.x","resRC.y","resRC.z","resRC.w"],n=[];for(let r=0;r<e;r++)n.push(`(${a[r]} % ${t}aShape[${r}])`);return n.join()}function Pg(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{reps:s}=n;if(a.shouldExecuteOnCPU([r])||r.dtype==="string"||r.shape.length>=5){let o=a.readSync(r.dataId),l=r.dtype==="string"?o.map(c=>v.decodeString(c)):o,u=$e(r.shape,r.dtype,l),p=jue(u,s);return a.makeTensorInfo(p.shape,p.dtype,p.values)}let i=new Gfe(r.shape,s);return a.runWebGPUProgram(i,[r],r.dtype)}var jfe={kernelName:ss,backendName:"webgpu",kernelFunc:Pg};function qfe(e){let{inputs:t,backend:a,attrs:n}=e,{sparseIndices:r,sparseValues:s,defaultValue:i}=t,{outputShape:o}=n,{sliceRank:l,numUpdates:u,sliceSize:p,strides:c,outputSize:d}=S.calculateShapes(s,r,o),h=!1;if(s.dtype==="string"){let N=a.bufferSync(r),M=a.bufferSync(s),F=v.decodeString(a.readSync(i.dataId)[0]),E=Wue(N,M,o,d,p,u,l,c,F,h);return a.makeTensorInfo(o,E.dtype,E.values)}let f=[d/p,p],m=we({inputs:{x:r},backend:a,attrs:{shape:[u,l]}}),g=s.shape.length?we({inputs:{x:s},backend:a,attrs:{shape:[u,p]}}):tn({inputs:{x:s},backend:a}),y=g.dtype,x=a.makeTensorInfo([],y,v.makeZerosTypedArray(1,y)),A=we({inputs:{x:i},backend:a,attrs:{shape:Array(f.length).fill(1)}}),b=Pg({inputs:{x:A},backend:a,attrs:{reps:f}}),w=v.sizeFromShape([u,p]),C=[{type:"int32",data:[l]},{type:"int32",data:c},{type:"int32",data:[w]}];switch(u){case 0:break;case 1:{let N=new Nd([u,p],l,m.shape.length,g.shape.length,c,f,y,h);a.runWebGPUProgram(N,[g,m],y,C,b)}break;default:{let N=new Nd([u,p],l,m.shape.length,x.shape.length,c,f,y,h);a.runWebGPUProgram(N,[x,m],y,C,b)}{let N=new Nd([u,p],l,m.shape.length,g.shape.length,c,f,y);a.runWebGPUProgram(N,[g,m],y,C,b)}}let T=we({inputs:{x:b},backend:a,attrs:{shape:o}});return a.disposeData(m.dataId),a.disposeData(g.dataId),a.disposeData(A.dataId),a.disposeData(x.dataId),a.disposeData(b.dataId),T}var Xfe={kernelName:Do,backendName:"webgpu",kernelFunc:qfe};function Kfe(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{numOrSizeSplits:s,axis:i}=n,o=v.parseAxisParam(i,r.shape)[0],l=S.prepareSplitSize(r,s,o),u=r.shape.length,p=new Array(u).fill(0),c=r.shape.slice();return l.map(d=>{let h=[...c];h[o]=d;let f=ju({inputs:{x:r},backend:a,attrs:{begin:p,size:h}});return p[o]+=d,f})}var Yfe={kernelName:Tu,backendName:"webgpu",kernelFunc:Kfe},Zfe=tt({opType:le.SQRT}),Jfe={kernelName:Fo,backendName:"webgpu",kernelFunc:Zfe},Qfe={kernelName:gp,backendName:"webgpu",kernelFunc:({inputs:e,backend:t})=>{let{x:a}=e,n=t,r=new Hu(a.shape,le.SQUARE);return n.runWebGPUProgram(r,[a],a.dtype)}},eme=ta({opType:_e.SQUARED_DIFFERENCE}),tme={kernelName:Po,backendName:"webgpu",kernelFunc:eme};function ame({inputs:e,attrs:t,backend:a}){let{x:n}=e,r=new Hu(n.shape,le.STEP,"stepAlpha : f32,"),s=[{type:"float32",data:[t.alpha]}];return a.runWebGPUProgram(r,[n],n.dtype,s)}var nme={kernelName:is,backendName:"webgpu",kernelFunc:ame},rme=class{constructor(e){this.variableNames=["x"],this.workPerThread=1,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize,[this.workPerThread,1,1]);let t=Jt(this.outputShape.length);this.uniforms=`begin : ${t}, strides : ${t}, `,this.shaderKey="stridedSlice"}getUserCode(){let e=this.outputShape.length,t="";if(e===1)t="coords * uniforms.strides + uniforms.begin";else{let a=0;t=this.outputShape.map((n,r)=>(a++,this.outputShape.length===1?`coords * uniforms.strides[${r}] + uniforms.begin[${r}]`:`coords[${a-1}] * uniforms.strides[${r}] + uniforms.begin[${r}]`)).join(",")}return`
|
|
${de("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
setOutputAtIndex(index, getX(${t}));
|
|
}
|
|
}
|
|
`}};function sme(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{begin:s,end:i,strides:o,beginMask:l,endMask:u,ellipsisMask:p,newAxisMask:c,shrinkAxisMask:d}=n,{finalShapeSparse:h,finalShape:f,isIdentity:m,sliceDim0:g,isSimpleSlice:y,begin:x,end:A,strides:b}=Nt.sliceInfo(r.shape,s,i,o,l,u,p,c,d),w;if(m)w=we({inputs:{x:r},backend:a,attrs:{shape:f}});else if(g||y){v.assert(r.shape.length>=1,()=>`Input must have rank at least 1, got: ${r.shape.length}`);let C=Nt.computeOutShape(x,A,b),T=ju({inputs:{x:r},backend:a,attrs:{begin:x,size:C}});w=we({inputs:{x:T},backend:a,attrs:{shape:f}}),a.disposeData(T.dataId)}else if(a.shouldExecuteOnCPU([r])){let C=a.readSync(r.dataId),T=$e(r.shape,r.dtype,C),N=Uue(h,T,b,x);w=a.makeTensorInfo(f,r.dtype,N.values)}else{let C=new rme(h),T=[{type:"int32",data:x},{type:"int32",data:b}],N=a.runWebGPUProgram(C,[r],r.dtype,T);w=we({inputs:{x:N},backend:a,attrs:{shape:f}}),a.disposeData(N.dataId)}return w}var ime={kernelName:Oo,backendName:"webgpu",kernelFunc:sme};function ome(e){let{inputs:t,backend:a,attrs:n}=e,{separator:r,nGramWidths:s,leftPad:i,rightPad:o,padWidth:l,preserveShortSequences:u}=n,{data:p,dataSplits:c}=t,d=a.readSync(p.dataId),h=a.readSync(c.dataId),[f,m]=Gue(d,h,r,s,i,o,l,u);return[a.makeTensorInfo([f.length],"string",f),a.makeTensorInfo(c.shape,"int32",m)]}var lme={kernelName:Ru,backendName:"webgpu",kernelFunc:ome},ume=ta({opType:_e.SUB,cpuKernelImpl:Hue,supportsComplex:!0}),dme={kernelName:zo,backendName:"webgpu",kernelFunc:ume},pme=tt({opType:le.TAN}),cme={kernelName:Lo,backendName:"webgpu",kernelFunc:pme},hme=tt({opType:le.TANH}),fme={kernelName:Wo,backendName:"webgpu",kernelFunc:hme};function mme(e){let{inputs:t,backend:a,attrs:n}=e,{tensor:r,indices:s,updates:i}=t,{}=n,{sliceRank:o,numUpdates:l,sliceSize:u,strides:p,outputSize:c}=S.calculateShapes(i,s,r.shape),d=[c/u,u];if(c===0)return a.makeTensorInfo(r.shape,s.dtype);let h=[],f=we({inputs:{x:s},backend:a,attrs:{shape:[l,o]}});h.push(f);let m=we({inputs:{x:i},backend:a,attrs:{shape:[l,u]}});h.push(m);let g=we({inputs:{x:r},backend:a,attrs:{shape:d}});h.push(g);let y=Pg({inputs:{x:g},backend:a,attrs:{reps:Array(d.length).fill(1)}}),x=new Nd([l,u],o,f.shape.length,m.shape.length,p,d,r.dtype,!1),A=v.sizeFromShape([l,u]),b=[{type:"int32",data:[o]},{type:"int32",data:p},{type:"int32",data:[A]}],w=a.runWebGPUProgram(x,[m,f],g.dtype,b,y);h.push(w);let C=we({inputs:{x:w},backend:a,attrs:{shape:r.shape}});return h.forEach(T=>a.disposeData(T.dataId)),C}var gme={kernelName:So,backendName:"webgpu",kernelFunc:mme},yme=class{constructor(e){this.variableNames=["x","indices"],this.workgroupSize=[256,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.uniforms=`inputSize : i32, firstPass : i32, negativeInf : f32,
|
|
dir : i32, inc : i32,`,this.shaderKey="swap"}getUserCode(){return`
|
|
${de("index")} {
|
|
if (index < uniforms.size) {
|
|
let outC = getCoordsFromIndex(index);
|
|
let batch = outC[0];
|
|
let elemIdx = outC[1];
|
|
// We compare elements pair-wise within a group of size 2 * inc.
|
|
// The comparing rule for each group alternates between ascending
|
|
// and descending. Within each group, we compare each pair at
|
|
// positions i and i+inc. To decide whether an element at position i
|
|
// is x0 or x1, we mod it by 2 * inc, if the result is smaller than
|
|
// inc, it is in the first half of the group, we denote it as x0,
|
|
// otherwise we denote it as x1.
|
|
// For example, as shown in the Bitonic top K paper referenced
|
|
// above, Figure5(a) shows that element[1] is in the second half of
|
|
// the group when group size is 2, but it is in the first half of
|
|
// the group when group size is 4.
|
|
let isFirstInPair = elemIdx % (2 * uniforms.inc) < uniforms.inc;
|
|
var i = 0;
|
|
if (isFirstInPair) {
|
|
i = elemIdx;
|
|
} else {
|
|
i = elemIdx - uniforms.inc;
|
|
}
|
|
|
|
var i0 = 0;
|
|
if (uniforms.firstPass == 1) {
|
|
i0 = i;
|
|
} else {
|
|
i0 = i32(getIndices(batch, i));
|
|
}
|
|
|
|
var i1 = 0;
|
|
if (uniforms.firstPass == 1) {
|
|
i1 = i + uniforms.inc;
|
|
} else {
|
|
i1 = i32(getIndices(batch, i + uniforms.inc));
|
|
}
|
|
|
|
var x0 = f32(0.0);
|
|
var x1 = f32(0.0);
|
|
if (i0 < uniforms.inputSize) {
|
|
x0 = getX(batch, i0);
|
|
} else {
|
|
x0 = uniforms.negativeInf;
|
|
}
|
|
if (i1 < uniforms.inputSize) {
|
|
x1 = getX(batch, i1);
|
|
} else {
|
|
x1 = uniforms.negativeInf;
|
|
}
|
|
|
|
let reverse = elemIdx % (2 * uniforms.dir) >= uniforms.dir;
|
|
let isGreater = x0 > x1 || (x0 == x1 && i1 > i0);
|
|
if (reverse == isGreater) {
|
|
// Elements in opposite order of direction
|
|
let iTemp = i0;
|
|
i0 = i1;
|
|
i1 = iTemp;
|
|
}
|
|
if (isFirstInPair) {
|
|
setOutputAtIndex(index, f32(i0));
|
|
} else {
|
|
setOutputAtIndex(index, f32(i1));
|
|
}
|
|
}
|
|
}
|
|
`}},xme=class{constructor(e){this.variableNames=["x","indices"],this.workgroupSize=[256,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.uniforms="inputSize : i32, firstPass : i32, k : i32,",this.shaderKey="merge"}getUserCode(){return`
|
|
${de("index")} {
|
|
if (index < uniforms.size) {
|
|
let outC = getCoordsFromIndex(index);
|
|
let batch = outC[0];
|
|
let elemIdx = outC[1];
|
|
// The output size is half of the previous size.
|
|
// If the previous sequence is | | | | _ _ _ _ | | | | _ _ _ _
|
|
// (k=4), we only need to output the indices at positions |, the
|
|
// indices at positions _ can be thrown away, see Figure5(b) After
|
|
// Phase 2 (Merge phase) in the Bitonic Top K paper referenced
|
|
// above.
|
|
// For example, the paper shows we only need to output the orange
|
|
// bars. The output sequence should look like this | | | | | | | |.
|
|
// Because the sequence is halved, to map the output index back to
|
|
// the previous sequence to find the corresponding value, we need
|
|
// to double the index. When we double the index, we basically
|
|
// interpolate a position, so 2i looks like
|
|
// | _ | _ | _ | _ | _ | _ | _. We move the | to the first k
|
|
// position of each 2k positions by - elemIdx % k. E.g. for output
|
|
// at index 4,5,6,7, we want to get the corresponding element at
|
|
// original index 8,9,10,11, for output at index 8,9,10,11,
|
|
// we want to get the corresponding element at original index
|
|
// 16,17,18,19, so on and so forth.
|
|
|
|
var i = 0;
|
|
if (elemIdx < uniforms.k) {
|
|
i = elemIdx;
|
|
} else {
|
|
i = elemIdx * 2 - elemIdx % uniforms.k;
|
|
}
|
|
var i0 = 0;
|
|
if (uniforms.firstPass == 1) {
|
|
i0 = i;
|
|
} else {
|
|
i0 = i32(getIndices(batch, i));
|
|
}
|
|
var i1 = 0;
|
|
if (uniforms.firstPass == 1) {
|
|
i1 = i + uniforms.k;
|
|
} else {
|
|
i1 = i32(getIndices(batch, i + uniforms.k));
|
|
}
|
|
|
|
let x0 = getX(batch, i0);
|
|
var x1 = f32(0.0);
|
|
if (i1 < uniforms.inputSize) {
|
|
x1 = getX(batch, i1);
|
|
} else {
|
|
x1 = x0;
|
|
}
|
|
|
|
if (x0 >= x1) {
|
|
setOutputAtIndex(index, f32(i0));
|
|
} else {
|
|
setOutputAtIndex(index, f32(i1));
|
|
}
|
|
}
|
|
}
|
|
`}};function Sl(e,t){t!==null&&e.disposeData(t.dataId)}function Gx(e){let t=1;for(;t<e;)t*=2;return t}function Ame(e){let{inputs:t,backend:a,attrs:n}=e,{x:r}=t,{k:s,sorted:i}=n,o=r.shape,l=o[o.length-1];if(a.shouldExecuteOnCPU([r])){let b=a.readSync(r.dataId),[w,C]=que(b,o,r.dtype,s,i);return[a.makeTensorInfo(w.shape,w.dtype,w.values),a.makeTensorInfo(C.shape,C.dtype,C.values)]}if(s===0)return o[o.length-1]=0,[a.makeTensorInfo(o,r.dtype,[]),a.makeTensorInfo(o,"int32",[])];if(l===1)return[r,Mn({attrs:{shape:o,dtype:"int32",value:0},backend:a})];let u=v.sizeFromShape(o)/l,p=we({inputs:{x:r},attrs:{shape:[u,l]},backend:a}),c=Gx(s),d=Gx(l),h=null,f=()=>h===null?[p,p]:[p,h],m=(b,w,C)=>{let T=f(),N=new yme(C),M=[{type:"int32",data:[l]},{type:"int32",data:[h===null?1:0]},{type:"float32",data:[Number.NEGATIVE_INFINITY]},{type:"int32",data:[b]},{type:"int32",data:[w]}],F=h;h=a.runWebGPUProgram(N,T,"int32",M),Sl(a,F)};for(let b=1;b<c;b*=2){let w=b*2;for(let C=b;C>=1;C/=2)m(w,C,[u,d])}for(let b=d;b>c;b/=2){let w=f(),C=new xme([u,b/2]),T=[{type:"int32",data:[l]},{type:"int32",data:[h===null?1:0]},{type:"int32",data:[c]}],N=h;h=a.runWebGPUProgram(C,w,"int32",T),Sl(a,N);let M=c/2,F=M*2;for(let E=M;E>=1;E/=2)m(F,E,h.shape)}let g=h;h=ju({inputs:{x:h},backend:a,attrs:{begin:0,size:[u,s]}}),Sl(a,g);let y=ck({inputs:{x:p,indices:h},backend:a,attrs:{axis:1,batchDims:1}});Sl(a,p);let x=o.slice(0,-1);x.push(s),g=h,h=we({inputs:{x:h},attrs:{shape:x},backend:a}),Sl(a,g);let A=y;return y=we({inputs:{x:y},attrs:{shape:x},backend:a}),Sl(a,A),[y,h]}var bme={kernelName:Bo,backendName:"webgpu",kernelFunc:Ame},vme=class{constructor(e){this.variableNames=["Image","Transforms"],this.uniforms="interpolationModeId : i32, fillModeId : i32, fillValue : f32,",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=be(this.outputShape),this.dispatch=ce(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="transform"}getUserCode(){return`
|
|
fn mapCoord(outCoord : f32, len : f32) -> f32{
|
|
var inCoord = outCoord;
|
|
if(uniforms.fillModeId == 2) {
|
|
if (inCoord < 0.0) {
|
|
if (len <= 1.0) {
|
|
inCoord = 0.0;
|
|
} else {
|
|
let sz2 = 2.0 * len;
|
|
if (inCoord < sz2) {
|
|
inCoord = sz2 * f32(i32(f32(-inCoord / sz2))) +
|
|
inCoord;
|
|
}
|
|
if (inCoord < -len) {
|
|
inCoord = inCoord + sz2;
|
|
} else {
|
|
inCoord = -inCoord - 1.0;
|
|
}
|
|
}
|
|
} else if (inCoord > len - 1.0) {
|
|
if (len <= 1.0) {
|
|
inCoord = 0.0;
|
|
} else {
|
|
let sz2 = 2.0 * len;
|
|
inCoord = inCoord - sz2 * f32(i32(f32(inCoord / sz2)));
|
|
if (inCoord >= len) {
|
|
inCoord = sz2 - inCoord - 1.0;
|
|
}
|
|
}
|
|
}
|
|
return clamp(inCoord, 0.0, len - 1.0);
|
|
} else if (uniforms.fillModeId == 3) {
|
|
if (inCoord < 0.0) {
|
|
if (len <= 1.0) {
|
|
inCoord = 0.0;
|
|
} else {
|
|
let sz = len - 1.0;
|
|
inCoord = inCoord + len * (f32(i32(f32(-inCoord / sz))) + 1.0);
|
|
}
|
|
} else if (inCoord > len - 1.0) {
|
|
if (len <= 1.0) {
|
|
inCoord = 0.0;
|
|
} else {
|
|
let sz = len - 1.0;
|
|
inCoord = inCoord - len * f32(i32(f32(inCoord / sz)));
|
|
}
|
|
}
|
|
return clamp(inCoord, 0.0, len - 1.0);
|
|
} else if (uniforms.fillModeId == 4) {
|
|
return clamp(outCoord, 0.0, len - 1.0);
|
|
}
|
|
return outCoord;
|
|
}
|
|
fn readWithFillValue(batch : i32, coordY : i32, coordX : i32,
|
|
channel : i32) -> f32 {
|
|
var outputValue : f32;
|
|
if (0 <= coordY && coordY < uniforms.imageShape[1] && 0 <= coordX && coordX < uniforms.imageShape[2]) {
|
|
outputValue = getImage(batch, coordY, coordX, channel);
|
|
} else {
|
|
outputValue = uniforms.fillValue;
|
|
}
|
|
return outputValue;
|
|
}
|
|
|
|
${de("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
var outputValue : f32;
|
|
let batch = coords[0];
|
|
let x = coords[2];
|
|
let y = coords[1];
|
|
let channel = coords[3];
|
|
let xf = f32(x);
|
|
let yf = f32(y);
|
|
let a1 = getTransforms(batch, 0);
|
|
let a2 = getTransforms(batch, 1);
|
|
let a3 = getTransforms(batch, 2);
|
|
let b1 = getTransforms(batch, 3);
|
|
let b2 = getTransforms(batch, 4);
|
|
let b3 = getTransforms(batch, 5);
|
|
let c1 = getTransforms(batch, 6);
|
|
let c2 = getTransforms(batch, 7);
|
|
let projection = c1 * xf + c2 * yf + 1.0;
|
|
if (projection == 0.0) {
|
|
outputValue = uniforms.fillValue;
|
|
} else {
|
|
let inX = (a1 * xf + a2 * yf + a3) / projection;
|
|
let inY = (b1 * xf + b2 * yf + b3) / projection;
|
|
let mapX = mapCoord(inX, f32(uniforms.imageShape[2]));
|
|
let mapY = mapCoord(inY, f32(uniforms.imageShape[1]));
|
|
|
|
if (uniforms.interpolationModeId == 1) {
|
|
let coordY = i32(round(mapY));
|
|
let coordX = i32(round(mapX));
|
|
outputValue = readWithFillValue(batch, coordY, coordX,
|
|
channel);
|
|
} else {
|
|
let yFloor = floor(mapY);
|
|
let xFloor = floor(mapX);
|
|
let yCeil = yFloor + 1.0;
|
|
let xCeil = xFloor + 1.0;
|
|
let valueYFloor = (xCeil - mapX) *
|
|
readWithFillValue(batch, i32(yFloor), i32(xFloor), channel) +
|
|
(mapX - xFloor) *
|
|
readWithFillValue(batch, i32(yFloor), i32(xCeil), channel);
|
|
let valueYCeil = (xCeil - mapX) *
|
|
readWithFillValue(batch, i32(yCeil), i32(xFloor), channel) +
|
|
(mapX - xFloor) *
|
|
readWithFillValue(batch, i32(yCeil), i32(xCeil), channel);
|
|
outputValue = (yCeil - mapY) * valueYFloor +
|
|
(mapY - yFloor) * valueYCeil;
|
|
}
|
|
}
|
|
setOutputAtIndex(index, outputValue);
|
|
}
|
|
}
|
|
`}};function wme(e){let{inputs:t,backend:a,attrs:n}=e,{image:r,transforms:s}=t,{interpolation:i,fillMode:o,fillValue:l,outputShape:u}=n,[p,c,d,h]=r.shape,[f,m]=u!=null?u:[c,d],g=[p,f,m,h],y=new vme(g),x=i==="nearest"?1:2,A;switch(o){case"constant":A=1;break;case"reflect":A=2;break;case"wrap":A=3;break;case"nearest":A=4;break;default:A=1;break}let b=[{type:"int32",data:[x]},{type:"int32",data:[A]},{type:"float32",data:[l]}];return a.runWebGPUProgram(y,[r,s],"float32",b)}var kme={kernelName:Vo,backendName:"webgpu",kernelFunc:wme};function Ime(e){let{inputs:t,backend:a,attrs:n}=e,{value:r}=t,{axis:s}=n;s<0&&(s+=r.shape.length);let i=r,o=i.shape.length,l=r.shape[s],u=new Array(o-1),p=0;for(let m=0;m<o;m++)m!==s&&(u[p++]=i.shape[m]);let c=[],d=new Array(o).fill(0),h=i.shape.slice();h[s]=1;let f=new Array(l);for(let m=0;m<f.length;m++){d[s]=m;let g=ju({inputs:{x:i},backend:a,attrs:{begin:d,size:h}}),y=we({inputs:{x:g},backend:a,attrs:{shape:u}});f[m]=y,c.push(g)}return c.forEach(m=>a.disposeData(m.dataId)),f}var Sme={kernelName:Eu,backendName:"webgpu",kernelFunc:Ime},Cme=class{constructor(e,t,a){if(this.outputShape=[],this.variableNames=["x","segmentIds"],this.uniforms="numSegments : i32, xSize: i32,",this.workgroupSize=[64,1,1],this.atomic=!0,this.outputShape=t,this.dispatchLayout=be(e),this.dispatch=ce(this.dispatchLayout,e,this.workgroupSize),a!=="float32"&&a!=="int32")throw new Error(`UnsortedSegmentSum only supports float32 and int32
|
|
types, does not support ${a} type.`);this.type=a,this.shaderKey="unsortedSegmentSum"}getUserCode(){return`
|
|
${de("index")} {
|
|
if (index < uniforms.xSize) {
|
|
let coords = getXCoordsFromIndex(index);
|
|
let b = coords[0];
|
|
let inCol = coords[1];
|
|
|
|
let segmentId = i32(getSegmentIds(inCol));
|
|
if (segmentId >= 0) {
|
|
let flatIndex = b * uniforms.numSegments + segmentId % uniforms.numSegments;
|
|
let value = getX(b, inCol);
|
|
|
|
${Gu("&result[flatIndex]","value",this.type)}
|
|
}
|
|
}
|
|
}
|
|
`}};function Tme(e){let{inputs:t,backend:a,attrs:n}=e,{x:r,segmentIds:s}=t,{numSegments:i}=n,o=r.shape.length,l=[],u=0,p=S.getAxesPermutation([u],o),c=r;p!=null&&(c=Bn({inputs:{x:r},backend:a,attrs:{perm:p}}),l.push(c),u=S.getInnerMostAxes(1,o)[0]);let d=S.segment_util.computeOutShape(c.shape,u,i),h=v.sizeFromShape([c.shape[u]]),f=we({inputs:{x:c},backend:a,attrs:{shape:[-1,h]}});l.push(f);let m=r.dtype,g=[f.shape[0],i],y=Mn({backend:a,attrs:{shape:g,value:0,dtype:m}}),x=new Cme(f.shape,g,m),A=[{type:"int32",data:[i]},{type:"int32",data:[v.sizeFromShape(f.shape)]}],b=a.runWebGPUProgram(x,[f,s],m,A,y),w=we({inputs:{x:b},backend:a,attrs:{shape:d}});l.push(b);let C=w;if(p!=null){l.push(w);let T=S.getUndoAxesPermutation(p);C=Bn({inputs:{x:C},backend:a,attrs:{perm:T}})}return l.forEach(T=>a.disposeData(T.dataId)),C}var Nme={kernelName:vp,backendName:"webgpu",kernelFunc:Tme},Rme=[cue,Yue,Jue,ede,ade,sde,cde,fde,gde,xde,bde,wde,Ide,Cde,Nde,_de,Dde,Lde,Bde,Ude,Xde,Jde,tpe,spe,ope,ppe,fue,fpe,xpe,Cpe,Fpe,Ppe,Lpe,Bpe,Upe,Hpe,qpe,Ype,Jpe,ece,ace,sce,cce,fce,lce,yce,bce,Ice,Cce,Ece,Fce,$ce,Pce,zce,Wce,Bce,Uce,Hce,uue,qce,Qce,Kce,Zce,ahe,rhe,ihe,uhe,che,fhe,ghe,hue,xhe,gpe,bhe,whe,Ihe,Che,Nhe,Ehe,_he,Ohe,Dhe,Lhe,Bhe,Uhe,qhe,Yhe,Ede,Jhe,e0e,l0e,a0e,i0e,d0e,Mde,c0e,f0e,g0e,x0e,k0e,Tce,S0e,T0e,R0e,ape,F0e,$0e,P0e,z0e,W0e,V0e,G0e,j0e,npe,X0e,Y0e,J0e,efe,due,nfe,ife,ufe,cfe,mfe,yfe,Afe,vfe,kfe,Cfe,Rfe,Mfe,_fe,Dfe,Ofe,Lfe,jde,nme,ime,lme,v0e,Bfe,Ufe,Xfe,Yfe,Jfe,Qfe,tme,dme,Nce,cme,fme,gme,jfe,bme,kme,ude,Sme,Nme,_0e];for(let e of Rme)yn(e);var Hx="4.5.0",Eme="4.5.0",Mme="4.5.0",Fme="4.5.0",_me="4.5.0",$me="0.0.1-alpha.20",qp={tfjs:Hx,"tfjs-core":Hx,"tfjs-converter":Eme,"tfjs-backend-cpu":Mme,"tfjs-backend-webgl":Fme,"tfjs-backend-wasm":_me,"tfjs-backend-webgpu":$me};function K(...e){let t=new Date,a=`${t.getHours().toString().padStart(2,"0")}:${t.getMinutes().toString().padStart(2,"0")}:${t.getSeconds().toString().padStart(2,"0")}.${t.getMilliseconds().toString().padStart(3,"0")}`;e&&console.log(a,"Human:",...e)}function yk(e,t){let a=e.endsWith("/")?"":"/",r=t.startsWith(".")||t.startsWith("/")||t.startsWith("http:")||t.startsWith("https:")||t.startsWith("file:")?`${t}`:`${e}${a}${t}`;if(!r.toLocaleLowerCase().includes(".json"))throw new Error(`modelpath error: expecting json file: ${r}`);return r}var ae=()=>typeof performance!="undefined"?performance.now():parseInt((Number(process.hrtime.bigint())/1e3/1e3).toString());function Og(e,t,a="config",n=[]){for(let r of Object.keys(t))if(typeof t[r]=="object")Og(e[r],t[r],r,n);else{let s=e&&typeof e[r]!="undefined";s||n.push({reason:"unknown property",where:`${a}.${r} = ${t[r]}`});let i=e&&typeof e[r]==typeof t[r];s&&!i&&n.push({reason:"property type mismatch",where:`${a}.${r} = ${t[r]}`,expected:typeof e[r]})}return t.debug&&a==="config"&&n.length>0&&K("invalid configuration",n),n}function Et(...e){let t=a=>a&&typeof a=="object";return e.reduce((a,n)=>(Object.keys(n||{}).forEach(r=>{let s=a[r],i=n[r];Array.isArray(s)&&Array.isArray(i)?a[r]=s.concat(...i):t(s)&&t(i)?a[r]=Et(s,i):a[r]=i}),a),{})}var Qo={backend:"",modelBasePath:"",cacheModels:!0,validateModels:!0,wasmPath:"",wasmPlatformFetch:!1,debug:!1,async:!0,warmup:"full",cacheSensitivity:.7,skipAllowed:!1,deallocate:!1,flags:{},softwareKernels:!1,filter:{enabled:!0,equalization:!1,width:0,height:0,flip:!1,return:!0,autoBrightness:!0,brightness:0,contrast:0,sharpness:0,blur:0,saturation:0,hue:0,negative:!1,sepia:!1,vintage:!1,kodachrome:!1,technicolor:!1,polaroid:!1,pixelate:0},gesture:{enabled:!0},face:{enabled:!0,detector:{modelPath:"blazeface.json",rotation:!1,maxDetected:1,skipFrames:99,skipTime:2500,minConfidence:.2,minSize:0,iouThreshold:.1,mask:!1,return:!1},mesh:{enabled:!0,modelPath:"facemesh.json",keepInvalid:!1},attention:{enabled:!1,modelPath:"facemesh-attention.json"},iris:{enabled:!0,modelPath:"iris.json"},emotion:{enabled:!0,minConfidence:.1,skipFrames:99,skipTime:1500,modelPath:"emotion.json"},description:{enabled:!0,modelPath:"faceres.json",skipFrames:99,skipTime:3e3,minConfidence:.1},antispoof:{enabled:!1,skipFrames:99,skipTime:4e3,modelPath:"antispoof.json"},liveness:{enabled:!1,skipFrames:99,skipTime:4e3,modelPath:"liveness.json"}},body:{enabled:!0,modelPath:"movenet-lightning.json",maxDetected:-1,minConfidence:.3,skipFrames:1,skipTime:200},hand:{enabled:!0,rotation:!0,skipFrames:99,skipTime:1e3,minConfidence:.5,iouThreshold:.2,maxDetected:-1,landmarks:!0,detector:{modelPath:"handtrack.json"},skeleton:{modelPath:"handlandmark-lite.json"}},object:{enabled:!1,modelPath:"centernet.json",minConfidence:.2,iouThreshold:.4,maxDetected:10,skipFrames:99,skipTime:2e3},segmentation:{enabled:!1,modelPath:"rvm.json",ratio:.5,mode:"default"}};var xk=`
|
|
precision highp float;
|
|
attribute vec2 pos;
|
|
attribute vec2 uv;
|
|
varying vec2 vUv;
|
|
uniform float flipY;
|
|
void main(void) {
|
|
vUv = uv;
|
|
gl_Position = vec4(pos.x, pos.y*flipY, 0.0, 1.);
|
|
}
|
|
`;var Ak=`
|
|
precision highp float;
|
|
varying vec2 vUv;
|
|
uniform sampler2D texture;
|
|
uniform float m[20];
|
|
void main(void) {
|
|
vec4 c = texture2D(texture, vUv);
|
|
gl_FragColor.r = m[0] * c.r + m[1] * c.g + m[2] * c.b + m[3] * c.a + m[4];
|
|
gl_FragColor.g = m[5] * c.r + m[6] * c.g + m[7] * c.b + m[8] * c.a + m[9];
|
|
gl_FragColor.b = m[10] * c.r + m[11] * c.g + m[12] * c.b + m[13] * c.a + m[14];
|
|
gl_FragColor.a = m[15] * c.r + m[16] * c.g + m[17] * c.b + m[18] * c.a + m[19];
|
|
}
|
|
`,bk=`
|
|
precision highp float;
|
|
varying vec2 vUv;
|
|
uniform sampler2D texture;
|
|
uniform float m[20];
|
|
void main(void) {
|
|
vec4 c = texture2D(texture, vUv);
|
|
gl_FragColor.r = m[0] * c.r + m[1] * c.g + m[2] * c.b + m[4];
|
|
gl_FragColor.g = m[5] * c.r + m[6] * c.g + m[7] * c.b + m[9];
|
|
gl_FragColor.b = m[10] * c.r + m[11] * c.g + m[12] * c.b + m[14];
|
|
gl_FragColor.a = c.a;
|
|
}
|
|
`,vk=`
|
|
precision highp float;
|
|
varying vec2 vUv;
|
|
uniform vec2 size;
|
|
uniform sampler2D texture;
|
|
vec2 pixelate(vec2 coord, vec2 size) {
|
|
return floor( coord / size ) * size;
|
|
}
|
|
void main(void) {
|
|
gl_FragColor = vec4(0.0);
|
|
vec2 coord = pixelate(vUv, size);
|
|
gl_FragColor += texture2D(texture, coord);
|
|
}
|
|
`,wk=`
|
|
precision highp float;
|
|
varying vec2 vUv;
|
|
uniform sampler2D texture;
|
|
uniform vec2 px;
|
|
void main(void) {
|
|
gl_FragColor = vec4(0.0);
|
|
gl_FragColor += texture2D(texture, vUv + vec2(-7.0*px.x, -7.0*px.y))*0.0044299121055113265;
|
|
gl_FragColor += texture2D(texture, vUv + vec2(-6.0*px.x, -6.0*px.y))*0.00895781211794;
|
|
gl_FragColor += texture2D(texture, vUv + vec2(-5.0*px.x, -5.0*px.y))*0.0215963866053;
|
|
gl_FragColor += texture2D(texture, vUv + vec2(-4.0*px.x, -4.0*px.y))*0.0443683338718;
|
|
gl_FragColor += texture2D(texture, vUv + vec2(-3.0*px.x, -3.0*px.y))*0.0776744219933;
|
|
gl_FragColor += texture2D(texture, vUv + vec2(-2.0*px.x, -2.0*px.y))*0.115876621105;
|
|
gl_FragColor += texture2D(texture, vUv + vec2(-1.0*px.x, -1.0*px.y))*0.147308056121;
|
|
gl_FragColor += texture2D(texture, vUv )*0.159576912161;
|
|
gl_FragColor += texture2D(texture, vUv + vec2( 1.0*px.x, 1.0*px.y))*0.147308056121;
|
|
gl_FragColor += texture2D(texture, vUv + vec2( 2.0*px.x, 2.0*px.y))*0.115876621105;
|
|
gl_FragColor += texture2D(texture, vUv + vec2( 3.0*px.x, 3.0*px.y))*0.0776744219933;
|
|
gl_FragColor += texture2D(texture, vUv + vec2( 4.0*px.x, 4.0*px.y))*0.0443683338718;
|
|
gl_FragColor += texture2D(texture, vUv + vec2( 5.0*px.x, 5.0*px.y))*0.0215963866053;
|
|
gl_FragColor += texture2D(texture, vUv + vec2( 6.0*px.x, 6.0*px.y))*0.00895781211794;
|
|
gl_FragColor += texture2D(texture, vUv + vec2( 7.0*px.x, 7.0*px.y))*0.0044299121055113265;
|
|
}
|
|
`,kk=`
|
|
precision highp float;
|
|
varying vec2 vUv;
|
|
uniform sampler2D texture;
|
|
uniform vec2 px;
|
|
uniform float m[9];
|
|
void main(void) {
|
|
vec4 c11 = texture2D(texture, vUv - px); // top left
|
|
vec4 c12 = texture2D(texture, vec2(vUv.x, vUv.y - px.y)); // top center
|
|
vec4 c13 = texture2D(texture, vec2(vUv.x + px.x, vUv.y - px.y)); // top right
|
|
vec4 c21 = texture2D(texture, vec2(vUv.x - px.x, vUv.y) ); // mid left
|
|
vec4 c22 = texture2D(texture, vUv); // mid center
|
|
vec4 c23 = texture2D(texture, vec2(vUv.x + px.x, vUv.y) ); // mid right
|
|
vec4 c31 = texture2D(texture, vec2(vUv.x - px.x, vUv.y + px.y) ); // bottom left
|
|
vec4 c32 = texture2D(texture, vec2(vUv.x, vUv.y + px.y) ); // bottom center
|
|
vec4 c33 = texture2D(texture, vUv + px ); // bottom right
|
|
gl_FragColor =
|
|
c11 * m[0] + c12 * m[1] + c22 * m[2] +
|
|
c21 * m[3] + c22 * m[4] + c23 * m[5] +
|
|
c31 * m[6] + c32 * m[7] + c33 * m[8];
|
|
gl_FragColor.a = c22.a;
|
|
}
|
|
`;var zg=(e,t,a)=>{let n=new RegExp("\\b"+t+" \\w+ (\\w+)","ig");e.replace(n,(r,s)=>(a[s]=0,r))},Lg=class{constructor(t,a,n){he(this,"uniform",{});he(this,"attribute",{});he(this,"gl");he(this,"id");he(this,"compile",(t,a)=>{let n=this.gl.createShader(a);return n?(this.gl.shaderSource(n,t),this.gl.compileShader(n),this.gl.getShaderParameter(n,this.gl.COMPILE_STATUS)?n:(K(`filter: gl compile failed: ${this.gl.getShaderInfoLog(n)||"unknown"}`),null)):(K("filter: could not create shader"),null)});this.gl=t;let r=this.compile(a,this.gl.VERTEX_SHADER),s=this.compile(n,this.gl.FRAGMENT_SHADER);if(this.id=this.gl.createProgram(),!(!r||!s)){if(!this.id){K("filter: could not create webgl program");return}if(this.gl.attachShader(this.id,r),this.gl.attachShader(this.id,s),this.gl.linkProgram(this.id),!this.gl.getProgramParameter(this.id,this.gl.LINK_STATUS)){K(`filter: gl link failed: ${this.gl.getProgramInfoLog(this.id)||"unknown"}`);return}this.gl.useProgram(this.id),zg(a,"attribute",this.attribute);for(let i in this.attribute)this.attribute[i]=this.gl.getAttribLocation(this.id,i);zg(a,"uniform",this.uniform),zg(n,"uniform",this.uniform);for(let i in this.uniform)this.uniform[i]=this.gl.getUniformLocation(this.id,i)}}};function Ik(){let e=0,t=null,a=!1,n=-1,r=[null,null],s=[],i=null,o=null,l=Fn(100,100),u={},p={INTERMEDIATE:1},c=l.getContext("webgl");if(!c){K("filter: cannot get webgl context");return}this.gl=c;function d(x,A){if(!(x===l.width&&A===l.height)){if(l.width=x,l.height=A,!i){let b=new Float32Array([-1,-1,0,1,1,-1,1,1,-1,1,0,0,-1,1,0,0,1,-1,1,1,1,1,1,0]);i=c.createBuffer(),c.bindBuffer(c.ARRAY_BUFFER,i),c.bufferData(c.ARRAY_BUFFER,b,c.STATIC_DRAW),c.pixelStorei(c.UNPACK_PREMULTIPLY_ALPHA_WEBGL,!0)}c.viewport(0,0,l.width,l.height),r=[null,null]}}function h(x,A){let b=c.createFramebuffer();c.bindFramebuffer(c.FRAMEBUFFER,b);let w=c.createRenderbuffer();c.bindRenderbuffer(c.RENDERBUFFER,w);let C=c.createTexture();return c.bindTexture(c.TEXTURE_2D,C),c.texImage2D(c.TEXTURE_2D,0,c.RGBA,x,A,0,c.RGBA,c.UNSIGNED_BYTE,null),c.texParameteri(c.TEXTURE_2D,c.TEXTURE_MAG_FILTER,c.LINEAR),c.texParameteri(c.TEXTURE_2D,c.TEXTURE_MIN_FILTER,c.LINEAR),c.texParameteri(c.TEXTURE_2D,c.TEXTURE_WRAP_S,c.CLAMP_TO_EDGE),c.texParameteri(c.TEXTURE_2D,c.TEXTURE_WRAP_T,c.CLAMP_TO_EDGE),c.framebufferTexture2D(c.FRAMEBUFFER,c.COLOR_ATTACHMENT0,c.TEXTURE_2D,C,0),c.bindTexture(c.TEXTURE_2D,null),c.bindFramebuffer(c.FRAMEBUFFER,null),{fbo:b,texture:C}}function f(x){return r[x]=r[x]||h(l.width,l.height),r[x]}function m(x=0){if(!o)return;let A=null,b=null,w=!1;e===0?A=t:A=f(n).texture||null,e++,a&&!(x&p.INTERMEDIATE)?(b=null,w=e%2===0):(n=(n+1)%2,b=f(n).fbo||null),c.bindTexture(c.TEXTURE_2D,A),c.bindFramebuffer(c.FRAMEBUFFER,b),c.uniform1f(o.uniform.flipY,w?-1:1),c.drawArrays(c.TRIANGLES,0,6)}function g(x){if(u[x])return o=u[x],c.useProgram((o?o.id:null)||null),o;if(o=new Lg(c,xk,x),!o)return K("filter: could not get webgl program"),null;let A=Float32Array.BYTES_PER_ELEMENT,b=4*A;return c.enableVertexAttribArray(o.attribute.pos),c.vertexAttribPointer(o.attribute.pos,2,c.FLOAT,!1,b,0*A),c.enableVertexAttribArray(o.attribute.uv),c.vertexAttribPointer(o.attribute.uv,2,c.FLOAT,!1,b,2*A),u[x]=o,o}let y={colorMatrix:x=>{let A=new Float32Array(x);A[4]/=255,A[9]/=255,A[14]/=255,A[19]/=255;let b=A[18]===1&&A[3]===0&&A[8]===0&&A[13]===0&&A[15]===0&&A[16]===0&&A[17]===0&&A[19]===0?bk:Ak,w=g(b);w&&(c.uniform1fv(w.uniform.m,A),m())},brightness:x=>{let A=(x||0)+1;y.colorMatrix([A,0,0,0,0,0,A,0,0,0,0,0,A,0,0,0,0,0,1,0])},saturation:x=>{let A=(x||0)*2/3+1,b=(A-1)*-.5;y.colorMatrix([A,b,b,0,0,b,A,b,0,0,b,b,A,0,0,0,0,0,1,0])},desaturate:()=>{y.saturation(-1)},contrast:x=>{let A=(x||0)+1,b=-128*(A-1);y.colorMatrix([A,0,0,0,b,0,A,0,0,b,0,0,A,0,b,0,0,0,1,0])},negative:()=>{y.contrast(-2)},hue:x=>{x=(x||0)/180*Math.PI;let A=Math.cos(x),b=Math.sin(x),w=.213,C=.715,T=.072;y.colorMatrix([w+A*(1-w)+b*-w,C+A*-C+b*-C,T+A*-T+b*(1-T),0,0,w+A*-w+b*.143,C+A*(1-C)+b*.14,T+A*-T+b*-.283,0,0,w+A*-w+b*-(1-w),C+A*-C+b*C,T+A*(1-T)+b*T,0,0,0,0,0,1,0])},desaturateLuminance:()=>{y.colorMatrix([.2764723,.929708,.0938197,0,-37.1,.2764723,.929708,.0938197,0,-37.1,.2764723,.929708,.0938197,0,-37.1,0,0,0,1,0])},sepia:()=>{y.colorMatrix([.393,.7689999,.18899999,0,0,.349,.6859999,.16799999,0,0,.272,.5339999,.13099999,0,0,0,0,0,1,0])},brownie:()=>{y.colorMatrix([.5997023498159715,.34553243048391263,-.2708298674538042,0,47.43192855600873,-.037703249837783157,.8609577587992641,.15059552388459913,0,-36.96841498319127,.24113635128153335,-.07441037908422492,.44972182064877153,0,-7.562075277591283,0,0,0,1,0])},vintagePinhole:()=>{y.colorMatrix([.6279345635605994,.3202183420819367,-.03965408211312453,0,9.651285835294123,.02578397704808868,.6441188644374771,.03259127616149294,0,7.462829176470591,.0466055556782719,-.0851232987247891,.5241648018700465,0,5.159190588235296,0,0,0,1,0])},kodachrome:()=>{y.colorMatrix([1.1285582396593525,-.3967382283601348,-.03992559172921793,0,63.72958762196502,-.16404339962244616,1.0835251566291304,-.05498805115633132,0,24.732407896706203,-.16786010706155763,-.5603416277695248,1.6014850761964943,0,35.62982807460946,0,0,0,1,0])},technicolor:()=>{y.colorMatrix([1.9125277891456083,-.8545344976951645,-.09155508482755585,0,11.793603434377337,-.3087833385928097,1.7658908555458428,-.10601743074722245,0,-70.35205161461398,-.231103377548616,-.7501899197440212,1.847597816108189,0,30.950940869491138,0,0,0,1,0])},polaroid:()=>{y.colorMatrix([1.438,-.062,-.062,0,0,-.122,1.378,-.122,0,0,-.016,-.016,1.483,0,0,0,0,0,1,0])},shiftToBGR:()=>{y.colorMatrix([0,0,1,0,0,0,1,0,0,0,1,0,0,0,0,0,0,0,1,0])},convolution:x=>{let A=new Float32Array(x),b=1/l.width,w=1/l.height,C=g(kk);C&&(c.uniform1fv(C.uniform.m,A),c.uniform2f(C.uniform.px,b,w),m())},detectEdges:()=>{y.convolution.call(this,[0,1,0,1,-4,1,0,1,0])},sobelX:()=>{y.convolution.call(this,[-1,0,1,-2,0,2,-1,0,1])},sobelY:()=>{y.convolution.call(this,[-1,-2,-1,0,0,0,1,2,1])},sharpen:x=>{let A=x||1;y.convolution.call(this,[0,-1*A,0,-1*A,1+4*A,-1*A,0,-1*A,0])},emboss:x=>{let A=x||1;y.convolution.call(this,[-2*A,-1*A,0,-1*A,1,1*A,0,1*A,2*A])},blur:x=>{let A=x/7/l.width,b=x/7/l.height,w=g(wk);w&&(c.uniform2f(w.uniform.px,0,b),m(p.INTERMEDIATE),c.uniform2f(w.uniform.px,A,0),m())},pixelate:x=>{let A=x/l.width,b=x/l.height,w=g(vk);w&&(c.uniform2f(w.uniform.size,A,b),m())}};this.add=function(x){let A=Array.prototype.slice.call(arguments,1),b=y[x];s.push({func:b,args:A})},this.reset=function(){s=[]},this.get=function(){return s},this.apply=function(x){d(x.width,x.height),e=0,t||(t=c.createTexture()),c.bindTexture(c.TEXTURE_2D,t),c.texParameteri(c.TEXTURE_2D,c.TEXTURE_WRAP_S,c.CLAMP_TO_EDGE),c.texParameteri(c.TEXTURE_2D,c.TEXTURE_WRAP_T,c.CLAMP_TO_EDGE),c.texParameteri(c.TEXTURE_2D,c.TEXTURE_MIN_FILTER,c.NEAREST),c.texParameteri(c.TEXTURE_2D,c.TEXTURE_MAG_FILTER,c.NEAREST),c.texImage2D(c.TEXTURE_2D,0,c.RGBA,c.RGBA,c.UNSIGNED_BYTE,x);for(let A=0;A<s.length;A++){a=A===s.length-1;let b=s[A];b.func.apply(this,b.args||[])}return l},this.draw=function(x){return this.add("brightness",0),this.apply(x)}}async function u0(e){let t=e.shape.length===4?Oe(e):e,a=Ca(t,3,2),n=[Jr(a[0]),Jr(a[1]),Jr(a[2])],r=[ga(a[0]),ga(a[1]),ga(a[2])],s=await Promise.all(r.map(p=>p.data())),i=Math.max(s[0][0],s[1][0],s[2][0]),l=(i>1?255:1)/i,u;if(l>1){let p=[ye(a[0],n[0]),ye(a[1],n[1]),ye(a[2],n[2])],c=[ye(r[0],n[0]),ye(r[1],n[1]),ye(r[2],n[2])],d=[te(p[0],l),te(p[1],l),te(p[2],l)],h=ca([d[0],d[1],d[2]],2);u=Q(h,[1,t.shape[0]||0,t.shape[1]||0,3]),J([...p,...c,...d,h])}else u=Ht(t,0);return J([...a,...n,...r,a,t,e]),u}var d0=3840,aa=null,na=null,qu=null,bt,xn={inputSum:0,cacheDiff:1,sumMethod:0,inputTensor:void 0};function Wg(){xn.inputSum=0,xn.cacheDiff=1,xn.sumMethod=0,xn.inputTensor=void 0}function Fn(e,t){let a;if(ne.browser)if(ne.worker){if(typeof OffscreenCanvas=="undefined")throw new Error("canvas error: attempted to run in web worker but OffscreenCanvas is not supported");a=new OffscreenCanvas(e,t)}else if(typeof document!="undefined")a=document.createElement("canvas"),a.width=e,a.height=t;else if(typeof navigator!="undefined"&&navigator.product==="ReactNative")if(typeof ne.Canvas!="undefined")a=new ne.Canvas(e,t);else if(typeof globalThis.Canvas!="undefined")a=new globalThis.Canvas(e,t);else throw new Error("canvas error: attempted to use canvas in react-native without canvas support installed");else throw new Error("canvas error: attempted to run in browser but DOM is not defined");else typeof ne.Canvas!="undefined"?a=new ne.Canvas(e,t):typeof globalThis.Canvas!="undefined"&&(a=new globalThis.Canvas(e,t));return a}function p0(e,t){let a=t||Fn(e.width,e.height);return a.getContext("2d").drawImage(e,0,0),a}async function c0(e,t,a=!0){var d,h,f;if(!e)return t.debug&&K("input error: input is missing"),{tensor:null,canvas:null};if(!(e instanceof gt)&&!(typeof Image!="undefined"&&e instanceof Image)&&!(typeof globalThis.Canvas!="undefined"&&e instanceof globalThis.Canvas)&&!(typeof ImageData!="undefined"&&e instanceof ImageData)&&!(typeof ImageBitmap!="undefined"&&e instanceof ImageBitmap)&&!(typeof HTMLImageElement!="undefined"&&e instanceof HTMLImageElement)&&!(typeof HTMLMediaElement!="undefined"&&e instanceof HTMLMediaElement)&&!(typeof HTMLVideoElement!="undefined"&&e instanceof HTMLVideoElement)&&!(typeof HTMLCanvasElement!="undefined"&&e instanceof HTMLCanvasElement)&&!(typeof OffscreenCanvas!="undefined"&&e instanceof OffscreenCanvas))throw new Error("input error: type not recognized");if(e instanceof gt){let m=null;if(e.isDisposedInternal)throw new Error("input error: attempted to use tensor but it is disposed");if(!e.shape)throw new Error("input error: attempted to use tensor without a shape");if(e.shape.length===3){if(e.shape[2]===3)m=Ht(e,0);else if(e.shape[2]===4){let g=zp(e,[0,0,0],[-1,-1,3]);m=Ht(g,0),J(g)}}else e.shape.length===4&&(e.shape[3]===3?m=Sa(e):e.shape[3]===4&&(m=Ph(e,[0,0,0,0],[-1,-1,-1,3])));if(m==null||m.shape.length!==4||m.shape[0]!==1||m.shape[3]!==3)throw new Error(`input error: attempted to use tensor with unrecognized shape: ${e.shape.toString()}`);if(m.dtype==="int32"){let g=Ke(m,"float32");J(m),m=g}return{tensor:m,canvas:t.filter.return?na:null}}if(typeof e.readyState!="undefined"&&e.readyState<=2)return t.debug&&K("input stream is not ready"),{tensor:null,canvas:aa};let n=e.naturalWidth||e.videoWidth||e.width||e.shape&&e.shape[1]>0,r=e.naturalHeight||e.videoHeight||e.height||e.shape&&e.shape[2]>0;if(!n||!r)return t.debug&&K("cannot determine input dimensions"),{tensor:null,canvas:aa};let s=n,i=r;if(s>d0&&(s=d0,i=Math.trunc(s*r/n)),i>d0&&(i=d0,s=Math.trunc(i*n/r)),(((d=t.filter)==null?void 0:d.width)||0)>0?s=t.filter.width:(((h=t.filter)==null?void 0:h.height)||0)>0&&(s=n*((t.filter.height||0)/r)),(t.filter.height||0)>0?i=t.filter.height:(t.filter.width||0)>0&&(i=r*((t.filter.width||0)/n)),!s||!i)throw new Error("input error: cannot determine dimension");(!aa||aa.width!==s||aa.height!==i)&&(aa=Fn(s,i));let o=aa.getContext("2d");if(typeof ImageData!="undefined"&&e instanceof ImageData?o.putImageData(e,0,0):t.filter.flip&&typeof o.translate!="undefined"?(o.translate(n,0),o.scale(-1,1),o.drawImage(e,0,0,n,r,0,0,aa.width,aa.height),o.setTransform(1,0,0,1,0,0)):o.drawImage(e,0,0,n,r,0,0,aa.width,aa.height),(!na||aa.width!==na.width||aa.height!==na.height)&&(na=Fn(aa.width,aa.height)),t.filter.enabled&&ne.webgl.supported?(bt||(bt=ne.browser?new Ik:null),ne.filter=!!bt,bt!=null&&bt.add?(bt.reset(),t.filter.brightness!==0&&bt.add("brightness",t.filter.brightness),t.filter.contrast!==0&&bt.add("contrast",t.filter.contrast),t.filter.sharpness!==0&&bt.add("sharpen",t.filter.sharpness),t.filter.blur!==0&&bt.add("blur",t.filter.blur),t.filter.saturation!==0&&bt.add("saturation",t.filter.saturation),t.filter.hue!==0&&bt.add("hue",t.filter.hue),t.filter.negative&&bt.add("negative"),t.filter.sepia&&bt.add("sepia"),t.filter.vintage&&bt.add("brownie"),t.filter.sepia&&bt.add("sepia"),t.filter.kodachrome&&bt.add("kodachrome"),t.filter.technicolor&&bt.add("technicolor"),t.filter.polaroid&&bt.add("polaroid"),t.filter.pixelate!==0&&bt.add("pixelate",t.filter.pixelate),((f=bt.get())==null?void 0:f.length)>1?na=bt.apply(aa):na=bt.draw(aa)):(t.debug&&K("input process error: cannot initialize filters"),ne.webgl.supported=!1,t.filter.enabled=!1,p0(aa,na))):(p0(aa,na),bt&&(bt=null),ne.filter=!!bt),!a)return{tensor:null,canvas:na};if(!na)throw new Error("canvas error: cannot create output");let l,u=3;if(typeof ImageData!="undefined"&&e instanceof ImageData||e.data&&e.width&&e.height)if(ne.browser&&Rr)l=Rr?Rr.fromPixels(e):null;else{u=e.data.length/e.height/e.width;let m=new Uint8Array(e.data.buffer);l=Ve(m,[e.height,e.width,u],"int32")}else if((!qu||na.width!==qu.width||na.height!==qu.height)&&(qu=Fn(na.width,na.height)),Rr&&ne.browser)t.backend==="webgl"||t.backend==="humangl"||t.backend==="webgpu"?l=Rr.fromPixels(na):(qu=p0(na),l=Rr.fromPixels(qu));else{let y=p0(na).getContext("2d").getImageData(0,0,s,i);u=y.data.length/s/i;let x=new Uint8Array(y.data.buffer);l=Ve(x,[s,i,u])}if(u===4){let m=zp(l,[0,0,0],[-1,-1,3]);J(l),l=m}if(!l)throw new Error("input error: cannot create tensor");let p=Ke(l,"float32"),c=t.filter.equalization?await u0(p):Ht(p,0);if(J([l,p]),t.filter.autoBrightness){let m=ga(c),g=await m.data();t.filter.brightness=g[0]>1?1-g[0]/255:1-g[0],J(m)}return{tensor:c,canvas:t.filter.return?na:null}}async function Sk(e,t){let a=!1;if(e.cacheSensitivity===0||!t.shape||t.shape.length!==4||t.shape[1]>3840||t.shape[2]>2160)return a;if(!xn.inputTensor)xn.inputTensor=Sa(t);else if(xn.inputTensor.shape[1]!==t.shape[1]||xn.inputTensor.shape[2]!==t.shape[2])J(xn.inputTensor),xn.inputTensor=Sa(t);else{let n={};n.diff=ye(t,xn.inputTensor),n.squared=te(n.diff,n.diff),n.sum=ot(n.squared);let s=(await n.sum.data())[0]/(t.shape[1]||1)/(t.shape[2]||1)/255/3;J([xn.inputTensor,n.diff,n.squared,n.sum]),xn.inputTensor=Sa(t),a=s<=(e.cacheSensitivity||0)}return a}async function Ck(e,t,a){let n={};if(!t||!a||t.shape.length!==4||t.shape.length!==a.shape.length)return e.debug||K("invalid input tensor or tensor shapes do not match:",t.shape,a.shape),0;if(t.shape[0]!==1||a.shape[0]!==1||t.shape[3]!==3||a.shape[3]!==3)return e.debug||K("input tensors must be of shape [1, height, width, 3]:",t.shape,a.shape),0;n.input1=Sa(t),n.input2=t.shape[1]!==a.shape[1]||t.shape[2]!==a.shape[2]?fe.resizeBilinear(a,[t.shape[1],t.shape[2]]):Sa(a),n.diff=ye(n.input1,n.input2),n.squared=te(n.diff,n.diff),n.sum=ot(n.squared);let s=(await n.sum.data())[0]/(t.shape[1]||1)/(t.shape[2]||1)/255/3;return J([n.input1,n.input2,n.diff,n.squared,n.sum]),s}var Kp,Yp,Zp,Xp=class{constructor(){he(this,"browser");he(this,"node");he(this,"worker");he(this,"platform","");he(this,"agent","");he(this,"backends",[]);he(this,"initial");he(this,"filter");he(this,"tfjs");he(this,"offscreen");he(this,"perfadd",!1);he(this,"tensorflow",{version:void 0,gpu:void 0});he(this,"wasm",{supported:void 0,backend:void 0,simd:void 0,multithread:void 0});he(this,"webgl",{supported:void 0,backend:void 0,version:void 0,renderer:void 0,shader:void 0,vendor:void 0});he(this,"webgpu",{supported:void 0,backend:void 0,adapter:void 0});he(this,"cpu",{model:void 0,flags:[]});he(this,"kernels",[]);qn(this,Kp,void 0);qn(this,Yp,void 0);qn(this,Zp,void 0);if(this.browser=typeof navigator!="undefined",this.node=typeof process!="undefined"&&typeof process.versions!="undefined"&&typeof process.versions.node!="undefined",this.tfjs={version:qp["tfjs-core"]},this.offscreen=typeof OffscreenCanvas!="undefined",this.initial=!0,this.worker=this.browser&&this.offscreen?typeof WorkerGlobalScope!="undefined":void 0,typeof navigator!="undefined"&&typeof navigator.userAgent!="undefined"){let t=navigator.userAgent||"",a=t.match(/\(([^()]+)\)/g);if(a!=null&&a[0]){let n=a[0].match(/\(([^()]+)\)/g);this.platform=n!=null&&n[0]?n[0].replace(/\(|\)/g,""):"",this.agent=t.replace(a[0],""),this.platform[1]&&(this.agent=this.agent.replace(a[1],"")),this.agent=this.agent.replace(/ /g," ")}}else typeof process!="undefined"&&(this.platform=`${process.platform} ${process.arch}`,this.agent=`NodeJS ${process.version}`)}get Canvas(){return Xa(this,Kp)}set Canvas(t){xr(this,Kp,t),globalThis.Canvas=t}get Image(){return Xa(this,Yp)}set Image(t){xr(this,Yp,t),globalThis.Image=t}get ImageData(){return Xa(this,Zp)}set ImageData(t){xr(this,Zp,t),globalThis.ImageData=t}async updateBackend(){this.backends=Object.keys(It().registryFactory);try{this.tensorflow={version:nr().binding?nr().binding.TF_Version:void 0,gpu:nr().binding?nr().binding.isUsingGpuDevice():void 0}}catch(n){}this.wasm.supported=typeof WebAssembly!="undefined",this.wasm.backend=this.backends.includes("wasm"),this.wasm.supported&&this.wasm.backend&&(this.wasm.simd=await W().getAsync("WASM_HAS_SIMD_SUPPORT"),this.wasm.multithread=await W().getAsync("WASM_HAS_MULTITHREAD_SUPPORT"));let t=Fn(100,100),a=t?t.getContext("webgl2"):void 0;this.webgl.supported=typeof a!="undefined",this.webgl.backend=this.backends.includes("webgl"),this.webgl.supported&&this.webgl.backend&&a&&(this.webgl.version=a.getParameter(a.VERSION),this.webgl.vendor=a.getParameter(a.VENDOR),this.webgl.renderer=a.getParameter(a.RENDERER),this.webgl.shader=a.getParameter(a.SHADING_LANGUAGE_VERSION)),this.webgpu.supported=this.browser&&typeof navigator!="undefined"&&typeof navigator.gpu!="undefined",this.webgpu.backend=this.backends.includes("webgpu");try{if(this.webgpu.supported){let n=await navigator.gpu.requestAdapter();this.webgpu.adapter=await(n==null?void 0:n.requestAdapterInfo())}}catch(n){this.webgpu.supported=!1}try{this.kernels=Jn(ha()).map(n=>n.kernelName.toLowerCase())}catch(n){}}updateCPU(){let t={model:"",flags:[]};this.node&&this.platform.startsWith("linux"),this.cpu?this.cpu=t:Object.defineProperty(this,"cpu",{value:t})}};Kp=new WeakMap,Yp=new WeakMap,Zp=new WeakMap;var ne=new Xp;var f0=class{constructor(){he(this,"config");he(this,"element");he(this,"stream");he(this,"devices",[]);he(this,"enumerate",async()=>{try{let t=await navigator.mediaDevices.enumerateDevices();this.devices=t.filter(a=>a.kind==="videoinput")}catch(t){this.devices=[]}return this.devices});he(this,"start",async t=>{var r,s;if(t!=null&&t.debug&&(this.config.debug=t==null?void 0:t.debug),t!=null&&t.crop&&(this.config.crop=t==null?void 0:t.crop),t!=null&&t.mode&&(this.config.mode=t==null?void 0:t.mode),t!=null&&t.width&&(this.config.width=t==null?void 0:t.width),t!=null&&t.height&&(this.config.height=t==null?void 0:t.height),t!=null&&t.id&&(this.config.id=t==null?void 0:t.id),t!=null&&t.element)if(typeof t.element=="string"){let i=document.getElementById(t.element);if(i&&i instanceof HTMLVideoElement)this.element=i;else{this.config.debug&&K("webcam","cannot get dom element",t.element);return}}else if(t.element instanceof HTMLVideoElement)this.element=t.element;else{this.config.debug&&K("webcam","unknown dom element",t.element);return}else this.element=document.createElement("video");let a={audio:!1,video:{facingMode:this.config.mode==="front"?"user":"environment",resizeMode:this.config.crop?"crop-and-scale":"none"}};if(((r=this.config)==null?void 0:r.width)>0&&(a.video.width={ideal:this.config.width}),((s=this.config)==null?void 0:s.height)>0&&(a.video.height={ideal:this.config.height}),this.config.id&&(a.video.deviceId=this.config.id),this.element.addEventListener("play",()=>{this.config.debug&&K("webcam","play")}),this.element.addEventListener("pause",()=>{this.config.debug&&K("webcam","pause")}),this.element.addEventListener("click",async()=>{!this.element||!this.stream||(this.element.paused?await this.element.play():this.element.pause())}),!(navigator!=null&&navigator.mediaDevices)){this.config.debug&&K("webcam","no devices");return}try{this.stream=await navigator.mediaDevices.getUserMedia(a)}catch(i){K("webcam",i);return}if(!this.stream){this.config.debug&&K("webcam","no stream");return}this.element.srcObject=this.stream,await new Promise(i=>{this.element?this.element.onloadeddata=()=>i(!0):i(!1)}),await this.element.play(),this.config.debug&&K("webcam",{width:this.width,height:this.height,label:this.label,stream:this.stream,track:this.track,settings:this.settings,constraints:this.constraints,capabilities:this.capabilities})});he(this,"pause",()=>{this.element&&this.element.pause()});he(this,"play",async()=>{this.element&&await this.element.play()});he(this,"stop",()=>{this.config.debug&&K("webcam","stop"),this.track&&this.track.stop()});this.config={element:void 0,debug:!0,mode:"front",crop:!1,width:0,height:0}}get track(){if(this.stream)return this.stream.getVideoTracks()[0]}get capabilities(){if(this.track)return this.track.getCapabilities?this.track.getCapabilities():void 0}get constraints(){if(this.track)return this.track.getConstraints?this.track.getConstraints():void 0}get settings(){if(!this.stream)return;let t=this.stream.getVideoTracks()[0];return t.getSettings?t.getSettings():void 0}get label(){return this.track?this.track.label:""}get paused(){var t;return((t=this.element)==null?void 0:t.paused)||!1}get width(){var t;return((t=this.element)==null?void 0:t.videoWidth)||0}get height(){var t;return((t=this.element)==null?void 0:t.videoHeight)||0}};var Bg={};yr(Bg,{"affectnet-mobilenet":()=>Yme,age:()=>Zme,"anti-spoofing":()=>R2e,antispoof:()=>zme,blazeface:()=>Lme,"blazeface-back":()=>Jme,"blazeface-front":()=>Qme,"blazepose-detector":()=>e2e,"blazepose-full":()=>t2e,"blazepose-heavy":()=>a2e,"blazepose-lite":()=>n2e,centernet:()=>Wme,default:()=>V2e,efficientpose:()=>r2e,"efficientpose-i-lite":()=>E2e,"efficientpose-ii-lite":()=>M2e,"efficientpose-iv":()=>F2e,emotion:()=>Bme,faceboxes:()=>s2e,facemesh:()=>Vme,"facemesh-attention":()=>o2e,"facemesh-attention-pinto":()=>i2e,"facemesh-detection-full":()=>l2e,"facemesh-detection-short":()=>u2e,faceres:()=>Ume,"faceres-deep":()=>d2e,gear:()=>h2e,"gear-e1":()=>p2e,"gear-e2":()=>c2e,gender:()=>m2e,"gender-ssrnet-imdb":()=>f2e,handdetect:()=>g2e,"handlandmark-full":()=>y2e,"handlandmark-lite":()=>Gme,"handlandmark-sparse":()=>x2e,handskeleton:()=>A2e,handtrack:()=>Hme,"insightface-efficientnet-b0":()=>_2e,"insightface-ghostnet-strides1":()=>$2e,"insightface-ghostnet-strides2":()=>D2e,"insightface-mobilenet-emore":()=>P2e,"insightface-mobilenet-swish":()=>O2e,iris:()=>jme,liveness:()=>qme,meet:()=>b2e,mobileface:()=>v2e,mobilefacenet:()=>w2e,models:()=>Xme,"movenet-lightning":()=>Kme,"movenet-multipose":()=>k2e,"movenet-thunder":()=>I2e,nanodet:()=>S2e,"nanodet-e":()=>z2e,"nanodet-g":()=>L2e,"nanodet-m":()=>W2e,"nanodet-t":()=>B2e,posenet:()=>C2e,rvm:()=>T2e,selfie:()=>N2e});var zme=853098,Lme=538928,Wme=4030290,Bme=820516,Vme=1477958,Ume=6978814,Gme=2023432,Hme=2964837,jme=2599092,qme=592976,Xme=0,Kme=4650216,Yme=6920630,Zme=161240,Jme=538928,Qme=402048,e2e=5928856,t2e=6339202,a2e=27502466,n2e=2726402,r2e=5651240,s2e=2013002,i2e=2387598,o2e=2382414,l2e=1026192,u2e=201268,d2e=13957620,p2e=112438,c2e=112438,h2e=1498916,f2e=161236,m2e=201808,g2e=3515612,y2e=5431368,x2e=5286322,A2e=5502280,b2e=372228,v2e=2183192,w2e=5171976,k2e=9448838,I2e=12477112,S2e=7574558,C2e=5032780,T2e=3739355,N2e=212886,R2e=853098,E2e=2269064,M2e=5651240,F2e=25643252,_2e=13013224,$2e=8093408,D2e=8049584,P2e=6938536,O2e=12168584,z2e=12319156,L2e=7574558,W2e=1887474,B2e=5294216,V2e={antispoof:zme,blazeface:Lme,centernet:Wme,emotion:Bme,facemesh:Vme,faceres:Ume,"handlandmark-lite":Gme,handtrack:Hme,iris:jme,liveness:qme,models:Xme,"movenet-lightning":Kme,"affectnet-mobilenet":Yme,age:Zme,"blazeface-back":Jme,"blazeface-front":Qme,"blazepose-detector":e2e,"blazepose-full":t2e,"blazepose-heavy":a2e,"blazepose-lite":n2e,efficientpose:r2e,faceboxes:s2e,"facemesh-attention-pinto":i2e,"facemesh-attention":o2e,"facemesh-detection-full":l2e,"facemesh-detection-short":u2e,"faceres-deep":d2e,"gear-e1":p2e,"gear-e2":c2e,gear:h2e,"gender-ssrnet-imdb":f2e,gender:m2e,handdetect:g2e,"handlandmark-full":y2e,"handlandmark-sparse":x2e,handskeleton:A2e,meet:b2e,mobileface:v2e,mobilefacenet:w2e,"movenet-multipose":k2e,"movenet-thunder":I2e,nanodet:S2e,posenet:C2e,rvm:T2e,selfie:N2e,"anti-spoofing":R2e,"efficientpose-i-lite":E2e,"efficientpose-ii-lite":M2e,"efficientpose-iv":F2e,"insightface-efficientnet-b0":_2e,"insightface-ghostnet-strides1":$2e,"insightface-ghostnet-strides2":D2e,"insightface-mobilenet-emore":P2e,"insightface-mobilenet-swish":O2e,"nanodet-e":z2e,"nanodet-g":L2e,"nanodet-m":W2e,"nanodet-t":B2e};var Ma={cacheModels:!0,cacheSupported:!0,verbose:!0,debug:!1,modelBasePath:""},xa={};async function U2e(e,t){return Ma.debug&&K("load model fetch:",e,t),fetch(e,t)}function Tk(e){Ma.cacheModels=e.cacheModels,Ma.verbose=e.debug,Ma.modelBasePath=e.modelBasePath}async function Fe(e){var u,p,c,d;let t=yk(Ma.modelBasePath,e||"");t.toLowerCase().endsWith(".json")||(t+=".json");let a=t.includes("/")?t.split("/"):t.split("\\"),n=a[a.length-1].replace(".json",""),r="indexeddb://"+n;xa[n]={name:n,sizeFromManifest:0,sizeLoadedWeights:0,sizeDesired:Bg[n],inCache:!1,url:""},Ma.cacheSupported=typeof indexedDB!="undefined";let s={};try{s=Ma.cacheSupported&&Ma.cacheModels?await Kn.listModels():{}}catch(h){Ma.cacheSupported=!1}xa[n].inCache=Ma.cacheSupported&&Ma.cacheModels&&Object.keys(s).includes(r),xa[n].url=xa[n].inCache?r:t;let i=typeof fetch=="undefined"?{}:{fetchFunc:(h,f)=>U2e(h,f)},o=new Lp(xa[n].url,i),l=!1;try{o.findIOHandler(),Ma.debug&&K("model load handler:",o.handler)}catch(h){K("error finding model i/o handler:",t,h)}try{let h=await((u=o.handler)==null?void 0:u.load())||null;xa[n].sizeFromManifest=((p=h==null?void 0:h.weightData)==null?void 0:p.byteLength)||0,h?o.loadSync(h):o=await Y3(xa[n].inCache?r:t,i),xa[n].sizeLoadedWeights=((d=(c=o.artifacts)==null?void 0:c.weightData)==null?void 0:d.byteLength)||0,Ma.verbose&&K("load:",{model:n,url:o.modelUrl,bytes:xa[n].sizeLoadedWeights}),l=!0}catch(h){K("error loading model:",t,h)}if(l&&Ma.cacheModels&&Ma.cacheSupported&&!xa[n].inCache)try{let h=await o.save(r);Ma.debug&&K("model saved:",r,h)}catch(h){K("error saving model:",t,h)}return o}var Vg="3.0.6";var St={name:"humangl",priority:999,canvas:null,gl:null,extensions:[],webGLattr:{alpha:!1,antialias:!1,premultipliedAlpha:!1,preserveDrawingBuffer:!1,depth:!1,stencil:!1,failIfMajorPerformanceCaveat:!1,desynchronized:!0}};function j2e(){let e=St.gl;e&&(St.extensions=e.getSupportedExtensions())}function Nk(e){var t;if(e.config.backend==="humangl"&&(St.name in It().registry&&!((t=St==null?void 0:St.gl)!=null&&t.getParameter(St.gl.VERSION))&&(K("humangl error: backend invalid context"),e.models.reset()),!Z1(St.name))){try{St.canvas=Fn(100,100)}catch(r){K("humangl error: cannot create canvas:",r);return}try{if(St.gl=St.canvas.getContext("webgl2",St.webGLattr),!St.gl){K("humangl error: cannot get webgl context");return}if(!St.gl.getParameter(St.gl.VERSION).includes("2.0")){K("backend override: using fallback webgl backend as webgl 2.0 is not detected"),e.config.backend="webgl";return}St.canvas&&(St.canvas.addEventListener("webglcontextlost",s=>{throw K("humangl error:",s.type),K("possible browser memory leak using webgl or conflict with multiple backend registrations"),e.emit("error"),new Error("backend error: webgl context lost")}),St.canvas.addEventListener("webglcontextrestored",s=>{K("humangl error: context restored:",s)}),St.canvas.addEventListener("webglcontextcreationerror",s=>{K("humangl error: context create:",s)}))}catch(r){K("humangl error: cannot get webgl context:",r);return}try{Zh(2,St.gl)}catch(r){K("humangl error: cannot set webgl context:",r);return}try{let r=new _l(St.gl);Go(St.name,()=>new Bu(r),St.priority)}catch(r){K("humangl error: cannot register webgl backend:",r);return}try{Jn("webgl").forEach(s=>{let i={...s,backendName:St.name};yn(i)})}catch(r){K("humangl error: cannot update webgl backend registration:",r);return}try{W().flagRegistry.WEBGL_VERSION&&W().set("WEBGL_VERSION",2)}catch(r){K("humangl error: cannot set WebGL backend flags:",r);return}j2e();let a=nr(),n=typeof a.gpgpu!="undefined"?a.getGPGPUContext().gl:null;n?e.config.debug&&K("humangl backend registered:",{webgl:n.getParameter(n.VERSION),renderer:n.getParameter(n.RENDERER)}):K("humangl error: no current gl context:",n,St.gl)}}var ze={tf255:255,tf1:1,tf2:2,tf05:.5,tf127:127.5,rgb:[.2989,.587,.114]};function Rk(){ze.tf255=Ue(255,"float32"),ze.tf1=Ue(1,"float32"),ze.tf2=Ue(2,"float32"),ze.tf05=Ue(.5,"float32"),ze.tf127=Ue(127.5,"float32"),ze.rgb=jt([.2989,.587,.114],"float32")}async function K2e(){var e;return await ne.updateBackend(),(e=ne.tensorflow)!=null&&e.version?"tensorflow":ne.webgpu.supported&&ne.webgpu.backend?"webgpu":ne.webgl.supported&&ne.webgl.backend?"webgl":ne.wasm.supported&&ne.wasm.backend?"wasm":"cpu"}function Y2e(e){let t=[];if(!ne.kernels.includes("mod")){let a={kernelName:"Mod",backendName:ha(),kernelFunc:n=>Pe(()=>ye(n.inputs.a,te(ve(n.inputs.a,n.inputs.b),n.inputs.b)))};yn(a),ne.kernels.push("mod"),t.push("mod")}if(!ne.kernels.includes("floormod")){let a={kernelName:"FloorMod",backendName:ha(),kernelFunc:n=>Pe(()=>ke(te(Rp(n.inputs.a,n.inputs.b),n.inputs.b),_u(n.inputs.a,n.inputs.b)))};yn(a),ne.kernels.push("floormod"),t.push("floormod")}if(!ne.kernels.includes("rotatewithoffset")&&e.softwareKernels){let a={kernelName:"RotateWithOffset",backendName:ha(),kernelFunc:n=>Pe(()=>{let r=ha();Tp("cpu");let s=fe.rotateWithOffset(n.inputs.image,n.attrs.radians,n.attrs.fillValue,n.attrs.center);return Tp(r),s})};yn(a),ne.kernels.push("rotatewithoffset"),t.push("rotatewithoffset")}t.length>0&&e.debug&&K("registered kernels:",t)}var Ek={};async function Jp(e,t=!1){var a,n;if(e.state="backend",((a=e.config.backend)==null?void 0:a.length)===0&&(e.config.backend=await K2e()),t||ne.initial||e.config.backend&&e.config.backend.length>0&&ha()!==e.config.backend){let r=ae();if(e.config.backend&&e.config.backend.length>0){typeof window=="undefined"&&typeof WorkerGlobalScope!="undefined"&&e.config.debug&&e.config.debug&&K("running inside web worker"),typeof navigator!="undefined"&&((n=navigator==null?void 0:navigator.userAgent)!=null&&n.toLowerCase().includes("electron"))&&e.config.debug&&K("running inside electron");let s=Object.keys(It().registryFactory);if(e.config.backend==="humangl"&&!s.includes("humangl")&&(Nk(e),s=Object.keys(It().registryFactory)),e.config.debug&&K("available backends:",s),ne.browser&&!ne.node&&e.config.backend==="tensorflow"&&s.includes("webgl")&&(e.config.debug&&K("override: backend set to tensorflow while running in browser"),e.config.backend="webgl"),ne.node&&!ne.browser&&(e.config.backend==="webgl"||e.config.backend==="humangl")&&s.includes("tensorflow")&&(e.config.debug&&K(`override: backend set to ${e.config.backend} while running in nodejs`),e.config.backend="tensorflow"),ne.browser&&e.config.backend==="webgpu")if(typeof navigator=="undefined"||typeof navigator.gpu=="undefined")K("override: backend set to webgpu but browser does not support webgpu"),e.config.backend="webgl";else{let i=await navigator.gpu.requestAdapter();if(e.config.debug&&K("enumerated webgpu adapter:",i),!i)K("override: backend set to webgpu but browser reports no available gpu"),e.config.backend="webgl";else{let o="requestAdapterInfo"in i?await i.requestAdapterInfo():void 0;K("webgpu adapter info:",o)}}if(s.includes(e.config.backend)||(K(`error: backend ${e.config.backend} not found in registry`),e.config.backend=ne.node?"tensorflow":"webgl",e.config.debug&&K(`override: setting backend ${e.config.backend}`)),e.config.debug&&K("setting backend:",[e.config.backend]),e.config.backend==="wasm"){if(W().flagRegistry.CANVAS2D_WILL_READ_FREQUENTLY&&W().set("CANVAS2D_WILL_READ_FREQUENTLY",!0),e.config.debug&&K("wasm path:",e.config.wasmPath),typeof n0!="undefined")n0(e.config.wasmPath,e.config.wasmPlatformFetch);else throw new Error("backend error: attempting to use wasm backend but wasm path is not set");let i=!1,o=!1;try{i=await W().getAsync("WASM_HAS_MULTITHREAD_SUPPORT"),o=await W().getAsync("WASM_HAS_SIMD_SUPPORT"),e.config.debug&&K(`wasm execution: ${o?"simd":"no simd"} ${i?"multithreaded":"singlethreaded"}`),e.config.debug&&!o&&K("warning: wasm simd support is not enabled")}catch(l){K("wasm detection failed")}}try{await Tp(e.config.backend),await Np()}catch(i){return K("error: cannot set backend:",e.config.backend,i),!1}e.config.debug&&(Ek=JSON.parse(JSON.stringify(W().flags)))}if((ha()==="humangl"||ha()==="webgl")&&(W().flagRegistry.WEBGL_USE_SHAPES_UNIFORMS&&W().set("WEBGL_USE_SHAPES_UNIFORMS",!0),W().flagRegistry.WEBGL_EXP_CONV&&W().set("WEBGL_EXP_CONV",!0),e.config.debug&&typeof e.config.deallocate!="undefined"&&e.config.deallocate&&(K("changing webgl: WEBGL_DELETE_TEXTURE_THRESHOLD:",!0),W().set("WEBGL_DELETE_TEXTURE_THRESHOLD",0))),ha(),e.config.debug){let s=W().flags,i={};for(let o of Object.keys(s))Ek[o]!==s[o]&&(i[o]=s[o]);e.config.debug&&Object.keys(i).length>0&&K("backend:",ha(),"flags:",i)}if(e.config.flags&&Object.keys(e.config.flags).length>0){e.config.debug&&K("flags:",e.config.flags);for(let[s,i]of Object.entries(e.config.flags))W().set(s,i)}K1(),Rk(),e.performance.initBackend=Math.trunc(ae()-r),e.config.backend=ha(),await ne.updateBackend(),Y2e(e.config)}return!0}function m0(e,t){for(let a of e){let n={kernelName:a,backendName:t.backend,kernelFunc:r=>{var s;return t.debug&&K("kernelFunc",a,t.backend,r),(s=r==null?void 0:r.inputs)==null?void 0:s.info}};yn(n)}ne.kernels=Jn(ha()).map(a=>a.kernelName.toLowerCase())}var v0={};yr(v0,{all:()=>T1e,body:()=>y0,canvas:()=>C1e,face:()=>g0,gesture:()=>b0,hand:()=>x0,init:()=>Kg,object:()=>A0,options:()=>Dt,person:()=>S1e});var An=e=>{if(!e)K("draw error: invalid canvas");else if(!e.getContext)K("draw error: canvas context not defined");else{let t=e.getContext("2d");if(!t)K("draw error: cannot get canvas context");else return t}return null},el=e=>Math.round(e*180/Math.PI),ut=(e,t,a)=>e.replace(t,typeof a=="number"?a.toFixed(1):a),tl=(e,t)=>{if(!t.useDepth||typeof e=="undefined")return t.color;let a=Uint8ClampedArray.from([127+2*e,127-2*e,255]);return`rgba(${a[0]}, ${a[1]}, ${a[2]}, ${t.alpha})`};function bn(e,t,a,n,r){let s=t.replace(/\[.*\]/g,"").split(`
|
|
`).map(o=>o.trim()),i=Math.max(0,a);for(let o=s.length-1;o>=0;o--){let l=o*r.lineHeight+n;r.shadowColor&&r.shadowColor!==""&&(e.fillStyle=r.shadowColor,e.fillText(s[o],i+5,l+16)),e.fillStyle=r.labelColor,e.fillText(s[o],i+4,l+15)}}function lr(e,t,a,n,r){e.fillStyle=tl(n,r),e.beginPath(),e.arc(t,a,r.pointSize,0,2*Math.PI),e.fill()}function ur(e,t,a,n,r,s){if(e.beginPath(),e.lineWidth=s.lineWidth,s.useCurves){let i=(t+t+n)/2,o=(a+a+r)/2;e.ellipse(i,o,n/2,r/2,0,0,2*Math.PI)}else e.moveTo(t+s.roundRect,a),e.lineTo(t+n-s.roundRect,a),e.quadraticCurveTo(t+n,a,t+n,a+s.roundRect),e.lineTo(t+n,a+r-s.roundRect),e.quadraticCurveTo(t+n,a+r,t+n-s.roundRect,a+r),e.lineTo(t+s.roundRect,a+r),e.quadraticCurveTo(t,a+r,t,a+r-s.roundRect),e.lineTo(t,a+s.roundRect),e.quadraticCurveTo(t,a,t+s.roundRect,a),e.closePath();e.stroke()}function Ug(e,t,a){if(!(t.length<2)){e.beginPath(),e.moveTo(t[0][0],t[0][1]);for(let n of t)e.strokeStyle=tl(n[2]||0,a),e.lineTo(Math.trunc(n[0]),Math.trunc(n[1]));e.stroke(),a.fillPolygons&&(e.closePath(),e.fill())}}function Fk(e,t,a){if(!(t.length<2)){if(e.lineWidth=a.lineWidth,!a.useCurves||t.length<=2){Ug(e,t,a);return}e.moveTo(t[0][0],t[0][1]);for(let n=0;n<t.length-2;n++){let r=(t[n][0]+t[n+1][0])/2,s=(t[n][1]+t[n+1][1])/2;e.quadraticCurveTo(t[n][0],t[n][1],r,s)}e.quadraticCurveTo(t[t.length-2][0],t[t.length-2][1],t[t.length-1][0],t[t.length-1][1]),e.stroke(),a.fillPolygons&&(e.closePath(),e.fill())}}function Gg(e,t,a,n=5){let r,s,i;e.beginPath(),e.moveTo(t[0],t[1]),e.lineTo(a[0],a[1]),r=Math.atan2(a[1]-t[1],a[0]-t[0]),s=n*Math.cos(r)+a[0],i=n*Math.sin(r)+a[1],e.moveTo(s,i),r+=1/3*(2*Math.PI),s=n*Math.cos(r)+a[0],i=n*Math.sin(r)+a[1],e.lineTo(s,i),r+=1/3*(2*Math.PI),s=n*Math.cos(r)+a[0],i=n*Math.sin(r)+a[1],e.lineTo(s,i),e.closePath(),e.stroke(),e.fill()}var Dt={color:"rgba(173, 216, 230, 0.6)",labelColor:"rgba(173, 216, 230, 1)",shadowColor:"black",alpha:.5,font:'small-caps 16px "Segoe UI"',lineHeight:18,lineWidth:4,pointSize:2,roundRect:8,drawPoints:!1,drawLabels:!0,drawBoxes:!0,drawAttention:!0,drawGestures:!0,drawPolygons:!0,drawGaze:!0,fillPolygons:!1,useDepth:!0,useCurves:!1,faceLabels:"",bodyLabels:"",bodyPartLabels:"",objectLabels:"",handLabels:"",fingerLabels:"",gestureLabels:""};var _n={silhouette:[10,338,297,332,284,251,389,356,454,323,361,288,397,365,379,378,400,377,152,148,176,149,150,136,172,58,132,93,234,127,162,21,54,103,67,109],lipsUpperOuter:[185,40,39,37,0,267,269,270,409],lipsLowerOuter:[61,146,91,181,84,17,314,405,321,375,291],lipsUpperInner:[191,80,81,82,13,312,311,310,415],lipsLowerInner:[78,95,88,178,87,14,317,402,318,324,308],lipsLowerSemiOuter:[76,77,90,180,85,16,315,404,320,307,306],lipsUpperSemiOuter:[184,74,73,72,11,302,303,304,408],lipsLowerSemiInner:[62,96,89,179,86,15,316,403,319,325,292],lipsUpperSemiInner:[183,42,41,38,12,268,271,272,407],rightEyeUpper0:[246,161,160,159,158,157,173],rightEyeLower0:[33,7,163,144,145,153,154,155,133],rightEyeUpper1:[247,30,29,27,28,56,190],rightEyeLower1:[130,25,110,24,23,22,26,112,243],rightEyeUpper2:[113,225,224,223,222,221,189],rightEyeLower2:[226,31,228,229,230,231,232,233,244],rightEyeLower3:[143,111,117,118,119,120,121,128,245],rightEyebrowUpper:[156,70,63,105,66,107,55,193],rightEyebrowLower:[35,124,46,53,52,65],rightEyeIris:[473,474,475,476,477],leftEyeUpper0:[466,388,387,386,385,384,398],leftEyeLower0:[263,249,390,373,374,380,381,382,362],leftEyeUpper1:[467,260,259,257,258,286,414],leftEyeLower1:[359,255,339,254,253,252,256,341,463],leftEyeUpper2:[342,445,444,443,442,441,413],leftEyeLower2:[446,261,448,449,450,451,452,453,464],leftEyeLower3:[372,340,346,347,348,349,350,357,465],leftEyebrowUpper:[383,300,293,334,296,336,285,417],leftEyebrowLower:[265,353,276,283,282,295],leftEyeIris:[468,469,470,471,472],midwayBetweenEyes:[168],noseTip:[1],noseBottom:[2],noseRightCorner:[98],noseLeftCorner:[327],rightCheek:[205],leftCheek:[425]},Hg={count:468,mouth:13,symmetryLine:[13,_n.midwayBetweenEyes[0]]},al={leftEye:0,rightEye:1,nose:2,mouth:3,leftEar:4,rightEar:5,symmetryLine:[3,2]},jg=[{key:"EyeUpper0",indices:[9,10,11,12,13,14,15]},{key:"EyeUpper1",indices:[25,26,27,28,29,30,31]},{key:"EyeUpper2",indices:[41,42,43,44,45,46,47]},{key:"EyeLower0",indices:[0,1,2,3,4,5,6,7,8]},{key:"EyeLower1",indices:[16,17,18,19,20,21,22,23,24]},{key:"EyeLower2",indices:[32,33,34,35,36,37,38,39,40]},{key:"EyeLower3",indices:[54,55,56,57,58,59,60,61,62]},{key:"EyebrowUpper",indices:[63,64,65,66,67,68,69,70]},{key:"EyebrowLower",indices:[48,49,50,51,52,53]}],Qp=[[.499976992607117,.652534008026123],[.500025987625122,.547487020492554],[.499974012374878,.602371990680695],[.482113003730774,.471979022026062],[.500150978565216,.527155995368958],[.499909996986389,.498252987861633],[.499523013830185,.40106201171875],[.289712011814117,.380764007568359],[.499954998493195,.312398016452789],[.499987006187439,.269918978214264],[.500023007392883,.107050001621246],[.500023007392883,.666234016418457],[.5000159740448,.679224014282227],[.500023007392883,.692348003387451],[.499976992607117,.695277988910675],[.499976992607117,.70593398809433],[.499976992607117,.719385027885437],[.499976992607117,.737019002437592],[.499967992305756,.781370997428894],[.499816000461578,.562981009483337],[.473773002624512,.573909997940063],[.104906998574734,.254140973091125],[.365929991006851,.409575998783112],[.338757991790771,.41302502155304],[.311120003461838,.409460008144379],[.274657994508743,.389131009578705],[.393361985683441,.403706014156342],[.345234006643295,.344011008739471],[.370094001293182,.346076011657715],[.319321990013123,.347265005111694],[.297903001308441,.353591024875641],[.24779200553894,.410809993743896],[.396889001131058,.842755019664764],[.280097991228104,.375599980354309],[.106310002505779,.399955987930298],[.2099249958992,.391353011131287],[.355807989835739,.534406006336212],[.471751004457474,.65040397644043],[.474155008792877,.680191993713379],[.439785003662109,.657229006290436],[.414617002010345,.66654098033905],[.450374007225037,.680860996246338],[.428770989179611,.682690978050232],[.374971002340317,.727805018424988],[.486716985702515,.547628998756409],[.485300987958908,.527395009994507],[.257764995098114,.314490020275116],[.401223003864288,.455172002315521],[.429818987846375,.548614978790283],[.421351999044418,.533740997314453],[.276895999908447,.532056987285614],[.483370006084442,.499586999416351],[.33721199631691,.282882988452911],[.296391993761063,.293242990970612],[.169294998049736,.193813979625702],[.447580009698868,.302609980106354],[.392390012741089,.353887975215912],[.354490011930466,.696784019470215],[.067304998636246,.730105042457581],[.442739009857178,.572826027870178],[.457098007202148,.584792017936707],[.381974011659622,.694710969924927],[.392388999462128,.694203019142151],[.277076005935669,.271932005882263],[.422551989555359,.563233017921448],[.385919004678726,.281364023685455],[.383103013038635,.255840003490448],[.331431001424789,.119714021682739],[.229923993349075,.232002973556519],[.364500999450684,.189113974571228],[.229622006416321,.299540996551514],[.173287004232407,.278747975826263],[.472878992557526,.666198015213013],[.446828007698059,.668527007102966],[.422762006521225,.673889994621277],[.445307999849319,.580065965652466],[.388103008270264,.693961024284363],[.403039008378983,.706539988517761],[.403629004955292,.693953037261963],[.460041999816895,.557139039039612],[.431158006191254,.692366003990173],[.452181994915009,.692366003990173],[.475387006998062,.692366003990173],[.465828001499176,.779190003871918],[.472328990697861,.736225962638855],[.473087012767792,.717857003211975],[.473122000694275,.704625964164734],[.473033010959625,.695277988910675],[.427942007780075,.695277988910675],[.426479011774063,.703539967536926],[.423162013292313,.711845993995667],[.4183090031147,.720062971115112],[.390094995498657,.639572978019714],[.013953999616206,.560034036636353],[.499913990497589,.58014702796936],[.413199990987778,.69539999961853],[.409626007080078,.701822996139526],[.468080013990402,.601534962654114],[.422728985548019,.585985004901886],[.463079988956451,.593783974647522],[.37211999297142,.47341400384903],[.334562003612518,.496073007583618],[.411671012639999,.546965003013611],[.242175996303558,.14767599105835],[.290776997804642,.201445996761322],[.327338010072708,.256527006626129],[.399509996175766,.748921036720276],[.441727995872498,.261676013469696],[.429764986038208,.187834024429321],[.412198007106781,.108901023864746],[.288955003023148,.398952007293701],[.218936994671822,.435410976409912],[.41278201341629,.398970007896423],[.257135003805161,.355440020561218],[.427684992551804,.437960982322693],[.448339998722076,.536936044692993],[.178560003638268,.45755398273468],[.247308000922203,.457193970680237],[.286267012357712,.467674970626831],[.332827985286713,.460712015628815],[.368755996227264,.447206974029541],[.398963987827301,.432654976844788],[.476410001516342,.405806005001068],[.189241006970406,.523923993110657],[.228962004184723,.348950982093811],[.490725994110107,.562400996685028],[.404670000076294,.485132992267609],[.019469000399113,.401564002037048],[.426243007183075,.420431017875671],[.396993011236191,.548797011375427],[.266469985246658,.376977026462555],[.439121007919312,.51895797252655],[.032313998788595,.644356966018677],[.419054001569748,.387154996395111],[.462783008813858,.505746960639954],[.238978996872902,.779744982719421],[.198220998048782,.831938028335571],[.107550002634525,.540755033493042],[.183610007166862,.740257024765015],[.134409993886948,.333683013916016],[.385764002799988,.883153975009918],[.490967005491257,.579378008842468],[.382384985685349,.508572995662689],[.174399003386497,.397670984268188],[.318785011768341,.39623498916626],[.343364000320435,.400596976280212],[.396100014448166,.710216999053955],[.187885001301765,.588537991046906],[.430987000465393,.944064974784851],[.318993002176285,.898285031318665],[.266247987747192,.869701027870178],[.500023007392883,.190576016902924],[.499976992607117,.954452991485596],[.366169989109039,.398822009563446],[.393207013607025,.39553701877594],[.410373002290726,.391080021858215],[.194993004202843,.342101991176605],[.388664990663528,.362284004688263],[.365961998701096,.355970978736877],[.343364000320435,.355356991291046],[.318785011768341,.35834002494812],[.301414996385574,.363156020641327],[.058132998645306,.319076001644135],[.301414996385574,.387449026107788],[.499987989664078,.618434011936188],[.415838003158569,.624195992946625],[.445681989192963,.566076993942261],[.465844005346298,.620640993118286],[.49992299079895,.351523995399475],[.288718998432159,.819945991039276],[.335278987884521,.852819979190826],[.440512001514435,.902418971061707],[.128294005990028,.791940987110138],[.408771991729736,.373893976211548],[.455606997013092,.451801002025604],[.499877005815506,.908990025520325],[.375436991453171,.924192011356354],[.11421000212431,.615022003650665],[.448662012815475,.695277988910675],[.4480200111866,.704632043838501],[.447111994028091,.715808033943176],[.444831997156143,.730794012546539],[.430011987686157,.766808986663818],[.406787008047104,.685672998428345],[.400738000869751,.681069016456604],[.392399996519089,.677703022956848],[.367855995893478,.663918972015381],[.247923001646996,.601333022117615],[.452769994735718,.420849978923798],[.43639200925827,.359887003898621],[.416164010763168,.368713974952698],[.413385987281799,.692366003990173],[.228018000721931,.683571994304657],[.468268007040024,.352671027183533],[.411361992359161,.804327011108398],[.499989002943039,.469825029373169],[.479153990745544,.442654013633728],[.499974012374878,.439637005329132],[.432112008333206,.493588984012604],[.499886006116867,.866917014122009],[.49991300702095,.821729004383087],[.456548988819122,.819200992584229],[.344549000263214,.745438992977142],[.37890899181366,.574010014533997],[.374292999505997,.780184984207153],[.319687992334366,.570737957954407],[.357154995203018,.604269981384277],[.295284003019333,.621580958366394],[.447750002145767,.862477004528046],[.410986006259918,.508723020553589],[.31395098567009,.775308012962341],[.354128003120422,.812552988529205],[.324548006057739,.703992962837219],[.189096003770828,.646299958229065],[.279776990413666,.71465802192688],[.1338230073452,.682700991630554],[.336768001317978,.644733011722565],[.429883986711502,.466521978378296],[.455527991056442,.548622965812683],[.437114000320435,.558896005153656],[.467287987470627,.529924988746643],[.414712011814117,.335219979286194],[.37704598903656,.322777986526489],[.344107985496521,.320150971412659],[.312875986099243,.32233202457428],[.283526003360748,.333190023899078],[.241245999932289,.382785975933075],[.102986000478268,.468762993812561],[.267612010240555,.424560010433197],[.297879010438919,.433175981044769],[.333433985710144,.433878004550934],[.366427004337311,.426115989685059],[.396012008190155,.416696012020111],[.420121014118195,.41022801399231],[.007561000064015,.480777025222778],[.432949006557465,.569517970085144],[.458638995885849,.479089021682739],[.473466008901596,.545744001865387],[.476087987422943,.563830018043518],[.468472003936768,.555056989192963],[.433990985155106,.582361996173859],[.483518004417419,.562983989715576],[.482482999563217,.57784903049469],[.42645001411438,.389798998832703],[.438998997211456,.39649498462677],[.450067013502121,.400434017181396],[.289712011814117,.368252992630005],[.276670008897781,.363372981548309],[.517862021923065,.471948027610779],[.710287988185883,.380764007568359],[.526226997375488,.573909997940063],[.895093023777008,.254140973091125],[.634069979190826,.409575998783112],[.661242008209229,.41302502155304],[.688880026340485,.409460008144379],[.725341975688934,.389131009578705],[.606630027294159,.40370500087738],[.654766023159027,.344011008739471],[.629905998706818,.346076011657715],[.680678009986877,.347265005111694],[.702096998691559,.353591024875641],[.75221198797226,.410804986953735],[.602918028831482,.842862963676453],[.719901978969574,.375599980354309],[.893692970275879,.399959981441498],[.790081977844238,.391354024410248],[.643998026847839,.534487962722778],[.528249025344849,.65040397644043],[.525849997997284,.680191040039062],[.560214996337891,.657229006290436],[.585384011268616,.66654098033905],[.549625992774963,.680860996246338],[.57122802734375,.682691991329193],[.624852001667023,.72809898853302],[.513050019741058,.547281980514526],[.51509702205658,.527251958847046],[.742246985435486,.314507007598877],[.598631024360657,.454979002475739],[.570338010787964,.548575043678284],[.578631997108459,.533622980117798],[.723087012767792,.532054007053375],[.516445994377136,.499638974666595],[.662801027297974,.282917976379395],[.70362401008606,.293271005153656],[.830704987049103,.193813979625702],[.552385985851288,.302568018436432],[.607609987258911,.353887975215912],[.645429015159607,.696707010269165],[.932694971561432,.730105042457581],[.557260990142822,.572826027870178],[.542901992797852,.584792017936707],[.6180260181427,.694710969924927],[.607590973377228,.694203019142151],[.722943007946014,.271963000297546],[.577413976192474,.563166975975037],[.614082992076874,.281386971473694],[.616907000541687,.255886018276215],[.668509006500244,.119913995265961],[.770092010498047,.232020974159241],[.635536015033722,.189248979091644],[.77039098739624,.299556016921997],[.826722025871277,.278755009174347],[.527121007442474,.666198015213013],[.553171992301941,.668527007102966],[.577238023281097,.673889994621277],[.554691970348358,.580065965652466],[.611896991729736,.693961024284363],[.59696102142334,.706539988517761],[.596370995044708,.693953037261963],[.539958000183105,.557139039039612],[.568841993808746,.692366003990173],[.547818005084991,.692366003990173],[.52461302280426,.692366003990173],[.534089982509613,.779141008853912],[.527670979499817,.736225962638855],[.526912987232208,.717857003211975],[.526877999305725,.704625964164734],[.526966989040375,.695277988910675],[.572058022022247,.695277988910675],[.573521018028259,.703539967536926],[.57683801651001,.711845993995667],[.581691026687622,.720062971115112],[.609944999217987,.639909982681274],[.986046016216278,.560034036636353],[.5867999792099,.69539999961853],[.590372025966644,.701822996139526],[.531915009021759,.601536989212036],[.577268004417419,.585934996604919],[.536915004253387,.593786001205444],[.627542972564697,.473352015018463],[.665585994720459,.495950996875763],[.588353991508484,.546862006187439],[.757824003696442,.14767599105835],[.709249973297119,.201507985591888],[.672684013843536,.256581008434296],[.600408971309662,.74900496006012],[.55826598405838,.261672019958496],[.570303976535797,.187870979309082],[.588165998458862,.109044015407562],[.711045026779175,.398952007293701],[.781069993972778,.435405015945435],[.587247014045715,.398931980133057],[.742869973182678,.355445981025696],[.572156012058258,.437651991844177],[.55186802148819,.536570012569427],[.821442008018494,.457556009292603],[.752701997756958,.457181990146637],[.71375697851181,.467626988887787],[.66711300611496,.460672974586487],[.631101012229919,.447153985500336],[.6008620262146,.432473003864288],[.523481011390686,.405627012252808],[.810747981071472,.523926019668579],[.771045982837677,.348959028720856],[.509127020835876,.562718033790588],[.595292985439301,.485023975372314],[.980530977249146,.401564002037048],[.573499977588654,.420000016689301],[.602994978427887,.548687994480133],[.733529984951019,.376977026462555],[.560611009597778,.519016981124878],[.967685997486115,.644356966018677],[.580985009670258,.387160003185272],[.537728011608124,.505385041236877],[.760966002941132,.779752969741821],[.801778972148895,.831938028335571],[.892440974712372,.54076099395752],[.816350996494293,.740260004997253],[.865594983100891,.333687007427216],[.614073991775513,.883246004581451],[.508952975273132,.579437971115112],[.617941975593567,.508316040039062],[.825608015060425,.397674977779388],[.681214988231659,.39623498916626],[.656635999679565,.400596976280212],[.603900015354156,.710216999053955],[.81208598613739,.588539004325867],[.56801301240921,.944564998149872],[.681007981300354,.898285031318665],[.733752012252808,.869701027870178],[.633830010890961,.398822009563446],[.606792986392975,.39553701877594],[.589659988880157,.391062021255493],[.805015981197357,.342108011245728],[.611334979534149,.362284004688263],[.634037971496582,.355970978736877],[.656635999679565,.355356991291046],[.681214988231659,.35834002494812],[.698584973812103,.363156020641327],[.941866993904114,.319076001644135],[.698584973812103,.387449026107788],[.584177017211914,.624107003211975],[.554318010807037,.566076993942261],[.534153997898102,.62064003944397],[.711217999458313,.819975018501282],[.664629995822906,.852871000766754],[.559099972248077,.902631998062134],[.871706008911133,.791940987110138],[.591234028339386,.373893976211548],[.544341027736664,.451583981513977],[.624562978744507,.924192011356354],[.88577002286911,.615028977394104],[.551338016986847,.695277988910675],[.551980018615723,.704632043838501],[.552887976169586,.715808033943176],[.555167973041534,.730794012546539],[.569944024085999,.767035007476807],[.593203008174896,.685675978660583],[.599261999130249,.681069016456604],[.607599973678589,.677703022956848],[.631937980651855,.663500010967255],[.752032995223999,.601315021514893],[.547226011753082,.420395016670227],[.563543975353241,.359827995300293],[.583841025829315,.368713974952698],[.586614012718201,.692366003990173],[.771915018558502,.683578014373779],[.531597018241882,.352482974529266],[.588370978832245,.804440975189209],[.52079701423645,.442565023899078],[.567984998226166,.493479013442993],[.543282985687256,.819254994392395],[.655317008495331,.745514988899231],[.621008992195129,.574018001556396],[.625559985637665,.78031200170517],[.680198013782501,.570719003677368],[.64276397228241,.604337990283966],[.704662978649139,.621529996395111],[.552012026309967,.862591981887817],[.589071989059448,.508637011051178],[.685944974422455,.775357007980347],[.645735025405884,.812640011310577],[.675342977046967,.703978002071381],[.810858011245728,.646304965019226],[.72012197971344,.714666962623596],[.866151988506317,.682704985141754],[.663187026977539,.644596993923187],[.570082008838654,.466325998306274],[.544561982154846,.548375964164734],[.562758982181549,.558784961700439],[.531987011432648,.530140042304993],[.585271000862122,.335177004337311],[.622952997684479,.32277899980545],[.655896008014679,.320163011550903],[.687132000923157,.322345972061157],[.716481983661652,.333200991153717],[.758756995201111,.382786989212036],[.897013008594513,.468769013881683],[.732392013072968,.424547016620636],[.70211398601532,.433162987232208],[.66652500629425,.433866024017334],[.633504986763,.426087975502014],[.603875994682312,.416586995124817],[.579657971858978,.409945011138916],[.992439985275269,.480777025222778],[.567192018032074,.569419980049133],[.54136598110199,.478899002075195],[.526564002037048,.546118021011353],[.523913025856018,.563830018043518],[.531529009342194,.555056989192963],[.566035985946655,.582329034805298],[.51631098985672,.563053965568542],[.5174720287323,.577877044677734],[.573594987392426,.389806985855103],[.560697972774506,.395331978797913],[.549755990505219,.399751007556915],[.710287988185883,.368252992630005],[.723330020904541,.363372981548309]],nl=[127,34,139,11,0,37,232,231,120,72,37,39,128,121,47,232,121,128,104,69,67,175,171,148,157,154,155,118,50,101,73,39,40,9,151,108,48,115,131,194,204,211,74,40,185,80,42,183,40,92,186,230,229,118,202,212,214,83,18,17,76,61,146,160,29,30,56,157,173,106,204,194,135,214,192,203,165,98,21,71,68,51,45,4,144,24,23,77,146,91,205,50,187,201,200,18,91,106,182,90,91,181,85,84,17,206,203,36,148,171,140,92,40,39,193,189,244,159,158,28,247,246,161,236,3,196,54,68,104,193,168,8,117,228,31,189,193,55,98,97,99,126,47,100,166,79,218,155,154,26,209,49,131,135,136,150,47,126,217,223,52,53,45,51,134,211,170,140,67,69,108,43,106,91,230,119,120,226,130,247,63,53,52,238,20,242,46,70,156,78,62,96,46,53,63,143,34,227,173,155,133,123,117,111,44,125,19,236,134,51,216,206,205,154,153,22,39,37,167,200,201,208,36,142,100,57,212,202,20,60,99,28,158,157,35,226,113,160,159,27,204,202,210,113,225,46,43,202,204,62,76,77,137,123,116,41,38,72,203,129,142,64,98,240,49,102,64,41,73,74,212,216,207,42,74,184,169,170,211,170,149,176,105,66,69,122,6,168,123,147,187,96,77,90,65,55,107,89,90,180,101,100,120,63,105,104,93,137,227,15,86,85,129,102,49,14,87,86,55,8,9,100,47,121,145,23,22,88,89,179,6,122,196,88,95,96,138,172,136,215,58,172,115,48,219,42,80,81,195,3,51,43,146,61,171,175,199,81,82,38,53,46,225,144,163,110,246,33,7,52,65,66,229,228,117,34,127,234,107,108,69,109,108,151,48,64,235,62,78,191,129,209,126,111,35,143,163,161,246,117,123,50,222,65,52,19,125,141,221,55,65,3,195,197,25,7,33,220,237,44,70,71,139,122,193,245,247,130,33,71,21,162,153,158,159,170,169,150,188,174,196,216,186,92,144,160,161,2,97,167,141,125,241,164,167,37,72,38,12,145,159,160,38,82,13,63,68,71,226,35,111,158,153,154,101,50,205,206,92,165,209,198,217,165,167,97,220,115,218,133,112,243,239,238,241,214,135,169,190,173,133,171,208,32,125,44,237,86,87,178,85,86,179,84,85,180,83,84,181,201,83,182,137,93,132,76,62,183,61,76,184,57,61,185,212,57,186,214,207,187,34,143,156,79,239,237,123,137,177,44,1,4,201,194,32,64,102,129,213,215,138,59,166,219,242,99,97,2,94,141,75,59,235,24,110,228,25,130,226,23,24,229,22,23,230,26,22,231,112,26,232,189,190,243,221,56,190,28,56,221,27,28,222,29,27,223,30,29,224,247,30,225,238,79,20,166,59,75,60,75,240,147,177,215,20,79,166,187,147,213,112,233,244,233,128,245,128,114,188,114,217,174,131,115,220,217,198,236,198,131,134,177,132,58,143,35,124,110,163,7,228,110,25,356,389,368,11,302,267,452,350,349,302,303,269,357,343,277,452,453,357,333,332,297,175,152,377,384,398,382,347,348,330,303,304,270,9,336,337,278,279,360,418,262,431,304,408,409,310,415,407,270,409,410,450,348,347,422,430,434,313,314,17,306,307,375,387,388,260,286,414,398,335,406,418,364,367,416,423,358,327,251,284,298,281,5,4,373,374,253,307,320,321,425,427,411,421,313,18,321,405,406,320,404,405,315,16,17,426,425,266,377,400,369,322,391,269,417,465,464,386,257,258,466,260,388,456,399,419,284,332,333,417,285,8,346,340,261,413,441,285,327,460,328,355,371,329,392,439,438,382,341,256,429,420,360,364,394,379,277,343,437,443,444,283,275,440,363,431,262,369,297,338,337,273,375,321,450,451,349,446,342,467,293,334,282,458,461,462,276,353,383,308,324,325,276,300,293,372,345,447,382,398,362,352,345,340,274,1,19,456,248,281,436,427,425,381,256,252,269,391,393,200,199,428,266,330,329,287,273,422,250,462,328,258,286,384,265,353,342,387,259,257,424,431,430,342,353,276,273,335,424,292,325,307,366,447,345,271,303,302,423,266,371,294,455,460,279,278,294,271,272,304,432,434,427,272,407,408,394,430,431,395,369,400,334,333,299,351,417,168,352,280,411,325,319,320,295,296,336,319,403,404,330,348,349,293,298,333,323,454,447,15,16,315,358,429,279,14,15,316,285,336,9,329,349,350,374,380,252,318,402,403,6,197,419,318,319,325,367,364,365,435,367,397,344,438,439,272,271,311,195,5,281,273,287,291,396,428,199,311,271,268,283,444,445,373,254,339,263,466,249,282,334,296,449,347,346,264,447,454,336,296,299,338,10,151,278,439,455,292,407,415,358,371,355,340,345,372,390,249,466,346,347,280,442,443,282,19,94,370,441,442,295,248,419,197,263,255,359,440,275,274,300,383,368,351,412,465,263,467,466,301,368,389,380,374,386,395,378,379,412,351,419,436,426,322,373,390,388,2,164,393,370,462,461,164,0,267,302,11,12,374,373,387,268,12,13,293,300,301,446,261,340,385,384,381,330,266,425,426,423,391,429,355,437,391,327,326,440,457,438,341,382,362,459,457,461,434,430,394,414,463,362,396,369,262,354,461,457,316,403,402,315,404,403,314,405,404,313,406,405,421,418,406,366,401,361,306,408,407,291,409,408,287,410,409,432,436,410,434,416,411,264,368,383,309,438,457,352,376,401,274,275,4,421,428,262,294,327,358,433,416,367,289,455,439,462,370,326,2,326,370,305,460,455,254,449,448,255,261,446,253,450,449,252,451,450,256,452,451,341,453,452,413,464,463,441,413,414,258,442,441,257,443,442,259,444,443,260,445,444,467,342,445,459,458,250,289,392,290,290,328,460,376,433,435,250,290,392,411,416,433,341,463,464,453,464,465,357,465,412,343,412,399,360,363,440,437,399,456,420,456,363,401,435,288,372,383,353,339,255,249,448,261,255,133,243,190,133,155,112,33,246,247,33,130,25,398,384,286,362,398,414,362,463,341,263,359,467,263,249,255,466,467,260,75,60,166,238,239,79,162,127,139,72,11,37,121,232,120,73,72,39,114,128,47,233,232,128,103,104,67,152,175,148,173,157,155,119,118,101,74,73,40,107,9,108,49,48,131,32,194,211,184,74,185,191,80,183,185,40,186,119,230,118,210,202,214,84,83,17,77,76,146,161,160,30,190,56,173,182,106,194,138,135,192,129,203,98,54,21,68,5,51,4,145,144,23,90,77,91,207,205,187,83,201,18,181,91,182,180,90,181,16,85,17,205,206,36,176,148,140,165,92,39,245,193,244,27,159,28,30,247,161,174,236,196,103,54,104,55,193,8,111,117,31,221,189,55,240,98,99,142,126,100,219,166,218,112,155,26,198,209,131,169,135,150,114,47,217,224,223,53,220,45,134,32,211,140,109,67,108,146,43,91,231,230,120,113,226,247,105,63,52,241,238,242,124,46,156,95,78,96,70,46,63,116,143,227,116,123,111,1,44,19,3,236,51,207,216,205,26,154,22,165,39,167,199,200,208,101,36,100,43,57,202,242,20,99,56,28,157,124,35,113,29,160,27,211,204,210,124,113,46,106,43,204,96,62,77,227,137,116,73,41,72,36,203,142,235,64,240,48,49,64,42,41,74,214,212,207,183,42,184,210,169,211,140,170,176,104,105,69,193,122,168,50,123,187,89,96,90,66,65,107,179,89,180,119,101,120,68,63,104,234,93,227,16,15,85,209,129,49,15,14,86,107,55,9,120,100,121,153,145,22,178,88,179,197,6,196,89,88,96,135,138,136,138,215,172,218,115,219,41,42,81,5,195,51,57,43,61,208,171,199,41,81,38,224,53,225,24,144,110,105,52,66,118,229,117,227,34,234,66,107,69,10,109,151,219,48,235,183,62,191,142,129,126,116,111,143,7,163,246,118,117,50,223,222,52,94,19,141,222,221,65,196,3,197,45,220,44,156,70,139,188,122,245,139,71,162,145,153,159,149,170,150,122,188,196,206,216,92,163,144,161,164,2,167,242,141,241,0,164,37,11,72,12,144,145,160,12,38,13,70,63,71,31,226,111,157,158,154,36,101,205,203,206,165,126,209,217,98,165,97,237,220,218,237,239,241,210,214,169,140,171,32,241,125,237,179,86,178,180,85,179,181,84,180,182,83,181,194,201,182,177,137,132,184,76,183,185,61,184,186,57,185,216,212,186,192,214,187,139,34,156,218,79,237,147,123,177,45,44,4,208,201,32,98,64,129,192,213,138,235,59,219,141,242,97,97,2,141,240,75,235,229,24,228,31,25,226,230,23,229,231,22,230,232,26,231,233,112,232,244,189,243,189,221,190,222,28,221,223,27,222,224,29,223,225,30,224,113,247,225,99,60,240,213,147,215,60,20,166,192,187,213,243,112,244,244,233,245,245,128,188,188,114,174,134,131,220,174,217,236,236,198,134,215,177,58,156,143,124,25,110,7,31,228,25,264,356,368,0,11,267,451,452,349,267,302,269,350,357,277,350,452,357,299,333,297,396,175,377,381,384,382,280,347,330,269,303,270,151,9,337,344,278,360,424,418,431,270,304,409,272,310,407,322,270,410,449,450,347,432,422,434,18,313,17,291,306,375,259,387,260,424,335,418,434,364,416,391,423,327,301,251,298,275,281,4,254,373,253,375,307,321,280,425,411,200,421,18,335,321,406,321,320,405,314,315,17,423,426,266,396,377,369,270,322,269,413,417,464,385,386,258,248,456,419,298,284,333,168,417,8,448,346,261,417,413,285,326,327,328,277,355,329,309,392,438,381,382,256,279,429,360,365,364,379,355,277,437,282,443,283,281,275,363,395,431,369,299,297,337,335,273,321,348,450,349,359,446,467,283,293,282,250,458,462,300,276,383,292,308,325,283,276,293,264,372,447,346,352,340,354,274,19,363,456,281,426,436,425,380,381,252,267,269,393,421,200,428,371,266,329,432,287,422,290,250,328,385,258,384,446,265,342,386,387,257,422,424,430,445,342,276,422,273,424,306,292,307,352,366,345,268,271,302,358,423,371,327,294,460,331,279,294,303,271,304,436,432,427,304,272,408,395,394,431,378,395,400,296,334,299,6,351,168,376,352,411,307,325,320,285,295,336,320,319,404,329,330,349,334,293,333,366,323,447,316,15,315,331,358,279,317,14,316,8,285,9,277,329,350,253,374,252,319,318,403,351,6,419,324,318,325,397,367,365,288,435,397,278,344,439,310,272,311,248,195,281,375,273,291,175,396,199,312,311,268,276,283,445,390,373,339,295,282,296,448,449,346,356,264,454,337,336,299,337,338,151,294,278,455,308,292,415,429,358,355,265,340,372,388,390,466,352,346,280,295,442,282,354,19,370,285,441,295,195,248,197,457,440,274,301,300,368,417,351,465,251,301,389,385,380,386,394,395,379,399,412,419,410,436,322,387,373,388,326,2,393,354,370,461,393,164,267,268,302,12,386,374,387,312,268,13,298,293,301,265,446,340,380,385,381,280,330,425,322,426,391,420,429,437,393,391,326,344,440,438,458,459,461,364,434,394,428,396,262,274,354,457,317,316,402,316,315,403,315,314,404,314,313,405,313,421,406,323,366,361,292,306,407,306,291,408,291,287,409,287,432,410,427,434,411,372,264,383,459,309,457,366,352,401,1,274,4,418,421,262,331,294,358,435,433,367,392,289,439,328,462,326,94,2,370,289,305,455,339,254,448,359,255,446,254,253,449,253,252,450,252,256,451,256,341,452,414,413,463,286,441,414,286,258,441,258,257,442,257,259,443,259,260,444,260,467,445,309,459,250,305,289,290,305,290,460,401,376,435,309,250,392,376,411,433,453,341,464,357,453,465,343,357,412,437,343,399,344,360,440,420,437,456,360,420,363,361,401,288,265,372,353,390,339,249,339,448,255];var Z2e=[127,234,132,58,172,150,149,148,152,377,378,379,397,288,361,454,356,70,63,105,66,107,336,296,334,293,300,168,6,195,4,98,97,2,326,327,33,160,158,133,153,144,362,385,387,263,373,380,57,40,37,0,267,270,287,321,314,17,84,91,78,81,13,311,308,402,14,178],J2e=[33,133,362,263,1,62,308,159,145,386,374,6,102,331,2,13,14,70,105,107,336,334,300,54,10,284,50,280,234,454,58,288,152],Q2e=[33,133,362,263,1,78,308],_Ae=Z2e.map(e=>Qp[e]),$Ae=J2e.map(e=>Qp[e]),DAe=Q2e.map(e=>Qp[e]);function hs(e){let t=e.map(a=>a[0]);return t.push(e[e.length-1][1]),t}var e1e=[[61,146],[146,91],[91,181],[181,84],[84,17],[17,314],[314,405],[405,321],[321,375],[375,291],[61,185],[185,40],[40,39],[39,37],[37,0],[0,267],[267,269],[269,270],[270,409],[409,291],[78,95],[95,88],[88,178],[178,87],[87,14],[14,317],[317,402],[402,318],[318,324],[324,308],[78,191],[191,80],[80,81],[81,82],[82,13],[13,312],[312,311],[311,310],[310,415],[415,308]],t1e=[[263,249],[249,390],[390,373],[373,374],[374,380],[380,381],[381,382],[382,362],[263,466],[466,388],[388,387],[387,386],[386,385],[385,384],[384,398],[398,362]],a1e=[[276,283],[283,282],[282,295],[295,285],[300,293],[293,334],[334,296],[296,336]],n1e=[[474,475],[475,476],[476,477],[477,474]],r1e=[[33,7],[7,163],[163,144],[144,145],[145,153],[153,154],[154,155],[155,133],[33,246],[246,161],[161,160],[160,159],[159,158],[158,157],[157,173],[173,133]],s1e=[[46,53],[53,52],[52,65],[65,55],[70,63],[63,105],[105,66],[66,107]],i1e=[[469,470],[470,471],[471,472],[472,469]],o1e=[[10,338],[338,297],[297,332],[332,284],[284,251],[251,389],[389,356],[356,454],[454,323],[323,361],[361,288],[288,397],[397,365],[365,379],[379,378],[378,400],[400,377],[377,152],[152,148],[148,176],[176,149],[149,150],[150,136],[136,172],[172,58],[58,132],[132,93],[93,234],[234,127],[127,162],[162,21],[21,54],[54,103],[103,67],[67,109],[109,10]],PAe={lips:hs(e1e),leftEye:hs(t1e),leftEyebrow:hs(a1e),leftIris:hs(n1e),rightEye:hs(r1e),rightEyebrow:hs(s1e),rightIris:hs(i1e),faceOval:hs(o1e)};var l1e=[[61,146],[146,91],[91,181],[181,84],[84,17],[17,314],[314,405],[405,321],[321,375],[375,291],[61,185],[185,40],[40,39],[39,37],[37,0],[0,267],[267,269],[269,270],[270,409],[409,291],[78,95],[95,88],[88,178],[178,87],[87,14],[14,317],[317,402],[402,318],[318,324],[324,308],[78,191],[191,80],[80,81],[81,82],[82,13],[13,312],[312,311],[311,310],[310,415],[415,308]],u1e=[[263,249],[249,390],[390,373],[373,374],[374,380],[380,381],[381,382],[382,362],[263,466],[466,388],[388,387],[387,386],[386,385],[385,384],[384,398],[398,362]],d1e=[[276,283],[283,282],[282,295],[295,285],[300,293],[293,334],[334,296],[296,336]],p1e=[[474,475],[475,476],[476,477],[477,474]],c1e=[[33,7],[7,163],[163,144],[144,145],[145,153],[153,154],[154,155],[155,133],[33,246],[246,161],[161,160],[160,159],[159,158],[158,157],[157,173],[173,133]],h1e=[[46,53],[53,52],[52,65],[65,55],[70,63],[63,105],[105,66],[66,107]],f1e=[[469,470],[470,471],[471,472],[472,469]],m1e=[[10,338],[338,297],[297,332],[332,284],[284,251],[251,389],[389,356],[356,454],[454,323],[323,361],[361,288],[288,397],[397,365],[365,379],[379,378],[378,400],[400,377],[377,152],[152,148],[148,176],[176,149],[149,150],[150,136],[136,172],[172,58],[58,132],[132,93],[93,234],[234,127],[127,162],[162,21],[21,54],[54,103],[103,67],[67,109],[109,10]];function fs(e){let t=e.map(a=>a[0]);return t.push(e[e.length-1][1]),t}var g1e={lips:fs(l1e),leftEye:fs(u1e),leftEyebrow:fs(d1e),leftIris:fs(p1e),rightEye:fs(c1e),rightEyebrow:fs(h1e),rightIris:fs(f1e),faceOval:fs(m1e)},y1e=Object.entries(g1e).map(([e,t])=>t.map(a=>[a,e])).flat(),OAe=new Map(y1e),ec=[61,146,91,181,84,17,314,405,321,375,291,185,40,39,37,0,267,269,270,409,78,95,88,178,87,14,317,402,318,324,308,191,80,81,82,13,312,311,310,415,76,77,90,180,85,16,315,404,320,307,306,184,74,73,72,11,302,303,304,408,62,96,89,179,86,15,316,403,319,325,292,183,42,41,38,12,268,271,272,407],rl=[33,7,163,144,145,153,154,155,133,246,161,160,159,158,157,173,130,25,110,24,23,22,26,112,243,247,30,29,27,28,56,190,226,31,228,229,230,231,232,233,244,113,225,224,223,222,221,189,35,124,46,53,52,65,143,111,117,118,119,120,121,128,245,156,70,63,105,66,107,55,193],sl=[263,249,390,373,374,380,381,382,362,466,388,387,386,385,384,398,359,255,339,254,253,252,256,341,463,467,260,259,257,258,286,414,446,261,448,449,450,451,452,453,464,342,445,444,443,442,441,413,265,353,276,283,282,295,372,340,346,347,348,349,350,357,465,383,300,293,334,296,336,285,417];var rt;function x1e(e,t){var n,r,s,i,o,l,u,p,c;if(!rt.drawLabels||((n=rt.faceLabels)==null?void 0:n.length)===0)return;let a=rt.faceLabels.slice();if(a=ut(a,"[id]",e.id.toFixed(0)),e.score&&(a=ut(a,"[score]",100*e.score)),e.gender&&(a=ut(a,"[gender]",e.gender)),e.genderScore&&(a=ut(a,"[genderScore]",100*e.genderScore)),e.age&&(a=ut(a,"[age]",e.age)),e.distance&&(a=ut(a,"[distance]",100*e.distance)),e.real&&(a=ut(a,"[real]",100*e.real)),e.live&&(a=ut(a,"[live]",100*e.live)),e.emotion&&e.emotion.length>0){let d=e.emotion.map(h=>`${Math.trunc(100*h.score)}% ${h.emotion}`);d.length>3&&(d.length=3),a=ut(a,"[emotions]",d.join(" "))}(s=(r=e.rotation)==null?void 0:r.angle)!=null&&s.roll&&(a=ut(a,"[roll]",el(e.rotation.angle.roll))),(o=(i=e.rotation)==null?void 0:i.angle)!=null&&o.yaw&&(a=ut(a,"[yaw]",el(e.rotation.angle.yaw))),(u=(l=e.rotation)==null?void 0:l.angle)!=null&&u.pitch&&(a=ut(a,"[pitch]",el(e.rotation.angle.pitch))),(c=(p=e.rotation)==null?void 0:p.gaze)!=null&&c.bearing&&(a=ut(a,"[gaze]",el(e.rotation.gaze.bearing))),bn(t,a,e.box[0],e.box[1],rt)}function A1e(e,t){var a,n,r,s;if((a=e.annotations)!=null&&a.leftEyeIris&&((n=e.annotations)!=null&&n.leftEyeIris[0])){t.strokeStyle=rt.useDepth?"rgba(255, 200, 255, 0.3)":rt.color,t.beginPath();let i=Math.abs(e.annotations.leftEyeIris[3][0]-e.annotations.leftEyeIris[1][0])/2,o=Math.abs(e.annotations.leftEyeIris[4][1]-e.annotations.leftEyeIris[2][1])/2;t.ellipse(e.annotations.leftEyeIris[0][0],e.annotations.leftEyeIris[0][1],i,o,0,0,2*Math.PI),t.stroke(),rt.fillPolygons&&(t.fillStyle=rt.useDepth?"rgba(255, 255, 200, 0.3)":rt.color,t.fill())}if((r=e.annotations)!=null&&r.rightEyeIris&&((s=e.annotations)!=null&&s.rightEyeIris[0])){t.strokeStyle=rt.useDepth?"rgba(255, 200, 255, 0.3)":rt.color,t.beginPath();let i=Math.abs(e.annotations.rightEyeIris[3][0]-e.annotations.rightEyeIris[1][0])/2,o=Math.abs(e.annotations.rightEyeIris[4][1]-e.annotations.rightEyeIris[2][1])/2;t.ellipse(e.annotations.rightEyeIris[0][0],e.annotations.rightEyeIris[0][1],i,o,0,0,2*Math.PI),t.stroke(),rt.fillPolygons&&(t.fillStyle=rt.useDepth?"rgba(255, 255, 200, 0.3)":rt.color,t.fill())}}function b1e(e,t){var a;if(rt.drawGaze&&((a=e.rotation)!=null&&a.angle)&&typeof Path2D!="undefined"){t.strokeStyle="pink";let n=e.box[0]+e.box[2]/2-e.box[3]*el(e.rotation.angle.yaw)/90,r=e.box[1]+e.box[3]/2+e.box[2]*el(e.rotation.angle.pitch)/90,s=new Path2D(`
|
|
M ${e.box[0]+e.box[2]/2} ${e.box[1]}
|
|
C
|
|
${n} ${e.box[1]},
|
|
${n} ${e.box[1]+e.box[3]},
|
|
${e.box[0]+e.box[2]/2} ${e.box[1]+e.box[3]}
|
|
`),i=new Path2D(`
|
|
M ${e.box[0]} ${e.box[1]+e.box[3]/2}
|
|
C
|
|
${e.box[0]} ${r},
|
|
${e.box[0]+e.box[2]} ${r},
|
|
${e.box[0]+e.box[2]} ${e.box[1]+e.box[3]/2}
|
|
`);t.stroke(i),t.stroke(s)}}function v1e(e,t){var a;if(rt.drawGaze&&((a=e.rotation)!=null&&a.gaze.strength)&&e.rotation.gaze.bearing&&e.annotations.leftEyeIris&&e.annotations.rightEyeIris&&e.annotations.leftEyeIris[0]&&e.annotations.rightEyeIris[0]){t.strokeStyle="pink",t.fillStyle="pink";let n=[e.annotations.leftEyeIris[0][0]+Math.sin(e.rotation.gaze.bearing)*e.rotation.gaze.strength*e.box[3],e.annotations.leftEyeIris[0][1]+Math.cos(e.rotation.gaze.bearing)*e.rotation.gaze.strength*e.box[2]];Gg(t,[e.annotations.leftEyeIris[0][0],e.annotations.leftEyeIris[0][1]],[n[0],n[1]],4);let r=[e.annotations.rightEyeIris[0][0]+Math.sin(e.rotation.gaze.bearing)*e.rotation.gaze.strength*e.box[3],e.annotations.rightEyeIris[0][1]+Math.cos(e.rotation.gaze.bearing)*e.rotation.gaze.strength*e.box[2]];Gg(t,[e.annotations.rightEyeIris[0][0],e.annotations.rightEyeIris[0][1]],[r[0],r[1]],4)}}function w1e(e,t){if(rt.drawPolygons&&e.mesh.length>=468){t.lineWidth=1;for(let a=0;a<nl.length/3;a++){let n=[nl[a*3+0],nl[a*3+1],nl[a*3+2]].map(r=>e.mesh[r]);Ug(t,n,rt)}A1e(e,t)}}function k1e(e,t){if(rt.drawPoints)if((e==null?void 0:e.mesh.length)>=468)for(let a=0;a<e.mesh.length;a++)lr(t,e.mesh[a][0],e.mesh[a][1],e.mesh[a][2],rt),rt.drawAttention&&(ec.includes(a)&&lr(t,e.mesh[a][0],e.mesh[a][1],e.mesh[a][2]+127,rt),rl.includes(a)&&lr(t,e.mesh[a][0],e.mesh[a][1],e.mesh[a][2]-127,rt),sl.includes(a)&&lr(t,e.mesh[a][0],e.mesh[a][1],e.mesh[a][2]-127,rt));else for(let[a,n]of Object.entries((e==null?void 0:e.annotations)||{})){if(!(n!=null&&n[0]))continue;let r=n[0];lr(t,r[0],r[1],0,rt),rt.drawLabels&&bn(t,a,r[0],r[1],rt)}}function I1e(e,t){rt.drawBoxes&&ur(t,e.box[0],e.box[1],e.box[2],e.box[3],rt)}function g0(e,t,a){if(rt=Et(Dt,a),!t||!e)return;let n=An(e);if(n){n.font=rt.font,n.strokeStyle=rt.color,n.fillStyle=rt.color;for(let r of t)I1e(r,n),x1e(r,n),r.mesh&&r.mesh.length>0&&(k1e(r,n),w1e(r,n),b1e(r,n),v1e(r,n))}}function y0(e,t,a){var s,i;let n=Et(Dt,a);if(!t||!e)return;let r=An(e);if(r){r.lineJoin="round";for(let o=0;o<t.length;o++){if(r.strokeStyle=n.color,r.fillStyle=n.color,r.lineWidth=n.lineWidth,r.font=n.font,n.drawBoxes&&t[o].box&&t[o].box.length===4&&(ur(r,t[o].box[0],t[o].box[1],t[o].box[2],t[o].box[3],n),n.drawLabels&&((s=n.bodyLabels)==null?void 0:s.length)>0)){let l=n.bodyLabels.slice();l=ut(l,"[id]",t[o].id.toFixed(0)),l=ut(l,"[score]",100*t[o].score),bn(r,l,t[o].box[0],t[o].box[1],n)}if(n.drawPoints&&t[o].keypoints)for(let l=0;l<t[o].keypoints.length;l++)!t[o].keypoints[l].score||t[o].keypoints[l].score===0||(r.fillStyle=tl(t[o].keypoints[l].position[2],n),lr(r,t[o].keypoints[l].position[0],t[o].keypoints[l].position[1],0,n));if(n.drawLabels&&((i=n.bodyPartLabels)==null?void 0:i.length)>0&&t[o].keypoints){r.font=n.font;for(let l of t[o].keypoints){if(!l.score||l.score===0)continue;let u=n.bodyPartLabels.slice();u=ut(u,"[label]",l.part),u=ut(u,"[score]",100*l.score),bn(r,u,l.position[0],l.position[1],n)}}if(n.drawPolygons&&t[o].keypoints&&t[o].annotations)for(let l of Object.values(t[o].annotations))for(let u of l)Fk(r,u,n)}}}function x0(e,t,a){var s,i;let n=Et(Dt,a);if(!t||!e)return;let r=An(e);if(r){r.lineJoin="round",r.font=n.font;for(let o of t){if(n.drawBoxes){if(r.strokeStyle=n.color,r.fillStyle=n.color,ur(r,o.box[0],o.box[1],o.box[2],o.box[3],n),n.drawLabels&&((s=n.handLabels)==null?void 0:s.length)>0){let l=n.handLabels.slice();l=ut(l,"[id]",o.id.toFixed(0)),l=ut(l,"[label]",o.label),l=ut(l,"[score]",100*o.score),bn(r,l,o.box[0],o.box[1],n)}r.stroke()}if(n.drawPoints&&o.keypoints&&o.keypoints.length>0)for(let l of o.keypoints)r.fillStyle=tl(l[2],n),lr(r,l[0],l[1],0,n);if(n.drawLabels&&o.annotations&&((i=n.fingerLabels)==null?void 0:i.length)>0)for(let[l,u]of Object.entries(o.annotations)){let p=n.fingerLabels.slice();p=ut(p,"[label]",l),bn(r,p,u[u.length-1][0],u[u.length-1][1],n)}if(n.drawPolygons&&o.annotations){let l=u=>{if(!(!u||u.length===0||!u[0]))for(let p=0;p<u.length;p++){r.beginPath();let c=u[p][2]||0;r.strokeStyle=tl(p*c,n),r.moveTo(u[p>0?p-1:0][0],u[p>0?p-1:0][1]),r.lineTo(u[p][0],u[p][1]),r.stroke()}};r.lineWidth=n.lineWidth,l(o.annotations.index),l(o.annotations.middle),l(o.annotations.ring),l(o.annotations.pinky),l(o.annotations.thumb)}}}}function A0(e,t,a){var s;let n=Et(Dt,a);if(!t||!e)return;let r=An(e);if(r){r.lineJoin="round",r.font=n.font;for(let i of t)if(n.drawBoxes){if(r.strokeStyle=n.color,r.fillStyle=n.color,ur(r,i.box[0],i.box[1],i.box[2],i.box[3],n),n.drawLabels&&((s=n.objectLabels)==null?void 0:s.length)>0){let o=n.objectLabels.slice();o=ut(o,"[id]",i.id.toFixed(0)),o=ut(o,"[label]",i.label),o=ut(o,"[score]",100*i.score),bn(r,o,i.box[0],i.box[1],n)}r.stroke()}}}function b0(e,t,a){var r;let n=Et(Dt,a);if(!(!t||!e)&&n.drawGestures&&((r=n.gestureLabels)==null?void 0:r.length)>0){let s=An(e);if(!s)return;s.font=n.font,s.fillStyle=n.color;let i=1;for(let o=0;o<t.length;o++){let[l,u]=Object.entries(t[o]);if(u.length>1&&u[1].length>0){let p=l[1]>0?`#${l[1]}`:"",c=n.gestureLabels.slice();c=ut(c,"[where]",l[0]),c=ut(c,"[who]",p),c=ut(c,"[what]",u[1]),bn(s,c,8,2+i*n.lineHeight,n),i+=1}}}}var ms={face:`face
|
|
confidence: [score]%
|
|
[gender] [genderScore]%
|
|
age: [age] years
|
|
distance: [distance]cm
|
|
real: [real]%
|
|
live: [live]%
|
|
[emotions]
|
|
roll: [roll]\xB0 yaw:[yaw]\xB0 pitch:[pitch]\xB0
|
|
gaze: [gaze]\xB0`,body:"body [score]%",bodyPart:"[label] [score]%",object:"[label] [score]%",hand:"[label] [score]%",finger:"[label]",gesture:"[where] [who]: [what]"};var Xg=0;function S1e(e,t,a){let n=Et(Dt,a);if(!t||!e)return;let r=An(e);if(r){r.lineJoin="round",r.font=n.font;for(let s=0;s<t.length;s++)if(n.drawBoxes){if(r.strokeStyle=n.color,r.fillStyle=n.color,ur(r,t[s].box[0],t[s].box[1],t[s].box[2],t[s].box[3],n),n.drawLabels){let i=`person #${s}`;n.shadowColor&&n.shadowColor!==""&&(r.fillStyle=n.shadowColor,r.fillText(i,t[s].box[0]+3,1+t[s].box[1]+n.lineHeight,t[s].box[2])),r.fillStyle=n.labelColor,r.fillText(i,t[s].box[0]+2,0+t[s].box[1]+n.lineHeight,t[s].box[2])}r.stroke()}}}function C1e(e,t){if(!e||!t)return;let a=An(t);a&&a.drawImage(e,0,0)}async function T1e(e,t,a){if(!(t!=null&&t.performance)||!e)return null;let n=ae(),r=Et(Dt,a),s=Promise.all([g0(e,t.face,r),y0(e,t.body,r),x0(e,t.hand,r),A0(e,t.object,r),b0(e,t.gesture,r)]);return Xg=ne.perfadd?Xg+Math.round(ae()-n):Math.round(ae()-n),t.performance.draw=Xg,s}function Kg(){Dt.faceLabels=ms.face,Dt.bodyLabels=ms.body,Dt.bodyPartLabels=ms.bodyPart,Dt.handLabels=ms.hand,Dt.fingerLabels=ms.finger,Dt.objectLabels=ms.object,Dt.gestureLabels=ms.gesture}var w0={};yr(w0,{connected:()=>Zg,kpt:()=>Yg});var Yg=["nose","leftEyeInside","leftEye","leftEyeOutside","rightEyeInside","rightEye","rightEyeOutside","leftEar","rightEar","leftMouth","rightMouth","leftShoulder","rightShoulder","leftElbow","rightElbow","leftWrist","rightWrist","leftPinky","rightPinky","leftIndex","rightIndex","leftThumb","rightThumb","leftHip","rightHip","leftKnee","rightKnee","leftAnkle","rightAnkle","leftHeel","rightHeel","leftFoot","rightFoot","bodyCenter","bodyTop","leftPalm","leftHand","rightPalm","rightHand"],Zg={shoulders:["leftShoulder","rightShoulder"],hips:["rightHip","leftHip"],mouth:["leftMouth","rightMouth"],leftLegUpper:["leftHip","leftKnee"],leftLegLower:["leftKnee","leftAnkle"],leftFoot:["leftAnkle","leftHeel","leftFoot"],leftTorso:["leftShoulder","leftHip"],leftArmUpper:["leftShoulder","leftElbow"],leftArmLower:["leftElbow","leftWrist"],leftHand:["leftWrist","leftPalm"],leftHandPinky:["leftPalm","leftPinky"],leftHandIndex:["leftPalm","leftIndex"],leftHandThumb:["leftPalm","leftThumb"],leftEyeOutline:["leftEyeInside","leftEyeOutside"],rightLegUpper:["rightHip","rightKnee"],rightLegLower:["rightKnee","rightAnkle"],rightFoot:["rightAnkle","rightHeel","rightFoot"],rightTorso:["rightShoulder","rightHip"],rightArmUpper:["rightShoulder","rightElbow"],rightArmLower:["rightElbow","rightWrist"],rightHand:["rightWrist","rightPalm"],rightHandPinky:["rightPalm","rightPinky"],rightHandIndex:["rightPalm","rightIndex"],rightHandThumb:["rightPalm","rightThumb"],rightEyeOutline:["rightEyeInside","rightEyeOutside"]};var vn,il=224,Dk,N1e=5,k0=[8,16,32,32,32];function R1e(){let e=[],t=0;for(;t<N1e;){let a=0,n=t;for(;n<k0.length&&k0[n]===k0[t];)a+=2,n++;let r=k0[t],s=Math.ceil(il/r),i=Math.ceil(il/r);for(let o=0;o<s;++o)for(let l=0;l<i;++l)for(let u=0;u<a;++u)e.push({x:(l+.5)/i,y:(o+.5)/s});t=n}Dk={x:jt(e.map(a=>a.x)),y:jt(e.map(a=>a.y))}}async function Pk(e){if(ne.initial&&(vn=null),!vn&&e.body.detector&&e.body.detector.modelPath){vn=await Fe(e.body.detector.modelPath);let t=vn!=null&&vn.executor?Object.values(vn.modelSignature.inputs):void 0;il=Array.isArray(t)?parseInt(t[0].tensorShape.dim[1].size):0}else e.debug&&vn&&K("cached model:",vn.modelUrl);return R1e(),vn}var $k=[5,5];function E1e(e,t){return Pe(()=>{let a=Ca(e,12,1),n=Oe(a[0]),r=Oe(a[1]),s=Oe(a[2]),i=Oe(a[3]);n=ke(ve(n,il),t.x),r=ke(ve(r,il),t.y),s=te(ve(s,il),$k[0]),i=te(ve(i,il),$k[1]);let o=ye(n,ve(s,2)),l=ye(r,ve(i,2)),u=ke(o,s),p=ke(l,i);return ca([o,l,u,p],1)})}async function M1e(e,t,a,n){var u,p;let r=[],s={};s.boxes=E1e(e,Dk),s.scores=Wa(t),s.nms=await fe.nonMaxSuppressionAsync(s.boxes,s.scores,1,((u=a.body.detector)==null?void 0:u.minConfidence)||.1,((p=a.body.detector)==null?void 0:p.iouThreshold)||.1);let i=await s.nms.data(),o=await s.scores.data(),l=await s.boxes.array();for(let c of Array.from(i)){let d=o[c],h=l[c],f=[Math.round(h[0]*n[0]),Math.round(h[1]*n[1]),Math.round(h[2]*n[0]),Math.round(h[3]*n[1])],m={score:d,boxRaw:h,box:f};r.push(m)}return Object.keys(s).forEach(c=>J(s[c])),r}async function Ok(e,t,a){let n={};n.res=vn==null?void 0:vn.execute(e,["Identity"]),n.logitsRaw=De(n.res,[0,0,0],[1,-1,1]),n.boxesRaw=De(n.res,[0,0,1],[1,-1,-1]),n.logits=Oe(n.logitsRaw),n.boxes=Oe(n.boxesRaw);let r=await M1e(n.boxes,n.logits,t,a);return Object.keys(n).forEach(s=>J(n[s])),r}function gs(e,t=[1,1]){let a=[e.map(o=>o[0]),e.map(o=>o[1])],n=[Math.min(...a[0]),Math.min(...a[1])],r=[Math.max(...a[0]),Math.max(...a[1])],s=[n[0],n[1],r[0]-n[0],r[1]-n[1]],i=[s[0]/t[0],s[1]/t[1],s[2]/t[0],s[3]/t[1]];return{box:s,boxRaw:i}}function zk(e,t=[1,1]){let a=[e.map(u=>u[0]),e.map(u=>u[1])],n=[Math.min(...a[0]),Math.min(...a[1])],r=[Math.max(...a[0]),Math.max(...a[1])],s=[(n[0]+r[0])/2,(n[1]+r[1])/2],i=Math.max(s[0]-n[0],s[1]-n[1],-s[0]+r[0],-s[1]+r[1]),o=[Math.trunc(s[0]-i),Math.trunc(s[1]-i),Math.trunc(2*i),Math.trunc(2*i)],l=[o[0]/t[0],o[1]/t[1],o[2]/t[0],o[3]/t[1]];return{box:o,boxRaw:l}}function I0(e,t){let a=[e[2]*t,e[3]*t];return[e[0]-(a[0]-e[2])/2,e[1]-(a[1]-e[3])/2,a[0],a[1]]}var Ga,Qg=256,Jg=Number.MAX_SAFE_INTEGER,F1e={landmarks:["ld_3d","activation_segmentation","activation_heatmap","world_3d","output_poseflag"],detector:[]},C0=[],ys=[[0,0],[0,0],[0,0],[0,0]],Lk=0,Wk=e=>1-1/(1+Math.exp(e)),Vk=e=>Pk(e);async function Uk(e){if(ne.initial&&(Ga=null),Ga)e.debug&&K("cached model:",Ga.modelUrl);else{Ga=await Fe(e.body.modelPath);let t=Ga!=null&&Ga.executor?Object.values(Ga.modelSignature.inputs):void 0;Qg=Array.isArray(t)?parseInt(t[0].tensorShape.dim[1].size):0}return Ga}function Bk(e,t,a){var s,i;let n={};if(!((s=e==null?void 0:e.shape)!=null&&s[1])||!((i=e==null?void 0:e.shape)!=null&&i[2]))return e;let r;if(a&&(n.cropped=fe.cropAndResize(e,[a],[0],[e.shape[1],e.shape[2]])),e.shape[1]!==e.shape[2]){let o=[e.shape[2]>e.shape[1]?Math.trunc((e.shape[2]-e.shape[1])/2):0,e.shape[2]>e.shape[1]?Math.trunc((e.shape[2]-e.shape[1])/2):0],l=[e.shape[1]>e.shape[2]?Math.trunc((e.shape[1]-e.shape[2])/2):0,e.shape[1]>e.shape[2]?Math.trunc((e.shape[1]-e.shape[2])/2):0];ys=[[0,0],o,l,[0,0]],n.pad=ir(n.cropped||e,ys),n.resize=fe.resizeBilinear(n.pad,[t,t]),r=ve(n.resize,ze.tf255)}else e.shape[1]!==t?(n.resize=fe.resizeBilinear(n.cropped||e,[t,t]),r=ve(n.resize,ze.tf255)):r=ve(n.cropped||e,ze.tf255);return Object.keys(n).forEach(o=>J(n[o])),r}function _1e(e,t,a){for(let n of e)n.position=[Math.trunc(n.position[0]*(t[0]+ys[2][0]+ys[2][1])/t[0]-ys[2][0]),Math.trunc(n.position[1]*(t[1]+ys[1][0]+ys[1][1])/t[1]-ys[1][0]),n.position[2]],n.positionRaw=[n.position[0]/t[0],n.position[1]/t[1],2*n.position[2]/(t[0]+t[1])];if(a){let n=a[2]-a[0],r=a[3]-a[1];for(let s of e)s.positionRaw=[s.positionRaw[0]/r+a[1],s.positionRaw[1]/n+a[0],s.positionRaw[2]],s.position=[Math.trunc(s.positionRaw[0]*t[0]),Math.trunc(s.positionRaw[1]*t[1]),s.positionRaw[2]]}return e}function $1e(e){let t=e.find(o=>o.part==="leftPalm"),a=e.find(o=>o.part==="leftWrist"),n=e.find(o=>o.part==="leftIndex");t.position[2]=((a.position[2]||0)+(n.position[2]||0))/2;let r=e.find(o=>o.part==="rightPalm"),s=e.find(o=>o.part==="rightWrist"),i=e.find(o=>o.part==="rightIndex");r.position[2]=((s.position[2]||0)+(i.position[2]||0))/2}async function D1e(e,t,a){if(!(Ga!=null&&Ga.executor))return null;let n={};[n.ld,n.segmentation,n.heatmap,n.world,n.poseflag]=Ga==null?void 0:Ga.execute(e,F1e.landmarks);let r=(await n.poseflag.data())[0],s=await n.ld.data(),i=await n.world.data();Object.keys(n).forEach(f=>J(n[f]));let o=[],l=5;for(let f=0;f<s.length/l;f++){let m=Wk(s[l*f+3]),g=Wk(s[l*f+4]),y=Math.trunc(100*m*g*r)/100,x=[s[l*f+0]/Qg,s[l*f+1]/Qg,s[l*f+2]+0],A=[Math.trunc(a[0]*x[0]),Math.trunc(a[1]*x[1]),x[2]],b=[i[l*f+0],i[l*f+1],i[l*f+2]+0];o.push({part:Yg[f],positionRaw:x,position:A,distance:b,score:y})}if(r<(t.body.minConfidence||0))return null;$1e(o);let u=_1e(o,a),p=u.map(f=>f.position),c=gs(p,[a[0],a[1]]),d={};for(let[f,m]of Object.entries(Zg)){let g=[];for(let y=0;y<m.length-1;y++){let x=u.find(b=>b.part===m[y]),A=u.find(b=>b.part===m[y+1]);x&&A&&g.push([x.position,A.position])}d[f]=g}return{id:0,score:Math.trunc(100*r)/100,box:c.box,boxRaw:c.boxRaw,keypoints:u,annotations:d}}async function ey(e,t){var s,i,o;let a=[e.shape[2]||0,e.shape[1]||0],n=(t.body.skipTime||0)>ae()-Lk,r=Jg<(t.body.skipFrames||0);if(t.skipAllowed&&n&&r&&C0!==null)Jg++;else{let l=[];if((i=(s=t.body)==null?void 0:s.detector)!=null&&i.enabled){let u=Bk(e,224);l=await Ok(u,t,a),J(u)}else l=[{box:[0,0,0,0],boxRaw:[0,0,1,1],score:0}];for(let u=0;u<l.length;u++){let p=Bk(e,256,(o=l[u])==null?void 0:o.boxRaw);C0.length=0;let c=await D1e(p,t,a);J(p),c&&(c.id=u,C0.push(c))}Lk=ae(),Jg=0}return C0}var Xu=[{class:1,label:"person"},{class:2,label:"bicycle"},{class:3,label:"car"},{class:4,label:"motorcycle"},{class:5,label:"airplane"},{class:6,label:"bus"},{class:7,label:"train"},{class:8,label:"truck"},{class:9,label:"boat"},{class:10,label:"traffic light"},{class:11,label:"fire hydrant"},{class:12,label:"stop sign"},{class:13,label:"parking meter"},{class:14,label:"bench"},{class:15,label:"bird"},{class:16,label:"cat"},{class:17,label:"dog"},{class:18,label:"horse"},{class:19,label:"sheep"},{class:20,label:"cow"},{class:21,label:"elephant"},{class:22,label:"bear"},{class:23,label:"zebra"},{class:24,label:"giraffe"},{class:25,label:"backpack"},{class:26,label:"umbrella"},{class:27,label:"handbag"},{class:28,label:"tie"},{class:29,label:"suitcase"},{class:30,label:"frisbee"},{class:31,label:"skis"},{class:32,label:"snowboard"},{class:33,label:"sports ball"},{class:34,label:"kite"},{class:35,label:"baseball bat"},{class:36,label:"baseball glove"},{class:37,label:"skateboard"},{class:38,label:"surfboard"},{class:39,label:"tennis racket"},{class:40,label:"bottle"},{class:41,label:"wine glass"},{class:42,label:"cup"},{class:43,label:"fork"},{class:44,label:"knife"},{class:45,label:"spoon"},{class:46,label:"bowl"},{class:47,label:"banana"},{class:48,label:"apple"},{class:49,label:"sandwich"},{class:50,label:"orange"},{class:51,label:"broccoli"},{class:52,label:"carrot"},{class:53,label:"hot dog"},{class:54,label:"pizza"},{class:55,label:"donut"},{class:56,label:"cake"},{class:57,label:"chair"},{class:58,label:"couch"},{class:59,label:"potted plant"},{class:60,label:"bed"},{class:61,label:"dining table"},{class:62,label:"toilet"},{class:63,label:"tv"},{class:64,label:"laptop"},{class:65,label:"mouse"},{class:66,label:"remote"},{class:67,label:"keyboard"},{class:68,label:"cell phone"},{class:69,label:"microwave"},{class:70,label:"oven"},{class:71,label:"toaster"},{class:72,label:"sink"},{class:73,label:"refrigerator"},{class:74,label:"book"},{class:75,label:"clock"},{class:76,label:"vase"},{class:77,label:"scissors"},{class:78,label:"teddy bear"},{class:79,label:"hair drier"},{class:80,label:"toothbrush"}];var Ha,ol=0,ty=[],Hk=0,ay=Number.MAX_SAFE_INTEGER;async function jk(e){if(ne.initial&&(Ha=null),Ha)e.debug&&K("cached model:",Ha.modelUrl);else{Ha=await Fe(e.object.modelPath);let t=Ha!=null&&Ha.executor?Object.values(Ha.modelSignature.inputs):void 0;ol=Array.isArray(t)?parseInt(t[0].tensorShape.dim[2].size):0}return Ha}async function P1e(e,t,a){if(!e)return[];let n={},r=[],s=await e.array();n.squeeze=Oe(e);let i=Ca(n.squeeze,6,1);n.stack=ca([i[1],i[0],i[3],i[2]],1),n.boxes=Oe(n.stack),n.scores=Oe(i[4]),n.classes=Oe(i[5]),J([e,...i]),n.nms=await fe.nonMaxSuppressionAsync(n.boxes,n.scores,a.object.maxDetected||0,a.object.iouThreshold,a.object.minConfidence||0);let o=await n.nms.data(),l=0;for(let u of Array.from(o)){let p=Math.trunc(100*s[0][u][4])/100,c=s[0][u][5];if(Number.isNaN(c))continue;let d=Xu[c].label,[h,f]=[s[0][u][0]/ol,s[0][u][1]/ol],m=[h,f,s[0][u][2]/ol-h,s[0][u][3]/ol-f],g=[Math.trunc(m[0]*t[0]),Math.trunc(m[1]*t[1]),Math.trunc(m[2]*t[0]),Math.trunc(m[3]*t[1])];r.push({id:l++,score:p,class:c,label:d,box:g,boxRaw:m})}return Object.keys(n).forEach(u=>J(n[u])),r}async function ny(e,t){if(!(Ha!=null&&Ha.executor))return[];let a=(t.object.skipTime||0)>ae()-Hk,n=ay<(t.object.skipFrames||0);return t.skipAllowed&&a&&n&&ty.length>0?(ay++,ty):(ay=0,new Promise(async r=>{let s=[e.shape[2]||0,e.shape[1]||0],i=fe.resizeBilinear(e,[ol,ol]),o=t.object.enabled?Ha==null?void 0:Ha.execute(i,["tower_0/detections"]):null;Hk=ae(),J(i);let l=await P1e(o,s,t);ty=l,r(l)}))}var T0={};yr(T0,{connected:()=>sy,kpt:()=>ry});var ry=["head","neck","rightShoulder","rightElbow","rightWrist","chest","leftShoulder","leftElbow","leftWrist","bodyCenter","rightHip","rightKnee","rightAnkle","leftHip","leftKnee","leftAnkle"],sy={leftLeg:["leftHip","leftKnee","leftAnkle"],rightLeg:["rightHip","rightKnee","rightAnkle"],torso:["leftShoulder","rightShoulder","rightHip","leftHip","leftShoulder"],leftArm:["leftShoulder","leftElbow","leftWrist"],rightArm:["rightShoulder","rightElbow","rightWrist"],head:[]};var Mt,Xk=0,Fa={id:0,keypoints:[],box:[0,0,0,0],boxRaw:[0,0,0,0],score:0,annotations:{}},iy=Number.MAX_SAFE_INTEGER;async function Kk(e){return ne.initial&&(Mt=null),Mt?e.debug&&K("cached model:",Mt.modelUrl):Mt=await Fe(e.body.modelPath),Mt}async function O1e(e,t){let[a,n]=e.shape,r=Q(e,[n*a]),s=ga(r,0),i=(await s.data())[0];if(i>t){let o=rr(r,0),l=_u(o,a),u=(await l.data())[0],p=ve(o,a),c=(await p.data())[0];return J([r,s,o,l,p]),[u,c,i]}return J([r,s]),[0,0,i]}async function oy(e,t){if(!(Mt!=null&&Mt.executor)||!(Mt!=null&&Mt.inputs[0].shape))return[];let a=(t.body.skipTime||0)>ae()-Xk,n=iy<(t.body.skipFrames||0);return t.skipAllowed&&a&&n&&Object.keys(Fa.keypoints).length>0?(iy++,[Fa]):(iy=0,new Promise(async r=>{let s=Pe(()=>{var f,m;let c=fe.resizeBilinear(e,[((f=Mt==null?void 0:Mt.inputs[0].shape)==null?void 0:f[2])||0,((m=Mt==null?void 0:Mt.inputs[0].shape)==null?void 0:m[1])||0],!1),d=te(c,ze.tf2);return ye(d,ze.tf1)}),i;if(t.body.enabled&&(i=Mt==null?void 0:Mt.execute(s)),Xk=ae(),J(s),i){Fa.keypoints.length=0;let c=Oe(i);J(i);let d=Ra(c,2);J(c);for(let h=0;h<d.length;h++){let[f,m,g]=await O1e(d[h],t.body.minConfidence);g>(t.body.minConfidence||0)&&Fa.keypoints.push({score:Math.round(100*g)/100,part:ry[h],positionRaw:[f/Mt.inputs[0].shape[2],m/Mt.inputs[0].shape[1]],position:[Math.round(e.shape[2]*f/Mt.inputs[0].shape[2]),Math.round(e.shape[1]*m/Mt.inputs[0].shape[1])]})}d.forEach(h=>J(h))}Fa.score=Fa.keypoints.reduce((c,d)=>d.score>c?d.score:c,0);let o=Fa.keypoints.map(c=>c.position[0]),l=Fa.keypoints.map(c=>c.position[1]);Fa.box=[Math.min(...o),Math.min(...l),Math.max(...o)-Math.min(...o),Math.max(...l)-Math.min(...l)];let u=Fa.keypoints.map(c=>c.positionRaw[0]),p=Fa.keypoints.map(c=>c.positionRaw[1]);Fa.boxRaw=[Math.min(...u),Math.min(...p),Math.max(...u)-Math.min(...u),Math.max(...p)-Math.min(...p)];for(let[c,d]of Object.entries(sy)){let h=[];for(let f=0;f<d.length-1;f++){let m=Fa.keypoints.find(y=>y.part===d[f]),g=Fa.keypoints.find(y=>y.part===d[f+1]);m&&g&&m.score>(t.body.minConfidence||0)&&g.score>(t.body.minConfidence||0)&&h.push([m.position,g.position])}Fa.annotations[c]=h}r([Fa])}))}var Ku=e=>[Math.abs(e.endPoint[0]-e.startPoint[0]),Math.abs(e.endPoint[1]-e.startPoint[1])],N0=e=>[e.startPoint[0]+(e.endPoint[0]-e.startPoint[0])/2,e.startPoint[1]+(e.endPoint[1]-e.startPoint[1])/2,1],R0=(e,t)=>e?[Math.trunc(Math.max(0,e.startPoint[0])),Math.trunc(Math.max(0,e.startPoint[1])),Math.trunc(Math.min(t.shape[2]||0,e.endPoint[0])-Math.max(0,e.startPoint[0])),Math.trunc(Math.min(t.shape[1]||0,e.endPoint[1])-Math.max(0,e.startPoint[1]))]:[0,0,0,0],E0=(e,t)=>e?[e.startPoint[0]/(t.shape[2]||0),e.startPoint[1]/(t.shape[1]||0),(e.endPoint[0]-e.startPoint[0])/(t.shape[2]||0),(e.endPoint[1]-e.startPoint[1])/(t.shape[1]||0)]:[0,0,0,0],Qk=(e,t,a)=>{let n=[e.startPoint[0]*t[0],e.startPoint[1]*t[1]],r=[e.endPoint[0]*t[0],e.endPoint[1]*t[1]],s=e.landmarks.map(i=>[(i[0]+a[0])*t[0],(i[1]+a[1])*t[1]]);return{startPoint:n,endPoint:r,landmarks:s,confidence:e.confidence}},ly=(e,t,a)=>{let n=t.shape[1],r=t.shape[2],s=[e.startPoint[1]/n,e.startPoint[0]/r,e.endPoint[1]/n,e.endPoint[0]/r],i=fe.cropAndResize(t,[s],[0],a),o=ve(i,ze.tf255);return J(i),o},M0=(e,t)=>{let a=N0(e),n=Ku(e),r=[t*n[0]/2,t*n[1]/2];return{startPoint:[a[0]-r[0],a[1]-r[1]],endPoint:[a[0]+r[0],a[1]+r[1]],landmarks:e.landmarks,confidence:e.confidence,size:n}},F0=e=>{let t=N0(e),a=Ku(e),n=Math.max(...a)/2;return{startPoint:[Math.round(t[0]-n),Math.round(t[1]-n)],endPoint:[Math.round(t[0]+n),Math.round(t[1]+n)],landmarks:e.landmarks,confidence:e.confidence,size:[Math.round(a[0]),Math.round(a[1])]}},e9=e=>{let t=e.map(n=>n[0]),a=e.map(n=>n[1]);return{startPoint:[Math.min(...t),Math.min(...a)],endPoint:[Math.max(...t),Math.max(...a)],landmarks:e}},uy=[[1,0,0],[0,1,0],[0,0,1]],z1e=e=>e-2*Math.PI*Math.floor((e+Math.PI)/(2*Math.PI)),L1e=(e,t)=>z1e(Math.PI/2-Math.atan2(-(t[1]-e[1]),t[0]-e[0]));var Zk=(e,t)=>[[1,0,e],[0,1,t],[0,0,1]],ll=(e,t)=>{let a=0;for(let n=0;n<e.length;n++)a+=e[n]*t[n];return a},W1e=(e,t)=>{let a=[];for(let n=0;n<e.length;n++)a.push(e[n][t]);return a},Jk=(e,t)=>{let a=[],n=e.length;for(let r=0;r<n;r++){a.push([]);for(let s=0;s<n;s++)a[r].push(ll(e[r],W1e(t,s)))}return a},t9=(e,t)=>{let a=Math.cos(e),n=Math.sin(e),r=[[a,-n,0],[n,a,0],[0,0,1]],s=Zk(t[0],t[1]),i=Jk(s,r),o=Zk(-t[0],-t[1]);return Jk(i,o)},B1e=e=>{let t=[[e[0][0],e[1][0]],[e[0][1],e[1][1]]],a=[e[0][2],e[1][2]],n=[-ll(t[0],a),-ll(t[1],a)];return[t[0].concat(n[0]),t[1].concat(n[1]),[0,0,1]]},V1e=(e,t)=>[ll(e,t[0]),ll(e,t[1])];function a9(e){let t=e===192?{strides:[4],anchors:[1]}:{strides:[e/16,e/8],anchors:[2,6]},a=[];for(let n=0;n<t.strides.length;n++){let r=t.strides[n],s=Math.floor((e+r-1)/r),i=Math.floor((e+r-1)/r),o=t.anchors[n];for(let l=0;l<s;l++){let u=r*(l+.5);for(let p=0;p<i;p++){let c=r*(p+.5);for(let d=0;d<o;d++)a.push([c,u])}}}return a}function n9(e,t,a,n,r){let s=Ku(t),i=e.map(h=>[s[0]/r*(h[0]-r/2),s[1]/r*(h[1]-r/2),h[2]||0]),o=a&&a!==0&&Math.abs(a)>.2,l=o?t9(a,[0,0]):uy,u=o?i.map(h=>[...V1e(h,l),h[2]]):i,p=o?B1e(n):uy,c=N0(t),d=[ll(c,p[0]),ll(c,p[1])];return u.map(h=>[Math.trunc(h[0]+d[0]),Math.trunc(h[1]+d[1]),Math.trunc(h[2]||0)])}function r9(e,t,a,n){let r=t.landmarks.length>=Hg.count?Hg.symmetryLine:al.symmetryLine,s=0,i=uy,o;if(e&&ne.kernels.includes("rotatewithoffset"))if(s=L1e(t.landmarks[r[0]],t.landmarks[r[1]]),s&&s!==0&&Math.abs(s)>.2){let u=N0(t),p=[u[0]/a.shape[2],u[1]/a.shape[1]],c=fe.rotateWithOffset(a,s,0,[p[0],p[1]]);i=t9(-s,u),o=ly(t,c,[n,n]),J(c)}else o=ly(t,a,[n,n]);else o=ly(t,a,[n,n]);return[s,i,o]}var U1e=e=>{let t=e.map(n=>n[0]),a=e.map(n=>n[1]);return[Math.min(...t)+(Math.max(...t)-Math.min(...t))/2,Math.min(...a)+(Math.max(...a)-Math.min(...a))/2]},s9=(e,t)=>{let a=U1e(e),n=Ku(t);return{startPoint:[a[0]-n[0]/2,a[1]-n[1]/2],endPoint:[a[0]+n[0]/2,a[1]+n[1]/2]}};var i9=6,G1e=1.4,Vn,_0=null,xs=0,Yu=null,o9=()=>xs;async function l9(e){var t;return ne.initial&&(Vn=null),Vn?e.debug&&K("cached model:",Vn.modelUrl):Vn=await Fe((t=e.face.detector)==null?void 0:t.modelPath),xs=Vn.executor&&Vn.inputs[0].shape?Vn.inputs[0].shape[2]:256,Yu=Ue(xs,"int32"),_0=Zn(a9(xs)),Vn}function H1e(e){if(!_0||!Yu)return gn([0,0]);let t={};t.boxStarts=De(e,[0,1],[-1,2]),t.centers=ke(t.boxStarts,_0),t.boxSizes=De(e,[0,3],[-1,2]),t.boxSizesNormalized=ve(t.boxSizes,Yu),t.centersNormalized=ve(t.centers,Yu),t.halfBoxSize=ve(t.boxSizesNormalized,ze.tf2),t.starts=ye(t.centersNormalized,t.halfBoxSize),t.ends=ke(t.centersNormalized,t.halfBoxSize),t.startNormalized=te(t.starts,Yu),t.endNormalized=te(t.ends,Yu);let a=Fu([t.startNormalized,t.endNormalized],1);return Object.keys(t).forEach(n=>J(t[n])),a}async function u9(e,t){var o,l,u,p,c,d;if(!e||e.isDisposedInternal||e.shape.length!==4||e.shape[1]<1||e.shape[2]<1)return[];let a={};a.resized=fe.resizeBilinear(e,[xs,xs]),a.div=ve(a.resized,ze.tf127),a.normalized=ye(a.div,ze.tf05);let n=Vn==null?void 0:Vn.execute(a.normalized);if(Array.isArray(n)&&n.length>2){let h=n.sort((f,m)=>f.size-m.size);a.concat384=lt([h[0],h[2]],2),a.concat512=lt([h[1],h[3]],2),a.concat=lt([a.concat512,a.concat384],1),a.batch=Oe(a.concat,[0])}else Array.isArray(n)?a.batch=Oe(n[0]):a.batch=Oe(n);J(n),a.boxes=H1e(a.batch),a.logits=De(a.batch,[0,0],[-1,1]),a.sigmoid=Wa(a.logits),a.scores=Oe(a.sigmoid),a.nms=await fe.nonMaxSuppressionAsync(a.boxes,a.scores,((o=t.face.detector)==null?void 0:o.maxDetected)||0,((l=t.face.detector)==null?void 0:l.iouThreshold)||0,((u=t.face.detector)==null?void 0:u.minConfidence)||0);let r=await a.nms.array(),s=[],i=await a.scores.data();for(let h=0;h<r.length;h++){let f=i[r[h]];if(f>(((p=t.face.detector)==null?void 0:p.minConfidence)||0)){let m={};m.bbox=De(a.boxes,[r[h],0],[1,-1]),m.slice=De(a.batch,[r[h],i9-1],[1,-1]),m.squeeze=Oe(m.slice),m.landmarks=Q(m.squeeze,[i9,-1]);let g=await m.bbox.data(),y={startPoint:[g[0],g[1]],endPoint:[g[2],g[3]],landmarks:await m.landmarks.array(),confidence:f};m.anchor=De(_0,[r[h],0],[1,2]);let x=await m.anchor.data(),A=Qk(y,[(e.shape[2]||0)/xs,(e.shape[1]||0)/xs],x),b=M0(A,t.face.scale||G1e),w=F0(b);w.size[0]>(((c=t.face.detector)==null?void 0:c.minSize)||0)&&w.size[1]>(((d=t.face.detector)==null?void 0:d.minSize)||0)&&s.push(w),Object.keys(m).forEach(C=>J(m[C]))}}return Object.keys(a).forEach(h=>J(a[h])),s}var nn,As=0,j1e=2.3,py=_n.leftEyeLower0,cy=_n.rightEyeLower0,Zu={leftBounds:[py[0],py[py.length-1]],rightBounds:[cy[0],cy[cy.length-1]]},Ju={upperCenter:3,lowerCenter:4,index:71,numCoordinates:76};async function f9(e){var t,a;return ne.initial&&(nn=null),nn?e.debug&&K("cached model:",nn.modelUrl):nn=await Fe((t=e.face.iris)==null?void 0:t.modelPath),As=nn!=null&&nn.executor&&((a=nn.inputs)!=null&&a[0].shape)?nn.inputs[0].shape[2]:0,As===-1&&(As=64),nn}function $0(e,t,a,n){for(let r=0;r<jg.length;r++){let{key:s,indices:i}=jg[r],o=_n[`${a}${s}`];if(!n||n.includes(s))for(let l=0;l<i.length;l++){let u=i[l];e[o[l]]=[t[u][0],t[u][1],(t[u][2]+e[o[l]][2])/2]}}}var q1e=e=>{let t=e[Zu.leftBounds[0]][2],a=e[Zu.rightBounds[0]][2];return t-a},p9=(e,t,a,n,r,s=!1)=>{let i=F0(M0(e9([e[a],e[n]]),j1e)),o=Ku(i),l=fe.cropAndResize(t,[[i.startPoint[1]/r,i.startPoint[0]/r,i.endPoint[1]/r,i.endPoint[0]/r]],[0],[As,As]);if(s&&ne.kernels.includes("flipleftright")){let u=fe.flipLeftRight(l);J(l),l=u}return{box:i,boxSize:o,crop:l}},c9=(e,t,a,n=!1)=>{let r=[];for(let s=0;s<Ju.numCoordinates;s++){let i=e[s*3],o=e[s*3+1],l=e[s*3+2];r.push([(n?1-i/As:i/As)*a[0]+t.startPoint[0],o/As*a[1]+t.startPoint[1],l])}return{rawCoords:r,iris:r.slice(Ju.index)}},h9=(e,t,a)=>{let n=e[_n[`${a}EyeUpper0`][Ju.upperCenter]][2],r=e[_n[`${a}EyeLower0`][Ju.lowerCenter]][2],s=(n+r)/2;return t.map((i,o)=>{let l=s;return o===2?l=n:o===4&&(l=r),[i[0],i[1],l]})};async function m9(e,t,a){if(!(nn!=null&&nn.executor))return e;let{box:n,boxSize:r,crop:s}=p9(e,t,Zu.leftBounds[0],Zu.leftBounds[1],a,!0),{box:i,boxSize:o,crop:l}=p9(e,t,Zu.rightBounds[0],Zu.rightBounds[1],a,!0),u=lt([s,l]);J(s),J(l);let p=nn.execute(u);J(u);let c=await p.data();J(p);let d=c.slice(0,Ju.numCoordinates*3),{rawCoords:h,iris:f}=c9(d,n,r,!0),m=c.slice(Ju.numCoordinates*3),{rawCoords:g,iris:y}=c9(m,i,o,!1),x=q1e(e);Math.abs(x)<30?($0(e,h,"left",null),$0(e,g,"right",null)):x<1?$0(e,h,"left",["EyeUpper0","EyeLower0"]):$0(e,g,"right",["EyeUpper0","EyeLower0"]);let A=h9(e,f,"left"),b=h9(e,y,"right");return e.concat(A).concat(b)}async function y9(e,t){var s,i,o,l,u,p,c,d,h,f;let a={lips:await((i=(s=t.filter(m=>m.size===160))==null?void 0:s[0])==null?void 0:i.data()),irisL:await((l=(o=t.filter(m=>m.size===10))==null?void 0:o[0])==null?void 0:l.data()),eyeL:await((p=(u=t.filter(m=>m.size===142))==null?void 0:u[0])==null?void 0:p.data()),irisR:await((d=(c=t.filter(m=>m.size===10))==null?void 0:c[1])==null?void 0:d.data()),eyeR:await((f=(h=t.filter(m=>m.size===142))==null?void 0:h[1])==null?void 0:f.data())};for(let m of Object.values(a))if(!m)return e;let n=rl.reduce((m,g)=>m+=e[g][2],0)/rl.length;for(let m=0;m<a.irisL.length/2;m++)e.push([a.irisL[2*m+0],a.irisL[2*m+1],n]);let r=sl.reduce((m,g)=>m+=e[g][2],0)/sl.length;for(let m=0;m<a.irisR.length/2;m++)e.push([a.irisR[2*m+0],a.irisR[2*m+1],r]);for(let m=0;m<a.eyeL.length/2;m++)e[rl[m]]=[a.eyeL[2*m+0],a.eyeL[2*m+1],e[rl[m]][2]];for(let m=0;m<a.eyeR.length/2;m++)e[sl[m]]=[a.eyeR[2*m+0],a.eyeR[2*m+1],e[sl[m]][2]];for(let m=0;m<a.lips.length/2;m++)e[ec[m]]=[a.lips[2*m+0],a.lips[2*m+1],e[ec[m]][2]];return e}var dr={boxes:[],skipped:Number.MAX_SAFE_INTEGER,timestamp:0},Ct=null,tc=0;async function x9(e,t){var l,u,p,c,d,h,f,m,g,y;let a=(((l=t.face.detector)==null?void 0:l.skipTime)||0)>ae()-dr.timestamp,n=dr.skipped<(((u=t.face.detector)==null?void 0:u.skipFrames)||0);!t.skipAllowed||!a||!n||dr.boxes.length===0?(dr.boxes=await u9(e,t),dr.timestamp=ae(),dr.skipped=0):dr.skipped++;let r=[],s=[],i=0,o=tc;for(let x=0;x<dr.boxes.length;x++){let A=dr.boxes[x],b=0,w,C={id:i++,mesh:[],meshRaw:[],box:[0,0,0,0],boxRaw:[0,0,0,0],score:0,boxScore:0,faceScore:0,size:[0,0],annotations:{}};if([b,w,C.tensor]=r9((p=t.face.detector)==null?void 0:p.rotation,A,e,(c=t.face.mesh)!=null&&c.enabled?tc:o9()),t.filter.equalization){let T=C.tensor?await u0(C.tensor):void 0;J(C.tensor),T&&(C.tensor=T)}if(C.boxScore=Math.round(100*A.confidence)/100,!((d=t.face.mesh)!=null&&d.enabled)||!(Ct!=null&&Ct.executor)){C.box=R0(A,e),C.boxRaw=E0(A,e),C.score=C.boxScore,C.size=A.size,C.mesh=A.landmarks,C.meshRaw=C.mesh.map(T=>[T[0]/(e.shape[2]||0),T[1]/(e.shape[1]||0),(T[2]||0)/o]);for(let T of Object.keys(al))C.annotations[T]=[C.mesh[al[T]]]}else if(!Ct)t.debug&&K("face mesh detection requested, but model is not loaded");else{if((h=t.face.attention)!=null&&h.enabled&&!ne.kernels.includes("atan2"))return t.face.attention.enabled=!1,J(C.tensor),r;let T=Ct.execute(C.tensor),M=await T.find(F=>F.shape[F.shape.length-1]===1).data();if(C.faceScore=Math.round(100*M[0])/100,C.faceScore<(((f=t.face.detector)==null?void 0:f.minConfidence)||1)){if(A.confidence=C.faceScore,t.face.mesh.keepInvalid){C.box=R0(A,e),C.boxRaw=E0(A,e),C.score=C.boxScore,C.mesh=A.landmarks,C.meshRaw=C.mesh.map(F=>[F[0]/(e.shape[2]||1),F[1]/(e.shape[1]||1),(F[2]||0)/o]);for(let F of Object.keys(al))C.annotations[F]=[C.mesh[al[F]]]}}else{let F=T.find(O=>O.shape[O.shape.length-1]===1404),E=Q(F,[-1,3]),I=await E.array();J(E),(m=t.face.attention)!=null&&m.enabled?I=await y9(I,T):(g=t.face.iris)!=null&&g.enabled&&(I=await m9(I,C.tensor,tc)),C.mesh=n9(I,A,b,w,tc),C.meshRaw=C.mesh.map(O=>[O[0]/(e.shape[2]||0),O[1]/(e.shape[1]||0),(O[2]||0)/o]);for(let O of Object.keys(_n))C.annotations[O]=_n[O].map(B=>C.mesh[B]);C.score=C.faceScore;let _={...s9(C.mesh,A),confidence:A.confidence,landmarks:A.landmarks,size:A.size};C.box=R0(_,e),C.boxRaw=E0(_,e),s.push(_)}J(T)}C.score>(((y=t.face.detector)==null?void 0:y.minConfidence)||1)?r.push(C):J(C.tensor)}return dr.boxes=s,r}async function A9(e){var t,a,n,r,s,i;return ne.initial&&(Ct=null),(t=e.face.attention)!=null&&t.enabled&&(Ct!=null&&Ct.signature)&&Object.keys(((a=Ct==null?void 0:Ct.signature)==null?void 0:a.outputs)||{}).length<6&&(Ct=null),Ct?e.debug&&K("cached model:",Ct.modelUrl):(n=e.face.attention)!=null&&n.enabled?Ct=await Fe(e.face.attention.modelPath):Ct=await Fe((r=e.face.mesh)==null?void 0:r.modelPath),tc=Ct.executor&&((s=Ct==null?void 0:Ct.inputs)!=null&&s[0].shape)?(i=Ct==null?void 0:Ct.inputs)==null?void 0:i[0].shape[2]:256,Ct}var b9=nl,v9=Qp;var my=[],ra,D0=[],w9=0,k9=0,fy=Number.MAX_SAFE_INTEGER,gy=!1;async function I9(e){var t,a,n;return ne.initial&&(ra=null),ra?e.debug&&K("cached model:",ra.modelUrl):(ra=await Fe((t=e.face.emotion)==null?void 0:t.modelPath),gy=((n=(a=ra==null?void 0:ra.inputs)==null?void 0:a[0].shape)==null?void 0:n[3])===3,gy?my=["angry","disgust","fear","happy","neutral","sad","surprise"]:my=["angry","disgust","fear","happy","sad","surprise","neutral"]),ra}async function yy(e,t,a,n){var i,o;if(!ra)return[];let r=fy<(((i=t.face.emotion)==null?void 0:i.skipFrames)||0),s=(((o=t.face.emotion)==null?void 0:o.skipTime)||0)>ae()-k9;return t.skipAllowed&&s&&r&&w9===n&&D0[a]&&D0[a].length>0?(fy++,D0[a]):(fy=0,new Promise(async l=>{var p,c,d;let u=[];if((p=t.face.emotion)!=null&&p.enabled){let h={},f=ra!=null&&ra.inputs[0].shape?ra.inputs[0].shape[2]:0;if(((c=t.face.emotion)==null?void 0:c.crop)>0){let g=(d=t.face.emotion)==null?void 0:d.crop,y=[[g,g,1-g,1-g]];h.resize=fe.cropAndResize(e,y,[0],[f,f])}else h.resize=fe.resizeBilinear(e,[f,f],!1);gy?(h.mul=te(h.resize,255),h.normalize=ye(h.mul,[103.939,116.779,123.68]),h.emotion=ra==null?void 0:ra.execute(h.normalize)):(h.channels=te(h.resize,ze.rgb),h.grayscale=ot(h.channels,3,!0),h.grayscaleSub=ye(h.grayscale,ze.tf05),h.grayscaleMul=te(h.grayscaleSub,ze.tf2),h.emotion=ra==null?void 0:ra.execute(h.grayscaleMul)),k9=ae();let m=await h.emotion.data();for(let g=0;g<m.length;g++)m[g]>(t.face.emotion.minConfidence||0)&&u.push({score:Math.min(.99,Math.trunc(100*m[g])/100),emotion:my[g]});u.sort((g,y)=>y.score-g.score),Object.keys(h).forEach(g=>J(h[g]))}D0[a]=u,w9=n,l(u)}))}var sa,bs=[],C9=0,T9=0,xy=Number.MAX_SAFE_INTEGER;async function N9(e){var t;return ne.initial&&(sa=null),sa?e.debug&&K("cached model:",sa.modelUrl):sa=await Fe((t=e.face.description)==null?void 0:t.modelPath),sa}function K1e(e,t){var s,i;let a=e.image||e.tensor||e;if(!(sa!=null&&sa.inputs[0].shape))return a;let n;if(((s=t.face.description)==null?void 0:s.crop)>0){let o=(i=t.face.description)==null?void 0:i.crop,l=[[o,o,1-o,1-o]];n=fe.cropAndResize(a,l,[0],[sa.inputs[0].shape[2],sa.inputs[0].shape[1]])}else n=fe.resizeBilinear(a,[sa.inputs[0].shape[2],sa.inputs[0].shape[1]],!1);let r=te(n,ze.tf255);return J(n),r}async function Ay(e,t,a,n){var o,l,u,p;let r={age:0,gender:"unknown",genderScore:0,descriptor:[]};if(!(sa!=null&&sa.executor))return r;let s=xy<(((o=t.face.description)==null?void 0:o.skipFrames)||0),i=(((l=t.face.description)==null?void 0:l.skipTime)||0)>ae()-C9;return t.skipAllowed&&s&&i&&T9===n&&((u=bs==null?void 0:bs[a])==null?void 0:u.age)>0&&((p=bs==null?void 0:bs[a])==null?void 0:p.genderScore)>0?(xy++,bs[a]):(xy=0,new Promise(async c=>{var d;if((d=t.face.description)!=null&&d.enabled){let h=K1e(e,t),f=sa==null?void 0:sa.execute(h);C9=ae(),J(h);let g=await f.find(N=>N.shape[1]===1).data(),y=Math.trunc(200*Math.abs(g[0]-.5))/100;y>(t.face.description.minConfidence||0)&&(r.gender=g[0]<=.5?"female":"male",r.genderScore=Math.min(.99,y));let x=rr(f.find(N=>N.shape[1]===100),1),A=(await x.data())[0];J(x);let w=await f.find(N=>N.shape[1]===100).data();r.age=Math.round(w[A-1]>w[A+1]?10*A-100*w[A-1]:10*A+100*w[A+1])/10,(Number.isNaN(g[0])||Number.isNaN(w[0]))&&K("faceres error:",{model:sa,result:f});let C=f.find(N=>N.shape[1]===1024),T=C?await C.data():[];r.descriptor=Array.from(T),f.forEach(N=>J(N))}bs[a]=r,T9=n,c(r)}))}var Qu=.1,by=.5;function Y1e(e,t,a){let n=!1,r=a.length-1;for(let s=0;s<a.length;r=s++)a[s].y>t!=a[r].y>t&&e<(a[r].x-a[s].x)*(t-a[s].y)/(a[r].y-a[s].y)+a[s].x&&(n=!n);return n}async function E9(e){if(!e.tensor||!e.mesh||e.mesh.length<100)return e.tensor;let t=e.tensor.shape[2]||0,a=e.tensor.shape[1]||0,n=await e.tensor.buffer(),r=[];for(let i of _n.silhouette)r.push({x:(e.mesh[i][0]-e.box[0])/e.box[2],y:(e.mesh[i][1]-e.box[1])/e.box[3]});Qu&&Qu>0&&(r=r.map(i=>({x:i.x>.5?i.x+Qu:i.x-Qu,y:i.y>.5?i.y+Qu:i.y-Qu})));for(let i=0;i<t;i++)for(let o=0;o<a;o++)Y1e(i/t,o/t,r)||(n.set(by*n.get(0,o,i,0),0,o,i,0),n.set(by*n.get(0,o,i,1),0,o,i,1),n.set(by*n.get(0,o,i,2),0,o,i,2));return n.toTensor()}var ia,P0=[],vy=Number.MAX_SAFE_INTEGER,M9=0,F9=0;async function _9(e){var t;return ne.initial&&(ia=null),ia?e.debug&&K("cached model:",ia.modelUrl):ia=await Fe((t=e.face.antispoof)==null?void 0:t.modelPath),ia}async function wy(e,t,a,n){var i,o;if(!(ia!=null&&ia.executor))return 0;let r=(((i=t.face.antispoof)==null?void 0:i.skipTime)||0)>ae()-F9,s=vy<(((o=t.face.antispoof)==null?void 0:o.skipFrames)||0);return t.skipAllowed&&r&&s&&M9===n&&P0[a]?(vy++,P0[a]):(vy=0,new Promise(async l=>{let u=fe.resizeBilinear(e,[ia!=null&&ia.inputs[0].shape?ia.inputs[0].shape[2]:0,ia!=null&&ia.inputs[0].shape?ia.inputs[0].shape[1]:0],!1),p=ia==null?void 0:ia.execute(u),c=(await p.data())[0];P0[a]=Math.round(100*c)/100,M9=n,F9=ae(),J([u,p]),l(P0[a])}))}var oa,O0=[],ky=Number.MAX_SAFE_INTEGER,D9=0,P9=0;async function O9(e){var t;return ne.initial&&(oa=null),oa?e.debug&&K("cached model:",oa.modelUrl):oa=await Fe((t=e.face.liveness)==null?void 0:t.modelPath),oa}async function Iy(e,t,a,n){var i,o;if(!(oa!=null&&oa.executor))return 0;let r=(((i=t.face.liveness)==null?void 0:i.skipTime)||0)>ae()-P9,s=ky<(((o=t.face.liveness)==null?void 0:o.skipFrames)||0);return t.skipAllowed&&r&&s&&D9===n&&O0[a]?(ky++,O0[a]):(ky=0,new Promise(async l=>{let u=fe.resizeBilinear(e,[oa!=null&&oa.inputs[0].shape?oa.inputs[0].shape[2]:0,oa!=null&&oa.inputs[0].shape?oa.inputs[0].shape[1]:0],!1),p=oa==null?void 0:oa.execute(u),c=(await p.data())[0];O0[a]=Math.round(100*c)/100,D9=n,P9=ae(),J([u,p]),l(O0[a])}))}var $n,Sy=[],J1e=["white","black","asian","indian","other"],Q1e=[15,23,28,35.5,45.5,55.5,65],L9=0,W9=0,Cy=Number.MAX_SAFE_INTEGER;async function B9(e){var t;return ne.initial&&($n=null),$n?e.debug&&K("cached model:",$n.modelUrl):$n=await Fe((t=e.face.gear)==null?void 0:t.modelPath),$n}async function Ty(e,t,a,n){var i,o;if(!$n)return{age:0,gender:"unknown",genderScore:0,race:[]};let r=Cy<(((i=t.face.gear)==null?void 0:i.skipFrames)||0),s=(((o=t.face.gear)==null?void 0:o.skipTime)||0)>ae()-W9;return t.skipAllowed&&s&&r&&L9===n&&Sy[a]?(Cy++,Sy[a]):(Cy=0,new Promise(async l=>{var y,x,A,b;if(!($n!=null&&$n.inputs[0].shape))return;let u={},p=[[0,.1,.9,.9]];if(((y=t.face.gear)==null?void 0:y.crop)>0){let w=(x=t.face.gear)==null?void 0:x.crop;p=[[w,w,1-w,1-w]]}u.resize=fe.cropAndResize(e,p,[0],[$n.inputs[0].shape[2],$n.inputs[0].shape[1]]);let c={age:0,gender:"unknown",genderScore:0,race:[]};(A=t.face.gear)!=null&&A.enabled&&([u.age,u.gender,u.race]=$n.execute(u.resize,["age_output","gender_output","race_output"]));let d=await u.gender.data();c.gender=d[0]>d[1]?"male":"female",c.genderScore=Math.round(100*(d[0]>d[1]?d[0]:d[1]))/100;let h=await u.race.data();for(let w=0;w<h.length;w++)h[w]>(((b=t.face.gear)==null?void 0:b.minConfidence)||.2)&&c.race.push({score:Math.round(100*h[w])/100,race:J1e[w]});c.race.sort((w,C)=>C.score-w.score);let m=Array.from(await u.age.data()).map((w,C)=>[Q1e[C],w]).sort((w,C)=>C[1]-w[1]),g=m[0][0];for(let w=1;w<m.length;w++)g+=m[w][1]*(m[w][0]-g);c.age=Math.round(10*g)/10,Object.keys(u).forEach(w=>J(u[w])),Sy[a]=c,L9=n,W9=ae(),l(c)}))}var _a,z0=[],U9=0,G9=0,Ny=Number.MAX_SAFE_INTEGER;async function H9(e){return ne.initial&&(_a=null),_a?e.debug&&K("cached model:",_a.modelUrl):_a=await Fe(e.face.ssrnet.modelPathAge),_a}async function Ry(e,t,a,n){var i,o,l,u;if(!_a)return{age:0};let r=Ny<(((i=t.face.ssrnet)==null?void 0:i.skipFrames)||0),s=(((o=t.face.ssrnet)==null?void 0:o.skipTime)||0)>ae()-G9;return t.skipAllowed&&r&&s&&U9===n&&((l=z0[a])!=null&&l.age)&&((u=z0[a])==null?void 0:u.age)>0?(Ny++,z0[a]):(Ny=0,new Promise(async p=>{var h,f,m;if(!(_a!=null&&_a.inputs)||!_a.inputs[0]||!_a.inputs[0].shape)return;let c={};if(((h=t.face.ssrnet)==null?void 0:h.crop)>0){let g=(f=t.face.ssrnet)==null?void 0:f.crop,y=[[g,g,1-g,1-g]];c.resize=fe.cropAndResize(e,y,[0],[_a.inputs[0].shape[2],_a.inputs[0].shape[1]])}else c.resize=fe.resizeBilinear(e,[_a.inputs[0].shape[2],_a.inputs[0].shape[1]],!1);c.enhance=te(c.resize,ze.tf255);let d={age:0};if((m=t.face.ssrnet)!=null&&m.enabled&&(c.age=_a.execute(c.enhance)),c.age){let g=await c.age.data();d.age=Math.trunc(10*g[0])/10}Object.keys(c).forEach(g=>J(c[g])),z0[a]=d,U9=n,G9=ae(),p(d)}))}var Aa,L0=[],q9=0,X9=0,Ey=Number.MAX_SAFE_INTEGER,My=[.2989,.587,.114];async function K9(e){var t;return ne.initial&&(Aa=null),Aa?e.debug&&K("cached model:",Aa.modelUrl):Aa=await Fe((t=e.face.ssrnet)==null?void 0:t.modelPathGender),Aa}async function Fy(e,t,a,n){var i,o,l,u;if(!Aa)return{gender:"unknown",genderScore:0};let r=Ey<(((i=t.face.ssrnet)==null?void 0:i.skipFrames)||0),s=(((o=t.face.ssrnet)==null?void 0:o.skipTime)||0)>ae()-X9;return t.skipAllowed&&r&&s&&q9===n&&((l=L0[a])!=null&&l.gender)&&((u=L0[a])==null?void 0:u.genderScore)>0?(Ey++,L0[a]):(Ey=0,new Promise(async p=>{var f,m,g;if(!(Aa!=null&&Aa.inputs[0].shape))return;let c={};if(((f=t.face.ssrnet)==null?void 0:f.crop)>0){let y=(m=t.face.ssrnet)==null?void 0:m.crop,x=[[y,y,1-y,1-y]];c.resize=fe.cropAndResize(e,x,[0],[Aa.inputs[0].shape[2],Aa.inputs[0].shape[1]])}else c.resize=fe.resizeBilinear(e,[Aa.inputs[0].shape[2],Aa.inputs[0].shape[1]],!1);c.enhance=Pe(()=>{var x,A;let y;if(((A=(x=Aa==null?void 0:Aa.inputs)==null?void 0:x[0].shape)==null?void 0:A[3])===1){let[b,w,C]=Ca(c.resize,3,3),T=te(b,My[0]),N=te(w,My[1]),M=te(C,My[2]),F=Eh([T,N,M]);y=te(ye(F,ze.tf05),2)}else y=te(ye(c.resize,ze.tf05),2);return y});let d={gender:"unknown",genderScore:0};(g=t.face.ssrnet)!=null&&g.enabled&&(c.gender=Aa.execute(c.enhance));let h=await c.gender.data();d.gender=h[0]>h[1]?"female":"male",d.genderScore=h[0]>h[1]?Math.trunc(100*h[0])/100:Math.trunc(100*h[1])/100,Object.keys(c).forEach(y=>J(c[y])),L0[a]=d,q9=n,X9=ae(),p(d)}))}var rn,_y=[],Z9=0,J9=0,Q9=Number.MAX_SAFE_INTEGER;async function eI(e){var t;return ne.initial&&(rn=null),rn?e.debug&&K("cached model:",rn.modelUrl):rn=await Fe((t=e.face.mobilefacenet)==null?void 0:t.modelPath),rn}async function $y(e,t,a,n){var i,o;if(!(rn!=null&&rn.executor))return[];let r=Q9<(((i=t.face.mobilefacenet)==null?void 0:i.skipFrames)||0),s=(((o=t.face.mobilefacenet)==null?void 0:o.skipTime)||0)>ae()-J9;return t.skipAllowed&&s&&r&&Z9===n&&_y[a]?(Q9++,_y[a]):new Promise(async l=>{var p;let u=[];if((p=t.face.mobilefacenet)!=null&&p.enabled&&(rn!=null&&rn.inputs[0].shape)){let c={};c.crop=fe.resizeBilinear(e,[rn.inputs[0].shape[2],rn.inputs[0].shape[1]],!1),c.data=rn.execute(c.crop);let d=await c.data.data();u=Array.from(d),Object.keys(c).forEach(h=>J(c[h]))}_y[a]=u,Z9=n,J9=ae(),l(u)})}var sn,Dy=[],aI=0,nI=0,rI=Number.MAX_SAFE_INTEGER;async function sI(e){return ne.initial&&(sn=null),sn?e.debug&&K("cached model:",sn.modelUrl):sn=await Fe(e.face.insightface.modelPath),sn}async function Py(e,t,a,n){var i,o;if(!(sn!=null&&sn.executor))return[];let r=rI<(((i=t.face.insightface)==null?void 0:i.skipFrames)||0),s=(((o=t.face.insightface)==null?void 0:o.skipTime)||0)>ae()-nI;return t.skipAllowed&&s&&r&&aI===n&&Dy[a]?(rI++,Dy[a]):new Promise(async l=>{var p;let u=[];if((p=t.face.insightface)!=null&&p.enabled&&(sn!=null&&sn.inputs[0].shape)){let c={};c.crop=fe.resizeBilinear(e,[sn.inputs[0].shape[2],sn.inputs[0].shape[1]],!1),c.data=sn.execute(c.crop);let d=await c.data.data();u=Array.from(d),Object.keys(c).forEach(h=>J(c[h]))}Dy[a]=u,aI=n,nI=ae(),l(u)})}var e3e=e=>{let t=(c,d)=>Math.atan2(c[1]-d[1],c[0]-d[0]);if(!e.annotations.rightEyeIris||!e.annotations.leftEyeIris)return{bearing:0,strength:0};let a=[0,-.1],n=1,r=(e.mesh[33][2]||0)>(e.mesh[263][2]||0),s=r?e.mesh[473]:e.mesh[468],i=r?[(e.mesh[133][0]+e.mesh[33][0])/2,(e.mesh[133][1]+e.mesh[33][1])/2]:[(e.mesh[263][0]+e.mesh[362][0])/2,(e.mesh[263][1]+e.mesh[362][1])/2],o=r?[e.mesh[133][0]-e.mesh[33][0],e.mesh[23][1]-e.mesh[27][1]]:[e.mesh[263][0]-e.mesh[362][0],e.mesh[253][1]-e.mesh[257][1]],l=[(i[0]-s[0])/o[0]-a[0],n*(s[1]-i[1])/o[1]-a[1]],u=Math.sqrt(l[0]*l[0]+l[1]*l[1]);return u=Math.min(u,e.boxRaw[2]/2,e.boxRaw[3]/2),{bearing:(t([0,0],l)+Math.PI/2)%Math.PI,strength:u}},oI=(e,t)=>{let a=m=>{let g=Math.sqrt(m[0]*m[0]+m[1]*m[1]+m[2]*m[2]);return m[0]/=g,m[1]/=g,m[2]/=g,m},n=(m,g)=>{let y=m[0]-g[0],x=m[1]-g[1],A=m[2]-g[2];return[y,x,A]},r=(m,g)=>{let y=m[1]*g[2]-m[2]*g[1],x=m[2]*g[0]-m[0]*g[2],A=m[0]*g[1]-m[1]*g[0];return[y,x,A]},s=m=>{let[g,y,x,A,b,w,C,T,N]=m,M,F,E;return A<1?A>-1?(E=Math.asin(A),F=Math.atan2(-C,g),M=Math.atan2(-w,b)):(E=-Math.PI/2,F=-Math.atan2(T,N),M=0):(E=Math.PI/2,F=Math.atan2(T,N),M=0),Number.isNaN(M)&&(M=0),Number.isNaN(F)&&(F=0),Number.isNaN(E)&&(E=0),{pitch:2*-M,yaw:2*-F,roll:2*-E}},i=e.meshRaw;if(!i||i.length<300)return{angle:{pitch:0,yaw:0,roll:0},matrix:[1,0,0,0,1,0,0,0,1],gaze:{bearing:0,strength:0}};let o=Math.max(e.boxRaw[2]*t[0],e.boxRaw[3]*t[1])/1.5,l=[i[10],i[152],i[234],i[454]].map(m=>[m[0]*t[0]/o,m[1]*t[1]/o,m[2]]),u=a(n(l[1],l[0])),p=a(n(l[3],l[2])),c=a(r(p,u));p=r(u,c);let d=[p[0],p[1],p[2],u[0],u[1],u[2],c[0],c[1],c[2]],h=s(d),f=i.length===478?e3e(e):{bearing:0,strength:0};return{angle:h,matrix:d,gaze:f}};function lI(e,t){let a=e==null?void 0:e.annotations;if(!(a!=null&&a.leftEyeIris)||!(a!=null&&a.rightEyeIris))return 0;let n=Math.max(Math.abs(a.leftEyeIris[3][0]-a.leftEyeIris[1][0]),Math.abs(a.rightEyeIris[3][0]-a.rightEyeIris[1][0]))/t;return Math.round(1.17/n)/100}var Oy=async(e,t)=>{var f,m,g,y,x,A,b,w,C,T,N,M,F,E,I,_,O,B,$,U,G,q,H;let a=ae(),n,r,s,i,o,l,u,p,c,d=[];e.state="run:face";let h=await x9(t,e.config);if(e.performance.face=ne.perfadd?(e.performance.face||0)+Math.trunc(ae()-a):Math.trunc(ae()-a),!t.shape||t.shape.length!==4)return[];if(!h)return[];for(let V=0;V<h.length;V++){if(e.analyze("Get Face"),!h[V].tensor||h[V].tensor.isDisposedInternal){K("Face object is disposed:",h[V].tensor);continue}if((f=e.config.face.detector)!=null&&f.mask){let me=await E9(h[V]);J(h[V].tensor),me&&(h[V].tensor=me)}let Z=h[V].mesh&&h[V].mesh.length>200?oI(h[V],[t.shape[2],t.shape[1]]):null;e.analyze("Start Emotion:"),e.config.async?i=(m=e.config.face.emotion)!=null&&m.enabled?yy(h[V].tensor||Ve([]),e.config,V,h.length):[]:(e.state="run:emotion",a=ae(),i=(g=e.config.face.emotion)!=null&&g.enabled?await yy(h[V].tensor||Ve([]),e.config,V,h.length):[],e.performance.emotion=ne.perfadd?(e.performance.emotion||0)+Math.trunc(ae()-a):Math.trunc(ae()-a)),e.analyze("End Emotion:"),e.analyze("Start AntiSpoof:"),e.config.async?u=(y=e.config.face.antispoof)!=null&&y.enabled?wy(h[V].tensor||Ve([]),e.config,V,h.length):0:(e.state="run:antispoof",a=ae(),u=(x=e.config.face.antispoof)!=null&&x.enabled?await wy(h[V].tensor||Ve([]),e.config,V,h.length):0,e.performance.antispoof=ne.perfadd?(e.performance.antispoof||0)+Math.trunc(ae()-a):Math.trunc(ae()-a)),e.analyze("End AntiSpoof:"),e.analyze("Start Liveness:"),e.config.async?p=(A=e.config.face.liveness)!=null&&A.enabled?Iy(h[V].tensor||Ve([]),e.config,V,h.length):0:(e.state="run:liveness",a=ae(),p=(b=e.config.face.liveness)!=null&&b.enabled?await Iy(h[V].tensor||Ve([]),e.config,V,h.length):0,e.performance.liveness=ne.perfadd?(e.performance.antispoof||0)+Math.trunc(ae()-a):Math.trunc(ae()-a)),e.analyze("End Liveness:"),e.analyze("Start GEAR:"),e.config.async?r=(w=e.config.face.gear)!=null&&w.enabled?Ty(h[V].tensor||Ve([]),e.config,V,h.length):null:(e.state="run:gear",a=ae(),r=(C=e.config.face.gear)!=null&&C.enabled?await Ty(h[V].tensor||Ve([]),e.config,V,h.length):null,e.performance.gear=Math.trunc(ae()-a)),e.analyze("End GEAR:"),e.analyze("Start SSRNet:"),e.config.async?(n=(T=e.config.face.ssrnet)!=null&&T.enabled?Ry(h[V].tensor||Ve([]),e.config,V,h.length):null,s=(N=e.config.face.ssrnet)!=null&&N.enabled?Fy(h[V].tensor||Ve([]),e.config,V,h.length):null):(e.state="run:ssrnet",a=ae(),n=(M=e.config.face.ssrnet)!=null&&M.enabled?await Ry(h[V].tensor||Ve([]),e.config,V,h.length):null,s=(F=e.config.face.ssrnet)!=null&&F.enabled?await Fy(h[V].tensor||Ve([]),e.config,V,h.length):null,e.performance.ssrnet=Math.trunc(ae()-a)),e.analyze("End SSRNet:"),e.analyze("Start MobileFaceNet:"),e.config.async?o=(E=e.config.face.mobilefacenet)!=null&&E.enabled?$y(h[V].tensor||Ve([]),e.config,V,h.length):null:(e.state="run:mobilefacenet",a=ae(),o=(I=e.config.face.mobilefacenet)!=null&&I.enabled?await $y(h[V].tensor||Ve([]),e.config,V,h.length):null,e.performance.mobilefacenet=Math.trunc(ae()-a)),e.analyze("End MobileFaceNet:"),e.analyze("Start InsightFace:"),e.config.async?l=(_=e.config.face.insightface)!=null&&_.enabled?Py(h[V].tensor||Ve([]),e.config,V,h.length):null:(e.state="run:mobilefacenet",a=ae(),l=(O=e.config.face.insightface)!=null&&O.enabled?await Py(h[V].tensor||Ve([]),e.config,V,h.length):null,e.performance.mobilefacenet=Math.trunc(ae()-a)),e.analyze("End InsightFace:"),e.analyze("Start Description:"),e.config.async?c=Ay(h[V].tensor||Ve([]),e.config,V,h.length):(e.state="run:description",a=ae(),c=await Ay(h[V].tensor||Ve([]),e.config,V,h.length),e.performance.description=ne.perfadd?(e.performance.description||0)+Math.trunc(ae()-a):Math.trunc(ae()-a)),e.analyze("End Description:"),e.config.async&&([n,s,i,o,l,c,r,u,p]=await Promise.all([n,s,i,o,l,c,r,u,p])),e.analyze("Finish Face:"),(B=e.config.face.ssrnet)!=null&&B.enabled&&n&&s&&(c={...c,age:n.age,gender:s.gender,genderScore:s.genderScore}),($=e.config.face.gear)!=null&&$.enabled&&r&&(c={...c,age:r.age,gender:r.gender,genderScore:r.genderScore,race:r.race}),(U=e.config.face.mobilefacenet)!=null&&U.enabled&&o&&(c.descriptor=o),(G=e.config.face.insightface)!=null&&G.enabled&&l&&(c.descriptor=l);let X=(q=e.config.face.iris)!=null&&q.enabled?lI(h[V],t.shape[2]):0,re=(H=e.config.face.detector)!=null&&H.return?Oe(h[V].tensor):null;J(h[V].tensor),h[V].tensor&&delete h[V].tensor;let ee={...h[V],id:V};c.age&&(ee.age=c.age),c.gender&&(ee.gender=c.gender),c.genderScore&&(ee.genderScore=c.genderScore),c.descriptor&&(ee.embedding=c.descriptor),c.race&&(ee.race=c.race),i&&(ee.emotion=i),u&&(ee.real=u),p&&(ee.live=p),X>0&&(ee.distance=X),Z&&(ee.rotation=Z),re&&(ee.tensor=re),d.push(ee),e.analyze("End Face")}return e.analyze("End FaceMesh:"),e.config.async&&(e.performance.face&&delete e.performance.face,e.performance.age&&delete e.performance.age,e.performance.gender&&delete e.performance.gender,e.performance.emotion&&delete e.performance.emotion),d};var $a={thumb:0,index:1,middle:2,ring:3,pinky:4,all:[0,1,2,3,4],nameMapping:{0:"thumb",1:"index",2:"middle",3:"ring",4:"pinky"},pointsMapping:{0:[[0,1],[1,2],[2,3],[3,4]],1:[[0,5],[5,6],[6,7],[7,8]],2:[[0,9],[9,10],[10,11],[11,12]],3:[[0,13],[13,14],[14,15],[15,16]],4:[[0,17],[17,18],[18,19],[19,20]]},getName:e=>$a.nameMapping[e],getPoints:e=>$a.pointsMapping[e]},ws={none:0,half:1,full:2,nameMapping:{0:"none",1:"half",2:"full"},getName:e=>ws.nameMapping[e]},Ft={verticalUp:0,verticalDown:1,horizontalLeft:2,horizontalRight:3,diagonalUpRight:4,diagonalUpLeft:5,diagonalDownRight:6,diagonalDownLeft:7,nameMapping:{0:"verticalUp",1:"verticalDown",2:"horizontalLeft",3:"horizontalRight",4:"diagonalUpRight",5:"diagonalUpLeft",6:"diagonalDownRight",7:"diagonalDownLeft"},getName:e=>Ft.nameMapping[e]},vs=class{constructor(t){he(this,"name");he(this,"curls");he(this,"directions");he(this,"weights");he(this,"weightsRelative");this.name=t,this.curls={},this.directions={},this.weights=[1,1,1,1,1],this.weightsRelative=[1,1,1,1,1]}curl(t,a,n){typeof this.curls[t]=="undefined"&&(this.curls[t]=[]),this.curls[t].push([a,n])}direction(t,a,n){this.directions[t]||(this.directions[t]=[]),this.directions[t].push([a,n])}weight(t,a){this.weights[t]=a;let n=this.weights.reduce((r,s)=>r+s,0);this.weightsRelative=this.weights.map(r=>r*5/n)}matchAgainst(t,a){let n=0;for(let r in t){let s=t[r],i=this.curls[r];if(typeof i=="undefined"){n+=this.weightsRelative[r];continue}for(let[o,l]of i)if(s===o){n+=l*this.weightsRelative[r];break}}for(let r in a){let s=a[r],i=this.directions[r];if(typeof i=="undefined"){n+=this.weightsRelative[r];continue}for(let[o,l]of i)if(s===o){n+=l*this.weightsRelative[r];break}}return n/10}};var{thumb:Un,index:Fr,middle:_r,ring:ul,pinky:dl}=$a,{none:Gn,half:a3e,full:Hn}=ws,{verticalUp:ed,verticalDown:B4e,horizontalLeft:zy,horizontalRight:n3e,diagonalUpRight:r3e,diagonalUpLeft:td,diagonalDownRight:V4e,diagonalDownLeft:U4e}=Ft,ks=new vs("thumbs up");ks.curl(Un,Gn,1);ks.direction(Un,ed,1);ks.direction(Un,td,.25);ks.direction(Un,r3e,.25);for(let e of[$a.index,$a.middle,$a.ring,$a.pinky])ks.curl(e,Hn,1),ks.direction(e,zy,1),ks.direction(e,n3e,1);var Vt=new vs("victory");Vt.curl(Un,a3e,.5);Vt.curl(Un,Gn,.5);Vt.direction(Un,ed,1);Vt.direction(Un,td,1);Vt.curl(Fr,Gn,1);Vt.direction(Fr,ed,.75);Vt.direction(Fr,td,1);Vt.curl(_r,Gn,1);Vt.direction(_r,ed,1);Vt.direction(_r,td,.75);Vt.curl(ul,Hn,1);Vt.direction(ul,ed,.2);Vt.direction(ul,td,1);Vt.direction(ul,zy,.2);Vt.curl(dl,Hn,1);Vt.direction(dl,ed,.2);Vt.direction(dl,td,1);Vt.direction(dl,zy,.2);Vt.weight(Fr,2);Vt.weight(_r,2);var Is=new vs("point");Is.curl(Un,Hn,1);Is.curl(Fr,Gn,.5);Is.curl(_r,Hn,.5);Is.curl(ul,Hn,.5);Is.curl(dl,Hn,.5);Is.weight(Fr,2);Is.weight(_r,2);var Ss=new vs("middle finger");Ss.curl(Un,Gn,1);Ss.curl(Fr,Hn,.5);Ss.curl(_r,Hn,.5);Ss.curl(ul,Hn,.5);Ss.curl(dl,Hn,.5);Ss.weight(Fr,2);Ss.weight(_r,2);var ad=new vs("open palm");ad.curl(Un,Gn,.75);ad.curl(Fr,Gn,.75);ad.curl(_r,Gn,.75);ad.curl(ul,Gn,.75);ad.curl(dl,Gn,.75);var uI=[ks,Vt,Is,Ss,ad];var s3e=.7,pl={HALF_CURL_START_LIMIT:60,NO_CURL_START_LIMIT:130,DISTANCE_VOTE_POWER:1.1,SINGLE_ANGLE_VOTE_POWER:.9,TOTAL_ANGLE_VOTE_POWER:1.6};function dI(e,t,a,n){let r=(t-n)/(e-a),s=Math.atan(r)*180/Math.PI;return s<=0?s=-s:s>0&&(s=180-s),s}function cI(e,t){if(!e||!t)return[0,0];let a=dI(e[0],e[1],t[0],t[1]);if(e.length===2)return a;let n=dI(e[1],e[2],t[1],t[2]);return[a,n]}function pI(e,t=1){let a=0,n=0,r=0;return e>=75&&e<=105?a=1*t:e>=25&&e<=155?n=1*t:r=1*t,[a,n,r]}function i3e(e,t,a){let n=e[0]-t[0],r=e[0]-a[0],s=t[0]-a[0],i=e[1]-t[1],o=e[1]-a[1],l=t[1]-a[1],u=e[2]-t[2],p=e[2]-a[2],c=t[2]-a[2],d=Math.sqrt(n*n+i*i+u*u),h=Math.sqrt(r*r+o*o+p*p),f=Math.sqrt(s*s+l*l+c*c),m=(f*f+d*d-h*h)/(2*f*d);m>1?m=1:m<-1&&(m=-1);let g=Math.acos(m);g=57.2958*g%180;let y;return g>pl.NO_CURL_START_LIMIT?y=ws.none:g>pl.HALF_CURL_START_LIMIT?y=ws.half:y=ws.full,y}function hI(e,t,a,n){let r;return n===Math.abs(e)?e>0?r=Ft.horizontalLeft:r=Ft.horizontalRight:n===Math.abs(t)?t>0?r=Ft.horizontalLeft:r=Ft.horizontalRight:a>0?r=Ft.horizontalLeft:r=Ft.horizontalRight,r}function fI(e,t,a,n){let r;return n===Math.abs(e)?e<0?r=Ft.verticalDown:r=Ft.verticalUp:n===Math.abs(t)?t<0?r=Ft.verticalDown:r=Ft.verticalUp:a<0?r=Ft.verticalDown:r=Ft.verticalUp,r}function o3e(e,t,a,n,r,s,i,o){let l,u=fI(e,t,a,n),p=hI(r,s,i,o);return u===Ft.verticalUp?p===Ft.horizontalLeft?l=Ft.diagonalUpLeft:l=Ft.diagonalUpRight:p===Ft.horizontalLeft?l=Ft.diagonalDownLeft:l=Ft.diagonalDownRight,l}function l3e(e,t,a,n){let r=e[0]-t[0],s=e[0]-a[0],i=t[0]-a[0],o=e[1]-t[1],l=e[1]-a[1],u=t[1]-a[1],p=Math.max(Math.abs(r),Math.abs(s),Math.abs(i)),c=Math.max(Math.abs(o),Math.abs(l),Math.abs(u)),d=0,h=0,f=0,m=c/(p+1e-5);m>1.5?d+=pl.DISTANCE_VOTE_POWER:m>.66?h+=pl.DISTANCE_VOTE_POWER:f+=pl.DISTANCE_VOTE_POWER;let g=Math.sqrt(r*r+o*o),y=Math.sqrt(s*s+l*l),x=Math.sqrt(i*i+u*u),A=Math.max(g,y,x),b=e[0],w=e[1],C=a[0],T=a[1];A===g?(C=a[0],T=a[1]):A===x&&(b=t[0],w=t[1]);let F=cI([b,w],[C,T]),E=pI(F,pl.TOTAL_ANGLE_VOTE_POWER);d+=E[0],h+=E[1],f+=E[2];for(let _ of n){let O=pI(_,pl.SINGLE_ANGLE_VOTE_POWER);d+=O[0],h+=O[1],f+=O[2]}let I;return d===Math.max(d,h,f)?I=fI(l,o,u,c):f===Math.max(h,f)?I=hI(s,r,i,p):I=o3e(l,o,u,c,s,r,i,p),I}function mI(e){let t=[],a=[],n=[],r=[];if(!e)return{curls:n,directions:r};for(let s of $a.all){let i=$a.getPoints(s),o=[],l=[];for(let u of i){let p=e[u[0]],c=e[u[1]],d=cI(p,c),h=d[0],f=d[1];o.push(h),l.push(f)}t.push(o),a.push(l)}for(let s of $a.all){let i=s===$a.thumb?1:0,o=$a.getPoints(s),l=e[o[i][0]],u=e[o[i+1][1]],p=e[o[3][1]],c=i3e(l,u,p),d=l3e(l,u,p,t[s].slice(i));n[s]=c,r[s]=d}return{curls:n,directions:r}}function W0(e){if(!e||e.length===0)return null;let t=mI(e),a={};for(let n of $a.all)a[$a.getName(n)]={curl:ws.getName(t.curls[n]),direction:Ft.getName(t.directions[n])};return a}function gI(e){let t=[];if(!e||e.length===0)return t;let a=mI(e);for(let n of uI){let r=n.matchAgainst(a.curls,a.directions);r>=s3e&&t.push({name:n.name,confidence:r})}return t}var yI=e=>{if(!e)return[];let t=[];for(let a=0;a<e.length;a++){let n=e[a].keypoints.find(l=>l.part==="leftWrist"),r=e[a].keypoints.find(l=>l.part==="rightWrist"),s=e[a].keypoints.find(l=>l.part==="nose");s&&n&&r&&n.position[1]<s.position[1]&&r.position[1]<s.position[1]?t.push({body:a,gesture:"i give up"}):s&&n&&n.position[1]<s.position[1]?t.push({body:a,gesture:"raise left hand"}):s&&r&&r.position[1]<s.position[1]&&t.push({body:a,gesture:"raise right hand"});let i=e[a].keypoints.find(l=>l.part==="leftShoulder"),o=e[a].keypoints.find(l=>l.part==="rightShoulder");i&&o&&Math.abs(i.positionRaw[1]-o.positionRaw[1])>.1&&t.push({body:a,gesture:`leaning ${i.position[1]>o.position[1]?"left":"right"}`})}return t},xI=e=>{if(!e)return[];let t=[];for(let a=0;a<e.length;a++)if(e[a].mesh&&e[a].mesh.length>450){let n=(e[a].mesh[33][2]||0)-(e[a].mesh[263][2]||0),r=e[a].mesh[33][0]-e[a].mesh[263][0];Math.abs(n/r)<=.15?t.push({face:a,gesture:"facing center"}):t.push({face:a,gesture:`facing ${n<0?"left":"right"}`}),Math.abs(e[a].mesh[374][1]-e[a].mesh[386][1])/Math.abs(e[a].mesh[443][1]-e[a].mesh[450][1])<.2&&t.push({face:a,gesture:"blink left eye"}),Math.abs(e[a].mesh[145][1]-e[a].mesh[159][1])/Math.abs(e[a].mesh[223][1]-e[a].mesh[230][1])<.2&&t.push({face:a,gesture:"blink right eye"});let o=Math.min(100,500*Math.abs(e[a].mesh[13][1]-e[a].mesh[14][1])/Math.abs(e[a].mesh[10][1]-e[a].mesh[152][1]));o>10&&t.push({face:a,gesture:`mouth ${Math.trunc(o)}% open`});let l=e[a].mesh[152][2]||0;Math.abs(l)>10&&t.push({face:a,gesture:`head ${l<0?"up":"down"}`})}return t},AI=e=>{var a,n,r,s;if(!e)return[];let t=[];for(let i=0;i<e.length;i++){if(!((n=(a=e[i].annotations)==null?void 0:a.leftEyeIris)!=null&&n[0])||!((s=(r=e[i].annotations)==null?void 0:r.rightEyeIris)!=null&&s[0]))continue;let o=e[i].annotations.leftEyeIris[3][0]-e[i].annotations.leftEyeIris[1][0],l=e[i].annotations.leftEyeIris[4][1]-e[i].annotations.leftEyeIris[2][1],u=Math.abs(o*l),p=e[i].annotations.rightEyeIris[3][0]-e[i].annotations.rightEyeIris[1][0],c=e[i].annotations.rightEyeIris[4][1]-e[i].annotations.rightEyeIris[2][1],d=Math.abs(p*c),h=!1;Math.abs(u-d)/Math.max(u,d)<.25&&(h=!0,t.push({iris:i,gesture:"facing center"}));let m=Math.abs(e[i].mesh[263][0]-e[i].annotations.leftEyeIris[0][0])/e[i].box[2],g=Math.abs(e[i].mesh[33][0]-e[i].annotations.rightEyeIris[0][0])/e[i].box[2];(m>.06||g>.06)&&(h=!1),m>g?g>.04&&t.push({iris:i,gesture:"looking right"}):m>.04&&t.push({iris:i,gesture:"looking left"});let y=Math.abs(e[i].mesh[145][1]-e[i].annotations.rightEyeIris[0][1])/e[i].box[3],x=Math.abs(e[i].mesh[374][1]-e[i].annotations.leftEyeIris[0][1])/e[i].box[3];(x<.01||y<.01||x>.022||y>.022)&&(h=!1),(x<.01||y<.01)&&t.push({iris:i,gesture:"looking down"}),(x>.022||y>.022)&&t.push({iris:i,gesture:"looking up"}),h&&t.push({iris:i,gesture:"looking center"})}return t},bI=e=>{if(!e)return[];let t=[];for(let a=0;a<e.length;a++){let n=[];if(e[a].annotations)for(let[r,s]of Object.entries(e[a].annotations))r!=="palmBase"&&Array.isArray(s)&&s[0]&&n.push({name:r.toLowerCase(),position:s[0]});if(n&&n.length>0){let r=n.reduce((i,o)=>(i.position[2]||0)<(o.position[2]||0)?i:o);t.push({hand:a,gesture:`${r.name} forward`});let s=n.reduce((i,o)=>i.position[1]<o.position[1]?i:o);t.push({hand:a,gesture:`${s.name} up`})}if(e[a].keypoints){let r=gI(e[a].keypoints);for(let s of r)t.push({hand:a,gesture:s.name})}}return t};function B0(e){return[Math.abs(e.endPoint[0]-e.startPoint[0]),Math.abs(e.endPoint[1]-e.startPoint[1])]}function ac(e){return[e.startPoint[0]+(e.endPoint[0]-e.startPoint[0])/2,e.startPoint[1]+(e.endPoint[1]-e.startPoint[1])/2]}function kI(e,t,a){let n=t.shape[1],r=t.shape[2],s=[[e.startPoint[1]/n,e.startPoint[0]/r,e.endPoint[1]/n,e.endPoint[0]/r]];return fe.cropAndResize(t,s,[0],a)}function II(e,t){let a=[e.startPoint[0]*t[0],e.startPoint[1]*t[1]],n=[e.endPoint[0]*t[0],e.endPoint[1]*t[1]],r=e.palmLandmarks.map(s=>[s[0]*t[0],s[1]*t[1]]);return{startPoint:a,endPoint:n,palmLandmarks:r,confidence:e.confidence}}function V0(e,t=1.5){let a=ac(e),n=B0(e),r=[t*n[0]/2,t*n[1]/2],s=[a[0]-r[0],a[1]-r[1]],i=[a[0]+r[0],a[1]+r[1]];return{startPoint:s,endPoint:i,palmLandmarks:e.palmLandmarks}}function U0(e){let t=ac(e),a=B0(e),r=Math.max(...a)/2,s=[t[0]-r,t[1]-r],i=[t[0]+r,t[1]+r];return{startPoint:s,endPoint:i,palmLandmarks:e.palmLandmarks}}function d3e(e){return e-2*Math.PI*Math.floor((e+Math.PI)/(2*Math.PI))}function SI(e,t){let a=Math.PI/2-Math.atan2(-(t[1]-e[1]),t[0]-e[0]);return d3e(a)}var vI=(e,t)=>[[1,0,e],[0,1,t],[0,0,1]];function Cs(e,t){let a=0;for(let n=0;n<e.length;n++)a+=e[n]*t[n];return a}function p3e(e,t){let a=[];for(let n=0;n<e.length;n++)a.push(e[n][t]);return a}function wI(e,t){let a=[],n=e.length;for(let r=0;r<n;r++){a.push([]);for(let s=0;s<n;s++)a[r].push(Cs(e[r],p3e(t,s)))}return a}function Wy(e,t){let a=Math.cos(e),n=Math.sin(e),r=[[a,-n,0],[n,a,0],[0,0,1]],s=vI(t[0],t[1]),i=wI(s,r),o=vI(-t[0],-t[1]);return wI(i,o)}function CI(e){let t=[[e[0][0],e[1][0]],[e[0][1],e[1][1]]],a=[e[0][2],e[1][2]],n=[-Cs(t[0],a),-Cs(t[1],a)];return[t[0].concat(n[0]),t[1].concat(n[1]),[0,0,1]]}function By(e,t){return[Cs(e,t[0]),Cs(e,t[1])]}var NI=[{x:.015625,y:.015625},{x:.015625,y:.015625},{x:.046875,y:.015625},{x:.046875,y:.015625},{x:.078125,y:.015625},{x:.078125,y:.015625},{x:.109375,y:.015625},{x:.109375,y:.015625},{x:.140625,y:.015625},{x:.140625,y:.015625},{x:.171875,y:.015625},{x:.171875,y:.015625},{x:.203125,y:.015625},{x:.203125,y:.015625},{x:.234375,y:.015625},{x:.234375,y:.015625},{x:.265625,y:.015625},{x:.265625,y:.015625},{x:.296875,y:.015625},{x:.296875,y:.015625},{x:.328125,y:.015625},{x:.328125,y:.015625},{x:.359375,y:.015625},{x:.359375,y:.015625},{x:.390625,y:.015625},{x:.390625,y:.015625},{x:.421875,y:.015625},{x:.421875,y:.015625},{x:.453125,y:.015625},{x:.453125,y:.015625},{x:.484375,y:.015625},{x:.484375,y:.015625},{x:.515625,y:.015625},{x:.515625,y:.015625},{x:.546875,y:.015625},{x:.546875,y:.015625},{x:.578125,y:.015625},{x:.578125,y:.015625},{x:.609375,y:.015625},{x:.609375,y:.015625},{x:.640625,y:.015625},{x:.640625,y:.015625},{x:.671875,y:.015625},{x:.671875,y:.015625},{x:.703125,y:.015625},{x:.703125,y:.015625},{x:.734375,y:.015625},{x:.734375,y:.015625},{x:.765625,y:.015625},{x:.765625,y:.015625},{x:.796875,y:.015625},{x:.796875,y:.015625},{x:.828125,y:.015625},{x:.828125,y:.015625},{x:.859375,y:.015625},{x:.859375,y:.015625},{x:.890625,y:.015625},{x:.890625,y:.015625},{x:.921875,y:.015625},{x:.921875,y:.015625},{x:.953125,y:.015625},{x:.953125,y:.015625},{x:.984375,y:.015625},{x:.984375,y:.015625},{x:.015625,y:.046875},{x:.015625,y:.046875},{x:.046875,y:.046875},{x:.046875,y:.046875},{x:.078125,y:.046875},{x:.078125,y:.046875},{x:.109375,y:.046875},{x:.109375,y:.046875},{x:.140625,y:.046875},{x:.140625,y:.046875},{x:.171875,y:.046875},{x:.171875,y:.046875},{x:.203125,y:.046875},{x:.203125,y:.046875},{x:.234375,y:.046875},{x:.234375,y:.046875},{x:.265625,y:.046875},{x:.265625,y:.046875},{x:.296875,y:.046875},{x:.296875,y:.046875},{x:.328125,y:.046875},{x:.328125,y:.046875},{x:.359375,y:.046875},{x:.359375,y:.046875},{x:.390625,y:.046875},{x:.390625,y:.046875},{x:.421875,y:.046875},{x:.421875,y:.046875},{x:.453125,y:.046875},{x:.453125,y:.046875},{x:.484375,y:.046875},{x:.484375,y:.046875},{x:.515625,y:.046875},{x:.515625,y:.046875},{x:.546875,y:.046875},{x:.546875,y:.046875},{x:.578125,y:.046875},{x:.578125,y:.046875},{x:.609375,y:.046875},{x:.609375,y:.046875},{x:.640625,y:.046875},{x:.640625,y:.046875},{x:.671875,y:.046875},{x:.671875,y:.046875},{x:.703125,y:.046875},{x:.703125,y:.046875},{x:.734375,y:.046875},{x:.734375,y:.046875},{x:.765625,y:.046875},{x:.765625,y:.046875},{x:.796875,y:.046875},{x:.796875,y:.046875},{x:.828125,y:.046875},{x:.828125,y:.046875},{x:.859375,y:.046875},{x:.859375,y:.046875},{x:.890625,y:.046875},{x:.890625,y:.046875},{x:.921875,y:.046875},{x:.921875,y:.046875},{x:.953125,y:.046875},{x:.953125,y:.046875},{x:.984375,y:.046875},{x:.984375,y:.046875},{x:.015625,y:.078125},{x:.015625,y:.078125},{x:.046875,y:.078125},{x:.046875,y:.078125},{x:.078125,y:.078125},{x:.078125,y:.078125},{x:.109375,y:.078125},{x:.109375,y:.078125},{x:.140625,y:.078125},{x:.140625,y:.078125},{x:.171875,y:.078125},{x:.171875,y:.078125},{x:.203125,y:.078125},{x:.203125,y:.078125},{x:.234375,y:.078125},{x:.234375,y:.078125},{x:.265625,y:.078125},{x:.265625,y:.078125},{x:.296875,y:.078125},{x:.296875,y:.078125},{x:.328125,y:.078125},{x:.328125,y:.078125},{x:.359375,y:.078125},{x:.359375,y:.078125},{x:.390625,y:.078125},{x:.390625,y:.078125},{x:.421875,y:.078125},{x:.421875,y:.078125},{x:.453125,y:.078125},{x:.453125,y:.078125},{x:.484375,y:.078125},{x:.484375,y:.078125},{x:.515625,y:.078125},{x:.515625,y:.078125},{x:.546875,y:.078125},{x:.546875,y:.078125},{x:.578125,y:.078125},{x:.578125,y:.078125},{x:.609375,y:.078125},{x:.609375,y:.078125},{x:.640625,y:.078125},{x:.640625,y:.078125},{x:.671875,y:.078125},{x:.671875,y:.078125},{x:.703125,y:.078125},{x:.703125,y:.078125},{x:.734375,y:.078125},{x:.734375,y:.078125},{x:.765625,y:.078125},{x:.765625,y:.078125},{x:.796875,y:.078125},{x:.796875,y:.078125},{x:.828125,y:.078125},{x:.828125,y:.078125},{x:.859375,y:.078125},{x:.859375,y:.078125},{x:.890625,y:.078125},{x:.890625,y:.078125},{x:.921875,y:.078125},{x:.921875,y:.078125},{x:.953125,y:.078125},{x:.953125,y:.078125},{x:.984375,y:.078125},{x:.984375,y:.078125},{x:.015625,y:.109375},{x:.015625,y:.109375},{x:.046875,y:.109375},{x:.046875,y:.109375},{x:.078125,y:.109375},{x:.078125,y:.109375},{x:.109375,y:.109375},{x:.109375,y:.109375},{x:.140625,y:.109375},{x:.140625,y:.109375},{x:.171875,y:.109375},{x:.171875,y:.109375},{x:.203125,y:.109375},{x:.203125,y:.109375},{x:.234375,y:.109375},{x:.234375,y:.109375},{x:.265625,y:.109375},{x:.265625,y:.109375},{x:.296875,y:.109375},{x:.296875,y:.109375},{x:.328125,y:.109375},{x:.328125,y:.109375},{x:.359375,y:.109375},{x:.359375,y:.109375},{x:.390625,y:.109375},{x:.390625,y:.109375},{x:.421875,y:.109375},{x:.421875,y:.109375},{x:.453125,y:.109375},{x:.453125,y:.109375},{x:.484375,y:.109375},{x:.484375,y:.109375},{x:.515625,y:.109375},{x:.515625,y:.109375},{x:.546875,y:.109375},{x:.546875,y:.109375},{x:.578125,y:.109375},{x:.578125,y:.109375},{x:.609375,y:.109375},{x:.609375,y:.109375},{x:.640625,y:.109375},{x:.640625,y:.109375},{x:.671875,y:.109375},{x:.671875,y:.109375},{x:.703125,y:.109375},{x:.703125,y:.109375},{x:.734375,y:.109375},{x:.734375,y:.109375},{x:.765625,y:.109375},{x:.765625,y:.109375},{x:.796875,y:.109375},{x:.796875,y:.109375},{x:.828125,y:.109375},{x:.828125,y:.109375},{x:.859375,y:.109375},{x:.859375,y:.109375},{x:.890625,y:.109375},{x:.890625,y:.109375},{x:.921875,y:.109375},{x:.921875,y:.109375},{x:.953125,y:.109375},{x:.953125,y:.109375},{x:.984375,y:.109375},{x:.984375,y:.109375},{x:.015625,y:.140625},{x:.015625,y:.140625},{x:.046875,y:.140625},{x:.046875,y:.140625},{x:.078125,y:.140625},{x:.078125,y:.140625},{x:.109375,y:.140625},{x:.109375,y:.140625},{x:.140625,y:.140625},{x:.140625,y:.140625},{x:.171875,y:.140625},{x:.171875,y:.140625},{x:.203125,y:.140625},{x:.203125,y:.140625},{x:.234375,y:.140625},{x:.234375,y:.140625},{x:.265625,y:.140625},{x:.265625,y:.140625},{x:.296875,y:.140625},{x:.296875,y:.140625},{x:.328125,y:.140625},{x:.328125,y:.140625},{x:.359375,y:.140625},{x:.359375,y:.140625},{x:.390625,y:.140625},{x:.390625,y:.140625},{x:.421875,y:.140625},{x:.421875,y:.140625},{x:.453125,y:.140625},{x:.453125,y:.140625},{x:.484375,y:.140625},{x:.484375,y:.140625},{x:.515625,y:.140625},{x:.515625,y:.140625},{x:.546875,y:.140625},{x:.546875,y:.140625},{x:.578125,y:.140625},{x:.578125,y:.140625},{x:.609375,y:.140625},{x:.609375,y:.140625},{x:.640625,y:.140625},{x:.640625,y:.140625},{x:.671875,y:.140625},{x:.671875,y:.140625},{x:.703125,y:.140625},{x:.703125,y:.140625},{x:.734375,y:.140625},{x:.734375,y:.140625},{x:.765625,y:.140625},{x:.765625,y:.140625},{x:.796875,y:.140625},{x:.796875,y:.140625},{x:.828125,y:.140625},{x:.828125,y:.140625},{x:.859375,y:.140625},{x:.859375,y:.140625},{x:.890625,y:.140625},{x:.890625,y:.140625},{x:.921875,y:.140625},{x:.921875,y:.140625},{x:.953125,y:.140625},{x:.953125,y:.140625},{x:.984375,y:.140625},{x:.984375,y:.140625},{x:.015625,y:.171875},{x:.015625,y:.171875},{x:.046875,y:.171875},{x:.046875,y:.171875},{x:.078125,y:.171875},{x:.078125,y:.171875},{x:.109375,y:.171875},{x:.109375,y:.171875},{x:.140625,y:.171875},{x:.140625,y:.171875},{x:.171875,y:.171875},{x:.171875,y:.171875},{x:.203125,y:.171875},{x:.203125,y:.171875},{x:.234375,y:.171875},{x:.234375,y:.171875},{x:.265625,y:.171875},{x:.265625,y:.171875},{x:.296875,y:.171875},{x:.296875,y:.171875},{x:.328125,y:.171875},{x:.328125,y:.171875},{x:.359375,y:.171875},{x:.359375,y:.171875},{x:.390625,y:.171875},{x:.390625,y:.171875},{x:.421875,y:.171875},{x:.421875,y:.171875},{x:.453125,y:.171875},{x:.453125,y:.171875},{x:.484375,y:.171875},{x:.484375,y:.171875},{x:.515625,y:.171875},{x:.515625,y:.171875},{x:.546875,y:.171875},{x:.546875,y:.171875},{x:.578125,y:.171875},{x:.578125,y:.171875},{x:.609375,y:.171875},{x:.609375,y:.171875},{x:.640625,y:.171875},{x:.640625,y:.171875},{x:.671875,y:.171875},{x:.671875,y:.171875},{x:.703125,y:.171875},{x:.703125,y:.171875},{x:.734375,y:.171875},{x:.734375,y:.171875},{x:.765625,y:.171875},{x:.765625,y:.171875},{x:.796875,y:.171875},{x:.796875,y:.171875},{x:.828125,y:.171875},{x:.828125,y:.171875},{x:.859375,y:.171875},{x:.859375,y:.171875},{x:.890625,y:.171875},{x:.890625,y:.171875},{x:.921875,y:.171875},{x:.921875,y:.171875},{x:.953125,y:.171875},{x:.953125,y:.171875},{x:.984375,y:.171875},{x:.984375,y:.171875},{x:.015625,y:.203125},{x:.015625,y:.203125},{x:.046875,y:.203125},{x:.046875,y:.203125},{x:.078125,y:.203125},{x:.078125,y:.203125},{x:.109375,y:.203125},{x:.109375,y:.203125},{x:.140625,y:.203125},{x:.140625,y:.203125},{x:.171875,y:.203125},{x:.171875,y:.203125},{x:.203125,y:.203125},{x:.203125,y:.203125},{x:.234375,y:.203125},{x:.234375,y:.203125},{x:.265625,y:.203125},{x:.265625,y:.203125},{x:.296875,y:.203125},{x:.296875,y:.203125},{x:.328125,y:.203125},{x:.328125,y:.203125},{x:.359375,y:.203125},{x:.359375,y:.203125},{x:.390625,y:.203125},{x:.390625,y:.203125},{x:.421875,y:.203125},{x:.421875,y:.203125},{x:.453125,y:.203125},{x:.453125,y:.203125},{x:.484375,y:.203125},{x:.484375,y:.203125},{x:.515625,y:.203125},{x:.515625,y:.203125},{x:.546875,y:.203125},{x:.546875,y:.203125},{x:.578125,y:.203125},{x:.578125,y:.203125},{x:.609375,y:.203125},{x:.609375,y:.203125},{x:.640625,y:.203125},{x:.640625,y:.203125},{x:.671875,y:.203125},{x:.671875,y:.203125},{x:.703125,y:.203125},{x:.703125,y:.203125},{x:.734375,y:.203125},{x:.734375,y:.203125},{x:.765625,y:.203125},{x:.765625,y:.203125},{x:.796875,y:.203125},{x:.796875,y:.203125},{x:.828125,y:.203125},{x:.828125,y:.203125},{x:.859375,y:.203125},{x:.859375,y:.203125},{x:.890625,y:.203125},{x:.890625,y:.203125},{x:.921875,y:.203125},{x:.921875,y:.203125},{x:.953125,y:.203125},{x:.953125,y:.203125},{x:.984375,y:.203125},{x:.984375,y:.203125},{x:.015625,y:.234375},{x:.015625,y:.234375},{x:.046875,y:.234375},{x:.046875,y:.234375},{x:.078125,y:.234375},{x:.078125,y:.234375},{x:.109375,y:.234375},{x:.109375,y:.234375},{x:.140625,y:.234375},{x:.140625,y:.234375},{x:.171875,y:.234375},{x:.171875,y:.234375},{x:.203125,y:.234375},{x:.203125,y:.234375},{x:.234375,y:.234375},{x:.234375,y:.234375},{x:.265625,y:.234375},{x:.265625,y:.234375},{x:.296875,y:.234375},{x:.296875,y:.234375},{x:.328125,y:.234375},{x:.328125,y:.234375},{x:.359375,y:.234375},{x:.359375,y:.234375},{x:.390625,y:.234375},{x:.390625,y:.234375},{x:.421875,y:.234375},{x:.421875,y:.234375},{x:.453125,y:.234375},{x:.453125,y:.234375},{x:.484375,y:.234375},{x:.484375,y:.234375},{x:.515625,y:.234375},{x:.515625,y:.234375},{x:.546875,y:.234375},{x:.546875,y:.234375},{x:.578125,y:.234375},{x:.578125,y:.234375},{x:.609375,y:.234375},{x:.609375,y:.234375},{x:.640625,y:.234375},{x:.640625,y:.234375},{x:.671875,y:.234375},{x:.671875,y:.234375},{x:.703125,y:.234375},{x:.703125,y:.234375},{x:.734375,y:.234375},{x:.734375,y:.234375},{x:.765625,y:.234375},{x:.765625,y:.234375},{x:.796875,y:.234375},{x:.796875,y:.234375},{x:.828125,y:.234375},{x:.828125,y:.234375},{x:.859375,y:.234375},{x:.859375,y:.234375},{x:.890625,y:.234375},{x:.890625,y:.234375},{x:.921875,y:.234375},{x:.921875,y:.234375},{x:.953125,y:.234375},{x:.953125,y:.234375},{x:.984375,y:.234375},{x:.984375,y:.234375},{x:.015625,y:.265625},{x:.015625,y:.265625},{x:.046875,y:.265625},{x:.046875,y:.265625},{x:.078125,y:.265625},{x:.078125,y:.265625},{x:.109375,y:.265625},{x:.109375,y:.265625},{x:.140625,y:.265625},{x:.140625,y:.265625},{x:.171875,y:.265625},{x:.171875,y:.265625},{x:.203125,y:.265625},{x:.203125,y:.265625},{x:.234375,y:.265625},{x:.234375,y:.265625},{x:.265625,y:.265625},{x:.265625,y:.265625},{x:.296875,y:.265625},{x:.296875,y:.265625},{x:.328125,y:.265625},{x:.328125,y:.265625},{x:.359375,y:.265625},{x:.359375,y:.265625},{x:.390625,y:.265625},{x:.390625,y:.265625},{x:.421875,y:.265625},{x:.421875,y:.265625},{x:.453125,y:.265625},{x:.453125,y:.265625},{x:.484375,y:.265625},{x:.484375,y:.265625},{x:.515625,y:.265625},{x:.515625,y:.265625},{x:.546875,y:.265625},{x:.546875,y:.265625},{x:.578125,y:.265625},{x:.578125,y:.265625},{x:.609375,y:.265625},{x:.609375,y:.265625},{x:.640625,y:.265625},{x:.640625,y:.265625},{x:.671875,y:.265625},{x:.671875,y:.265625},{x:.703125,y:.265625},{x:.703125,y:.265625},{x:.734375,y:.265625},{x:.734375,y:.265625},{x:.765625,y:.265625},{x:.765625,y:.265625},{x:.796875,y:.265625},{x:.796875,y:.265625},{x:.828125,y:.265625},{x:.828125,y:.265625},{x:.859375,y:.265625},{x:.859375,y:.265625},{x:.890625,y:.265625},{x:.890625,y:.265625},{x:.921875,y:.265625},{x:.921875,y:.265625},{x:.953125,y:.265625},{x:.953125,y:.265625},{x:.984375,y:.265625},{x:.984375,y:.265625},{x:.015625,y:.296875},{x:.015625,y:.296875},{x:.046875,y:.296875},{x:.046875,y:.296875},{x:.078125,y:.296875},{x:.078125,y:.296875},{x:.109375,y:.296875},{x:.109375,y:.296875},{x:.140625,y:.296875},{x:.140625,y:.296875},{x:.171875,y:.296875},{x:.171875,y:.296875},{x:.203125,y:.296875},{x:.203125,y:.296875},{x:.234375,y:.296875},{x:.234375,y:.296875},{x:.265625,y:.296875},{x:.265625,y:.296875},{x:.296875,y:.296875},{x:.296875,y:.296875},{x:.328125,y:.296875},{x:.328125,y:.296875},{x:.359375,y:.296875},{x:.359375,y:.296875},{x:.390625,y:.296875},{x:.390625,y:.296875},{x:.421875,y:.296875},{x:.421875,y:.296875},{x:.453125,y:.296875},{x:.453125,y:.296875},{x:.484375,y:.296875},{x:.484375,y:.296875},{x:.515625,y:.296875},{x:.515625,y:.296875},{x:.546875,y:.296875},{x:.546875,y:.296875},{x:.578125,y:.296875},{x:.578125,y:.296875},{x:.609375,y:.296875},{x:.609375,y:.296875},{x:.640625,y:.296875},{x:.640625,y:.296875},{x:.671875,y:.296875},{x:.671875,y:.296875},{x:.703125,y:.296875},{x:.703125,y:.296875},{x:.734375,y:.296875},{x:.734375,y:.296875},{x:.765625,y:.296875},{x:.765625,y:.296875},{x:.796875,y:.296875},{x:.796875,y:.296875},{x:.828125,y:.296875},{x:.828125,y:.296875},{x:.859375,y:.296875},{x:.859375,y:.296875},{x:.890625,y:.296875},{x:.890625,y:.296875},{x:.921875,y:.296875},{x:.921875,y:.296875},{x:.953125,y:.296875},{x:.953125,y:.296875},{x:.984375,y:.296875},{x:.984375,y:.296875},{x:.015625,y:.328125},{x:.015625,y:.328125},{x:.046875,y:.328125},{x:.046875,y:.328125},{x:.078125,y:.328125},{x:.078125,y:.328125},{x:.109375,y:.328125},{x:.109375,y:.328125},{x:.140625,y:.328125},{x:.140625,y:.328125},{x:.171875,y:.328125},{x:.171875,y:.328125},{x:.203125,y:.328125},{x:.203125,y:.328125},{x:.234375,y:.328125},{x:.234375,y:.328125},{x:.265625,y:.328125},{x:.265625,y:.328125},{x:.296875,y:.328125},{x:.296875,y:.328125},{x:.328125,y:.328125},{x:.328125,y:.328125},{x:.359375,y:.328125},{x:.359375,y:.328125},{x:.390625,y:.328125},{x:.390625,y:.328125},{x:.421875,y:.328125},{x:.421875,y:.328125},{x:.453125,y:.328125},{x:.453125,y:.328125},{x:.484375,y:.328125},{x:.484375,y:.328125},{x:.515625,y:.328125},{x:.515625,y:.328125},{x:.546875,y:.328125},{x:.546875,y:.328125},{x:.578125,y:.328125},{x:.578125,y:.328125},{x:.609375,y:.328125},{x:.609375,y:.328125},{x:.640625,y:.328125},{x:.640625,y:.328125},{x:.671875,y:.328125},{x:.671875,y:.328125},{x:.703125,y:.328125},{x:.703125,y:.328125},{x:.734375,y:.328125},{x:.734375,y:.328125},{x:.765625,y:.328125},{x:.765625,y:.328125},{x:.796875,y:.328125},{x:.796875,y:.328125},{x:.828125,y:.328125},{x:.828125,y:.328125},{x:.859375,y:.328125},{x:.859375,y:.328125},{x:.890625,y:.328125},{x:.890625,y:.328125},{x:.921875,y:.328125},{x:.921875,y:.328125},{x:.953125,y:.328125},{x:.953125,y:.328125},{x:.984375,y:.328125},{x:.984375,y:.328125},{x:.015625,y:.359375},{x:.015625,y:.359375},{x:.046875,y:.359375},{x:.046875,y:.359375},{x:.078125,y:.359375},{x:.078125,y:.359375},{x:.109375,y:.359375},{x:.109375,y:.359375},{x:.140625,y:.359375},{x:.140625,y:.359375},{x:.171875,y:.359375},{x:.171875,y:.359375},{x:.203125,y:.359375},{x:.203125,y:.359375},{x:.234375,y:.359375},{x:.234375,y:.359375},{x:.265625,y:.359375},{x:.265625,y:.359375},{x:.296875,y:.359375},{x:.296875,y:.359375},{x:.328125,y:.359375},{x:.328125,y:.359375},{x:.359375,y:.359375},{x:.359375,y:.359375},{x:.390625,y:.359375},{x:.390625,y:.359375},{x:.421875,y:.359375},{x:.421875,y:.359375},{x:.453125,y:.359375},{x:.453125,y:.359375},{x:.484375,y:.359375},{x:.484375,y:.359375},{x:.515625,y:.359375},{x:.515625,y:.359375},{x:.546875,y:.359375},{x:.546875,y:.359375},{x:.578125,y:.359375},{x:.578125,y:.359375},{x:.609375,y:.359375},{x:.609375,y:.359375},{x:.640625,y:.359375},{x:.640625,y:.359375},{x:.671875,y:.359375},{x:.671875,y:.359375},{x:.703125,y:.359375},{x:.703125,y:.359375},{x:.734375,y:.359375},{x:.734375,y:.359375},{x:.765625,y:.359375},{x:.765625,y:.359375},{x:.796875,y:.359375},{x:.796875,y:.359375},{x:.828125,y:.359375},{x:.828125,y:.359375},{x:.859375,y:.359375},{x:.859375,y:.359375},{x:.890625,y:.359375},{x:.890625,y:.359375},{x:.921875,y:.359375},{x:.921875,y:.359375},{x:.953125,y:.359375},{x:.953125,y:.359375},{x:.984375,y:.359375},{x:.984375,y:.359375},{x:.015625,y:.390625},{x:.015625,y:.390625},{x:.046875,y:.390625},{x:.046875,y:.390625},{x:.078125,y:.390625},{x:.078125,y:.390625},{x:.109375,y:.390625},{x:.109375,y:.390625},{x:.140625,y:.390625},{x:.140625,y:.390625},{x:.171875,y:.390625},{x:.171875,y:.390625},{x:.203125,y:.390625},{x:.203125,y:.390625},{x:.234375,y:.390625},{x:.234375,y:.390625},{x:.265625,y:.390625},{x:.265625,y:.390625},{x:.296875,y:.390625},{x:.296875,y:.390625},{x:.328125,y:.390625},{x:.328125,y:.390625},{x:.359375,y:.390625},{x:.359375,y:.390625},{x:.390625,y:.390625},{x:.390625,y:.390625},{x:.421875,y:.390625},{x:.421875,y:.390625},{x:.453125,y:.390625},{x:.453125,y:.390625},{x:.484375,y:.390625},{x:.484375,y:.390625},{x:.515625,y:.390625},{x:.515625,y:.390625},{x:.546875,y:.390625},{x:.546875,y:.390625},{x:.578125,y:.390625},{x:.578125,y:.390625},{x:.609375,y:.390625},{x:.609375,y:.390625},{x:.640625,y:.390625},{x:.640625,y:.390625},{x:.671875,y:.390625},{x:.671875,y:.390625},{x:.703125,y:.390625},{x:.703125,y:.390625},{x:.734375,y:.390625},{x:.734375,y:.390625},{x:.765625,y:.390625},{x:.765625,y:.390625},{x:.796875,y:.390625},{x:.796875,y:.390625},{x:.828125,y:.390625},{x:.828125,y:.390625},{x:.859375,y:.390625},{x:.859375,y:.390625},{x:.890625,y:.390625},{x:.890625,y:.390625},{x:.921875,y:.390625},{x:.921875,y:.390625},{x:.953125,y:.390625},{x:.953125,y:.390625},{x:.984375,y:.390625},{x:.984375,y:.390625},{x:.015625,y:.421875},{x:.015625,y:.421875},{x:.046875,y:.421875},{x:.046875,y:.421875},{x:.078125,y:.421875},{x:.078125,y:.421875},{x:.109375,y:.421875},{x:.109375,y:.421875},{x:.140625,y:.421875},{x:.140625,y:.421875},{x:.171875,y:.421875},{x:.171875,y:.421875},{x:.203125,y:.421875},{x:.203125,y:.421875},{x:.234375,y:.421875},{x:.234375,y:.421875},{x:.265625,y:.421875},{x:.265625,y:.421875},{x:.296875,y:.421875},{x:.296875,y:.421875},{x:.328125,y:.421875},{x:.328125,y:.421875},{x:.359375,y:.421875},{x:.359375,y:.421875},{x:.390625,y:.421875},{x:.390625,y:.421875},{x:.421875,y:.421875},{x:.421875,y:.421875},{x:.453125,y:.421875},{x:.453125,y:.421875},{x:.484375,y:.421875},{x:.484375,y:.421875},{x:.515625,y:.421875},{x:.515625,y:.421875},{x:.546875,y:.421875},{x:.546875,y:.421875},{x:.578125,y:.421875},{x:.578125,y:.421875},{x:.609375,y:.421875},{x:.609375,y:.421875},{x:.640625,y:.421875},{x:.640625,y:.421875},{x:.671875,y:.421875},{x:.671875,y:.421875},{x:.703125,y:.421875},{x:.703125,y:.421875},{x:.734375,y:.421875},{x:.734375,y:.421875},{x:.765625,y:.421875},{x:.765625,y:.421875},{x:.796875,y:.421875},{x:.796875,y:.421875},{x:.828125,y:.421875},{x:.828125,y:.421875},{x:.859375,y:.421875},{x:.859375,y:.421875},{x:.890625,y:.421875},{x:.890625,y:.421875},{x:.921875,y:.421875},{x:.921875,y:.421875},{x:.953125,y:.421875},{x:.953125,y:.421875},{x:.984375,y:.421875},{x:.984375,y:.421875},{x:.015625,y:.453125},{x:.015625,y:.453125},{x:.046875,y:.453125},{x:.046875,y:.453125},{x:.078125,y:.453125},{x:.078125,y:.453125},{x:.109375,y:.453125},{x:.109375,y:.453125},{x:.140625,y:.453125},{x:.140625,y:.453125},{x:.171875,y:.453125},{x:.171875,y:.453125},{x:.203125,y:.453125},{x:.203125,y:.453125},{x:.234375,y:.453125},{x:.234375,y:.453125},{x:.265625,y:.453125},{x:.265625,y:.453125},{x:.296875,y:.453125},{x:.296875,y:.453125},{x:.328125,y:.453125},{x:.328125,y:.453125},{x:.359375,y:.453125},{x:.359375,y:.453125},{x:.390625,y:.453125},{x:.390625,y:.453125},{x:.421875,y:.453125},{x:.421875,y:.453125},{x:.453125,y:.453125},{x:.453125,y:.453125},{x:.484375,y:.453125},{x:.484375,y:.453125},{x:.515625,y:.453125},{x:.515625,y:.453125},{x:.546875,y:.453125},{x:.546875,y:.453125},{x:.578125,y:.453125},{x:.578125,y:.453125},{x:.609375,y:.453125},{x:.609375,y:.453125},{x:.640625,y:.453125},{x:.640625,y:.453125},{x:.671875,y:.453125},{x:.671875,y:.453125},{x:.703125,y:.453125},{x:.703125,y:.453125},{x:.734375,y:.453125},{x:.734375,y:.453125},{x:.765625,y:.453125},{x:.765625,y:.453125},{x:.796875,y:.453125},{x:.796875,y:.453125},{x:.828125,y:.453125},{x:.828125,y:.453125},{x:.859375,y:.453125},{x:.859375,y:.453125},{x:.890625,y:.453125},{x:.890625,y:.453125},{x:.921875,y:.453125},{x:.921875,y:.453125},{x:.953125,y:.453125},{x:.953125,y:.453125},{x:.984375,y:.453125},{x:.984375,y:.453125},{x:.015625,y:.484375},{x:.015625,y:.484375},{x:.046875,y:.484375},{x:.046875,y:.484375},{x:.078125,y:.484375},{x:.078125,y:.484375},{x:.109375,y:.484375},{x:.109375,y:.484375},{x:.140625,y:.484375},{x:.140625,y:.484375},{x:.171875,y:.484375},{x:.171875,y:.484375},{x:.203125,y:.484375},{x:.203125,y:.484375},{x:.234375,y:.484375},{x:.234375,y:.484375},{x:.265625,y:.484375},{x:.265625,y:.484375},{x:.296875,y:.484375},{x:.296875,y:.484375},{x:.328125,y:.484375},{x:.328125,y:.484375},{x:.359375,y:.484375},{x:.359375,y:.484375},{x:.390625,y:.484375},{x:.390625,y:.484375},{x:.421875,y:.484375},{x:.421875,y:.484375},{x:.453125,y:.484375},{x:.453125,y:.484375},{x:.484375,y:.484375},{x:.484375,y:.484375},{x:.515625,y:.484375},{x:.515625,y:.484375},{x:.546875,y:.484375},{x:.546875,y:.484375},{x:.578125,y:.484375},{x:.578125,y:.484375},{x:.609375,y:.484375},{x:.609375,y:.484375},{x:.640625,y:.484375},{x:.640625,y:.484375},{x:.671875,y:.484375},{x:.671875,y:.484375},{x:.703125,y:.484375},{x:.703125,y:.484375},{x:.734375,y:.484375},{x:.734375,y:.484375},{x:.765625,y:.484375},{x:.765625,y:.484375},{x:.796875,y:.484375},{x:.796875,y:.484375},{x:.828125,y:.484375},{x:.828125,y:.484375},{x:.859375,y:.484375},{x:.859375,y:.484375},{x:.890625,y:.484375},{x:.890625,y:.484375},{x:.921875,y:.484375},{x:.921875,y:.484375},{x:.953125,y:.484375},{x:.953125,y:.484375},{x:.984375,y:.484375},{x:.984375,y:.484375},{x:.015625,y:.515625},{x:.015625,y:.515625},{x:.046875,y:.515625},{x:.046875,y:.515625},{x:.078125,y:.515625},{x:.078125,y:.515625},{x:.109375,y:.515625},{x:.109375,y:.515625},{x:.140625,y:.515625},{x:.140625,y:.515625},{x:.171875,y:.515625},{x:.171875,y:.515625},{x:.203125,y:.515625},{x:.203125,y:.515625},{x:.234375,y:.515625},{x:.234375,y:.515625},{x:.265625,y:.515625},{x:.265625,y:.515625},{x:.296875,y:.515625},{x:.296875,y:.515625},{x:.328125,y:.515625},{x:.328125,y:.515625},{x:.359375,y:.515625},{x:.359375,y:.515625},{x:.390625,y:.515625},{x:.390625,y:.515625},{x:.421875,y:.515625},{x:.421875,y:.515625},{x:.453125,y:.515625},{x:.453125,y:.515625},{x:.484375,y:.515625},{x:.484375,y:.515625},{x:.515625,y:.515625},{x:.515625,y:.515625},{x:.546875,y:.515625},{x:.546875,y:.515625},{x:.578125,y:.515625},{x:.578125,y:.515625},{x:.609375,y:.515625},{x:.609375,y:.515625},{x:.640625,y:.515625},{x:.640625,y:.515625},{x:.671875,y:.515625},{x:.671875,y:.515625},{x:.703125,y:.515625},{x:.703125,y:.515625},{x:.734375,y:.515625},{x:.734375,y:.515625},{x:.765625,y:.515625},{x:.765625,y:.515625},{x:.796875,y:.515625},{x:.796875,y:.515625},{x:.828125,y:.515625},{x:.828125,y:.515625},{x:.859375,y:.515625},{x:.859375,y:.515625},{x:.890625,y:.515625},{x:.890625,y:.515625},{x:.921875,y:.515625},{x:.921875,y:.515625},{x:.953125,y:.515625},{x:.953125,y:.515625},{x:.984375,y:.515625},{x:.984375,y:.515625},{x:.015625,y:.546875},{x:.015625,y:.546875},{x:.046875,y:.546875},{x:.046875,y:.546875},{x:.078125,y:.546875},{x:.078125,y:.546875},{x:.109375,y:.546875},{x:.109375,y:.546875},{x:.140625,y:.546875},{x:.140625,y:.546875},{x:.171875,y:.546875},{x:.171875,y:.546875},{x:.203125,y:.546875},{x:.203125,y:.546875},{x:.234375,y:.546875},{x:.234375,y:.546875},{x:.265625,y:.546875},{x:.265625,y:.546875},{x:.296875,y:.546875},{x:.296875,y:.546875},{x:.328125,y:.546875},{x:.328125,y:.546875},{x:.359375,y:.546875},{x:.359375,y:.546875},{x:.390625,y:.546875},{x:.390625,y:.546875},{x:.421875,y:.546875},{x:.421875,y:.546875},{x:.453125,y:.546875},{x:.453125,y:.546875},{x:.484375,y:.546875},{x:.484375,y:.546875},{x:.515625,y:.546875},{x:.515625,y:.546875},{x:.546875,y:.546875},{x:.546875,y:.546875},{x:.578125,y:.546875},{x:.578125,y:.546875},{x:.609375,y:.546875},{x:.609375,y:.546875},{x:.640625,y:.546875},{x:.640625,y:.546875},{x:.671875,y:.546875},{x:.671875,y:.546875},{x:.703125,y:.546875},{x:.703125,y:.546875},{x:.734375,y:.546875},{x:.734375,y:.546875},{x:.765625,y:.546875},{x:.765625,y:.546875},{x:.796875,y:.546875},{x:.796875,y:.546875},{x:.828125,y:.546875},{x:.828125,y:.546875},{x:.859375,y:.546875},{x:.859375,y:.546875},{x:.890625,y:.546875},{x:.890625,y:.546875},{x:.921875,y:.546875},{x:.921875,y:.546875},{x:.953125,y:.546875},{x:.953125,y:.546875},{x:.984375,y:.546875},{x:.984375,y:.546875},{x:.015625,y:.578125},{x:.015625,y:.578125},{x:.046875,y:.578125},{x:.046875,y:.578125},{x:.078125,y:.578125},{x:.078125,y:.578125},{x:.109375,y:.578125},{x:.109375,y:.578125},{x:.140625,y:.578125},{x:.140625,y:.578125},{x:.171875,y:.578125},{x:.171875,y:.578125},{x:.203125,y:.578125},{x:.203125,y:.578125},{x:.234375,y:.578125},{x:.234375,y:.578125},{x:.265625,y:.578125},{x:.265625,y:.578125},{x:.296875,y:.578125},{x:.296875,y:.578125},{x:.328125,y:.578125},{x:.328125,y:.578125},{x:.359375,y:.578125},{x:.359375,y:.578125},{x:.390625,y:.578125},{x:.390625,y:.578125},{x:.421875,y:.578125},{x:.421875,y:.578125},{x:.453125,y:.578125},{x:.453125,y:.578125},{x:.484375,y:.578125},{x:.484375,y:.578125},{x:.515625,y:.578125},{x:.515625,y:.578125},{x:.546875,y:.578125},{x:.546875,y:.578125},{x:.578125,y:.578125},{x:.578125,y:.578125},{x:.609375,y:.578125},{x:.609375,y:.578125},{x:.640625,y:.578125},{x:.640625,y:.578125},{x:.671875,y:.578125},{x:.671875,y:.578125},{x:.703125,y:.578125},{x:.703125,y:.578125},{x:.734375,y:.578125},{x:.734375,y:.578125},{x:.765625,y:.578125},{x:.765625,y:.578125},{x:.796875,y:.578125},{x:.796875,y:.578125},{x:.828125,y:.578125},{x:.828125,y:.578125},{x:.859375,y:.578125},{x:.859375,y:.578125},{x:.890625,y:.578125},{x:.890625,y:.578125},{x:.921875,y:.578125},{x:.921875,y:.578125},{x:.953125,y:.578125},{x:.953125,y:.578125},{x:.984375,y:.578125},{x:.984375,y:.578125},{x:.015625,y:.609375},{x:.015625,y:.609375},{x:.046875,y:.609375},{x:.046875,y:.609375},{x:.078125,y:.609375},{x:.078125,y:.609375},{x:.109375,y:.609375},{x:.109375,y:.609375},{x:.140625,y:.609375},{x:.140625,y:.609375},{x:.171875,y:.609375},{x:.171875,y:.609375},{x:.203125,y:.609375},{x:.203125,y:.609375},{x:.234375,y:.609375},{x:.234375,y:.609375},{x:.265625,y:.609375},{x:.265625,y:.609375},{x:.296875,y:.609375},{x:.296875,y:.609375},{x:.328125,y:.609375},{x:.328125,y:.609375},{x:.359375,y:.609375},{x:.359375,y:.609375},{x:.390625,y:.609375},{x:.390625,y:.609375},{x:.421875,y:.609375},{x:.421875,y:.609375},{x:.453125,y:.609375},{x:.453125,y:.609375},{x:.484375,y:.609375},{x:.484375,y:.609375},{x:.515625,y:.609375},{x:.515625,y:.609375},{x:.546875,y:.609375},{x:.546875,y:.609375},{x:.578125,y:.609375},{x:.578125,y:.609375},{x:.609375,y:.609375},{x:.609375,y:.609375},{x:.640625,y:.609375},{x:.640625,y:.609375},{x:.671875,y:.609375},{x:.671875,y:.609375},{x:.703125,y:.609375},{x:.703125,y:.609375},{x:.734375,y:.609375},{x:.734375,y:.609375},{x:.765625,y:.609375},{x:.765625,y:.609375},{x:.796875,y:.609375},{x:.796875,y:.609375},{x:.828125,y:.609375},{x:.828125,y:.609375},{x:.859375,y:.609375},{x:.859375,y:.609375},{x:.890625,y:.609375},{x:.890625,y:.609375},{x:.921875,y:.609375},{x:.921875,y:.609375},{x:.953125,y:.609375},{x:.953125,y:.609375},{x:.984375,y:.609375},{x:.984375,y:.609375},{x:.015625,y:.640625},{x:.015625,y:.640625},{x:.046875,y:.640625},{x:.046875,y:.640625},{x:.078125,y:.640625},{x:.078125,y:.640625},{x:.109375,y:.640625},{x:.109375,y:.640625},{x:.140625,y:.640625},{x:.140625,y:.640625},{x:.171875,y:.640625},{x:.171875,y:.640625},{x:.203125,y:.640625},{x:.203125,y:.640625},{x:.234375,y:.640625},{x:.234375,y:.640625},{x:.265625,y:.640625},{x:.265625,y:.640625},{x:.296875,y:.640625},{x:.296875,y:.640625},{x:.328125,y:.640625},{x:.328125,y:.640625},{x:.359375,y:.640625},{x:.359375,y:.640625},{x:.390625,y:.640625},{x:.390625,y:.640625},{x:.421875,y:.640625},{x:.421875,y:.640625},{x:.453125,y:.640625},{x:.453125,y:.640625},{x:.484375,y:.640625},{x:.484375,y:.640625},{x:.515625,y:.640625},{x:.515625,y:.640625},{x:.546875,y:.640625},{x:.546875,y:.640625},{x:.578125,y:.640625},{x:.578125,y:.640625},{x:.609375,y:.640625},{x:.609375,y:.640625},{x:.640625,y:.640625},{x:.640625,y:.640625},{x:.671875,y:.640625},{x:.671875,y:.640625},{x:.703125,y:.640625},{x:.703125,y:.640625},{x:.734375,y:.640625},{x:.734375,y:.640625},{x:.765625,y:.640625},{x:.765625,y:.640625},{x:.796875,y:.640625},{x:.796875,y:.640625},{x:.828125,y:.640625},{x:.828125,y:.640625},{x:.859375,y:.640625},{x:.859375,y:.640625},{x:.890625,y:.640625},{x:.890625,y:.640625},{x:.921875,y:.640625},{x:.921875,y:.640625},{x:.953125,y:.640625},{x:.953125,y:.640625},{x:.984375,y:.640625},{x:.984375,y:.640625},{x:.015625,y:.671875},{x:.015625,y:.671875},{x:.046875,y:.671875},{x:.046875,y:.671875},{x:.078125,y:.671875},{x:.078125,y:.671875},{x:.109375,y:.671875},{x:.109375,y:.671875},{x:.140625,y:.671875},{x:.140625,y:.671875},{x:.171875,y:.671875},{x:.171875,y:.671875},{x:.203125,y:.671875},{x:.203125,y:.671875},{x:.234375,y:.671875},{x:.234375,y:.671875},{x:.265625,y:.671875},{x:.265625,y:.671875},{x:.296875,y:.671875},{x:.296875,y:.671875},{x:.328125,y:.671875},{x:.328125,y:.671875},{x:.359375,y:.671875},{x:.359375,y:.671875},{x:.390625,y:.671875},{x:.390625,y:.671875},{x:.421875,y:.671875},{x:.421875,y:.671875},{x:.453125,y:.671875},{x:.453125,y:.671875},{x:.484375,y:.671875},{x:.484375,y:.671875},{x:.515625,y:.671875},{x:.515625,y:.671875},{x:.546875,y:.671875},{x:.546875,y:.671875},{x:.578125,y:.671875},{x:.578125,y:.671875},{x:.609375,y:.671875},{x:.609375,y:.671875},{x:.640625,y:.671875},{x:.640625,y:.671875},{x:.671875,y:.671875},{x:.671875,y:.671875},{x:.703125,y:.671875},{x:.703125,y:.671875},{x:.734375,y:.671875},{x:.734375,y:.671875},{x:.765625,y:.671875},{x:.765625,y:.671875},{x:.796875,y:.671875},{x:.796875,y:.671875},{x:.828125,y:.671875},{x:.828125,y:.671875},{x:.859375,y:.671875},{x:.859375,y:.671875},{x:.890625,y:.671875},{x:.890625,y:.671875},{x:.921875,y:.671875},{x:.921875,y:.671875},{x:.953125,y:.671875},{x:.953125,y:.671875},{x:.984375,y:.671875},{x:.984375,y:.671875},{x:.015625,y:.703125},{x:.015625,y:.703125},{x:.046875,y:.703125},{x:.046875,y:.703125},{x:.078125,y:.703125},{x:.078125,y:.703125},{x:.109375,y:.703125},{x:.109375,y:.703125},{x:.140625,y:.703125},{x:.140625,y:.703125},{x:.171875,y:.703125},{x:.171875,y:.703125},{x:.203125,y:.703125},{x:.203125,y:.703125},{x:.234375,y:.703125},{x:.234375,y:.703125},{x:.265625,y:.703125},{x:.265625,y:.703125},{x:.296875,y:.703125},{x:.296875,y:.703125},{x:.328125,y:.703125},{x:.328125,y:.703125},{x:.359375,y:.703125},{x:.359375,y:.703125},{x:.390625,y:.703125},{x:.390625,y:.703125},{x:.421875,y:.703125},{x:.421875,y:.703125},{x:.453125,y:.703125},{x:.453125,y:.703125},{x:.484375,y:.703125},{x:.484375,y:.703125},{x:.515625,y:.703125},{x:.515625,y:.703125},{x:.546875,y:.703125},{x:.546875,y:.703125},{x:.578125,y:.703125},{x:.578125,y:.703125},{x:.609375,y:.703125},{x:.609375,y:.703125},{x:.640625,y:.703125},{x:.640625,y:.703125},{x:.671875,y:.703125},{x:.671875,y:.703125},{x:.703125,y:.703125},{x:.703125,y:.703125},{x:.734375,y:.703125},{x:.734375,y:.703125},{x:.765625,y:.703125},{x:.765625,y:.703125},{x:.796875,y:.703125},{x:.796875,y:.703125},{x:.828125,y:.703125},{x:.828125,y:.703125},{x:.859375,y:.703125},{x:.859375,y:.703125},{x:.890625,y:.703125},{x:.890625,y:.703125},{x:.921875,y:.703125},{x:.921875,y:.703125},{x:.953125,y:.703125},{x:.953125,y:.703125},{x:.984375,y:.703125},{x:.984375,y:.703125},{x:.015625,y:.734375},{x:.015625,y:.734375},{x:.046875,y:.734375},{x:.046875,y:.734375},{x:.078125,y:.734375},{x:.078125,y:.734375},{x:.109375,y:.734375},{x:.109375,y:.734375},{x:.140625,y:.734375},{x:.140625,y:.734375},{x:.171875,y:.734375},{x:.171875,y:.734375},{x:.203125,y:.734375},{x:.203125,y:.734375},{x:.234375,y:.734375},{x:.234375,y:.734375},{x:.265625,y:.734375},{x:.265625,y:.734375},{x:.296875,y:.734375},{x:.296875,y:.734375},{x:.328125,y:.734375},{x:.328125,y:.734375},{x:.359375,y:.734375},{x:.359375,y:.734375},{x:.390625,y:.734375},{x:.390625,y:.734375},{x:.421875,y:.734375},{x:.421875,y:.734375},{x:.453125,y:.734375},{x:.453125,y:.734375},{x:.484375,y:.734375},{x:.484375,y:.734375},{x:.515625,y:.734375},{x:.515625,y:.734375},{x:.546875,y:.734375},{x:.546875,y:.734375},{x:.578125,y:.734375},{x:.578125,y:.734375},{x:.609375,y:.734375},{x:.609375,y:.734375},{x:.640625,y:.734375},{x:.640625,y:.734375},{x:.671875,y:.734375},{x:.671875,y:.734375},{x:.703125,y:.734375},{x:.703125,y:.734375},{x:.734375,y:.734375},{x:.734375,y:.734375},{x:.765625,y:.734375},{x:.765625,y:.734375},{x:.796875,y:.734375},{x:.796875,y:.734375},{x:.828125,y:.734375},{x:.828125,y:.734375},{x:.859375,y:.734375},{x:.859375,y:.734375},{x:.890625,y:.734375},{x:.890625,y:.734375},{x:.921875,y:.734375},{x:.921875,y:.734375},{x:.953125,y:.734375},{x:.953125,y:.734375},{x:.984375,y:.734375},{x:.984375,y:.734375},{x:.015625,y:.765625},{x:.015625,y:.765625},{x:.046875,y:.765625},{x:.046875,y:.765625},{x:.078125,y:.765625},{x:.078125,y:.765625},{x:.109375,y:.765625},{x:.109375,y:.765625},{x:.140625,y:.765625},{x:.140625,y:.765625},{x:.171875,y:.765625},{x:.171875,y:.765625},{x:.203125,y:.765625},{x:.203125,y:.765625},{x:.234375,y:.765625},{x:.234375,y:.765625},{x:.265625,y:.765625},{x:.265625,y:.765625},{x:.296875,y:.765625},{x:.296875,y:.765625},{x:.328125,y:.765625},{x:.328125,y:.765625},{x:.359375,y:.765625},{x:.359375,y:.765625},{x:.390625,y:.765625},{x:.390625,y:.765625},{x:.421875,y:.765625},{x:.421875,y:.765625},{x:.453125,y:.765625},{x:.453125,y:.765625},{x:.484375,y:.765625},{x:.484375,y:.765625},{x:.515625,y:.765625},{x:.515625,y:.765625},{x:.546875,y:.765625},{x:.546875,y:.765625},{x:.578125,y:.765625},{x:.578125,y:.765625},{x:.609375,y:.765625},{x:.609375,y:.765625},{x:.640625,y:.765625},{x:.640625,y:.765625},{x:.671875,y:.765625},{x:.671875,y:.765625},{x:.703125,y:.765625},{x:.703125,y:.765625},{x:.734375,y:.765625},{x:.734375,y:.765625},{x:.765625,y:.765625},{x:.765625,y:.765625},{x:.796875,y:.765625},{x:.796875,y:.765625},{x:.828125,y:.765625},{x:.828125,y:.765625},{x:.859375,y:.765625},{x:.859375,y:.765625},{x:.890625,y:.765625},{x:.890625,y:.765625},{x:.921875,y:.765625},{x:.921875,y:.765625},{x:.953125,y:.765625},{x:.953125,y:.765625},{x:.984375,y:.765625},{x:.984375,y:.765625},{x:.015625,y:.796875},{x:.015625,y:.796875},{x:.046875,y:.796875},{x:.046875,y:.796875},{x:.078125,y:.796875},{x:.078125,y:.796875},{x:.109375,y:.796875},{x:.109375,y:.796875},{x:.140625,y:.796875},{x:.140625,y:.796875},{x:.171875,y:.796875},{x:.171875,y:.796875},{x:.203125,y:.796875},{x:.203125,y:.796875},{x:.234375,y:.796875},{x:.234375,y:.796875},{x:.265625,y:.796875},{x:.265625,y:.796875},{x:.296875,y:.796875},{x:.296875,y:.796875},{x:.328125,y:.796875},{x:.328125,y:.796875},{x:.359375,y:.796875},{x:.359375,y:.796875},{x:.390625,y:.796875},{x:.390625,y:.796875},{x:.421875,y:.796875},{x:.421875,y:.796875},{x:.453125,y:.796875},{x:.453125,y:.796875},{x:.484375,y:.796875},{x:.484375,y:.796875},{x:.515625,y:.796875},{x:.515625,y:.796875},{x:.546875,y:.796875},{x:.546875,y:.796875},{x:.578125,y:.796875},{x:.578125,y:.796875},{x:.609375,y:.796875},{x:.609375,y:.796875},{x:.640625,y:.796875},{x:.640625,y:.796875},{x:.671875,y:.796875},{x:.671875,y:.796875},{x:.703125,y:.796875},{x:.703125,y:.796875},{x:.734375,y:.796875},{x:.734375,y:.796875},{x:.765625,y:.796875},{x:.765625,y:.796875},{x:.796875,y:.796875},{x:.796875,y:.796875},{x:.828125,y:.796875},{x:.828125,y:.796875},{x:.859375,y:.796875},{x:.859375,y:.796875},{x:.890625,y:.796875},{x:.890625,y:.796875},{x:.921875,y:.796875},{x:.921875,y:.796875},{x:.953125,y:.796875},{x:.953125,y:.796875},{x:.984375,y:.796875},{x:.984375,y:.796875},{x:.015625,y:.828125},{x:.015625,y:.828125},{x:.046875,y:.828125},{x:.046875,y:.828125},{x:.078125,y:.828125},{x:.078125,y:.828125},{x:.109375,y:.828125},{x:.109375,y:.828125},{x:.140625,y:.828125},{x:.140625,y:.828125},{x:.171875,y:.828125},{x:.171875,y:.828125},{x:.203125,y:.828125},{x:.203125,y:.828125},{x:.234375,y:.828125},{x:.234375,y:.828125},{x:.265625,y:.828125},{x:.265625,y:.828125},{x:.296875,y:.828125},{x:.296875,y:.828125},{x:.328125,y:.828125},{x:.328125,y:.828125},{x:.359375,y:.828125},{x:.359375,y:.828125},{x:.390625,y:.828125},{x:.390625,y:.828125},{x:.421875,y:.828125},{x:.421875,y:.828125},{x:.453125,y:.828125},{x:.453125,y:.828125},{x:.484375,y:.828125},{x:.484375,y:.828125},{x:.515625,y:.828125},{x:.515625,y:.828125},{x:.546875,y:.828125},{x:.546875,y:.828125},{x:.578125,y:.828125},{x:.578125,y:.828125},{x:.609375,y:.828125},{x:.609375,y:.828125},{x:.640625,y:.828125},{x:.640625,y:.828125},{x:.671875,y:.828125},{x:.671875,y:.828125},{x:.703125,y:.828125},{x:.703125,y:.828125},{x:.734375,y:.828125},{x:.734375,y:.828125},{x:.765625,y:.828125},{x:.765625,y:.828125},{x:.796875,y:.828125},{x:.796875,y:.828125},{x:.828125,y:.828125},{x:.828125,y:.828125},{x:.859375,y:.828125},{x:.859375,y:.828125},{x:.890625,y:.828125},{x:.890625,y:.828125},{x:.921875,y:.828125},{x:.921875,y:.828125},{x:.953125,y:.828125},{x:.953125,y:.828125},{x:.984375,y:.828125},{x:.984375,y:.828125},{x:.015625,y:.859375},{x:.015625,y:.859375},{x:.046875,y:.859375},{x:.046875,y:.859375},{x:.078125,y:.859375},{x:.078125,y:.859375},{x:.109375,y:.859375},{x:.109375,y:.859375},{x:.140625,y:.859375},{x:.140625,y:.859375},{x:.171875,y:.859375},{x:.171875,y:.859375},{x:.203125,y:.859375},{x:.203125,y:.859375},{x:.234375,y:.859375},{x:.234375,y:.859375},{x:.265625,y:.859375},{x:.265625,y:.859375},{x:.296875,y:.859375},{x:.296875,y:.859375},{x:.328125,y:.859375},{x:.328125,y:.859375},{x:.359375,y:.859375},{x:.359375,y:.859375},{x:.390625,y:.859375},{x:.390625,y:.859375},{x:.421875,y:.859375},{x:.421875,y:.859375},{x:.453125,y:.859375},{x:.453125,y:.859375},{x:.484375,y:.859375},{x:.484375,y:.859375},{x:.515625,y:.859375},{x:.515625,y:.859375},{x:.546875,y:.859375},{x:.546875,y:.859375},{x:.578125,y:.859375},{x:.578125,y:.859375},{x:.609375,y:.859375},{x:.609375,y:.859375},{x:.640625,y:.859375},{x:.640625,y:.859375},{x:.671875,y:.859375},{x:.671875,y:.859375},{x:.703125,y:.859375},{x:.703125,y:.859375},{x:.734375,y:.859375},{x:.734375,y:.859375},{x:.765625,y:.859375},{x:.765625,y:.859375},{x:.796875,y:.859375},{x:.796875,y:.859375},{x:.828125,y:.859375},{x:.828125,y:.859375},{x:.859375,y:.859375},{x:.859375,y:.859375},{x:.890625,y:.859375},{x:.890625,y:.859375},{x:.921875,y:.859375},{x:.921875,y:.859375},{x:.953125,y:.859375},{x:.953125,y:.859375},{x:.984375,y:.859375},{x:.984375,y:.859375},{x:.015625,y:.890625},{x:.015625,y:.890625},{x:.046875,y:.890625},{x:.046875,y:.890625},{x:.078125,y:.890625},{x:.078125,y:.890625},{x:.109375,y:.890625},{x:.109375,y:.890625},{x:.140625,y:.890625},{x:.140625,y:.890625},{x:.171875,y:.890625},{x:.171875,y:.890625},{x:.203125,y:.890625},{x:.203125,y:.890625},{x:.234375,y:.890625},{x:.234375,y:.890625},{x:.265625,y:.890625},{x:.265625,y:.890625},{x:.296875,y:.890625},{x:.296875,y:.890625},{x:.328125,y:.890625},{x:.328125,y:.890625},{x:.359375,y:.890625},{x:.359375,y:.890625},{x:.390625,y:.890625},{x:.390625,y:.890625},{x:.421875,y:.890625},{x:.421875,y:.890625},{x:.453125,y:.890625},{x:.453125,y:.890625},{x:.484375,y:.890625},{x:.484375,y:.890625},{x:.515625,y:.890625},{x:.515625,y:.890625},{x:.546875,y:.890625},{x:.546875,y:.890625},{x:.578125,y:.890625},{x:.578125,y:.890625},{x:.609375,y:.890625},{x:.609375,y:.890625},{x:.640625,y:.890625},{x:.640625,y:.890625},{x:.671875,y:.890625},{x:.671875,y:.890625},{x:.703125,y:.890625},{x:.703125,y:.890625},{x:.734375,y:.890625},{x:.734375,y:.890625},{x:.765625,y:.890625},{x:.765625,y:.890625},{x:.796875,y:.890625},{x:.796875,y:.890625},{x:.828125,y:.890625},{x:.828125,y:.890625},{x:.859375,y:.890625},{x:.859375,y:.890625},{x:.890625,y:.890625},{x:.890625,y:.890625},{x:.921875,y:.890625},{x:.921875,y:.890625},{x:.953125,y:.890625},{x:.953125,y:.890625},{x:.984375,y:.890625},{x:.984375,y:.890625},{x:.015625,y:.921875},{x:.015625,y:.921875},{x:.046875,y:.921875},{x:.046875,y:.921875},{x:.078125,y:.921875},{x:.078125,y:.921875},{x:.109375,y:.921875},{x:.109375,y:.921875},{x:.140625,y:.921875},{x:.140625,y:.921875},{x:.171875,y:.921875},{x:.171875,y:.921875},{x:.203125,y:.921875},{x:.203125,y:.921875},{x:.234375,y:.921875},{x:.234375,y:.921875},{x:.265625,y:.921875},{x:.265625,y:.921875},{x:.296875,y:.921875},{x:.296875,y:.921875},{x:.328125,y:.921875},{x:.328125,y:.921875},{x:.359375,y:.921875},{x:.359375,y:.921875},{x:.390625,y:.921875},{x:.390625,y:.921875},{x:.421875,y:.921875},{x:.421875,y:.921875},{x:.453125,y:.921875},{x:.453125,y:.921875},{x:.484375,y:.921875},{x:.484375,y:.921875},{x:.515625,y:.921875},{x:.515625,y:.921875},{x:.546875,y:.921875},{x:.546875,y:.921875},{x:.578125,y:.921875},{x:.578125,y:.921875},{x:.609375,y:.921875},{x:.609375,y:.921875},{x:.640625,y:.921875},{x:.640625,y:.921875},{x:.671875,y:.921875},{x:.671875,y:.921875},{x:.703125,y:.921875},{x:.703125,y:.921875},{x:.734375,y:.921875},{x:.734375,y:.921875},{x:.765625,y:.921875},{x:.765625,y:.921875},{x:.796875,y:.921875},{x:.796875,y:.921875},{x:.828125,y:.921875},{x:.828125,y:.921875},{x:.859375,y:.921875},{x:.859375,y:.921875},{x:.890625,y:.921875},{x:.890625,y:.921875},{x:.921875,y:.921875},{x:.921875,y:.921875},{x:.953125,y:.921875},{x:.953125,y:.921875},{x:.984375,y:.921875},{x:.984375,y:.921875},{x:.015625,y:.953125},{x:.015625,y:.953125},{x:.046875,y:.953125},{x:.046875,y:.953125},{x:.078125,y:.953125},{x:.078125,y:.953125},{x:.109375,y:.953125},{x:.109375,y:.953125},{x:.140625,y:.953125},{x:.140625,y:.953125},{x:.171875,y:.953125},{x:.171875,y:.953125},{x:.203125,y:.953125},{x:.203125,y:.953125},{x:.234375,y:.953125},{x:.234375,y:.953125},{x:.265625,y:.953125},{x:.265625,y:.953125},{x:.296875,y:.953125},{x:.296875,y:.953125},{x:.328125,y:.953125},{x:.328125,y:.953125},{x:.359375,y:.953125},{x:.359375,y:.953125},{x:.390625,y:.953125},{x:.390625,y:.953125},{x:.421875,y:.953125},{x:.421875,y:.953125},{x:.453125,y:.953125},{x:.453125,y:.953125},{x:.484375,y:.953125},{x:.484375,y:.953125},{x:.515625,y:.953125},{x:.515625,y:.953125},{x:.546875,y:.953125},{x:.546875,y:.953125},{x:.578125,y:.953125},{x:.578125,y:.953125},{x:.609375,y:.953125},{x:.609375,y:.953125},{x:.640625,y:.953125},{x:.640625,y:.953125},{x:.671875,y:.953125},{x:.671875,y:.953125},{x:.703125,y:.953125},{x:.703125,y:.953125},{x:.734375,y:.953125},{x:.734375,y:.953125},{x:.765625,y:.953125},{x:.765625,y:.953125},{x:.796875,y:.953125},{x:.796875,y:.953125},{x:.828125,y:.953125},{x:.828125,y:.953125},{x:.859375,y:.953125},{x:.859375,y:.953125},{x:.890625,y:.953125},{x:.890625,y:.953125},{x:.921875,y:.953125},{x:.921875,y:.953125},{x:.953125,y:.953125},{x:.953125,y:.953125},{x:.984375,y:.953125},{x:.984375,y:.953125},{x:.015625,y:.984375},{x:.015625,y:.984375},{x:.046875,y:.984375},{x:.046875,y:.984375},{x:.078125,y:.984375},{x:.078125,y:.984375},{x:.109375,y:.984375},{x:.109375,y:.984375},{x:.140625,y:.984375},{x:.140625,y:.984375},{x:.171875,y:.984375},{x:.171875,y:.984375},{x:.203125,y:.984375},{x:.203125,y:.984375},{x:.234375,y:.984375},{x:.234375,y:.984375},{x:.265625,y:.984375},{x:.265625,y:.984375},{x:.296875,y:.984375},{x:.296875,y:.984375},{x:.328125,y:.984375},{x:.328125,y:.984375},{x:.359375,y:.984375},{x:.359375,y:.984375},{x:.390625,y:.984375},{x:.390625,y:.984375},{x:.421875,y:.984375},{x:.421875,y:.984375},{x:.453125,y:.984375},{x:.453125,y:.984375},{x:.484375,y:.984375},{x:.484375,y:.984375},{x:.515625,y:.984375},{x:.515625,y:.984375},{x:.546875,y:.984375},{x:.546875,y:.984375},{x:.578125,y:.984375},{x:.578125,y:.984375},{x:.609375,y:.984375},{x:.609375,y:.984375},{x:.640625,y:.984375},{x:.640625,y:.984375},{x:.671875,y:.984375},{x:.671875,y:.984375},{x:.703125,y:.984375},{x:.703125,y:.984375},{x:.734375,y:.984375},{x:.734375,y:.984375},{x:.765625,y:.984375},{x:.765625,y:.984375},{x:.796875,y:.984375},{x:.796875,y:.984375},{x:.828125,y:.984375},{x:.828125,y:.984375},{x:.859375,y:.984375},{x:.859375,y:.984375},{x:.890625,y:.984375},{x:.890625,y:.984375},{x:.921875,y:.984375},{x:.921875,y:.984375},{x:.953125,y:.984375},{x:.953125,y:.984375},{x:.984375,y:.984375},{x:.984375,y:.984375},{x:.03125,y:.03125},{x:.03125,y:.03125},{x:.09375,y:.03125},{x:.09375,y:.03125},{x:.15625,y:.03125},{x:.15625,y:.03125},{x:.21875,y:.03125},{x:.21875,y:.03125},{x:.28125,y:.03125},{x:.28125,y:.03125},{x:.34375,y:.03125},{x:.34375,y:.03125},{x:.40625,y:.03125},{x:.40625,y:.03125},{x:.46875,y:.03125},{x:.46875,y:.03125},{x:.53125,y:.03125},{x:.53125,y:.03125},{x:.59375,y:.03125},{x:.59375,y:.03125},{x:.65625,y:.03125},{x:.65625,y:.03125},{x:.71875,y:.03125},{x:.71875,y:.03125},{x:.78125,y:.03125},{x:.78125,y:.03125},{x:.84375,y:.03125},{x:.84375,y:.03125},{x:.90625,y:.03125},{x:.90625,y:.03125},{x:.96875,y:.03125},{x:.96875,y:.03125},{x:.03125,y:.09375},{x:.03125,y:.09375},{x:.09375,y:.09375},{x:.09375,y:.09375},{x:.15625,y:.09375},{x:.15625,y:.09375},{x:.21875,y:.09375},{x:.21875,y:.09375},{x:.28125,y:.09375},{x:.28125,y:.09375},{x:.34375,y:.09375},{x:.34375,y:.09375},{x:.40625,y:.09375},{x:.40625,y:.09375},{x:.46875,y:.09375},{x:.46875,y:.09375},{x:.53125,y:.09375},{x:.53125,y:.09375},{x:.59375,y:.09375},{x:.59375,y:.09375},{x:.65625,y:.09375},{x:.65625,y:.09375},{x:.71875,y:.09375},{x:.71875,y:.09375},{x:.78125,y:.09375},{x:.78125,y:.09375},{x:.84375,y:.09375},{x:.84375,y:.09375},{x:.90625,y:.09375},{x:.90625,y:.09375},{x:.96875,y:.09375},{x:.96875,y:.09375},{x:.03125,y:.15625},{x:.03125,y:.15625},{x:.09375,y:.15625},{x:.09375,y:.15625},{x:.15625,y:.15625},{x:.15625,y:.15625},{x:.21875,y:.15625},{x:.21875,y:.15625},{x:.28125,y:.15625},{x:.28125,y:.15625},{x:.34375,y:.15625},{x:.34375,y:.15625},{x:.40625,y:.15625},{x:.40625,y:.15625},{x:.46875,y:.15625},{x:.46875,y:.15625},{x:.53125,y:.15625},{x:.53125,y:.15625},{x:.59375,y:.15625},{x:.59375,y:.15625},{x:.65625,y:.15625},{x:.65625,y:.15625},{x:.71875,y:.15625},{x:.71875,y:.15625},{x:.78125,y:.15625},{x:.78125,y:.15625},{x:.84375,y:.15625},{x:.84375,y:.15625},{x:.90625,y:.15625},{x:.90625,y:.15625},{x:.96875,y:.15625},{x:.96875,y:.15625},{x:.03125,y:.21875},{x:.03125,y:.21875},{x:.09375,y:.21875},{x:.09375,y:.21875},{x:.15625,y:.21875},{x:.15625,y:.21875},{x:.21875,y:.21875},{x:.21875,y:.21875},{x:.28125,y:.21875},{x:.28125,y:.21875},{x:.34375,y:.21875},{x:.34375,y:.21875},{x:.40625,y:.21875},{x:.40625,y:.21875},{x:.46875,y:.21875},{x:.46875,y:.21875},{x:.53125,y:.21875},{x:.53125,y:.21875},{x:.59375,y:.21875},{x:.59375,y:.21875},{x:.65625,y:.21875},{x:.65625,y:.21875},{x:.71875,y:.21875},{x:.71875,y:.21875},{x:.78125,y:.21875},{x:.78125,y:.21875},{x:.84375,y:.21875},{x:.84375,y:.21875},{x:.90625,y:.21875},{x:.90625,y:.21875},{x:.96875,y:.21875},{x:.96875,y:.21875},{x:.03125,y:.28125},{x:.03125,y:.28125},{x:.09375,y:.28125},{x:.09375,y:.28125},{x:.15625,y:.28125},{x:.15625,y:.28125},{x:.21875,y:.28125},{x:.21875,y:.28125},{x:.28125,y:.28125},{x:.28125,y:.28125},{x:.34375,y:.28125},{x:.34375,y:.28125},{x:.40625,y:.28125},{x:.40625,y:.28125},{x:.46875,y:.28125},{x:.46875,y:.28125},{x:.53125,y:.28125},{x:.53125,y:.28125},{x:.59375,y:.28125},{x:.59375,y:.28125},{x:.65625,y:.28125},{x:.65625,y:.28125},{x:.71875,y:.28125},{x:.71875,y:.28125},{x:.78125,y:.28125},{x:.78125,y:.28125},{x:.84375,y:.28125},{x:.84375,y:.28125},{x:.90625,y:.28125},{x:.90625,y:.28125},{x:.96875,y:.28125},{x:.96875,y:.28125},{x:.03125,y:.34375},{x:.03125,y:.34375},{x:.09375,y:.34375},{x:.09375,y:.34375},{x:.15625,y:.34375},{x:.15625,y:.34375},{x:.21875,y:.34375},{x:.21875,y:.34375},{x:.28125,y:.34375},{x:.28125,y:.34375},{x:.34375,y:.34375},{x:.34375,y:.34375},{x:.40625,y:.34375},{x:.40625,y:.34375},{x:.46875,y:.34375},{x:.46875,y:.34375},{x:.53125,y:.34375},{x:.53125,y:.34375},{x:.59375,y:.34375},{x:.59375,y:.34375},{x:.65625,y:.34375},{x:.65625,y:.34375},{x:.71875,y:.34375},{x:.71875,y:.34375},{x:.78125,y:.34375},{x:.78125,y:.34375},{x:.84375,y:.34375},{x:.84375,y:.34375},{x:.90625,y:.34375},{x:.90625,y:.34375},{x:.96875,y:.34375},{x:.96875,y:.34375},{x:.03125,y:.40625},{x:.03125,y:.40625},{x:.09375,y:.40625},{x:.09375,y:.40625},{x:.15625,y:.40625},{x:.15625,y:.40625},{x:.21875,y:.40625},{x:.21875,y:.40625},{x:.28125,y:.40625},{x:.28125,y:.40625},{x:.34375,y:.40625},{x:.34375,y:.40625},{x:.40625,y:.40625},{x:.40625,y:.40625},{x:.46875,y:.40625},{x:.46875,y:.40625},{x:.53125,y:.40625},{x:.53125,y:.40625},{x:.59375,y:.40625},{x:.59375,y:.40625},{x:.65625,y:.40625},{x:.65625,y:.40625},{x:.71875,y:.40625},{x:.71875,y:.40625},{x:.78125,y:.40625},{x:.78125,y:.40625},{x:.84375,y:.40625},{x:.84375,y:.40625},{x:.90625,y:.40625},{x:.90625,y:.40625},{x:.96875,y:.40625},{x:.96875,y:.40625},{x:.03125,y:.46875},{x:.03125,y:.46875},{x:.09375,y:.46875},{x:.09375,y:.46875},{x:.15625,y:.46875},{x:.15625,y:.46875},{x:.21875,y:.46875},{x:.21875,y:.46875},{x:.28125,y:.46875},{x:.28125,y:.46875},{x:.34375,y:.46875},{x:.34375,y:.46875},{x:.40625,y:.46875},{x:.40625,y:.46875},{x:.46875,y:.46875},{x:.46875,y:.46875},{x:.53125,y:.46875},{x:.53125,y:.46875},{x:.59375,y:.46875},{x:.59375,y:.46875},{x:.65625,y:.46875},{x:.65625,y:.46875},{x:.71875,y:.46875},{x:.71875,y:.46875},{x:.78125,y:.46875},{x:.78125,y:.46875},{x:.84375,y:.46875},{x:.84375,y:.46875},{x:.90625,y:.46875},{x:.90625,y:.46875},{x:.96875,y:.46875},{x:.96875,y:.46875},{x:.03125,y:.53125},{x:.03125,y:.53125},{x:.09375,y:.53125},{x:.09375,y:.53125},{x:.15625,y:.53125},{x:.15625,y:.53125},{x:.21875,y:.53125},{x:.21875,y:.53125},{x:.28125,y:.53125},{x:.28125,y:.53125},{x:.34375,y:.53125},{x:.34375,y:.53125},{x:.40625,y:.53125},{x:.40625,y:.53125},{x:.46875,y:.53125},{x:.46875,y:.53125},{x:.53125,y:.53125},{x:.53125,y:.53125},{x:.59375,y:.53125},{x:.59375,y:.53125},{x:.65625,y:.53125},{x:.65625,y:.53125},{x:.71875,y:.53125},{x:.71875,y:.53125},{x:.78125,y:.53125},{x:.78125,y:.53125},{x:.84375,y:.53125},{x:.84375,y:.53125},{x:.90625,y:.53125},{x:.90625,y:.53125},{x:.96875,y:.53125},{x:.96875,y:.53125},{x:.03125,y:.59375},{x:.03125,y:.59375},{x:.09375,y:.59375},{x:.09375,y:.59375},{x:.15625,y:.59375},{x:.15625,y:.59375},{x:.21875,y:.59375},{x:.21875,y:.59375},{x:.28125,y:.59375},{x:.28125,y:.59375},{x:.34375,y:.59375},{x:.34375,y:.59375},{x:.40625,y:.59375},{x:.40625,y:.59375},{x:.46875,y:.59375},{x:.46875,y:.59375},{x:.53125,y:.59375},{x:.53125,y:.59375},{x:.59375,y:.59375},{x:.59375,y:.59375},{x:.65625,y:.59375},{x:.65625,y:.59375},{x:.71875,y:.59375},{x:.71875,y:.59375},{x:.78125,y:.59375},{x:.78125,y:.59375},{x:.84375,y:.59375},{x:.84375,y:.59375},{x:.90625,y:.59375},{x:.90625,y:.59375},{x:.96875,y:.59375},{x:.96875,y:.59375},{x:.03125,y:.65625},{x:.03125,y:.65625},{x:.09375,y:.65625},{x:.09375,y:.65625},{x:.15625,y:.65625},{x:.15625,y:.65625},{x:.21875,y:.65625},{x:.21875,y:.65625},{x:.28125,y:.65625},{x:.28125,y:.65625},{x:.34375,y:.65625},{x:.34375,y:.65625},{x:.40625,y:.65625},{x:.40625,y:.65625},{x:.46875,y:.65625},{x:.46875,y:.65625},{x:.53125,y:.65625},{x:.53125,y:.65625},{x:.59375,y:.65625},{x:.59375,y:.65625},{x:.65625,y:.65625},{x:.65625,y:.65625},{x:.71875,y:.65625},{x:.71875,y:.65625},{x:.78125,y:.65625},{x:.78125,y:.65625},{x:.84375,y:.65625},{x:.84375,y:.65625},{x:.90625,y:.65625},{x:.90625,y:.65625},{x:.96875,y:.65625},{x:.96875,y:.65625},{x:.03125,y:.71875},{x:.03125,y:.71875},{x:.09375,y:.71875},{x:.09375,y:.71875},{x:.15625,y:.71875},{x:.15625,y:.71875},{x:.21875,y:.71875},{x:.21875,y:.71875},{x:.28125,y:.71875},{x:.28125,y:.71875},{x:.34375,y:.71875},{x:.34375,y:.71875},{x:.40625,y:.71875},{x:.40625,y:.71875},{x:.46875,y:.71875},{x:.46875,y:.71875},{x:.53125,y:.71875},{x:.53125,y:.71875},{x:.59375,y:.71875},{x:.59375,y:.71875},{x:.65625,y:.71875},{x:.65625,y:.71875},{x:.71875,y:.71875},{x:.71875,y:.71875},{x:.78125,y:.71875},{x:.78125,y:.71875},{x:.84375,y:.71875},{x:.84375,y:.71875},{x:.90625,y:.71875},{x:.90625,y:.71875},{x:.96875,y:.71875},{x:.96875,y:.71875},{x:.03125,y:.78125},{x:.03125,y:.78125},{x:.09375,y:.78125},{x:.09375,y:.78125},{x:.15625,y:.78125},{x:.15625,y:.78125},{x:.21875,y:.78125},{x:.21875,y:.78125},{x:.28125,y:.78125},{x:.28125,y:.78125},{x:.34375,y:.78125},{x:.34375,y:.78125},{x:.40625,y:.78125},{x:.40625,y:.78125},{x:.46875,y:.78125},{x:.46875,y:.78125},{x:.53125,y:.78125},{x:.53125,y:.78125},{x:.59375,y:.78125},{x:.59375,y:.78125},{x:.65625,y:.78125},{x:.65625,y:.78125},{x:.71875,y:.78125},{x:.71875,y:.78125},{x:.78125,y:.78125},{x:.78125,y:.78125},{x:.84375,y:.78125},{x:.84375,y:.78125},{x:.90625,y:.78125},{x:.90625,y:.78125},{x:.96875,y:.78125},{x:.96875,y:.78125},{x:.03125,y:.84375},{x:.03125,y:.84375},{x:.09375,y:.84375},{x:.09375,y:.84375},{x:.15625,y:.84375},{x:.15625,y:.84375},{x:.21875,y:.84375},{x:.21875,y:.84375},{x:.28125,y:.84375},{x:.28125,y:.84375},{x:.34375,y:.84375},{x:.34375,y:.84375},{x:.40625,y:.84375},{x:.40625,y:.84375},{x:.46875,y:.84375},{x:.46875,y:.84375},{x:.53125,y:.84375},{x:.53125,y:.84375},{x:.59375,y:.84375},{x:.59375,y:.84375},{x:.65625,y:.84375},{x:.65625,y:.84375},{x:.71875,y:.84375},{x:.71875,y:.84375},{x:.78125,y:.84375},{x:.78125,y:.84375},{x:.84375,y:.84375},{x:.84375,y:.84375},{x:.90625,y:.84375},{x:.90625,y:.84375},{x:.96875,y:.84375},{x:.96875,y:.84375},{x:.03125,y:.90625},{x:.03125,y:.90625},{x:.09375,y:.90625},{x:.09375,y:.90625},{x:.15625,y:.90625},{x:.15625,y:.90625},{x:.21875,y:.90625},{x:.21875,y:.90625},{x:.28125,y:.90625},{x:.28125,y:.90625},{x:.34375,y:.90625},{x:.34375,y:.90625},{x:.40625,y:.90625},{x:.40625,y:.90625},{x:.46875,y:.90625},{x:.46875,y:.90625},{x:.53125,y:.90625},{x:.53125,y:.90625},{x:.59375,y:.90625},{x:.59375,y:.90625},{x:.65625,y:.90625},{x:.65625,y:.90625},{x:.71875,y:.90625},{x:.71875,y:.90625},{x:.78125,y:.90625},{x:.78125,y:.90625},{x:.84375,y:.90625},{x:.84375,y:.90625},{x:.90625,y:.90625},{x:.90625,y:.90625},{x:.96875,y:.90625},{x:.96875,y:.90625},{x:.03125,y:.96875},{x:.03125,y:.96875},{x:.09375,y:.96875},{x:.09375,y:.96875},{x:.15625,y:.96875},{x:.15625,y:.96875},{x:.21875,y:.96875},{x:.21875,y:.96875},{x:.28125,y:.96875},{x:.28125,y:.96875},{x:.34375,y:.96875},{x:.34375,y:.96875},{x:.40625,y:.96875},{x:.40625,y:.96875},{x:.46875,y:.96875},{x:.46875,y:.96875},{x:.53125,y:.96875},{x:.53125,y:.96875},{x:.59375,y:.96875},{x:.59375,y:.96875},{x:.65625,y:.96875},{x:.65625,y:.96875},{x:.71875,y:.96875},{x:.71875,y:.96875},{x:.78125,y:.96875},{x:.78125,y:.96875},{x:.84375,y:.96875},{x:.84375,y:.96875},{x:.90625,y:.96875},{x:.90625,y:.96875},{x:.96875,y:.96875},{x:.96875,y:.96875},{x:.0625,y:.0625},{x:.0625,y:.0625},{x:.0625,y:.0625},{x:.0625,y:.0625},{x:.0625,y:.0625},{x:.0625,y:.0625},{x:.1875,y:.0625},{x:.1875,y:.0625},{x:.1875,y:.0625},{x:.1875,y:.0625},{x:.1875,y:.0625},{x:.1875,y:.0625},{x:.3125,y:.0625},{x:.3125,y:.0625},{x:.3125,y:.0625},{x:.3125,y:.0625},{x:.3125,y:.0625},{x:.3125,y:.0625},{x:.4375,y:.0625},{x:.4375,y:.0625},{x:.4375,y:.0625},{x:.4375,y:.0625},{x:.4375,y:.0625},{x:.4375,y:.0625},{x:.5625,y:.0625},{x:.5625,y:.0625},{x:.5625,y:.0625},{x:.5625,y:.0625},{x:.5625,y:.0625},{x:.5625,y:.0625},{x:.6875,y:.0625},{x:.6875,y:.0625},{x:.6875,y:.0625},{x:.6875,y:.0625},{x:.6875,y:.0625},{x:.6875,y:.0625},{x:.8125,y:.0625},{x:.8125,y:.0625},{x:.8125,y:.0625},{x:.8125,y:.0625},{x:.8125,y:.0625},{x:.8125,y:.0625},{x:.9375,y:.0625},{x:.9375,y:.0625},{x:.9375,y:.0625},{x:.9375,y:.0625},{x:.9375,y:.0625},{x:.9375,y:.0625},{x:.0625,y:.1875},{x:.0625,y:.1875},{x:.0625,y:.1875},{x:.0625,y:.1875},{x:.0625,y:.1875},{x:.0625,y:.1875},{x:.1875,y:.1875},{x:.1875,y:.1875},{x:.1875,y:.1875},{x:.1875,y:.1875},{x:.1875,y:.1875},{x:.1875,y:.1875},{x:.3125,y:.1875},{x:.3125,y:.1875},{x:.3125,y:.1875},{x:.3125,y:.1875},{x:.3125,y:.1875},{x:.3125,y:.1875},{x:.4375,y:.1875},{x:.4375,y:.1875},{x:.4375,y:.1875},{x:.4375,y:.1875},{x:.4375,y:.1875},{x:.4375,y:.1875},{x:.5625,y:.1875},{x:.5625,y:.1875},{x:.5625,y:.1875},{x:.5625,y:.1875},{x:.5625,y:.1875},{x:.5625,y:.1875},{x:.6875,y:.1875},{x:.6875,y:.1875},{x:.6875,y:.1875},{x:.6875,y:.1875},{x:.6875,y:.1875},{x:.6875,y:.1875},{x:.8125,y:.1875},{x:.8125,y:.1875},{x:.8125,y:.1875},{x:.8125,y:.1875},{x:.8125,y:.1875},{x:.8125,y:.1875},{x:.9375,y:.1875},{x:.9375,y:.1875},{x:.9375,y:.1875},{x:.9375,y:.1875},{x:.9375,y:.1875},{x:.9375,y:.1875},{x:.0625,y:.3125},{x:.0625,y:.3125},{x:.0625,y:.3125},{x:.0625,y:.3125},{x:.0625,y:.3125},{x:.0625,y:.3125},{x:.1875,y:.3125},{x:.1875,y:.3125},{x:.1875,y:.3125},{x:.1875,y:.3125},{x:.1875,y:.3125},{x:.1875,y:.3125},{x:.3125,y:.3125},{x:.3125,y:.3125},{x:.3125,y:.3125},{x:.3125,y:.3125},{x:.3125,y:.3125},{x:.3125,y:.3125},{x:.4375,y:.3125},{x:.4375,y:.3125},{x:.4375,y:.3125},{x:.4375,y:.3125},{x:.4375,y:.3125},{x:.4375,y:.3125},{x:.5625,y:.3125},{x:.5625,y:.3125},{x:.5625,y:.3125},{x:.5625,y:.3125},{x:.5625,y:.3125},{x:.5625,y:.3125},{x:.6875,y:.3125},{x:.6875,y:.3125},{x:.6875,y:.3125},{x:.6875,y:.3125},{x:.6875,y:.3125},{x:.6875,y:.3125},{x:.8125,y:.3125},{x:.8125,y:.3125},{x:.8125,y:.3125},{x:.8125,y:.3125},{x:.8125,y:.3125},{x:.8125,y:.3125},{x:.9375,y:.3125},{x:.9375,y:.3125},{x:.9375,y:.3125},{x:.9375,y:.3125},{x:.9375,y:.3125},{x:.9375,y:.3125},{x:.0625,y:.4375},{x:.0625,y:.4375},{x:.0625,y:.4375},{x:.0625,y:.4375},{x:.0625,y:.4375},{x:.0625,y:.4375},{x:.1875,y:.4375},{x:.1875,y:.4375},{x:.1875,y:.4375},{x:.1875,y:.4375},{x:.1875,y:.4375},{x:.1875,y:.4375},{x:.3125,y:.4375},{x:.3125,y:.4375},{x:.3125,y:.4375},{x:.3125,y:.4375},{x:.3125,y:.4375},{x:.3125,y:.4375},{x:.4375,y:.4375},{x:.4375,y:.4375},{x:.4375,y:.4375},{x:.4375,y:.4375},{x:.4375,y:.4375},{x:.4375,y:.4375},{x:.5625,y:.4375},{x:.5625,y:.4375},{x:.5625,y:.4375},{x:.5625,y:.4375},{x:.5625,y:.4375},{x:.5625,y:.4375},{x:.6875,y:.4375},{x:.6875,y:.4375},{x:.6875,y:.4375},{x:.6875,y:.4375},{x:.6875,y:.4375},{x:.6875,y:.4375},{x:.8125,y:.4375},{x:.8125,y:.4375},{x:.8125,y:.4375},{x:.8125,y:.4375},{x:.8125,y:.4375},{x:.8125,y:.4375},{x:.9375,y:.4375},{x:.9375,y:.4375},{x:.9375,y:.4375},{x:.9375,y:.4375},{x:.9375,y:.4375},{x:.9375,y:.4375},{x:.0625,y:.5625},{x:.0625,y:.5625},{x:.0625,y:.5625},{x:.0625,y:.5625},{x:.0625,y:.5625},{x:.0625,y:.5625},{x:.1875,y:.5625},{x:.1875,y:.5625},{x:.1875,y:.5625},{x:.1875,y:.5625},{x:.1875,y:.5625},{x:.1875,y:.5625},{x:.3125,y:.5625},{x:.3125,y:.5625},{x:.3125,y:.5625},{x:.3125,y:.5625},{x:.3125,y:.5625},{x:.3125,y:.5625},{x:.4375,y:.5625},{x:.4375,y:.5625},{x:.4375,y:.5625},{x:.4375,y:.5625},{x:.4375,y:.5625},{x:.4375,y:.5625},{x:.5625,y:.5625},{x:.5625,y:.5625},{x:.5625,y:.5625},{x:.5625,y:.5625},{x:.5625,y:.5625},{x:.5625,y:.5625},{x:.6875,y:.5625},{x:.6875,y:.5625},{x:.6875,y:.5625},{x:.6875,y:.5625},{x:.6875,y:.5625},{x:.6875,y:.5625},{x:.8125,y:.5625},{x:.8125,y:.5625},{x:.8125,y:.5625},{x:.8125,y:.5625},{x:.8125,y:.5625},{x:.8125,y:.5625},{x:.9375,y:.5625},{x:.9375,y:.5625},{x:.9375,y:.5625},{x:.9375,y:.5625},{x:.9375,y:.5625},{x:.9375,y:.5625},{x:.0625,y:.6875},{x:.0625,y:.6875},{x:.0625,y:.6875},{x:.0625,y:.6875},{x:.0625,y:.6875},{x:.0625,y:.6875},{x:.1875,y:.6875},{x:.1875,y:.6875},{x:.1875,y:.6875},{x:.1875,y:.6875},{x:.1875,y:.6875},{x:.1875,y:.6875},{x:.3125,y:.6875},{x:.3125,y:.6875},{x:.3125,y:.6875},{x:.3125,y:.6875},{x:.3125,y:.6875},{x:.3125,y:.6875},{x:.4375,y:.6875},{x:.4375,y:.6875},{x:.4375,y:.6875},{x:.4375,y:.6875},{x:.4375,y:.6875},{x:.4375,y:.6875},{x:.5625,y:.6875},{x:.5625,y:.6875},{x:.5625,y:.6875},{x:.5625,y:.6875},{x:.5625,y:.6875},{x:.5625,y:.6875},{x:.6875,y:.6875},{x:.6875,y:.6875},{x:.6875,y:.6875},{x:.6875,y:.6875},{x:.6875,y:.6875},{x:.6875,y:.6875},{x:.8125,y:.6875},{x:.8125,y:.6875},{x:.8125,y:.6875},{x:.8125,y:.6875},{x:.8125,y:.6875},{x:.8125,y:.6875},{x:.9375,y:.6875},{x:.9375,y:.6875},{x:.9375,y:.6875},{x:.9375,y:.6875},{x:.9375,y:.6875},{x:.9375,y:.6875},{x:.0625,y:.8125},{x:.0625,y:.8125},{x:.0625,y:.8125},{x:.0625,y:.8125},{x:.0625,y:.8125},{x:.0625,y:.8125},{x:.1875,y:.8125},{x:.1875,y:.8125},{x:.1875,y:.8125},{x:.1875,y:.8125},{x:.1875,y:.8125},{x:.1875,y:.8125},{x:.3125,y:.8125},{x:.3125,y:.8125},{x:.3125,y:.8125},{x:.3125,y:.8125},{x:.3125,y:.8125},{x:.3125,y:.8125},{x:.4375,y:.8125},{x:.4375,y:.8125},{x:.4375,y:.8125},{x:.4375,y:.8125},{x:.4375,y:.8125},{x:.4375,y:.8125},{x:.5625,y:.8125},{x:.5625,y:.8125},{x:.5625,y:.8125},{x:.5625,y:.8125},{x:.5625,y:.8125},{x:.5625,y:.8125},{x:.6875,y:.8125},{x:.6875,y:.8125},{x:.6875,y:.8125},{x:.6875,y:.8125},{x:.6875,y:.8125},{x:.6875,y:.8125},{x:.8125,y:.8125},{x:.8125,y:.8125},{x:.8125,y:.8125},{x:.8125,y:.8125},{x:.8125,y:.8125},{x:.8125,y:.8125},{x:.9375,y:.8125},{x:.9375,y:.8125},{x:.9375,y:.8125},{x:.9375,y:.8125},{x:.9375,y:.8125},{x:.9375,y:.8125},{x:.0625,y:.9375},{x:.0625,y:.9375},{x:.0625,y:.9375},{x:.0625,y:.9375},{x:.0625,y:.9375},{x:.0625,y:.9375},{x:.1875,y:.9375},{x:.1875,y:.9375},{x:.1875,y:.9375},{x:.1875,y:.9375},{x:.1875,y:.9375},{x:.1875,y:.9375},{x:.3125,y:.9375},{x:.3125,y:.9375},{x:.3125,y:.9375},{x:.3125,y:.9375},{x:.3125,y:.9375},{x:.3125,y:.9375},{x:.4375,y:.9375},{x:.4375,y:.9375},{x:.4375,y:.9375},{x:.4375,y:.9375},{x:.4375,y:.9375},{x:.4375,y:.9375},{x:.5625,y:.9375},{x:.5625,y:.9375},{x:.5625,y:.9375},{x:.5625,y:.9375},{x:.5625,y:.9375},{x:.5625,y:.9375},{x:.6875,y:.9375},{x:.6875,y:.9375},{x:.6875,y:.9375},{x:.6875,y:.9375},{x:.6875,y:.9375},{x:.6875,y:.9375},{x:.8125,y:.9375},{x:.8125,y:.9375},{x:.8125,y:.9375},{x:.8125,y:.9375},{x:.8125,y:.9375},{x:.8125,y:.9375},{x:.9375,y:.9375},{x:.9375,y:.9375},{x:.9375,y:.9375},{x:.9375,y:.9375},{x:.9375,y:.9375},{x:.9375,y:.9375}];var G0=class{constructor(t){he(this,"model");he(this,"anchors");he(this,"anchorsTensor");he(this,"inputSize");he(this,"inputSizeTensor");he(this,"doubleInputSizeTensor");var a,n,r,s;this.model=t,this.anchors=NI.map(i=>[i.x,i.y]),this.anchorsTensor=Zn(this.anchors),this.inputSize=((s=(r=(n=(a=this==null?void 0:this.model)==null?void 0:a.inputs)==null?void 0:n[0])==null?void 0:r.shape)==null?void 0:s[2])||0,this.inputSizeTensor=jt([this.inputSize,this.inputSize]),this.doubleInputSizeTensor=jt([this.inputSize*2,this.inputSize*2])}normalizeBoxes(t){let a={};a.boxOffsets=De(t,[0,0],[-1,2]),a.boxSizes=De(t,[0,2],[-1,2]),a.div=ve(a.boxOffsets,this.inputSizeTensor),a.boxCenterPoints=ke(a.div,this.anchorsTensor),a.halfBoxSizes=ve(a.boxSizes,this.doubleInputSizeTensor),a.sub=ye(a.boxCenterPoints,a.halfBoxSizes),a.startPoints=te(a.sub,this.inputSizeTensor),a.add=ke(a.boxCenterPoints,a.halfBoxSizes),a.endPoints=te(a.add,this.inputSizeTensor);let n=Fu([a.startPoints,a.endPoints],1);return Object.keys(a).forEach(r=>J(a[r])),n}normalizeLandmarks(t,a){let n={};n.reshape=Q(t,[-1,7,2]),n.div=ve(n.reshape,this.inputSizeTensor),n.landmarks=ke(n.div,this.anchors[a]?this.anchors[a]:0);let r=te(n.landmarks,this.inputSizeTensor);return Object.keys(n).forEach(s=>J(n[s])),r}async predict(t,a){var o;let n={};n.resize=fe.resizeBilinear(t,[this.inputSize,this.inputSize]),n.div=ve(n.resize,ze.tf127),n.image=ye(n.div,ze.tf1),n.batched=this.model.execute(n.image),n.predictions=Oe(n.batched),n.slice=De(n.predictions,[0,0],[-1,1]),n.sigmoid=Wa(n.slice),n.scores=Oe(n.sigmoid);let r=await n.scores.data();n.boxes=De(n.predictions,[0,1],[-1,4]),n.norm=this.normalizeBoxes(n.boxes),n.nms=await fe.nonMaxSuppressionAsync(n.norm,n.scores,3*(((o=a.hand)==null?void 0:o.maxDetected)||1),a.hand.iouThreshold,a.hand.minConfidence);let s=await n.nms.array(),i=[];for(let l of s){let u={};u.box=De(n.norm,[l,0],[1,-1]),u.slice=De(n.predictions,[l,5],[1,14]),u.norm=this.normalizeLandmarks(u.slice,l),u.palmLandmarks=Q(u.norm,[-1,2]);let p=await u.box.data(),c=p.slice(0,2),d=p.slice(2,4),h=await u.palmLandmarks.array(),f={startPoint:c,endPoint:d,palmLandmarks:h,confidence:r[l]},m=II(f,[(t.shape[2]||1)/this.inputSize,(t.shape[1]||0)/this.inputSize]);i.push(m),Object.keys(u).forEach(g=>J(u[g]))}return Object.keys(n).forEach(l=>J(n[l])),i}};var f3e=5,RI=1.65,EI=[0,5,9,13,17,1,2],m3e=0,g3e=2,MI=0,H0=class{constructor(t,a){he(this,"handDetector");he(this,"handPoseModel");he(this,"inputSize");he(this,"storedBoxes");he(this,"skipped");he(this,"detectedHands");var n,r,s;this.handDetector=t,this.handPoseModel=a,this.inputSize=((s=(r=(n=this.handPoseModel)==null?void 0:n.inputs)==null?void 0:r[0].shape)==null?void 0:s[2])||0,this.storedBoxes=[],this.skipped=Number.MAX_SAFE_INTEGER,this.detectedHands=0}calculateLandmarksBoundingBox(t){let a=t.map(i=>i[0]),n=t.map(i=>i[1]),r=[Math.min(...a),Math.min(...n)],s=[Math.max(...a),Math.max(...n)];return{startPoint:r,endPoint:s}}getBoxForPalmLandmarks(t,a){let n=t.map(s=>By([...s,1],a)),r=this.calculateLandmarksBoundingBox(n);return V0(U0(r),f3e)}getBoxForHandLandmarks(t){let a=this.calculateLandmarksBoundingBox(t),n=V0(U0(a),RI);n.palmLandmarks=[];for(let r=0;r<EI.length;r++)n.palmLandmarks.push(t[EI[r]].slice(0,2));return n}transformRawCoords(t,a,n,r){let s=B0(a),i=[s[0]/this.inputSize,s[1]/this.inputSize,(s[0]+s[1])/this.inputSize/2],o=t.map(h=>[i[0]*(h[0]-this.inputSize/2),i[1]*(h[1]-this.inputSize/2),i[2]*h[2]]),l=Wy(n,[0,0]),u=o.map(h=>[...By(h,l),h[2]]),p=CI(r),c=[...ac(a),1],d=[Cs(c,p[0]),Cs(c,p[1])];return u.map(h=>[Math.trunc(h[0]+d[0]),Math.trunc(h[1]+d[1]),Math.trunc(h[2])])}async estimateHands(t,a){let n=!1,r,s=(a.hand.skipTime||0)>ae()-MI,i=this.skipped<(a.hand.skipFrames||0);a.skipAllowed&&s&&i&&(r=await this.handDetector.predict(t,a),this.skipped=0),a.skipAllowed&&this.skipped++,r&&r.length>0&&(r.length!==this.detectedHands&&this.detectedHands!==a.hand.maxDetected||!a.hand.landmarks)&&(this.detectedHands=0,this.storedBoxes=[...r],this.storedBoxes.length>0&&(n=!0));let o=[];for(let l=0;l<this.storedBoxes.length;l++){let u=this.storedBoxes[l];if(u)if(a.hand.landmarks){let p=a.hand.rotation?SI(u.palmLandmarks[m3e],u.palmLandmarks[g3e]):0,c=ac(u),d=[c[0]/t.shape[2],c[1]/t.shape[1]],h=a.hand.rotation&&ne.kernels.includes("rotatewithoffset")?fe.rotateWithOffset(t,p,0,d):t.clone(),f=Wy(-p,c),m=n?this.getBoxForPalmLandmarks(u.palmLandmarks,f):u,g=kI(m,h,[this.inputSize,this.inputSize]),y=ve(g,ze.tf255);J(g),J(h);let[x,A]=this.handPoseModel.execute(y);MI=ae(),J(y);let b=(await x.data())[0];if(J(x),b>=a.hand.minConfidence/4){let w=Q(A,[-1,3]),C=await w.array();J(A),J(w);let T=this.transformRawCoords(C,m,p,f),N=this.getBoxForHandLandmarks(T);this.storedBoxes[l]={...N,confidence:b};let M={landmarks:T,confidence:b,boxConfidence:u.confidence,fingerConfidence:b,box:{topLeft:N.startPoint,bottomRight:N.endPoint}};o.push(M)}else this.storedBoxes[l]=null;J(A)}else{let p=V0(U0(u),RI),c={confidence:u.confidence,boxConfidence:u.confidence,fingerConfidence:0,box:{topLeft:p.startPoint,bottomRight:p.endPoint},landmarks:[]};o.push(c)}}return this.storedBoxes=this.storedBoxes.filter(l=>l!==null),this.detectedHands=o.length,o.length>a.hand.maxDetected&&(o.length=a.hand.maxDetected),o}};var FI={thumb:[1,2,3,4],index:[5,6,7,8],middle:[9,10,11,12],ring:[13,14,15,16],pinky:[17,18,19,20],palm:[0]},cl,hl,_I;async function Vy(e,t){let a=await _I.estimateHands(e,t);if(!a)return[];let n=[];for(let r=0;r<a.length;r++){let s={};if(a[r].landmarks)for(let p of Object.keys(FI))s[p]=FI[p].map(c=>a[r].landmarks[c]);let i=a[r].landmarks,o=[Number.MAX_SAFE_INTEGER,Number.MAX_SAFE_INTEGER,0,0],l=[0,0,0,0];if(i&&i.length>0){for(let p of i)p[0]<o[0]&&(o[0]=p[0]),p[1]<o[1]&&(o[1]=p[1]),p[0]>o[2]&&(o[2]=p[0]),p[1]>o[3]&&(o[3]=p[1]);o[2]-=o[0],o[3]-=o[1],l=[o[0]/(e.shape[2]||0),o[1]/(e.shape[1]||0),o[2]/(e.shape[2]||0),o[3]/(e.shape[1]||0)]}else o=a[r].box?[Math.trunc(Math.max(0,a[r].box.topLeft[0])),Math.trunc(Math.max(0,a[r].box.topLeft[1])),Math.trunc(Math.min(e.shape[2]||0,a[r].box.bottomRight[0])-Math.max(0,a[r].box.topLeft[0])),Math.trunc(Math.min(e.shape[1]||0,a[r].box.bottomRight[1])-Math.max(0,a[r].box.topLeft[1]))]:[0,0,0,0],l=[a[r].box.topLeft[0]/(e.shape[2]||0),a[r].box.topLeft[1]/(e.shape[1]||0),(a[r].box.bottomRight[0]-a[r].box.topLeft[0])/(e.shape[2]||0),(a[r].box.bottomRight[1]-a[r].box.topLeft[1])/(e.shape[1]||0)];let u=W0(i);n.push({id:r,score:Math.round(100*a[r].confidence)/100,boxScore:Math.round(100*a[r].boxConfidence)/100,fingerScore:Math.round(100*a[r].fingerConfidence)/100,label:"hand",box:o,boxRaw:l,keypoints:i,annotations:s,landmarks:u})}return n}async function $I(e){var a,n;ne.initial&&(cl=null,hl=null),!cl||!hl?[cl,hl]=await Promise.all([e.hand.enabled?Fe((a=e.hand.detector)==null?void 0:a.modelPath):null,e.hand.landmarks?Fe((n=e.hand.skeleton)==null?void 0:n.modelPath):null]):(e.debug&&K("cached model:",cl.modelUrl),e.debug&&K("cached model:",hl.modelUrl));let t=cl?new G0(cl):void 0;return t&&hl&&(_I=new H0(t,hl)),[cl,hl]}var Ot=[null,null],x3e=["StatefulPartitionedCall/Postprocessor/Slice","StatefulPartitionedCall/Postprocessor/ExpandDims_1"],Ts=[[0,0],[0,0]],A3e=["hand","fist","pinch","point","face","tip","pinchtip"],PI=4,OI=1.6,b3e=512,v3e=1.4,j0=Number.MAX_SAFE_INTEGER,Uy=0,$r=[0,0],Pt={boxes:[],hands:[]},zI={thumb:[1,2,3,4],index:[5,6,7,8],middle:[9,10,11,12],ring:[13,14,15,16],pinky:[17,18,19,20],base:[0],palm:[0,17,13,9,5,1,0]};async function LI(e){var t;if(ne.initial&&(Ot[0]=null),Ot[0])e.debug&&K("cached model:",Ot[0].modelUrl);else{m0(["tensorlistreserve","enter","tensorlistfromtensor","merge","loopcond","switch","exit","tensorliststack","nextiteration","tensorlistsetitem","tensorlistgetitem","reciprocal","shape","split","where"],e),Ot[0]=await Fe((t=e.hand.detector)==null?void 0:t.modelPath);let a=Ot[0].executor?Object.values(Ot[0].modelSignature.inputs):void 0;Ts[0][0]=Array.isArray(a)?parseInt(a[0].tensorShape.dim[1].size):0,Ts[0][1]=Array.isArray(a)?parseInt(a[0].tensorShape.dim[2].size):0}return Ot[0]}async function WI(e){var t;if(ne.initial&&(Ot[1]=null),Ot[1])e.debug&&K("cached model:",Ot[1].modelUrl);else{Ot[1]=await Fe((t=e.hand.skeleton)==null?void 0:t.modelPath);let a=Ot[1].executor?Object.values(Ot[1].modelSignature.inputs):void 0;Ts[1][0]=Array.isArray(a)?parseInt(a[0].tensorShape.dim[1].size):0,Ts[1][1]=Array.isArray(a)?parseInt(a[0].tensorShape.dim[2].size):0}return Ot[1]}async function w3e(e,t){let a=[];if(!e||!Ot[0])return a;let n={},r=(e.shape[2]||1)/(e.shape[1]||1),s=Math.min(Math.round((e.shape[1]||0)/8)*8,b3e),i=Math.round(s*r/8)*8;n.resize=fe.resizeBilinear(e,[s,i]),n.cast=Ke(n.resize,"int32"),[n.rawScores,n.rawBoxes]=await Ot[0].executeAsync(n.cast,x3e),n.boxes=Oe(n.rawBoxes,[0,2]),n.scores=Oe(n.rawScores,[0]);let o=Ra(n.scores,1);J(o[PI]),o.splice(PI,1),n.filtered=ca(o,1),J(o),n.max=ga(n.filtered,1),n.argmax=rr(n.filtered,1);let l=0;n.nms=await fe.nonMaxSuppressionAsync(n.boxes,n.max,(t.hand.maxDetected||0)+1,t.hand.iouThreshold||0,t.hand.minConfidence||1);let u=await n.nms.data(),p=await n.max.data(),c=await n.argmax.data();for(let d of Array.from(u)){let h=De(n.boxes,d,1),f=await h.data();J(h);let m=[f[1],f[0],f[3]-f[1],f[2]-f[0]],g=I0(m,v3e),y=[Math.trunc(m[0]*$r[0]),Math.trunc(m[1]*$r[1]),Math.trunc(m[2]*$r[0]),Math.trunc(m[3]*$r[1])],x=p[d],A=A3e[c[d]],b={id:l++,score:x,box:y,boxRaw:g,label:A};a.push(b)}return Object.keys(n).forEach(d=>J(n[d])),a.sort((d,h)=>h.score-d.score),a.length>(t.hand.maxDetected||1)&&(a.length=t.hand.maxDetected||1),a}async function Gy(e,t,a){let n={id:t.id,score:Math.round(100*t.score)/100,boxScore:Math.round(100*t.score)/100,fingerScore:0,box:t.box,boxRaw:t.boxRaw,label:t.label,keypoints:[],landmarks:{},annotations:{}};if(e&&Ot[1]&&a.hand.landmarks&&t.score>(a.hand.minConfidence||0)){let r={},s=[t.boxRaw[1],t.boxRaw[0],t.boxRaw[3]+t.boxRaw[1],t.boxRaw[2]+t.boxRaw[0]];r.crop=fe.cropAndResize(e,[s],[0],[Ts[1][0],Ts[1][1]],"bilinear"),r.div=ve(r.crop,ze.tf255),[r.score,r.keypoints]=Ot[1].execute(r.div,["Identity_1","Identity"]);let i=(await r.score.data())[0],o=(100-Math.trunc(100/(1+Math.exp(i))))/100;if(o>=(a.hand.minConfidence||0)){n.fingerScore=o,r.reshaped=Q(r.keypoints,[-1,3]);let p=(await r.reshaped.array()).map(c=>[c[0]/Ts[1][1],c[1]/Ts[1][0],c[2]||0]).map(c=>[c[0]*t.boxRaw[2],c[1]*t.boxRaw[3],c[2]||0]);n.keypoints=p.map(c=>[$r[0]*(c[0]+t.boxRaw[0]),$r[1]*(c[1]+t.boxRaw[1]),c[2]||0]),n.landmarks=W0(n.keypoints);for(let c of Object.keys(zI))n.annotations[c]=zI[c].map(d=>n.landmarks&&n.keypoints[d]?n.keypoints[d]:null)}Object.keys(r).forEach(l=>J(r[l]))}return n}async function Hy(e,t){var r,s;if(!((r=Ot[0])!=null&&r.executor)||!((s=Ot[1])!=null&&s.executor)||!Ot[0].inputs[0].shape||!Ot[1].inputs[0].shape)return[];$r=[e.shape[2]||0,e.shape[1]||0],j0++;let a=(t.hand.skipTime||0)>ae()-Uy,n=j0<(t.hand.skipFrames||0);return t.skipAllowed&&a&&n?Pt.hands:new Promise(async i=>{let o=3*(t.hand.skipTime||0)>ae()-Uy,l=j0<3*(t.hand.skipFrames||0);t.skipAllowed&&Pt.hands.length===t.hand.maxDetected?Pt.hands=await Promise.all(Pt.boxes.map(p=>Gy(e,p,t))):t.skipAllowed&&o&&l&&Pt.hands.length>0?Pt.hands=await Promise.all(Pt.boxes.map(p=>Gy(e,p,t))):(Pt.boxes=await w3e(e,t),Uy=ae(),Pt.hands=await Promise.all(Pt.boxes.map(p=>Gy(e,p,t))),j0=0);let u=[...Pt.boxes];if(Pt.boxes.length=0,t.cacheSensitivity>0)for(let p=0;p<Pt.hands.length;p++){let c=zk(Pt.hands[p].keypoints,$r);if(c.box[2]/(e.shape[2]||1)>.05&&c.box[3]/(e.shape[1]||1)>.05&&Pt.hands[p].fingerScore&&Pt.hands[p].fingerScore>(t.hand.minConfidence||0)){let d=I0(c.box,OI),h=I0(c.boxRaw,OI);Pt.boxes.push({...u[p],box:d,boxRaw:h})}}for(let p=0;p<Pt.hands.length;p++){let c=gs(Pt.hands[p].keypoints,$r);Pt.hands[p].box=c.box,Pt.hands[p].boxRaw=c.boxRaw}i(Pt.hands)})}var pr=(e=null)=>({face:[],body:[],hand:[],gesture:[],object:[],persons:[],performance:{},timestamp:0,width:0,height:0,error:e});var nc={};yr(nc,{connected:()=>X0,horizontal:()=>jy,kpt:()=>q0,relative:()=>Xy,vertical:()=>qy});var q0=["nose","leftEye","rightEye","leftEar","rightEar","leftShoulder","rightShoulder","leftElbow","rightElbow","leftWrist","rightWrist","leftHip","rightHip","leftKnee","rightKnee","leftAnkle","rightAnkle"],jy=[["leftEye","rightEye"],["leftEar","rightEar"],["leftShoulder","rightShoulder"],["leftElbow","rightElbow"],["leftWrist","rightWrist"],["leftHip","rightHip"],["leftKnee","rightKnee"],["leftAnkle","rightAnkle"]],qy=[["leftKnee","leftShoulder"],["rightKnee","rightShoulder"],["leftAnkle","leftKnee"],["rightAnkle","rightKnee"]],Xy=[[["leftHip","rightHip"],["leftShoulder","rightShoulder"]],[["leftElbow","rightElbow"],["leftShoulder","rightShoulder"]]],X0={leftLeg:["leftHip","leftKnee","leftAnkle"],rightLeg:["rightHip","rightKnee","rightAnkle"],torso:["leftShoulder","rightShoulder","rightHip","leftHip","leftShoulder"],leftArm:["leftShoulder","leftElbow","leftWrist"],rightArm:["rightShoulder","rightElbow","rightWrist"],head:[]};var xe=pr(),Ky=0;function VI(e,t){var i,o,l,u,p,c,d,h,f,m,g,y,x,A,b,w,C,T,N,M,F,E,I,_,O,B;let a=ae();if(!e)return pr();let n=Date.now()-e.timestamp,r=n<1e3?8-Math.log(n+1):1;if(e.canvas&&(xe.canvas=e.canvas),e.error&&(xe.error=e.error),!xe.body||e.body.length!==xe.body.length)xe.body=JSON.parse(JSON.stringify(e.body));else for(let $=0;$<e.body.length;$++){let U=e.body[$].box.map((Z,X)=>((r-1)*xe.body[$].box[X]+Z)/r),G=e.body[$].boxRaw.map((Z,X)=>((r-1)*xe.body[$].boxRaw[X]+Z)/r),q=e.body[$].keypoints.map((Z,X)=>{var re,ee,me,ie,Ae,Ce,Re,Le,qe;return{score:Z.score,part:Z.part,position:[xe.body[$].keypoints[X]?((r-1)*(xe.body[$].keypoints[X].position[0]||0)+(Z.position[0]||0))/r:Z.position[0],xe.body[$].keypoints[X]?((r-1)*(xe.body[$].keypoints[X].position[1]||0)+(Z.position[1]||0))/r:Z.position[1],xe.body[$].keypoints[X]?((r-1)*(xe.body[$].keypoints[X].position[2]||0)+(Z.position[2]||0))/r:Z.position[2]],positionRaw:[xe.body[$].keypoints[X]?((r-1)*(xe.body[$].keypoints[X].positionRaw[0]||0)+(Z.positionRaw[0]||0))/r:Z.positionRaw[0],xe.body[$].keypoints[X]?((r-1)*(xe.body[$].keypoints[X].positionRaw[1]||0)+(Z.positionRaw[1]||0))/r:Z.positionRaw[1],xe.body[$].keypoints[X]?((r-1)*(xe.body[$].keypoints[X].positionRaw[2]||0)+(Z.positionRaw[2]||0))/r:Z.positionRaw[2]],distance:[xe.body[$].keypoints[X]?((r-1)*(((re=xe.body[$].keypoints[X].distance)==null?void 0:re[0])||0)+(((ee=Z.distance)==null?void 0:ee[0])||0))/r:(me=Z.distance)==null?void 0:me[0],xe.body[$].keypoints[X]?((r-1)*(((ie=xe.body[$].keypoints[X].distance)==null?void 0:ie[1])||0)+(((Ae=Z.distance)==null?void 0:Ae[1])||0))/r:(Ce=Z.distance)==null?void 0:Ce[1],xe.body[$].keypoints[X]?((r-1)*(((Re=xe.body[$].keypoints[X].distance)==null?void 0:Re[2])||0)+(((Le=Z.distance)==null?void 0:Le[2])||0))/r:(qe=Z.distance)==null?void 0:qe[2]]}}),H={},V={connected:{}};(i=t.body.modelPath)!=null&&i.includes("efficientpose")?V=T0:(o=t.body.modelPath)!=null&&o.includes("blazepose")?V=w0:(l=t.body.modelPath)!=null&&l.includes("movenet")&&(V=nc);for(let[Z,X]of Object.entries(V.connected)){let re=[];for(let ee=0;ee<X.length-1;ee++){let me=q.find(Ae=>Ae.part===X[ee]),ie=q.find(Ae=>Ae.part===X[ee+1]);me&&ie&&re.push([me.position,ie.position])}H[Z]=re}xe.body[$]={...e.body[$],box:U,boxRaw:G,keypoints:q,annotations:H}}if(!xe.hand||e.hand.length!==xe.hand.length)xe.hand=JSON.parse(JSON.stringify(e.hand));else for(let $=0;$<e.hand.length;$++){let U=e.hand[$].box.map((V,Z)=>((r-1)*xe.hand[$].box[Z]+V)/r),G=e.hand[$].boxRaw.map((V,Z)=>((r-1)*xe.hand[$].boxRaw[Z]+V)/r);xe.hand[$].keypoints.length!==e.hand[$].keypoints.length&&(xe.hand[$].keypoints=e.hand[$].keypoints);let q=e.hand[$].keypoints&&e.hand[$].keypoints.length>0?e.hand[$].keypoints.map((V,Z)=>V.map((X,re)=>((r-1)*(xe.hand[$].keypoints[Z][re]||1)+(X||0))/r)):[],H={};if(Object.keys(xe.hand[$].annotations).length!==Object.keys(e.hand[$].annotations).length)xe.hand[$].annotations=e.hand[$].annotations,H=xe.hand[$].annotations;else if(e.hand[$].annotations)for(let V of Object.keys(e.hand[$].annotations))H[V]=(c=(p=(u=e.hand[$])==null?void 0:u.annotations)==null?void 0:p[V])!=null&&c[0]?e.hand[$].annotations[V].map((Z,X)=>Z.map((re,ee)=>((r-1)*xe.hand[$].annotations[V][X][ee]+re)/r)):null;xe.hand[$]={...e.hand[$],box:U,boxRaw:G,keypoints:q,annotations:H}}if(!xe.face||e.face.length!==xe.face.length)xe.face=JSON.parse(JSON.stringify(e.face));else for(let $=0;$<e.face.length;$++){let U=e.face[$].box.map((H,V)=>((r-1)*xe.face[$].box[V]+H)/r),G=e.face[$].boxRaw.map((H,V)=>((r-1)*xe.face[$].boxRaw[V]+H)/r),q=e.face[$].annotations;if(Object.keys(xe.face[$].annotations).length!==Object.keys(e.face[$].annotations).length)xe.face[$].annotations=e.face[$].annotations,q=xe.face[$].annotations;else if(e.face[$].annotations)for(let H of Object.keys(e.face[$].annotations))q[H]=(f=(h=(d=e.face[$])==null?void 0:d.annotations)==null?void 0:h[H])!=null&&f[0]?e.face[$].annotations[H].map((V,Z)=>V.map((X,re)=>((r-1)*xe.face[$].annotations[H][Z][re]+X)/r)):null;if(e.face[$].rotation){let H={matrix:[0,0,0,0,0,0,0,0,0],angle:{roll:0,yaw:0,pitch:0},gaze:{bearing:0,strength:0}};H.matrix=(m=e.face[$].rotation)==null?void 0:m.matrix,H.angle={roll:((r-1)*(((y=(g=xe.face[$].rotation)==null?void 0:g.angle)==null?void 0:y.roll)||0)+(((A=(x=e.face[$].rotation)==null?void 0:x.angle)==null?void 0:A.roll)||0))/r,yaw:((r-1)*(((w=(b=xe.face[$].rotation)==null?void 0:b.angle)==null?void 0:w.yaw)||0)+(((T=(C=e.face[$].rotation)==null?void 0:C.angle)==null?void 0:T.yaw)||0))/r,pitch:((r-1)*(((M=(N=xe.face[$].rotation)==null?void 0:N.angle)==null?void 0:M.pitch)||0)+(((E=(F=e.face[$].rotation)==null?void 0:F.angle)==null?void 0:E.pitch)||0))/r},H.gaze={bearing:((r-1)*(((I=xe.face[$].rotation)==null?void 0:I.gaze.bearing)||0)+(((_=e.face[$].rotation)==null?void 0:_.gaze.bearing)||0))/r,strength:((r-1)*(((O=xe.face[$].rotation)==null?void 0:O.gaze.strength)||0)+(((B=e.face[$].rotation)==null?void 0:B.gaze.strength)||0))/r},xe.face[$]={...e.face[$],rotation:H,box:U,boxRaw:G,annotations:q}}else xe.face[$]={...e.face[$],box:U,boxRaw:G,annotations:q}}if(!xe.object||e.object.length!==xe.object.length)xe.object=JSON.parse(JSON.stringify(e.object));else for(let $=0;$<e.object.length;$++){let U=e.object[$].box.map((q,H)=>((r-1)*xe.object[$].box[H]+q)/r),G=e.object[$].boxRaw.map((q,H)=>((r-1)*xe.object[$].boxRaw[H]+q)/r);xe.object[$]={...e.object[$],box:U,boxRaw:G}}if(e.persons){let $=e.persons;if(!xe.persons||$.length!==xe.persons.length)xe.persons=JSON.parse(JSON.stringify($));else for(let U=0;U<$.length;U++)xe.persons[U].box=$[U].box.map((G,q)=>((r-1)*xe.persons[U].box[q]+G)/r)}e.gesture&&(xe.gesture=e.gesture),xe.width=e.width,xe.height=e.height;let s=ae();return Ky=ne.perfadd?Ky+Math.round(s-a):Math.round(s-a),e.performance&&(xe.performance={...e.performance,interpolate:Ky}),xe}var ba;async function Yy(e){return!ba||ne.initial?ba=await Fe(e.segmentation.modelPath):e.debug&&K("cached model:",ba.modelUrl),ba}async function UI(e,t){var r;if(ba||(ba=await Yy(t)),!(ba!=null&&ba.executor)||!((r=ba==null?void 0:ba.inputs)!=null&&r[0].shape))return null;let a={};a.resize=fe.resizeBilinear(e,[ba.inputs[0].shape?ba.inputs[0].shape[1]:0,ba.inputs[0].shape?ba.inputs[0].shape[2]:0],!1),a.norm=ve(a.resize,ze.tf255),a.res=ba.execute(a.norm),a.squeeze=Oe(a.res,[0]),[a.bgRaw,a.fgRaw]=Ra(a.squeeze,2),a.fg=Oh(a.fgRaw),a.mul=te(a.fg,ze.tf255),a.expand=Ht(a.mul,2),a.output=fe.resizeBilinear(a.expand,[e.shape[1]||0,e.shape[2]||0]);let n;switch(t.segmentation.mode||"default"){case"default":a.input=Oe(e),a.concat=lt([a.input,a.output],-1),n=Ke(a.concat,"int32");break;case"alpha":n=Ke(a.output,"int32");break;default:n=Ve(0)}return Object.keys(a).forEach(s=>J(a[s])),n}var K0={};yr(K0,{distance:()=>Zy,find:()=>S3e,similarity:()=>I3e});function Zy(e,t,a={order:2,multiplier:25}){if(!e||!e)return Number.MAX_SAFE_INTEGER;let n=0;for(let r=0;r<e.length;r++){let s=!a.order||a.order===2?e[r]-t[r]:Math.abs(e[r]-t[r]);n+=!a.order||a.order===2?s*s:s**a.order}return(a.multiplier||20)*n}var HI=(e,t,a,n)=>{if(e===0)return 1;let s=(1-(t===2?Math.sqrt(e):e**(1/t))/100-a)/(n-a);return Math.max(Math.min(s,1),0)};function I3e(e,t,a={order:2,multiplier:25,min:.2,max:.8}){let n=Zy(e,t,a);return HI(n,a.order||2,a.min||0,a.max||1)}function S3e(e,t,a={order:2,multiplier:25,threshold:0,min:.2,max:.8}){if(!Array.isArray(e)||!Array.isArray(t)||e.length<64||t.length===0)return{index:-1,distance:Number.POSITIVE_INFINITY,similarity:0};let n=Number.MAX_SAFE_INTEGER,r=-1;for(let i=0;i<t.length;i++){let o=t[i].length===e.length?Zy(e,t[i],a):Number.MAX_SAFE_INTEGER;if(o<n&&(n=o,r=i),n<(a.threshold||0))break}let s=HI(n,a.order||2,a.min||0,a.max||1);return{index:r,distance:n,similarity:s}}var c5={};yr(c5,{Models:()=>ic,validateModel:()=>af});var jI=.005,on={keypoints:[],padding:[[0,0],[0,0],[0,0],[0,0]]};function Jy(e){for(let t of jy){let a=e.keypoints.findIndex(r=>r.part===t[0]),n=e.keypoints.findIndex(r=>r.part===t[1]);if(e.keypoints[a]&&e.keypoints[n]&&e.keypoints[a].position[0]<e.keypoints[n].position[0]){let r=e.keypoints[a];e.keypoints[a]=e.keypoints[n],e.keypoints[n]=r}}for(let t of qy){let a=e.keypoints.findIndex(r=>r&&r.part===t[0]),n=e.keypoints.findIndex(r=>r&&r.part===t[1]);e.keypoints[a]&&e.keypoints[n]&&e.keypoints[a].position[1]<e.keypoints[n].position[1]&&e.keypoints.splice(a,1)}for(let[t,a]of Xy){let n=e.keypoints.findIndex(u=>u&&u.part===t[0]),r=e.keypoints.findIndex(u=>u&&u.part===t[1]),s=e.keypoints.findIndex(u=>u&&u.part===a[0]),i=e.keypoints.findIndex(u=>u&&u.part===a[1]);if(!e.keypoints[s]||!e.keypoints[i])continue;let o=e.keypoints[n]?[Math.abs(e.keypoints[s].position[0]-e.keypoints[n].position[0]),Math.abs(e.keypoints[i].position[0]-e.keypoints[n].position[0])]:[0,0],l=e.keypoints[r]?[Math.abs(e.keypoints[i].position[0]-e.keypoints[r].position[0]),Math.abs(e.keypoints[s].position[0]-e.keypoints[r].position[0])]:[0,0];if(o[0]>o[1]||l[0]>l[1]){let u=e.keypoints[n];e.keypoints[n]=e.keypoints[r],e.keypoints[r]=u}}}function qI(e){for(let t=0;t<e.length;t++)if(e[t]&&on.keypoints[t]){let a=[Math.abs(e[t].positionRaw[0]-on.keypoints[t].positionRaw[0]),Math.abs(e[t].positionRaw[1]-on.keypoints[t].positionRaw[1])];a[0]<jI&&a[1]<jI?e[t]=on.keypoints[t]:on.keypoints[t]=e[t]}else on.keypoints[t]=e[t];return e}function XI(e,t){var r,s;let a={};if(!((r=e==null?void 0:e.shape)!=null&&r[1])||!((s=e==null?void 0:e.shape)!=null&&s[2]))return e;on.padding=[[0,0],[e.shape[2]>e.shape[1]?Math.trunc((e.shape[2]-e.shape[1])/2):0,e.shape[2]>e.shape[1]?Math.trunc((e.shape[2]-e.shape[1])/2):0],[e.shape[1]>e.shape[2]?Math.trunc((e.shape[1]-e.shape[2])/2):0,e.shape[1]>e.shape[2]?Math.trunc((e.shape[1]-e.shape[2])/2):0],[0,0]],a.pad=ir(e,on.padding),a.resize=fe.resizeBilinear(a.pad,[t,t]);let n=Ke(a.resize,"int32");return Object.keys(a).forEach(i=>J(a[i])),n}function KI(e,t){e.keypoints=e.keypoints.filter(n=>n==null?void 0:n.position);for(let n of e.keypoints)n.position=[n.position[0]*(t[0]+on.padding[2][0]+on.padding[2][1])/t[0]-on.padding[2][0],n.position[1]*(t[1]+on.padding[1][0]+on.padding[1][1])/t[1]-on.padding[1][0]],n.positionRaw=[n.position[0]/t[0],n.position[1]/t[1]];let a=gs(e.keypoints.map(n=>n.position),t);return e.box=a.box,e.boxRaw=a.boxRaw,e}var Ut,Y0=0,Qy=Number.MAX_SAFE_INTEGER,fl={boxes:[],bodies:[],last:0};async function YI(e){var t;return ne.initial&&(Ut=null),Ut?e.debug&&K("cached model:",Ut.modelUrl):(m0(["size"],e),Ut=await Fe(e.body.modelPath)),Y0=Ut!=null&&Ut.executor&&((t=Ut==null?void 0:Ut.inputs)!=null&&t[0].shape)?Ut.inputs[0].shape[2]:0,Y0<64&&(Y0=256),W().flagRegistry.WEBGL_USE_SHAPES_UNIFORMS&&W().set("WEBGL_USE_SHAPES_UNIFORMS",!1),Ut}function T3e(e,t,a){let n=e[0][0],r=[],s=0;for(let p=0;p<n.length;p++)if(s=n[p][2],s>t.body.minConfidence){let c=[n[p][1],n[p][0]];r.push({score:Math.round(100*s)/100,part:q0[p],positionRaw:c,position:[Math.round((a.shape[2]||0)*c[0]),Math.round((a.shape[1]||0)*c[1])]})}s=r.reduce((p,c)=>c.score>p?c.score:p,0);let i=[],o=gs(r.map(p=>p.position),[a.shape[2],a.shape[1]]),l={};for(let[p,c]of Object.entries(X0)){let d=[];for(let h=0;h<c.length-1;h++){let f=r.find(g=>g.part===c[h]),m=r.find(g=>g.part===c[h+1]);f&&m&&f.score>(t.body.minConfidence||0)&&m.score>(t.body.minConfidence||0)&&d.push([f.position,m.position])}l[p]=d}let u={id:0,score:s,box:o.box,boxRaw:o.boxRaw,keypoints:r,annotations:l};return Jy(u),i.push(u),i}function N3e(e,t,a){let n=[];for(let r=0;r<e[0].length;r++){let s=e[0][r],i=Math.round(100*s[51+4])/100;if(i>t.body.minConfidence){let o=[];for(let d=0;d<17;d++){let h=s[3*d+2];if(h>t.body.minConfidence){let f=[s[3*d+1],s[3*d+0]];o.push({part:q0[d],score:Math.round(100*h)/100,positionRaw:f,position:[Math.round((a.shape[2]||0)*f[0]),Math.round((a.shape[1]||0)*f[1])]})}}let l=[s[51+1],s[51+0],s[51+3]-s[51+1],s[51+2]-s[51+0]],u=[Math.trunc(l[0]*(a.shape[2]||0)),Math.trunc(l[1]*(a.shape[1]||0)),Math.trunc(l[2]*(a.shape[2]||0)),Math.trunc(l[3]*(a.shape[1]||0))],p={};for(let[d,h]of Object.entries(X0)){let f=[];for(let m=0;m<h.length-1;m++){let g=o.find(x=>x.part===h[m]),y=o.find(x=>x.part===h[m+1]);g&&y&&g.score>(t.body.minConfidence||0)&&y.score>(t.body.minConfidence||0)&&f.push([g.position,y.position])}p[d]=f}let c={id:r,score:i,box:u,boxRaw:l,keypoints:[...o],annotations:p};Jy(c),n.push(c)}}return n.sort((r,s)=>s.score-r.score),n.length>t.body.maxDetected&&(n.length=t.body.maxDetected),n}async function e5(e,t){var r;if(!(Ut!=null&&Ut.executor)||!((r=Ut==null?void 0:Ut.inputs)!=null&&r[0].shape))return[];t.skipAllowed||(fl.boxes.length=0),Qy++;let a=(t.body.skipTime||0)>ae()-fl.last,n=Qy<(t.body.skipFrames||0);return t.skipAllowed&&a&&n?fl.bodies:new Promise(async s=>{let i={};Qy=0,i.input=XI(e,Y0),i.res=Ut==null?void 0:Ut.execute(i.input),fl.last=ae();let o=await i.res.array();fl.bodies=i.res.shape[2]===17?T3e(o,t,e):N3e(o,t,e);for(let l of fl.bodies)KI(l,[e.shape[2]||1,e.shape[1]||1]),qI(l.keypoints);Object.keys(i).forEach(l=>J(i[l])),s(fl.bodies)})}var Dn,Z0=[],JI=0,t5=Number.MAX_SAFE_INTEGER,Q0=0,J0=2.5;async function QI(e){if(!Dn||ne.initial){Dn=await Fe(e.object.modelPath);let t=Dn!=null&&Dn.executor?Object.values(Dn.modelSignature.inputs):void 0;Q0=Array.isArray(t)?parseInt(t[0].tensorShape.dim[2].size):416}else e.debug&&K("cached model:",Dn.modelUrl);return Dn}async function R3e(e,t,a){var u,p;let n=0,r=[],s=Q0;for(let c of[1,2,4]){let d=c*13,h=Oe(e.find(A=>A.shape[1]===d**2&&(A.shape[2]||0)===Xu.length)),f=await h.array(),m=Oe(e.find(A=>A.shape[1]===d**2&&(A.shape[2]||0)<Xu.length)),g=Q(m,[-1,4,(((u=m.shape)==null?void 0:u[1])||0)/4]),y=rr(g,2),x=await y.array();for(let A=0;A<h.shape[0];A++)for(let b=0;b<(((p=h.shape)==null?void 0:p[1])||0);b++){let w=f[A][b];if(w>(a.object.minConfidence||0)&&b!==61){let C=(.5+Math.trunc(A%d))/d,T=(.5+Math.trunc(A/d))/d,N=x[A].map($=>$*(d/c/s)),[M,F]=[C-J0/c*N[0],T-J0/c*N[1]],[E,I]=[C+J0/c*N[2]-M,T+J0/c*N[3]-F],_=[M,F,E,I];_=_.map($=>Math.max(0,Math.min($,1)));let O=[_[0]*t[0],_[1]*t[1],_[2]*t[0],_[3]*t[1]],B={id:n++,score:Math.round(100*w)/100,class:b+1,label:Xu[b].label,box:O.map($=>Math.trunc($)),boxRaw:_};r.push(B)}}J([h,m,g,y])}let i=r.map(c=>[c.boxRaw[1],c.boxRaw[0],c.boxRaw[3],c.boxRaw[2]]),o=r.map(c=>c.score),l=[];if(i&&i.length>0){let c=await fe.nonMaxSuppressionAsync(i,o,a.object.maxDetected||0,a.object.iouThreshold,a.object.minConfidence);l=Array.from(await c.data()),J(c)}return r=r.filter((c,d)=>l.includes(d)).sort((c,d)=>d.score-c.score),r}async function a5(e,t){if(!(Dn!=null&&Dn.executor))return[];let a=(t.object.skipTime||0)>ae()-JI,n=t5<(t.object.skipFrames||0);return t.skipAllowed&&a&&n&&Z0.length>0?(t5++,Z0):(t5=0,!ne.kernels.includes("mod")||!ne.kernels.includes("sparsetodense")?Z0:new Promise(async r=>{let s=[e.shape[2]||0,e.shape[1]||0],i=fe.resizeBilinear(e,[Q0,Q0],!1),o=ve(i,ze.tf255),l=js(o,[0,3,1,2]),u;t.object.enabled&&(u=Dn.execute(l)),JI=ae();let p=await R3e(u,s,t);Z0=p,J([i,o,l,...u]),r(p)}))}var sc=["nose","leftEye","rightEye","leftEar","rightEar","leftShoulder","rightShoulder","leftElbow","rightElbow","leftWrist","rightWrist","leftHip","rightHip","leftKnee","rightKnee","leftAnkle","rightAnkle"],E3e=sc.length,rc=sc.reduce((e,t,a)=>(e[t]=a,e),{}),M3e=[["leftHip","leftShoulder"],["leftElbow","leftShoulder"],["leftElbow","leftWrist"],["leftHip","leftKnee"],["leftKnee","leftAnkle"],["rightHip","rightShoulder"],["rightElbow","rightShoulder"],["rightElbow","rightWrist"],["rightHip","rightKnee"],["rightKnee","rightAnkle"],["leftShoulder","rightShoulder"],["leftHip","rightHip"]],D7e=M3e.map(([e,t])=>[rc[e],rc[t]]),tS=[["nose","leftEye"],["leftEye","leftEar"],["nose","rightEye"],["rightEye","rightEar"],["nose","leftShoulder"],["leftShoulder","leftElbow"],["leftElbow","leftWrist"],["leftShoulder","leftHip"],["leftHip","leftKnee"],["leftKnee","leftAnkle"],["nose","rightShoulder"],["rightShoulder","rightElbow"],["rightElbow","rightWrist"],["rightShoulder","rightHip"],["rightHip","rightKnee"],["rightKnee","rightAnkle"]];function aS(e){let t=e.reduce(({maxX:a,maxY:n,minX:r,minY:s},{position:{x:i,y:o}})=>({maxX:Math.max(a,i),maxY:Math.max(n,o),minX:Math.min(r,i),minY:Math.min(s,o)}),{maxX:Number.NEGATIVE_INFINITY,maxY:Number.NEGATIVE_INFINITY,minX:Number.POSITIVE_INFINITY,minY:Number.POSITIVE_INFINITY});return[t.minX,t.minY,t.maxX-t.minX,t.maxY-t.minY]}function nS(e,[t,a],[n,r]){let s=t/n,i=a/r,o=(u,p)=>({id:p,score:u.score,boxRaw:[u.box[0]/r,u.box[1]/n,u.box[2]/r,u.box[3]/n],box:[Math.trunc(u.box[0]*i),Math.trunc(u.box[1]*s),Math.trunc(u.box[2]*i),Math.trunc(u.box[3]*s)],keypoints:u.keypoints.map(({score:c,part:d,position:h})=>({score:c,part:d,position:[Math.trunc(h.x*i),Math.trunc(h.y*s)],positionRaw:[h.x/n,h.y/n]})),annotations:{}});return e.map((u,p)=>o(u,p))}var ef=class{constructor(t,a){he(this,"priorityQueue");he(this,"numberOfElements");he(this,"getElementValue");this.priorityQueue=new Array(t),this.numberOfElements=-1,this.getElementValue=a}enqueue(t){this.priorityQueue[++this.numberOfElements]=t,this.swim(this.numberOfElements)}dequeue(){let t=this.priorityQueue[0];return this.exchange(0,this.numberOfElements--),this.sink(0),this.priorityQueue[this.numberOfElements+1]=null,t}empty(){return this.numberOfElements===-1}size(){return this.numberOfElements+1}all(){return this.priorityQueue.slice(0,this.numberOfElements+1)}max(){return this.priorityQueue[0]}swim(t){for(;t>0&&this.less(Math.floor(t/2),t);)this.exchange(t,Math.floor(t/2)),t=Math.floor(t/2)}sink(t){for(;2*t<=this.numberOfElements;){let a=2*t;if(a<this.numberOfElements&&this.less(a,a+1)&&a++,!this.less(t,a))break;this.exchange(t,a),t=a}}getValueAt(t){return this.getElementValue(this.priorityQueue[t])}less(t,a){return this.getValueAt(t)<this.getValueAt(a)}exchange(t,a){let n=this.priorityQueue[t];this.priorityQueue[t]=this.priorityQueue[a],this.priorityQueue[a]=n}};function n5(e,t,a,n){return{y:n.get(e,t,a),x:n.get(e,t,a+E3e)}}function r5(e,t,a){let{heatmapY:n,heatmapX:r,id:s}=e,{y:i,x:o}=n5(n,r,s,a);return{x:e.heatmapX*t+o,y:e.heatmapY*t+i}}function s5(e,t,a){return e<t?t:e>a?a:e}function rS(e,t,a,n){let r=a-e,s=n-t;return r*r+s*s}function i5(e,t){return{x:e.x+t.x,y:e.y+t.y}}var ln,_3e=["MobilenetV1/offset_2/BiasAdd","MobilenetV1/heatmap_2/BiasAdd","MobilenetV1/displacement_fwd_2/BiasAdd","MobilenetV1/displacement_bwd_2/BiasAdd"],tf=1,nd=16,$3e=50**2;function sS(e,t,a,n,r,s,i=2){let o=y=>({y:s.get(y.y,y.x,e),x:s.get(y.y,y.x,s.shape[2]/2+e)}),l=(y,x,A)=>({y:s5(Math.round(y.y/nd),0,x-1),x:s5(Math.round(y.x/nd),0,A-1)}),[u,p]=n.shape,c=l(t.position,u,p),d=o(c),f=i5(t.position,d);for(let y=0;y<i;y++){let x=l(f,u,p),A=n5(x.y,x.x,a,r);f=i5({x:x.x*nd,y:x.y*nd},{x:A.x,y:A.y})}let m=l(f,u,p),g=n.get(m.y,m.x,a);return{position:f,part:sc[a],score:g}}function D3e(e,t,a,n,r){let s=tS.map(([d,h])=>[rc[d],rc[h]]),i=s.map(([,d])=>d),o=s.map(([d])=>d),l=t.shape[2],u=i.length,p=new Array(l),c=r5(e.part,nd,a);p[e.part.id]={score:e.score,part:sc[e.part.id],position:c};for(let d=u-1;d>=0;--d){let h=i[d],f=o[d];p[h]&&!p[f]&&(p[f]=sS(d,p[h],f,t,a,r))}for(let d=0;d<u;++d){let h=o[d],f=i[d];p[h]&&!p[f]&&(p[f]=sS(d,p[h],f,t,a,n))}return p}function P3e(e,t,a,n,r){let[s,i]=r.shape,o=!0,l=Math.max(a-tf,0),u=Math.min(a+tf+1,s);for(let p=l;p<u;++p){let c=Math.max(n-tf,0),d=Math.min(n+tf+1,i);for(let h=c;h<d;++h)if(r.get(p,h,e)>t){o=!1;break}if(!o)break}return o}function O3e(e,t){let[a,n,r]=t.shape,s=new ef(a*n*r,({score:i})=>i);for(let i=0;i<a;++i)for(let o=0;o<n;++o)for(let l=0;l<r;++l){let u=t.get(i,o,l);u<e||P3e(l,u,i,o,t)&&s.enqueue({score:u,part:{heatmapY:i,heatmapX:o,id:l}})}return s}function iS(e,{x:t,y:a},n){return e.some(({keypoints:r})=>{var i;let s=(i=r[n])==null?void 0:i.position;return s?rS(a,t,s.y,s.x)<=$3e:!1})}function z3e(e,t){return t.reduce((n,{position:r,score:s},i)=>(iS(e,r,i)||(n+=s),n),0)/t.length}function L3e(e,t,a,n,r,s){let i=[],o=O3e(s,t);for(;i.length<r&&!o.empty();){let l=o.dequeue(),u=r5(l.part,nd,e);if(iS(i,u,l.part.id))continue;let p=D3e(l,t,e,a,n);p=p.filter(h=>h.score>s);let c=z3e(i,p),d=aS(p);c>s&&i.push({keypoints:p,box:d,score:Math.round(100*c)/100})}return i}async function o5(e,t){if(!(ln!=null&&ln.executor))return[];let a=Pe(()=>{if(!ln.inputs[0].shape)return[];let i=fe.resizeBilinear(e,[ln.inputs[0].shape[2],ln.inputs[0].shape[1]]),o=ye(ve(Ke(i,"float32"),127.5),1),u=ln.execute(o,_3e).map(p=>Oe(p,[0]));return u[1]=Wa(u[1]),u}),n=await Promise.all(a.map(i=>i.buffer()));for(let i of a)J(i);let r=L3e(n[0],n[1],n[2],n[3],t.body.maxDetected,t.body.minConfidence);return ln.inputs[0].shape?nS(r,[e.shape[1],e.shape[2]],[ln.inputs[0].shape[2],ln.inputs[0].shape[1]]):[]}async function oS(e){return!ln||ne.initial?ln=await Fe(e.body.modelPath):e.debug&&K("cached model:",ln.modelUrl),ln}var cr,W3e=["fgr","pha","r1o","r2o","r3o","r4o"],Gt={},u5=0;function dS(e){J([Gt.r1i,Gt.r2i,Gt.r3i,Gt.r4i,Gt.downsample_ratio]),Gt.r1i=Ve(0),Gt.r2i=Ve(0),Gt.r3i=Ve(0),Gt.r4i=Ve(0),u5=e.segmentation.ratio||.5,Gt.downsample_ratio=Ve(u5)}async function d5(e){return!cr||ne.initial?cr=await Fe(e.segmentation.modelPath):e.debug&&K("cached model:",cr.modelUrl),dS(e),cr}var uS=e=>Pe(()=>{let t=Oe(e,[0]),a=te(t,ze.tf255);return Ke(a,"int32")});function l5(e,t){let a=e?uS(e):sr([t.shape[1]||0,t.shape[2]||0,3],255,"int32"),n=t?uS(t):sr([e.shape[1]||0,e.shape[2]||0,1],255,"int32"),r=lt([a,n],-1);return J([a,n]),r}function B3e(e){return Pe(()=>{let t={};return t.unstack=Ra(e,-1),t.concat=lt(t.unstack,1),t.split=Ca(t.concat,4,1),t.stack=lt(t.split,2),t.squeeze=Oe(t.stack,[0]),t.expand=Ht(t.squeeze,-1),t.add=ke(t.expand,1),t.mul=te(t.add,127.5),t.cast=Ke(t.mul,"int32"),t.tile=jr(t.cast,[1,1,3]),t.alpha=sr([t.tile.shape[0]||0,t.tile.shape[1]||0,1],255,"int32"),lt([t.tile,t.alpha],-1)})}async function pS(e,t){if(cr||(cr=await d5(t)),!(cr!=null&&cr.executor))return null;Gt.src=ve(e,255),u5!==t.segmentation.ratio&&dS(t);let[a,n,r,s,i,o]=await cr.executeAsync(Gt,W3e),l;switch(t.segmentation.mode||"default"){case"default":l=l5(a,n);break;case"alpha":l=l5(null,n);break;case"foreground":l=l5(a,null);break;case"state":l=B3e(r);break;default:l=Ve(0)}return J([Gt.src,a,n,Gt.r1i,Gt.r2i,Gt.r3i,Gt.r4i]),[Gt.r1i,Gt.r2i,Gt.r3i,Gt.r4i]=[r,s,i,o],l}var va;async function p5(e){return!va||ne.initial?va=await Fe(e.segmentation.modelPath):e.debug&&K("cached model:",va.modelUrl),va}async function hS(e,t){var r;if(va||(va=await p5(t)),!(va!=null&&va.executor)||!((r=va==null?void 0:va.inputs)!=null&&r[0].shape))return null;let a={};a.resize=fe.resizeBilinear(e,[va.inputs[0].shape?va.inputs[0].shape[1]:0,va.inputs[0].shape?va.inputs[0].shape[2]:0],!1),a.norm=ve(a.resize,ze.tf255),a.res=va.execute(a.norm),a.squeeze=Oe(a.res,[0]),a.alpha=fe.resizeBilinear(a.squeeze,[e.shape[1]||0,e.shape[2]||0]),a.mul=te(a.alpha,ze.tf255);let n;switch(t.segmentation.mode||"default"){case"default":a.input=Oe(e),a.concat=lt([a.input,a.mul],-1),n=Ke(a.concat,"int32");break;case"alpha":n=Ke(a.mul,"int32");break;default:n=Ve(0)}return Object.keys(a).forEach(s=>J(a[s])),n}function af(e,t,a){var u,p;if(!t||!((u=e==null?void 0:e.config)!=null&&u.validateModels))return null;let n=["const","placeholder","noop","pad","squeeze","add","sub","mul","div"],r=["biasadd","fusedbatchnormv3","matmul","switch","shape","merge","split","broadcastto"],s=[],i=[],o=t.modelUrl,l=t.executor;if((p=l==null?void 0:l.graph)!=null&&p.nodes)for(let c of Object.values(l.graph.nodes)){let d=c.op.toLowerCase();s.includes(d)||s.push(d)}else!l&&e.config.debug&&K("model not loaded",a);for(let c of s)!n.includes(c)&&!r.includes(c)&&!e.env.kernels.includes(c)&&!e.env.kernels.includes(c.replace("_",""))&&!e.env.kernels.includes(c.replace("native",""))&&!e.env.kernels.includes(c.replace("v2",""))&&i.push(c);return e.config.debug&&i.length>0&&K("model validation failed:",a,i),i.length>0?{name:a,missing:i,ops:s,url:o}:null}var ic=class{constructor(t){he(this,"instance");he(this,"models",{});this.models={},this.instance=t}stats(){let t=0,a=0,n=0;for(let s of Object.values(xa))t+=s.sizeFromManifest,a+=s.sizeLoadedWeights,n+=s.sizeDesired;let r=n>0?a/n:0;return{numLoadedModels:Object.values(xa).length,numDefinedModels:Object.keys(this.models).length,percentageLoaded:r,totalSizeFromManifest:t,totalSizeWeights:a,totalSizeLoading:n,modelStats:Object.values(xa)}}reset(){for(let t of Object.keys(this.models))this.models[t]=null}async load(t){var n,r,s,i,o,l,u,p,c,d,h,f,m,g,y,x,A,b,w,C,T,N,M,F,E,I,_;ne.initial&&this.reset(),t&&(this.instance=t);let a={};a.blazeface=this.instance.config.face.enabled&&!this.models.blazeface?l9(this.instance.config):null,a.antispoof=this.instance.config.face.enabled&&((n=this.instance.config.face.antispoof)!=null&&n.enabled)&&!this.models.antispoof?_9(this.instance.config):null,a.liveness=this.instance.config.face.enabled&&((r=this.instance.config.face.liveness)!=null&&r.enabled)&&!this.models.liveness?O9(this.instance.config):null,a.faceres=this.instance.config.face.enabled&&((s=this.instance.config.face.description)!=null&&s.enabled)&&!this.models.faceres?N9(this.instance.config):null,a.emotion=this.instance.config.face.enabled&&((i=this.instance.config.face.emotion)!=null&&i.enabled)&&!this.models.emotion?I9(this.instance.config):null,a.iris=this.instance.config.face.enabled&&((o=this.instance.config.face.iris)!=null&&o.enabled)&&!((l=this.instance.config.face.attention)!=null&&l.enabled)&&!this.models.iris?f9(this.instance.config):null,a.facemesh=this.instance.config.face.enabled&&((u=this.instance.config.face.mesh)!=null&&u.enabled)&&!this.models.facemesh?A9(this.instance.config):null,a.gear=this.instance.config.face.enabled&&((p=this.instance.config.face.gear)!=null&&p.enabled)&&!this.models.gear?B9(this.instance.config):null,a.ssrnetage=this.instance.config.face.enabled&&((c=this.instance.config.face.ssrnet)!=null&&c.enabled)&&!this.models.ssrnetage?H9(this.instance.config):null,a.ssrnetgender=this.instance.config.face.enabled&&((d=this.instance.config.face.ssrnet)!=null&&d.enabled)&&!this.models.ssrnetgender?K9(this.instance.config):null,a.mobilefacenet=this.instance.config.face.enabled&&((h=this.instance.config.face.mobilefacenet)!=null&&h.enabled)&&!this.models.mobilefacenet?eI(this.instance.config):null,a.insightface=this.instance.config.face.enabled&&((f=this.instance.config.face.insightface)!=null&&f.enabled)&&!this.models.insightface?sI(this.instance.config):null,a.blazepose=this.instance.config.body.enabled&&!this.models.blazepose&&((m=this.instance.config.body.modelPath)!=null&&m.includes("blazepose"))?Uk(this.instance.config):null,a.blazeposedetect=this.instance.config.body.enabled&&!this.models.blazeposedetect&&this.instance.config.body.detector&&this.instance.config.body.detector.modelPath?Vk(this.instance.config):null,a.efficientpose=this.instance.config.body.enabled&&!this.models.efficientpose&&((g=this.instance.config.body.modelPath)!=null&&g.includes("efficientpose"))?Kk(this.instance.config):null,a.movenet=this.instance.config.body.enabled&&!this.models.movenet&&((y=this.instance.config.body.modelPath)!=null&&y.includes("movenet"))?YI(this.instance.config):null,a.posenet=this.instance.config.body.enabled&&!this.models.posenet&&((x=this.instance.config.body.modelPath)!=null&&x.includes("posenet"))?oS(this.instance.config):null,a.handtrack=this.instance.config.hand.enabled&&!this.models.handtrack&&((b=(A=this.instance.config.hand.detector)==null?void 0:A.modelPath)!=null&&b.includes("handtrack"))?LI(this.instance.config):null,a.handskeleton=this.instance.config.hand.enabled&&this.instance.config.hand.landmarks&&!this.models.handskeleton&&((C=(w=this.instance.config.hand.detector)==null?void 0:w.modelPath)!=null&&C.includes("handtrack"))?WI(this.instance.config):null,(N=(T=this.instance.config.hand.detector)==null?void 0:T.modelPath)!=null&&N.includes("handdetect")&&([a.handpose,a.handskeleton]=this.models.handpose?[null,null]:await $I(this.instance.config)),a.centernet=this.instance.config.object.enabled&&!this.models.centernet&&((M=this.instance.config.object.modelPath)!=null&&M.includes("centernet"))?jk(this.instance.config):null,a.nanodet=this.instance.config.object.enabled&&!this.models.nanodet&&((F=this.instance.config.object.modelPath)!=null&&F.includes("nanodet"))?QI(this.instance.config):null,a.selfie=this.instance.config.segmentation.enabled&&!this.models.selfie&&((E=this.instance.config.segmentation.modelPath)!=null&&E.includes("selfie"))?p5(this.instance.config):null,a.meet=this.instance.config.segmentation.enabled&&!this.models.meet&&((I=this.instance.config.segmentation.modelPath)!=null&&I.includes("meet"))?Yy(this.instance.config):null,a.rvm=this.instance.config.segmentation.enabled&&!this.models.rvm&&((_=this.instance.config.segmentation.modelPath)!=null&&_.includes("rvm"))?d5(this.instance.config):null;for(let[O,B]of Object.entries(a))B!=null&&B.then&&B.then($=>this.models[O]=$);await Promise.all(Object.values(a))}list(){let t=Object.keys(this.models).map(a=>{var n;return{name:a,loaded:this.models[a]!==null,size:0,url:this.models[a]?(n=this.models[a])==null?void 0:n.modelUrl:null}});for(let a of t){let n=Object.keys(xa).find(r=>r.startsWith(a.name));n&&(a.size=xa[n].sizeLoadedWeights,a.url=xa[n].url)}return t}loaded(){return this.list().filter(n=>n.loaded).map(n=>n.name)}validate(){let t=[];for(let a of Object.keys(this.models)){let n=this.models[a];if(!n)continue;let r=af(this.instance,n,a);r&&t.push(r)}return t}};function mS(e,t,a,n,r){var o,l,u,p,c,d;let s=0,i=[];for(let h of e){let f={id:s++,face:h,body:null,hands:{left:null,right:null},gestures:[],box:[0,0,0,0]};for(let b of t)h.box[0]>b.box[0]&&h.box[0]<b.box[0]+b.box[2]&&h.box[1]+h.box[3]>b.box[1]&&h.box[1]+h.box[3]<b.box[1]+b.box[3]&&(f.body=b);if(f.body)for(let b of a)b.box[0]+b.box[2]>f.body.box[0]&&b.box[0]+b.box[2]<f.body.box[0]+f.body.box[2]&&b.box[1]+b.box[3]>f.body.box[1]&&b.box[1]+b.box[3]<f.body.box[1]+f.body.box[3]&&f.hands&&(f.hands.left=b),b.box[0]<f.body.box[0]+f.body.box[2]&&b.box[0]>f.body.box[0]&&b.box[1]+b.box[3]>f.body.box[1]&&b.box[1]+b.box[3]<f.body.box[1]+f.body.box[3]&&f.hands&&(f.hands.right=b);for(let b of n)(b.face!==void 0&&b.face===h.id||b.iris!==void 0&&b.iris===h.id||b.body!==void 0&&b.body===((o=f.body)==null?void 0:o.id)||b.hand!==void 0&&b.hand===((l=f.hands.left)==null?void 0:l.id)||b.hand!==void 0&&b.hand===((u=f.hands.right)==null?void 0:u.id))&&f.gestures.push(b);let m=[],g=[],y=b=>{b&&b.length===4&&(m.push(b[0],b[0]+b[2]),g.push(b[1],b[1]+b[3]))};y(f.face.box),y((p=f.body)==null?void 0:p.box),y((c=f.hands.left)==null?void 0:c.box),y((d=f.hands.right)==null?void 0:d.box);let x=Math.min(...m),A=Math.min(...g);f.box=[x,A,Math.max(...m)-x,Math.max(...g)-A],r!=null&&r[1]&&(r!=null&&r[2])&&(f.boxRaw=[f.box[0]/r[2],f.box[1]/r[1],f.box[2]/r[2],f.box[3]/r[1]]),i.push(f)}return i}var nf=`
|
|
/9j/4AAQSkZJRgABAQEAYABgAAD/4QBoRXhpZgAATU0AKgAAAAgABAEaAAUAAAABAAAAPgEbAAUA
|
|
AAABAAAARgEoAAMAAAABAAIAAAExAAIAAAARAAAATgAAAAAAAABgAAAAAQAAAGAAAAABcGFpbnQu
|
|
bmV0IDQuMi4xMwAA/9sAQwAGBAUGBQQGBgUGBwcGCAoQCgoJCQoUDg8MEBcUGBgXFBYWGh0lHxob
|
|
IxwWFiAsICMmJykqKRkfLTAtKDAlKCko/9sAQwEHBwcKCAoTCgoTKBoWGigoKCgoKCgoKCgoKCgo
|
|
KCgoKCgoKCgoKCgoKCgoKCgoKCgoKCgoKCgoKCgoKCgoKCgo/8AAEQgBAAEAAwEhAAIRAQMRAf/E
|
|
AB8AAAEFAQEBAQEBAAAAAAAAAAABAgMEBQYHCAkKC//EALUQAAIBAwMCBAMFBQQEAAABfQECAwAE
|
|
EQUSITFBBhNRYQcicRQygZGhCCNCscEVUtHwJDNicoIJChYXGBkaJSYnKCkqNDU2Nzg5OkNERUZH
|
|
SElKU1RVVldYWVpjZGVmZ2hpanN0dXZ3eHl6g4SFhoeIiYqSk5SVlpeYmZqio6Slpqeoqaqys7S1
|
|
tre4ubrCw8TFxsfIycrS09TV1tfY2drh4uPk5ebn6Onq8fLz9PX29/j5+v/EAB8BAAMBAQEBAQEB
|
|
AQEAAAAAAAABAgMEBQYHCAkKC//EALURAAIBAgQEAwQHBQQEAAECdwABAgMRBAUhMQYSQVEHYXET
|
|
IjKBCBRCkaGxwQkjM1LwFWJy0QoWJDThJfEXGBkaJicoKSo1Njc4OTpDREVGR0hJSlNUVVZXWFla
|
|
Y2RlZmdoaWpzdHV2d3h5eoKDhIWGh4iJipKTlJWWl5iZmqKjpKWmp6ipqrKztLW2t7i5usLDxMXG
|
|
x8jJytLT1NXW19jZ2uLj5OXm5+jp6vLz9PX29/j5+v/aAAwDAQACEQMRAD8A+qaKACigApGOKAML
|
|
Xp8xlF5A7V4X8RtYs7PzfNImnx8sa8Kp9z3q2tEgp6angWs62ZZ5CTGoJ6DArGNz5p+UrID6EUrF
|
|
PUlW1EuN0XNW7PQ2L5j3JnoKXN0KijqNP0eYoqXBdgPuuo+ZPeupisWn2Jd4+0r924XgsQOCff3/
|
|
AJ1FzRKxDqGii6m3siiQ8F1XGfXI6YNWLfRbiRQMkcZI9fpTDluT2/h6Qy8gDPbtmtG38JeY480Z
|
|
5zSLUTZg8M28YwYxjAArXtdPt402qgHbpSaLWhma3o0Uqk7Nx9DWLaaVblgPs6qRyds2M/gRSQp9
|
|
zZOni2iWS2hlQ+kjYz9OMGrdjq89vIPPVhj+8M/lQyDq9P1WOYBlMZz1AOD+VdDaTiReOKulK0jO
|
|
tHmi0WDTlr0TyxRVhT8tJjIX+9SUxHXUV553BRQAVBcPhSBTSuxPY86+IGti0s5I7dsORy9fM3i6
|
|
8e8mfDO5P90ZrWWiJicNPpZZtxV/xrW0jQt4DOv6Vk2dEEdTY6BHuB25rpbPSo0QARjP0qTRI17W
|
|
wA/hFaMWmoQMgflQXYsDS142rU9tpqqenfNA7GgtihxkdKuRW6qMY/GkDZY8sY4Ap4hXbyB+VArk
|
|
EtuH4wPyrk/EGkOm+a3jw3suRQLc5i38SX9hJ9nnY+XnBUdPyNdFY6pa3KkkAE9l6f8AfJ/pSJT6
|
|
GhDmI+Zb4ZRycdv6ium0nUhKFydrelTsNnS2829RnrVgV6NKXNG55lWPLIM81Op+WrZkRMfmNNzT
|
|
A7GivPO4KKAEY4XNYWt3vkwPg4OK0giJdjw/xrqhm87Zs8tc7pX5A+leSajf6aHYJ50kn4AZpTep
|
|
rBWRm2Vobm4BXfyehPFdnpmnBFUY5rI2SN63tlToK0YI+KZpFF+3QdavwoKTLtoW0Toaswpk5pCb
|
|
LCxipAhoIuP2dKevHXoaYDylRyxhlwRQI4nxVoCXWZI1GfpXGtbSWjYPGP73+NIGupt6TqMsLruZ
|
|
ih4xnP5V09mQ+JLd8gn0xSYJnVaVdkook69K34zuUGunDS3Rx4qOzHVIp4rrOMY3NJQI7GivPO8K
|
|
KAILt9kZrz3xlebYiu8KCCWb0XvW0NFch6ysfO3jLVjfXLIn+pQkKorl7WxNxIPl71g2dUUdpo+l
|
|
pBGvHPet23iC8ihFosrxirkHQUFo0IF4FXI1O726CpKLacCrMJoJLYHAPpTwucHpSRJJ5e4AZI9x
|
|
UqpxzVpCuOC8cUpQUMRnXttuB4rjNdsYyeVwfXpmpGmcvcQyafMCFJjPY10eg34BUg4DcZP8jUO4
|
|
HaRq3lLNF+IHet7R7jz7c56rwa2wz9+xhiVeFy/T1PFegeaNPWigDsc0ZrzzvDNIaAM7VpNqdegr
|
|
xL4l6kywyRhseZ19lrdfAZL4jxYg3Fw20d63tJsdrDI5rm3Z3R0R0Mce1eKnQYAplIkWrMJ45oZS
|
|
NO3PHbNXIyfpSGWowSOasxLUiZdjFSqtNEMkUemKlAGKsRJjAppFAiORMjmsTVrNZEO4cfSoZSOD
|
|
1eJ7WXBUzQZ+7nkfSo7e2Ei+ZaMzxntjBX2NSU1Y6/wxqojiEFzkA8KTXYaUoWRyv3W5rSjpNHPX
|
|
+BmpSg8V6J5gUUAdhRXnneFFAGHrTfu5PpXzj8S70/aZtxzztXFbv4DKHxHI+H4GZiz9zxXXW8G3
|
|
GBXMjvLRXAx0oPGPSmMVeOnWrMTYpFI0bcg1fh54xmgovRcD3qxETSIZcRvzp+/BpEkqsBUqsM9K
|
|
q4Em4Gkxk0yRGXrVW6i8yFhkg+tJjRxGsWrxllkUMh9eK5uMz6bcebbnfG33kPcVkay2OntPKuo0
|
|
nhXI67c8qa7Lw3c+adjcEDGK1paSRhVV4s6A0or0jyRRQ1AHX0V553hRQBz+vNtt5z3xXzX8Qbdm
|
|
uic5YnOMdK3l8JnTXvlbwpYl+WySOgrp5YfLOOB9O1c62O7qQkc+9RsKChFPWp4DluOlSykaNruH
|
|
ArUgHShFNF2NT1qxGO3NBmyxGcE1N2560CFzjrUysO9JAPDDjFOVuKoQuSRTWouBkazbCa3cd8cV
|
|
wF7IISQccHBzUSWpV9C3o1x5b5GAjdQD1rs9DjC3kckbEhqKfxIzn8LOupRXqnkPccBSkUAzraK8
|
|
87wooA5rxMSI3HqK8B8bQl9Q8sffY5b/AAraXwkUviNrw9pH2W1ViMMRTdRjw4HpWNtDti9TPc4P
|
|
FQs2M5qdyyMHLcfjV63HTAoBGtap0wK0YxigpsuRDtVhVYd6GQydVwwIqdRnqKCR23I5pCMUW6gD
|
|
YNKuetAEise9KTxQBWuFyhrznxNZkXjFeN3I+tTIZg2OqmzmxNF0PO3vXp/g2+hukVl4zyPanTXv
|
|
JmVR+60dpThXpnlPceopWFAbnV0V553hSGgRynjC5FujOey14Ssp1HxNmTnc+a3kvcIpv37HoEYQ
|
|
QmMdVHSsnVbYJF5jVk0dsNzlruVIsl2wKxbjWrVHILjg1CRbZJb+ILHPzyhfStODWLQgFJFYd+el
|
|
UJM27HUIXxhga1Y5lLVLKLkMnoauxnPPrSEx7ShF+Y/n2qrc6xBbhizDAqkK1zJuvG9nbg8ZA681
|
|
ly/Ei052RO3uKAsZlx8QGd8xxvt9Aa1NH8dK7AXMcip64zigdkdrZX8F7EJLdwwNXMkrz1qRMRly
|
|
CK4TxmpidWI49felPYSOMmi80NIoOV6qRzXYeA5SskYPfirpfEjGr8LPWVHyD6U4CvQPL3ZItOYc
|
|
UDOoNFeed4Uhpks4H4iE/Z5MeleMeGULeLgjds10S+BGdL+Jc9OSBU2Huc5Nc74yvUtrcDBrJnZF
|
|
63PJdXvLy/lKWw46bvQVz82jXhkLO5Y+9ZlsYthcRnbIjY9R3q3awTRkEM3WmJI6C0ea3dGRsr1x
|
|
XY6TqW9FLHnjrUs0izpLK5DDjofSta3ckH09KRUkZuuTvFGdvPauE1Y3U6Mqbssf/rUxHPTaJPK2
|
|
ZmJPbBqzY6DCZh5xJC9s9aBJHU6dpemJjfEmfetJtI0+VPkUr/unFOxdiextHs33W07YHQHk11mk
|
|
Xb3KbZ1xIvcd6LEyWho4Nct41sTPYb16ipexCPPZN+wYGCvH1rrPAEJmvkPoc1VL4kZVvgZ6yFwK
|
|
cBXoHkkqinFaVyzo80GuE7WJRQSziPiGdthK5HQV4x4J/wBI8WPIewNdEvgRNL42emO/yj1UHNef
|
|
eNpRczbC+I17DvWT2OqJxc0sMK4TCisy41q0hfEkqj8aixdwTXNOlwvmqD9anS9tXH7uVG+hosO4
|
|
/wC0oOhrR0+6G4YNIEzsNEuCxAPNdjZruA4xxUmjINSjURksOlcbqFykbnjFA1sYGoassaknCqO5
|
|
rl7rxhGm7yBnBxuJq0rkSlYpw+NLlsfd5P8AerVsvHEqSBHwPVgcgVpyMyVXU3rXxcHYETAk+hru
|
|
/DWti6ZSTyOKzZqndHaxvvUGq2rQ+dYyqR24qWI8dvbr7LqDxyDAzXpvw6FvIxePGSM06Xxoyr/A
|
|
zviKFHNegeX1J41zUhXioGbuaSuM6wpCaBHG/EcA6HN/exxXjXw2jL67cv8A3Qa6H8CFR+NnoWpO
|
|
I4XI44rxLxrqjQzSEsQM1gdSPM9U1uR1YbmWIdXHf2rmpIb67YS28UrRlsLI3c/jW0VZGUpO5pW1
|
|
jfLNOjahawzwReYI5cjzMkDavHJ5/SrVv9uhtPtVxCPLBwzxnlT9KGghLU3tKvvPjHzbl7EGuisJ
|
|
GRxWLOg7nRXJEbDjmvSNK+aFSfSoZr0KutRkphc4NcRrdkVjL9aVio7Hk3iqS8ubhrWzUlsZY9kG
|
|
cZNc5D4aee5MclzJIFTzHAO0MfatqSOWu7bFS1srDUZEis0vIZoUxPvfcC+4/dx2xjr712XiTwXb
|
|
WmlQ6hol3cRhoFd4rlg3zY5wR0GelavQwjq7GD4etdVvSnk2wAB+9v8A8mvcfA2kXiRo0/UdcDis
|
|
ZnTTulqeoWqbUAJqWUb42X1FZlnjfjSwlGrr5S/eNdD4RkvLAAQ4yRyaUZcruVKl7TQ9I0G+mnzH
|
|
ckFwM8VuIK7ac3KF2eXiKapz5UWYxipNtMyNejNch0jSar3cjR27uoyQCRVRWom9DxTx54gu5fMi
|
|
lbKdMVjfCZPNlv5v9rFbVHpYqjGzbOn8SzFI9o715L4u0r7arYzk+lYdTqSujy7U/C0u4vHk+WwO
|
|
xuh9q3J9dgvbdVukMV1EwbDDgn04rZMwlHoZ+orZ6hfQ3RWVnQYCgZAq+8U0ln5NtBsV2yxYcfgK
|
|
JtW0CnB31LlroVwJ1nQLGDjeP7w+lb0dsFxjrWB0tHS6NuWPJ6A16ToUm63T3Gallr4S7cxiTjrX
|
|
PaxaF7dlVeSMUhxZ5jd+H7qCa4eF3DSE5x3zXN3Wk6jbyeaiFWUY6ZyPStYS5SalPmVipFbX0E4c
|
|
W0alvmPHJrag0rVvEE6LdljGpG2NRtQD+tW5XMI0uU9M8NeFo9PiQhecDIIrtrOMIoG3H4VlJm9t
|
|
C6CB06VPGM1IHLeItGS6uw+ORT7e3jsbQvj7gzUNam0JaWE+HN7NqOqX80n3FO1RXo8YzXdS+BHk
|
|
4z+KyzGPapcU2YIv7qQtiuaxvcaWqG4O6FwfSrS1JbPnrxoxkv7qIfejcitj4V2f2exumI+8+aKn
|
|
xHTT+G5d8Txlm4rjLxMsQwzWT3OiK0Mm6sEkVsAcjFc1d+FEmlGwEDPQVopaEuOpr6f4ZWNAu3tW
|
|
vHpAj5ZQcUFIWaDjGMVUMQ3cVDBmvbhY7QAV2nh+T/R1yeKhlrY31+b61FcQK6nIoJMi401WblRi
|
|
qr6PCw5UYq9y+YgOgWzNkRrx3xWjp+nx2v3FQcelAbmko9anQ4GBUNisPHWr1qMrQhS2K11HvmYV
|
|
hamcxSRZ5xRIqluS/DKAQQXZxyXrvo2FdlL4EeZjH+/ZbjNSZpswLNBrE1Gt7VE4ODVIlnh/j61F
|
|
j4lmeTGyUbq6LwdEqWbeX0YbhSqfEddP4Bddj4JIrhL5d8h7VjI6oLQqKNzelWre3yc4/ClFjaL6
|
|
wqBxxUUxwCKu5BmXRA6c+9ZjP83FSBoQuPs4BrsNBlUW659KmRrDY6G1lyQtW3Hy0lqQ1qVJnAbm
|
|
oy3b9KYJCqRj3o4zRctIlhjLHmpSuOBRbQOpLGpPFaES7UqkZzKN1KsEc87/AHUUmvPLTVGv72aQ
|
|
k7WJwKmRrQ3ud74Ltilgz4++2a6iNDXdS0gjyMU71my7GpqTbxSbMki3SViajTTHqkSeR/GeyZmg
|
|
nQHkEE1S+F+oPPavBL96I4/Cia1udVF+4dVrkW+Fq8+v4tjMDWUkdVJ6WM0cNV+F+MVmjUcZgqnP
|
|
1qpNNnkcVRLiZtxIS1UzzIF7mghlxUZpVQdq6nTVdAoAOKzkbQWhvwM6gMM1twOJYx3NOJE11Kt1
|
|
H1/pVVlwBkk+9NocXoOQ45FPj+fkUJFF2NSB700v/hTEty5ZpkjvVyUgcCq6GM9zC14/8Se6GcZQ
|
|
1574Xs5WkI2HBPHFQ1dm1KSSZ7Rotn9l0+KPHIHNacae1dy0Vjxaj5ptlhVp+2s2CJ9ppCKzuWNx
|
|
zSFc1SYrHNeNdIGpaYw25ZeRXmvheyk0jVpEdcLJ0q3ZxNKTa0O3vQHg/DNcHrsJDmsmjspnNzNt
|
|
fFIJ24GazOhC+azDmgZIOOKBsp3J2qSaZodubq58yQ4QAnmhGT3NO18pb7BORmu205LfYpyKVkWp
|
|
Oxr5gKYWoIZWgfGfloFq1qTPLubnGO1RPtxg4P0oBAkY/hBz6VNDDkZ6AU0W2WSdqkdKr9ZOaGSj
|
|
VtcLHmnOcgmmYvcz7mBLy3MbdD1q9ouiRK6bUAVeelOC1InPlidSsWMDFOCEdq3uefykqrinYqGy
|
|
rFvApMVka2DAowKAsMkRXQqwyDXn/iWyitNQ3qPl6itIvRoF8RXinW4tQ6HI6GuW8SIVBPalc6qe
|
|
5x9x97r3qruwTjrWZ0ksZ9TUmcDNAmZ9/wAoao63rR0+w22MLPtAzt6mghmfofiB76LdJBJBIp5D
|
|
d/oa7bSdWLIPnpDi9TM8TeKdas51XTbIyxd3J/pXS+E/EFxqNoFu7do5OmD60maHWrnZyDRkn/69
|
|
MlEyOR0xntVoNx+FUgYjPxg4FLCuWDZyKQr2RoRnP0qO+nEFpJITgAUzLqZnhu6+0rknOTXpOmwJ
|
|
Fbrt5yMmnHYyr6Oxb2ijaKLnPYMClwKQWK3n0hn+lachHOJ9pNNN0apQFzsY10a4v4hXQh0xpieQ
|
|
MA1XLZNjhK80cT8OdV+3Wl3A7ZZJCw+hrR1qLcjZ/CsbnfHRnFXseHJArOYYbrUs1uPhYbuatqFP
|
|
ByfSkMq3UIINYkto+87Tx6GkSxfsDbflGD7CtTw/pk4nzITtPIFMFudsukh4Rxz71paTpKwP5jcn
|
|
0qTRy0NORMDgVCqewoJTJgAoxjntTiTu7fWmFxAcnn1q3EPl+X8KZMi4gKqB1Peob/Tv7Us5bfeU
|
|
yOoq4R5nYxqT5I8xieH9J1DTbvyJELRg8ODwa9Ms5mSFV9BWiptbnNVrKdmif7Q1KLg96XIZc5Is
|
|
pNL5pqeUrmMtZs0jzV08phchaY00zH1p2ZNxjS1g+LdJOt6U9ssmxjyGp2urDjLlaZzng/wUPDqz
|
|
TSTmWeTrjpVjVk3Rvjr2rnqQ5dDvo1XUd2cTqSNk9OKxXGCeKxZ1DAxHTr2q5C/y8GokUhsz54qu
|
|
uCxzSQjQ0+FZblR2ro4bZYiMVQ0dBb7Qi5x0qzuG5QOh71LYErDufpSeWrHnimIXbjkUjLkH1Hem
|
|
gGxryc+tXI19KYmWegq9YLiLJ7mtqS945cS7QsWehqxA9dEjz4krPSxyZqbFFhGxUm6smjRM55Lk
|
|
HvSvNxXTY57kLT+9MNwKdhXGm5FIbkU7Bca1wMEVhaiuQcVhXWiZ14R6tHGanGBI2OtYkqEHjgVy
|
|
s9ErEeo6UBsHipKEZs5qpPdRxcbhx70NCSuybTNWihc5brW9Fq6vjMnFSdEIdDRi8RRKygZbHFbu
|
|
m6nb3RA3gMegNJhOm0jbXGOoxTuCc1Rz3FyoGKawz9KaAVcZqeMgCmIkB4FaUTbYwB6V00Fuzixb
|
|
0SFMuDU8Mlbs4UPeXHeiOXkUrDuXYnyKk3cVk0ap6HMxxketSMhrcwRC0dMMZFMQ3yzSeVQAeUaz
|
|
9Vj8uPd271nVV4m+GdpnHX67pCeKyLtBtNcR6xlk9RVeWTb3qRnO6trgttyIfm71z7ai8j7/AJmN
|
|
DNqUVa5Yi1AnjynHuBV+11YJhWWXcP8AZNSzqgmaEerSsf3NtIQP4mGKtRavdRgMIpVI9KjU0a7n
|
|
R6T43uYQI7qN2Tpkqciu503VVuQGAYZHQjFVc4alPlZrpKGAznpTwxOc9+lWjIlUACnM4XApiLNk
|
|
nmvnsK0NvpXZRVonmYqV52GsmanhXitTmFkSiJTSAvwrxUxXIrJ7miOfjf1pzNWxkRlqYWpgJupu
|
|
6gQbuahvIxPA6eo4pNXVioS5WmefakGhndH4INZs5DJXA10PaTurmLO21uKpSZqGMoXGnRzBiyjd
|
|
9Kx5rcQS428fSkjanLoaOliHGZFB56VswW+mtPufcBsGOAfmxz+tFkd8HpoaUx09FAtFY8DO71qb
|
|
Sms/Nb7RbecG6AEjFLS5c78t+p0djpVs9wsyQiJAdyr1rW+zqjErzSe559Sbk9S3C+MA1bjbgE1S
|
|
MSXzMVG0vNUI2tPKrAuCMnrVzNd0PhR49W/O2xrHmp4TxVMzQshpIzzQBehqesnuaI5VGzT2bitz
|
|
FEbNTC1ADS1JupgG6l3UAc14s04yR/aYRll+8BXCtLncDXFWjys9TCz5oW7GddH5qqNzWDOgQnC8
|
|
VSuo1kHzAGkPYopEY2+RWxV23Vzj5G/Kg3jWaNazhZuqNXS6TaKhB2c0jR1nJWOlhOxRxU4YkCgx
|
|
Y0OQatQyDbyaaFYe8uF4NY3iC9ltbVGj43NTIL3h7WzMihjzXVQXYYDdW9Cf2WcOJpfaRZ3g9KsQ
|
|
mupnCLIabGeaAL0LcVY3cVmzRHIxtUhetzEjZqjLUAIWpN1ArhupwagAfDKQ3Q1594v0c2bm6tx+
|
|
5Y8j+6ayrR5onThp8s7dzkZjuqAAmuBnqC7c0iwgtzSA0rWzjfGRW3ZadDu4AoNYo2rfS4v7orSh
|
|
05UA2r0pDbsTm29KRottBNyJ0wpJ9KhD7f6U0ikNWffIFBz60zVUW52ow4UcUN6EPcx44WsbgOmd
|
|
ua7TT5Bd24KHnFKnLlZFSN4koluLdueRWvp14swweG9DXoxldHlTjYtzGoo25qzEvwtUxas2jRPQ
|
|
5CNqkLVsYoYzUzdQA3dSFqBBmnqaBhuqhriCXTpVIzxUz+Fl03aSPI9QTypW2/dz0qKNw3SvOPZR
|
|
Mqin8VLKRcs3O4Cuk0w/MDjt1NBtHY6O2IIHY1pxgFaETIRwMkjtVSUEk4570MlFW5bap6dKzWm8
|
|
1tqH8aY+hp2FvGoGayNevVt7/ap4xzUvYjqTLtvLPcvJxSaVcyWsxTnFZlnT2t15xHmCtOBYwQy4
|
|
B9q7cPO+jPPxFO2qLEj5HWo42+aus4HpoX4W4FTF+KlotbHII9SFuK0MUNZqiLUDE3UbqBBupwag
|
|
Bc1DefPbyD/ZND2KjujyPWlKzuPesRZjHJXms9lMuw3StjnmphKDSLTJ7OfE3JrpbO4GQc9qlnRA
|
|
3LO82k5NbFvdADkjBoCSHyXIIIzgVQvdRigT7wzjgUzO1jHknlvG7qnp61etYFQDIpCZoqVijzXn
|
|
3iC8EmsOuaCGb/heR/s0ijkVv6fbxy3QMg5xmsnuX0Ldzut3+UYTPWk+2GJSe+M1pFtamcldalmx
|
|
1eO4XaThhWnC+TXqR2PHqL3maUJ4qRjxSEjj42qXdxVmaGs1MJoATfSbqBAG5p6mgAzTJTmNvpQU
|
|
tzzHXY83D/U1zF5FhjgV5r3Pa6FMsV5HWnLe7RhqBRdmTwagN2d2K2rPU1C5LAnPrUs6Iysbdrq6
|
|
f3gK0BrUKj/WClY05iM6xLOcQAj3NT29uznfKSzHuadzNu7NSBFjHNSm5VO9IRnajqoWMhTzXFtA
|
|
bvUfMduSeg702Qz0rS7FbTToQFwzjJqaGTFyfK5PQViyzUuFmuIdgGABya5u/vTaN5cnUHFUmLoZ
|
|
zyskwlgJweSK6zQdUEwVJeGr0aUrxPLxEfe0OrhPAqVjxWhznGRtUwatDK4jNxURbmkAm6jNABup
|
|
6tQAFqhupNtu59qUnZFwV5JHnWsHdIx96w5lz15rzT2uhRmt85xWbcxMnUGmZlB0bdxmrNvFIcfM
|
|
350mWjbs7YkDJY/jW5ZWW4jikWkdNp9mqYJFaJdEHHakUULu/VB1rLn1Ld/FgetMGYd/qWSQmSa0
|
|
/AemS32pfa7piLeLkg9z6UmQtz0W7uQ2cZx0A9BVzR7cAea6j2rPqX0L99KRat5A6Dk1wOoKZ52a
|
|
YfMORTYRLujiGWEq6/NWza2yKQVHNdOHerRy4laJo6TTnbbtb8KuM3Fdh5z3OJjbmpt3FaMxAtUZ
|
|
agBN1GaQBzTwaAAms3VbjERUGsa07RsdeFpuUuY4jUjljWTKK4j02RE4IpJYFk6imQkVl0xWarsO
|
|
mAEcUi0bNnZBR0rWtoguMCkUi21wI161mXuocEKaYXMS4u+pY/hVCSWSY4HT0pEmlouiSahdpEBl
|
|
mOceleiwWcNjClvHgJH97Hc1EmVFFi3Czy7mwIl/WtJbjP7uLgd/apQ2VNVvtsBhiPzdK5S4nAuR
|
|
nqOCaTGi9pcytPlU+XpmumtWII44rah8ZjiNIXRuWeNvvViQ/LXpJWPJbu7nCRvVkNxVsxBmqJmo
|
|
EPiXca0YLMuOlJsuKuPlsSi5IrNuG8s4HWs5VEkbwoOTKsk+FJY4rC1K53k1xTk5O7PSpwVNWRzt
|
|
4cms+WpKICtSLTETQj5q0YeBSGiys23pUguGxQMq3E59ayrm4x3yaAKiRtO2WPHcmhruKFxFajzZ
|
|
ScA44qRHoXhuMaLpxaUg6hcDLMf4F9KlhuDeXGASIl+8azZslYma68y48m1+7nFW5rtbRNhb5z1p
|
|
iMKbUg0zuW4A4rPgb7VdKXOMmpA7HRbMS7nUYiUda0lkQOBngVrS+JGdbWLRt2bAx5BqeQ/LXpnj
|
|
PQ4GJ+ashuK0MhWaoWcA0AaOmASMK7jRNPWYBmHyiuepO2x10qfcv6vYxCzYqoGK4HVYVTJrmb5l
|
|
c6oaM5TUJ8EgGsG4kLNUHT0M64OaqMMikSRsuKbnFMRLG3zVehOaGNE445NNlnVFpDMu6uie9Vo1
|
|
8z5mOAOST2pDK91cNN+5tsrH3PrW54a06KxT7fdrlh/q1Pc+tJ6IUdZGvHPLezMcnBOWbsPap5r3
|
|
ylFtbdT1xUWNWzU0/Zbwlgfmx8zGsHWtRHmMqE59aAMyNifvHPc1f0gtPdqkY5JosJHeNci2tktY
|
|
euPnNY+oXWZEVJNrZ9aun8SIq/CzodHuriIokhDIR1ronbKZr0o6o8ipoz//2Q==`,rf=`
|
|
/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAsICAoIBwsKCQoNDAsNERwSEQ8PESIZGhQcKSQrKigk
|
|
JyctMkA3LTA9MCcnOEw5PUNFSElIKzZPVU5GVEBHSEX/2wBDAQwNDREPESESEiFFLicuRUVFRUVF
|
|
RUVFRUVFRUVFRUVFRUVFRUVFRUVFRUVFRUVFRUVFRUVFRUVFRUVFRUVFRUX/wAARCASwBLADASIA
|
|
AhEBAxEB/8QAGwABAAIDAQEAAAAAAAAAAAAAAAEDAgQFBgf/xABDEAEAAgECBAMECQIDBgUFAQAA
|
|
AQIDBBEFEiExE0FRBiJhcRQjMkJSgZGhsWLBJDNyFSVTY3OSNEPR4fAHFjWCokT/xAAYAQEAAwEA
|
|
AAAAAAAAAAAAAAAAAQIDBP/EACARAQEBAQADAQEBAQEBAAAAAAABAhEDITFBEjJRIhP/2gAMAwEA
|
|
AhEDEQA/APqYAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAKNTq8OkxzfNkisQC8eb1XtRNbzXT4q7eU2nu0MntRq/D8StMccvW29ZmdvgjsTyvZjxOLj
|
|
+s8WLxn8TFPXs6Oj9oct7c14rkxz22nrB2I49KOdTjelmszfmpMeUxv/AA28OqwZ4icWWtt/SUi4
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAmdo3nsPNe0Pt
|
|
Fh09Z0+DNWL7+9O/7A3eJcZppsV5raI27esvH6jX5ddM25p79Ilo59VbUZOe2Tm/PeGvfPfT2iKR
|
|
PLv1+DO678XmW/a97U6TtOyzTbTF538/T9WjTNecm9a7126tqk3rSYxY5ta1plRZqZNXGjyZcPXl
|
|
mZmsx+qjBrsuO16xM7eXRt04JrdTltk5OWJnfaWf0a2lty5MdZnfzSn+WOHiOutFpjHa9e8bQ2fp
|
|
+alYy462pk7zXbuxjPesbRS0f6ZZV1ET1tErzXFLHo+A+1ddZf6NrI8PJHa1vN6iJi0bxMTHwfOa
|
|
zhzd61v1846utwniM6DUdb3nBaNrVmd9vjC/ZVePYirBqMWppz4rxaPgtEAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAItaK1m09ojcHnvarjM8P0vh49+a/eY8ng9D
|
|
h1fGM1rxjtGPfvbzdbjuTJxHX48cTPNltM/KsS9Dw7S49Jp6UpHaGe2vjz1y9J7LYK13vHWe7bj2
|
|
ex1tvM80ekuxW3RnW3Vm6P5jRx8H0+OYmMcb+bapo8GKPdpC6bQwtdHU8JpWkdJ/JweL6e23iU67
|
|
d4dubSqyVi9Zi0bwIs68XGp36TtEq7ZJmZmevzdbifCKWtbJinkt6eTgZPFw32t+sRurbWVzxs1y
|
|
Rv6T8V1NZNPtfq0seTm+Kevr+SZuxXjvaPiV8N4viycto9HseG6+uu08W6Rkj7UPmFck1tE1nlmP
|
|
Ld3eA8V8HVVi1pjq6Ma/pnqce/ERMTETHaUrKgAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAADW19+TQ5p/p2bLS4v04Zmt5VjeQeJ4bjnLqsupv+Ka1+ERLv4reTmcNxcuC
|
|
vy3l0qdI2hlr66sT02ot0ZV7qqrInruzrVZLGSZ37JjqgYTG0K5lbaFVhDT1Ub456RPweY4hixWi
|
|
eSdpjvD1eWejz3FNHWYtkpvFo9EIseb3tS3SerOms22rfpPqZKzvvHSYUz70TExG6Gdbs2rljeJ/
|
|
Mx5L0vEzPaelnOi98c9J2bFNTFpit47+a+PVUvx9T9nOIfT+GV5p3yY/ds67wvsXqpxau+G09Lx+
|
|
r3TqrEAQAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAADV4ljnLw3U0jvO
|
|
O0fs2lWqyUw6XLkyfYrWZkHldBEV09eveG3Fq1mI3jd4vPrOIaid8G9MP3Y38k6fNrt/rMk9Ou8s
|
|
tfXXn49rGWInuy8SO/k5Gl1E3rG/fzbOe94wTy99mbRvTrMOOvNfJWsesywniukrG/jU6fF43WYN
|
|
TmtEeJtEQ06aSmK2+bNtEd+qfSO17unF9Hmvy1y13XWyVmN4tExLxVK8PmNq5NrT58zawam+m/yc
|
|
0Xj8NpRYSvQZ7xEOdqI3rPozxayNRXe0ct/ON03jmrKB5nV4q1yTO20Obmv4c+cx8HoeI6WZpNoj
|
|
q83niYmYscU0r8aJ6T1n49zeJ+Meqm1drb9J+Kd5p136StGVem9l9TbHxLDFp7W7+sS+q1nesT6w
|
|
+PcAzVjiGHftzQ+v4f8AJpv6On8jH9ZgIAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAABp8VrW/C9TW0ztOO3b5Nxp8VmI4bn37TWYB8f1HFtTfUfR9FWJmsdZ9I7MtJxDX5s
|
|
d8ta1y0xzteaR2277rcuhycP12SceLxMeWNpjttHwlu8I0mfQ1y+D7k5YmJmY36T36Ka43z/AF1t
|
|
cI1ds+qxVj7/AEej19PCw9HJ4NoK4OIU5Y35YmZdzVTGebVZabx5jJS+Tmns81rNLm1Wrzc9rVw4
|
|
Yibbem72mXTTS0w0M3BvEta1bWrM95ie5EanY87wXgNOL6XPfxraXLhra/W28bR/dzYzarBqJxRe
|
|
bzE7Rt5vWU9n8mPHOGmS0Ypnea1naJb+k9ncNLR7u2y/WcxXO4TOoyUrN6zD0FaW5Y3hu49FiwUi
|
|
KxCvLMR0hlW0jn6ukWw3iXjOJzbDlneOj3GaN6zDzfFOH+LE7SRGo83XNSZ2lbG2/WfdlvaT2cy6
|
|
rNFInlrv1mfJ37cK4PwTTxOoidRm2+/2/KFuyMp47XB4LivXiunrH2b2iH2qn2K/J8x4fGDNxTSZ
|
|
9Nh8OviRvTyfT6xtWI+DeXs9MNZubypASqAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAOZx6/LoOWPvWiHTcf2hiZ0e8fc2mf1E5+vP/AEeuSd7RC2uKtI6QjHfeINTfwtPf
|
|
Jvty9WPfbt/lucP03gxfJf7d/wBoReYpm97zaNeLb4Ims9Nt94auDjem1Wo5PFi1onylS+1o7l8V
|
|
bxvtupjDMdNkYtXS1+Stt+m63xImEJ4xjHER2ZxMUjeUTO3VRmydBbjLJqPi08mbeVOXJPq1sl5Q
|
|
Vbkz9+rRy35rxHqzmZlVEe/Ez5LRlW5iyfR6zffaIjq1OSNZps2a21rZInafSPJhxGMl9LStLRWM
|
|
lorM/A4dkrWbYfLZC2W/7K6eubX6b4RzT+W76K8b7G6X62cu3Sten59nsm3j+OXz3/0ANGIAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA0OIYfpOHPijvNNo+fdvtXJO18k/
|
|
/OwPFYbz2ls3jx8VqW6xMdWPEdP9D4lkx/dt79flLLHbkxTPwY6nt2512ORTRzE2x4/dpE7cvkme
|
|
E4IrW3hRMxO8THRtU1FKWtvtvK2upx22rzRCtXkqzh2jtF7ZbT122b01ndnpuWuP3Z3+Ky20qDVv
|
|
fauzVy3mejZzNK8dVjqi87KLRLYtXruqvXzkQp7Qoid88R6rcl+WGlW0/Sa22mfhCZOq2x082ix6
|
|
jkm822pO8VrPdr4dNObVeDo8XW3uzMbzK+mvxT7szE27cvnu9j7PcNjSaXx8mOIzZevbrEeic5tN
|
|
+SZnpt8J4fHD9HXHO3PPW0x/DeBtJxx29vaAJQAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAKNRim9Z5e89Nl4DzXtVh5babURHrSf7f3ec1+qnDorWrvvt5Pccb0n0zhmWk
|
|
Rvevv1+cPE2rGTFNZU26PFfxwa5dVkjelI2772nZnX6bbrEUq3o0d678u8wmuDL2ittvVjXdneeK
|
|
cGv4jpJ6U56+kS7+j118+GLXpakzHaWlp9NNY3tv+bbiYiNoQy1y30uyZJlrWmZnuym6q1iIJnop
|
|
yW2Te8bdWnnypQqzZOadokiIpSZntWN5lrxki19vNRxrUeBwnNNd+fJEY6/OejXLn3Xe/wDp9wyn
|
|
E8uo4lqqxblv7lJ26T6vpD5X7G8QycKzeBMbzMRM1/FH/wA/h9QwZ6ajDXLitvWzRgsAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAeL45w+dDrZvWv1OWd4+E+j2jX
|
|
12jx67TWw5Y6T2nzifU+rZ1y9eHwzDYxxEy18+DJodXfT5o96vafWPVbjyxDn1OOzHudbM0rt2UW
|
|
iI69mVtRXZq5tREb9VUoy2iIlRbJ0UX1VZ6btTLrI7V6yk62M2oisT1c7JmtkttVMUyZp6x0beDS
|
|
RWOvdKijDimvWd3G9pNRMfRcNfvZOb9Hpb0itJeP47k/3hgjaZnbaP1XxWW3T0movbNS0W645nbf
|
|
0nrMPpXs3xamoxdJiLbe/X1n8Uf3fKsOTw4jbaXo+EarJhtGTHMxeJ6xH7Sti9Zaj6x3HM4NxXFx
|
|
DS1mtoi8dJrv2l011QAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AGjxLhODieOIye7kr9m8d4eM4to9RwjPXFa0ZIvG9bR0fQXmPbDFvTTZPOJmEWS/V8bs9R43NxLL
|
|
G8eFbePg1bajU5/s0l1ceKLx1hbjwRE9mOpx0y2uRTSZsm3PMw2aaKtIjo6kYo9EXpET0hVLXxYK
|
|
xC6MZvyx1lFs0RHfaPiCnU12pLyHGNDbUajBekWma2npWN3p8+opa20e9LSyZLxExTlpM+vdOdcZ
|
|
a9tPS8MyUvFrzWlI6727u1pYxYrbVmb7x+TQx6au3Nqcl7/0rcmW9axGnwZJj1novmxnZXV0fFp4
|
|
ZxLBPgTGK8xzXr5fOH0bFlpmxVyY7Rato3iYfNuG2x56Wrqa8s2jz+7Lu8O12bS6jkwzN6THNNI6
|
|
tvrN68Y4rxlx1vHa0bskAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAA4XtTTm0OKfTJ/aXdcL2pyRGjwU362yb7fkJz9eTxxyZJjyltRXzUZK7TFtl9Lbwy06YzrHwa+
|
|
fJFd/wCVt8m0bQ0eS2qzcm+1K/an+zNZFL5M1pjFXeI72ky48eGnPkvNp27+TPU6nHpMfLXaIjpE
|
|
erk5dRMxOfN1mPeisfshW1ne1a1577Y6x5R3U0zze31FOWI6ze0byU098kRlzbxM9qrMlPDpyRMR
|
|
Md5Vt/Ihp5898mWZm1pjftE91uCt7fCI7dWeHDEW3t723l6rslqxWZnasR+SYhFbzhnfxJ2jyeq9
|
|
lcGXWZcmW0zWKxHLaI7794eJx5fpfEKabT8t8l5isddo3l9S4VjrwrRUwzSJt3tav3pdOL6Y6dXD
|
|
j8HFWm+/KsU4NRXPvtWazHquWVAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAa+fXYNP9u8b+kdZBsDkZOO135cWOZn4y5Wu4xqctbe9y19Kp4njt6vi+PDm8DFMWybbzPlV
|
|
5PiGtz67UxbNbeKTtWIjaIXYpnwuaftT5tXJT3vmi1pMsrU5qIrG1V1a+5DCa7b9GFbRr5J6Wnbt
|
|
Cu+Wmk0m8956z8ZWZNorbfzcbX5rZslazPux3hUt41NTntktObJ13+zX1bek01r4/HzVm0bxPXy/
|
|
+bNfDgjVa2uOY92kdfg6ufJOKvLXtttVVSqbcta2vM7zXtHpLQy5ZtMd+vWd+7Zy3mdJHXra3f0c
|
|
vUarw7zFY5rT2hH1Lavnrgx81p3U49Pk4nE5L35MO/StfNRXR5tXnrS8W67WvfyiPSPi7uLHFK1p
|
|
jrtSsbR5Lc4RzsXBaYreP4l45esRD2HD9fnw6evvWvO3Tfr0aGk0U55ra0TFInv6uzgrXFXlx0i0
|
|
77RPlC83Yj+JW7oddqr6vHzTTw9/f6dod+L1t9m0T8pcbFSmPHER3892W0zPuz+jSbVvidkcqmfP
|
|
Sel7bekrI4n4dZnPWIrHeYnZee2Wpy8dEaml4npNZblw5qzb8M9JbYgAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAABEzFYmZnaI7yCXL1XGa0jJXT0571nbee27DiXEprp8nhbxG20W8
|
|
5cbD0ikfnKO+urTPvjoZdXqctdsmTaPSvRpWmsdZ6yztfaGplvv3lWW1tyRlz1x0vkn7Vo5atTNe
|
|
Y0+1o79V2KsZsvX7Ne5mwxnyTNvsx2iGneM/rCdRSuOsTasTt5kRFtpjqmOH4t4nk7estiMNa97R
|
|
Hwhna0iuKTEdmGWa4672nZtRele1N59Zlq6vLOSsYorEc07qcW65euzRvtXvPZy52naZ7ujr6fXV
|
|
rWdukREK8+njHgmZmPc67bq6ivVWhxxgxZLztNrT1mZ/SP4VZs0zaOvfp84WUtNsXLvtv3699+rU
|
|
z7+Jtt5qURqMnPpctaR1rMSw4ZoK57eNk6xHaJRh97Ltt7lo5Z+L1HAPZvVauZ2nFTSzMTzeJEz8
|
|
to6xPfvsZntPZ9rXxabmxzefdrv0j1dXh/BcmstW1qxTHHasR3+b0GPhGl+kWmd64dNEVjf73T7X
|
|
y8vy+Ddx6O3iRakxTH5RXrMw1/lX+3Itw2MFIraN48qRHdZi0cUjmmPen9noox1iO0fNzdXEYrTt
|
|
stcmd9aX0bJ+HePmiKTitO8TMLZ1cVjrMfqpz6ys4pjfrPRWZ9rXXptUit6zO+23VyaRHEc05L1/
|
|
w9J9ys/en1ljqdVbwYw452tlnl3jyjzbmmiMeKtYjpEbLeTXPUU8ee/+qjJpsV5rbkrFqzE1tEbT
|
|
DpYNbW21Mnu29fKWna0KbqTdjXXjld0cvQ63ltGHNPSfs2n+HUbS9c2s2UASqAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAOVxPWe99HpP8ArmP4b+r1EabT3yT3iOkesvMVtN7za07zad5l
|
|
XV5GmM9vVfEstvDx0jtaVVMlq+UJ18b5cMRvPeSuK87bUt+i2Z3PtG7zXpjkzXt6R+TXyTMzvM7t
|
|
ydHqZ+zhv1+Cv/ZuqvPTHMfOYaTMil1a1K2vHSLTELq2v+KWzThGo84rH5rq8JzedqR+ZeI7WnOS
|
|
34pYTafWXR/2Pln/AMyrKOCWnvmiPyR6O1y9585lhWJvl557Q6eo4T4dYiMvW3b3UanhldHpJtGX
|
|
e09unmjsT7eb1l4trI2t0hsZfrdNO0bzy+nzU20/+NmkzO9esz+TZxWis9dttvPv+Tn21jjaW8zn
|
|
26bTG3mp1M/Wzv3t0jyWXiKZJmsTERaZhXXDbNl8WaztWenxZLstPp5pau8frDtVrNMM5cfTfpMf
|
|
3aunxxbes9d/R09Dp8ebJi09ptFr3jtt2WyrW9wy1Jx132mK+Xq9PotT0iIU19ntLtExa3T47T+q
|
|
6nBaYvsZstZ+cT/LeMnUi0TXffo1s2m8Ws2/OIMWk5Jib5L328rS2t94Sh5TV4ppklpW6PT6rh+P
|
|
NbebTHyas8E081mZy5P2W6OFhjxNTE/hr/LoRO0Kvo9dPqctKzMxEx1la5t3tdnjnMs4noievcrO
|
|
yZjeFF1OSnNV0OG62cn1GWffj7Mz5w05joovzY7xes7TE7w0xrjPeex6Ua+j1UarBFu1o6Wj0lsN
|
|
3JfQAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAACrU5o0+nvlt92P3BxuM6nxNRGCs+7Tv8
|
|
2hToxm1r3m9utrTvMsonqyt7XTmcja0u3O6FMfi5t/u0/lzdJM81p9O3zdvHTwsUR5+bfPqOfX1h
|
|
dqV+3O7bs1+T31oqmI3TEM4rvCdkDGIIhlFd2daboS0NXG2bD6bufxXU1vlmu/u4us/N0+L1tTSx
|
|
kr9qk7w89j1FNZMV3jxLzvaJ8mer+LSOZqK2xZotbvljfr/89U453rXt9lse081xZtNjx7TGKu0t
|
|
DHlrevSevaN5Y6+tJ8c7VRNMt63n3ub+6/R54rERMztDYy4a5omclYmfxKcenrjtHLvtPrCnVmdb
|
|
eFe3JXmjy6eS/DrMuLVYsta9Mdt++6qLxO+0dEc8UmInr18iUfReHcXrqccb9Z27Q61Lb13eJ9nc
|
|
1Z35rTvE9avY4bTkpG8xEfB05vYxqybc07R281naGMREdoT5JQqy9mply7Q3bV3iXG1eXw7TWSka
|
|
c258t7+tpT5/BjT7MfHqndz12Z+M4lMMKyziUJJiN1WSu9fku23RaOgKNJqbaTU1t9yelo+D0cTE
|
|
xEx1iXmM1Nt3W4PqvFweDaffx9vjDbGvxz+TP66QDRiAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAOJxzU73rp6z296zsZMkYsdr2naKxvLyObNOfNfJbvad1dXkaeOdpvsc2yuZVzfbfqybutwu
|
|
s5s8R92J3dvJb3tnO4HSMegtmt3nfZvYp8SZl0z45NfSK7onH1bNcfRFqnUKJr0Y7dVtq7prjEsK
|
|
0XVpEM6028mW20IHK41aPo3J6zs4ODhdcvPnvExFevNXpMOrxi/PlrTee7PLX6Pwa09uaNlKtHg9
|
|
dM3z5d7ReOu02nu0JzZMfblrv5R5uvrcdImZ26T1mYhxs1Os7RH93PZ7axuafNfLitvbaYU3yZYt
|
|
PXs9NwHhui1HBa5LVicsb81onrEuVqNNSuS8Y67dZ6xPZa59Il9uX41vEitImZme3q2Kxbxora0T
|
|
Md/ROSa4Ztkj7c9OafL5LuGYubmyX3iu/TfbdSfVnpvZLT/XZK233+Mbbva1xRXyiPk8pwbH4N6T
|
|
adq5a71n0tD1WDL4tPe6Xr0tDpz8YVnJHWEXYxbqlBedoef4tW0XraO09HdyztSZcbUz43C+ee9b
|
|
SVMaeOfqq7+jGckQ1Yz7+7v2RN/WXPXZPjci2+2yyJaVMuy+uSJlA2d+pNoVRbeDcSxyTE+TDDlt
|
|
pdRXLTynrHrDOyiyZeVFnY9TjvXJjres71tG8MnJ4Nqt4tp7T1jrV1nRL1x2cvABKAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAHJ49qfD09cNZ97JPX5PPw2uI6j6Vrsl/ux7tfk1mWr7dOM8iLdm
|
|
vfebREefRsWldw7SxqNbWbR7lPesrn3Vteo7dYjDpMGCvfbeXQ0uLlxRLRxROfUc34p6fCHYrXlr
|
|
EejqrjY8uzCYW7MZjdVKqK9VlaxCYrsnYExBMRMJRPZA8/xPHtmpP9W2xx76vhWOInvt/C7ike7N
|
|
vwzE9kcapGfhlevTaFbFo8RqJ5vy8/RoW09ek0msxHfp3dzNoLzp4zUmZpMbT8HJyYJi20X2n0lh
|
|
ZY1li/RaidBF4w2mK3jrHaFGp1lN+tptPp5IjBkid5mIp16TKu0abBPv33vPlM7z+iPdFNcWXU5I
|
|
tkrNce/b1W5db1nTaf3ax9q0fxDW1ebNk2phty1mOu09VOm8W19orEz23j1TwfSeERFuEYMddptW
|
|
d43dvBn21eKJ75KbW+cf/JcTgMxXTb3nbljz+TpcPmc2uyZO1KRtVtGVdi0bx07qJnllsRO6rNTe
|
|
N4XVamsy8mnvPwc3R2jPwe8TPbdlxXNOPSZfhWWpwO85OFzv57qrODkzeHntSe8Sn6Rv0a3EZ218
|
|
8nXekfr1a0ZLVnqx19dWb6demXybOO7lYMvNMdW9S/VVLo0us7tPHdtUtEwJiZU3jq2Jhham8CVG
|
|
PNODNTJXvWd3qcWSubFXJWd4tG8PK3pPd1OB6veLaa89Y61/u2xfxh5c/rsgNHOAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAANLimq+i6O0xPv392rdeZ4rq/pOqnlnelOkIt5F8Z7Wj27I2I6sb25YY
|
|
V1ImY3dbQ08LRc23vZp2j5OJG+XJWle9p2h6HHtbJXFT7OOIpX+7TxT31j5rycdTh+Dpz+XaG/sw
|
|
w18PHWseULN2trBE9UcrJKBhFU7JAQi0dEomegNDUYovM7x3jb5tO1ZvpbaTLtzRExWfWPJ08kbT
|
|
Ex5NXWYYyV5omYtHWJieyeDzuizfRs19Jn6TM7Ru1uMcJxZqTkw+5f4ebqa7SV1MR4tdrx2vEfy1
|
|
axqsNOTLjnLXytVXi3Xj8+nmsxTLM16d5npPyUzpekTtSK+U7vS6vQ/SYmK1vWPS1HOn2dvvvvE/
|
|
tDO5XlcO+LbfHSd/W3o6/BdDOXPTnj3Kz38rS6Wm4FNrRyRzTH3p6RH/AKvR8L4dXSzE3jmtHn5I
|
|
mbfqLV+m4dbLSsZInHjr3iI6zLpYaxS01rHuxHRHiT9mv6s67Vj1aqL6326MrWiYa+/Q54BxPaGe
|
|
XRZpj8MquB4+Xg8zPnB7SX30to379GxpK1xcHiKz5IS8xr8PLPixH2bftLTy05o6dHYyVjLhy0t1
|
|
izjZa3pMVv3iO/qz1G2L+NbSajbNyW7xLsY8kTDz+fJXFqKZN4iZnafi6WHL0iYlStI7OO+7axW2
|
|
crFl7dW9jvE9ULN+J3ZbdFGOy+AYWpEqN7afNXLj+1Wd23KrJVMvCzseh0+auow1yU7WhY4fCdV4
|
|
OadPefcvPuz6S7jol649Tl4AJVAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAV581NPhtkvO0R+4NPi2
|
|
r8DB4dJ9+/7Q83Po2NTqLanNbLfvPaPSFDHV66sZ5ET0hRknyW2lTtMyouz0c8usx2n7s7vScKwx
|
|
zc1vu/y85p+maJh6Th+SOWeveXR4/wDLm8v+nX5mUWa9bbrInolmu5jdTNkxYFk2Isr3TuCzeGMz
|
|
+THdEyDDJO9Ja823rt2XWnya946pGvktDXta0ztWu/ybvLE9dkcoOf4GbJPWK1j49VmLh9JtE33v
|
|
Mevb9G7WsW8l1ccREISophiJ2jpDYpijbaOjOuOJ8ujOdqxsgVcsUjaETYvbaFFrgu5lVsm0yUtu
|
|
ryg43H5m+GIj1XcJzePoL4pnrWGtxmfchr8JvfHS1622if3QljzTTLes+qrNjrkiYtCzPMxnm095
|
|
YZJ6boS5teB49Tqscza97VtvWvlv8V/FOF34RrIxTM2xXjelp/eHoeA6XnzReY3ivX/0dfivDcfE
|
|
9HbDbaLx1pb0lOs+jO7K8Lis3cN+0NKcd9PmthzV5clJ2mF9J9GHHVL108dm1SznYr/Ft0tuhLb8
|
|
mNohFbMhLWy0mJ3rPXvDvcO1karBG8/WV6Wj+7kWrvDDBlvpdRGSnbzj1hpjX4z8mOx6UYYstc2O
|
|
uSk71tG7Ns5AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAACZ2jeXneJ62dVl5KT9VTt8Z9W9xbWclPo+O
|
|
fft9qfSHEU1pv48ftYST23ZTDC/p0YtlVuvVjMbM5+LCZjYGWGdrTPxiHY4ffaf3cjTxz1v6xMS6
|
|
Olty2iXVj/Dk8n+ndrkhnGRo1v8AFdW3RCrZ5uiYsqrboncSu508yjmZRYQt50TfowYTbYGVrKrT
|
|
uTZjvukQnYhMIGVY2ZxPVWyrHVCWzXpVXkt3TE7Va+W4K7X3jv1auTNy3jdba0RZpamfroQN7Hk3
|
|
6wr1GTaN2OOJiu6Mu98NvgDi8Wy74d/yZ8PiPAiO2zU4nb6qIn1bugjfFE/ASp1ke9u15mbbRDZ1
|
|
Mb823kx0Ontn1OOkedoJCvT8I03gaKsz9q/WW+isRWsVjtHRKyrhe0XCfpWL6Vgr9fjjrEfeh5fF
|
|
feH0V5Dj3DPoOo+k4a/U5J6xH3ZZ7z3228evytOk7NvFbo0cdols47bSybt7HbddHVqUs2aW3Qnq
|
|
xVeu8LILR3SlZw3V/R8nhXn6u0/pLuPMXjeHT4Zruf6jLPvR9mZ8/g1xrvpz+TH7HUAaMAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAABRq9VXSYJyW79qx6yvmdo3l5viGs+maqYrO+OnSvx+KLeLZz2te1rZL2v
|
|
ed7WneZYWnZl5K72YV1xEyxmeqJljzIEWlVkszvbZp5soN3h2SJz3pP3odCnuWmPRxuERfJrZmtZ
|
|
mtY96fR28kbX3dXj/wAuTyf6bmK+9YX1s0cNtm3Sd4LFY2K23W1s16StiUJW7bp22RW3RluBuruz
|
|
mWEgrmCGWyNkoExKE1QlPmsqRDKeyBjaejWy2W3ttDUyz1QKslvehVqKTNosyyTvELabXptIJpaP
|
|
B39Ia2mz+JGpr51jdZefDx2hzuHZObNq58poJaGtjxJ2+LoaKP8ADRPo5+T3skx5OhpOmC0fBNQ0
|
|
5yTbn+bt8A0u9raiY6RHLVwY62mI6zMvaaHBGn0mPHt1iN5+aYVsACBXqMFNTgviyxvW0bSsAeE1
|
|
mkvw7V2w5Ote9besJx2er4rw2nEdNNekZa9aW9JeQjnxZLYskTW9Z2mJY7zz26fHrrdpbZsY7NGt
|
|
mxjvso1b9NmUwpx33XRO4K7VUTE1nmrvEx1bVo2VWiJE/XY4frY1WPlt0y17x6/FuPM0m+HJGTHO
|
|
1qu9pNVXVYt46Xj7VfRtnXXL5MfzexsALsgAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAHM4jxOMFJphmJv529Dq
|
|
ZLfjDjPEIx450+K3v2+1MeUOHSOWFc3nJkmZnf4yujpVlqunOeFpV2nctLCZUXRM7MJtsWlRkv3Q
|
|
ky5NmpWt9RnrixVm17TtEQnJabXisRMzPSIew9n+CRoccajURvqLx5/chfOest642OGcIpoOG2w7
|
|
ROW9d72+LQvXevyejcPUU5M+SvpLeOataraw2a0dLbLqTtK1G3Es4lVWWUSoldFtmcXUbpidgXzK
|
|
GEW3TuCUSncnsDFMMLSms9EC6J6FpVzbZE5ALy0809ZbFr9GtfrEoFMzuuwz0Ueey3HbaBLDXe7i
|
|
tMOfwWnP9I+NZbuttvhs1uBRtXPb4SDm3iIvf57N7Dbl0VrS5+XrltEd+Z1Jx7cNms9N4TURRw3T
|
|
+PrcO3WszEvZOD7P6aYiMlvu16S7y1QAIAABxOPcLnUY/pWCv1tI96I+9DtgmXl68Biy7/NtUu3+
|
|
O8HnFa2s0tfd75KR5fFyMWTdhrPHVnX9R0cd21S3Rzsdm1iuqs256wrmGcT0RYSx5d047X02SMmO
|
|
esd49YRE9WcdSXhZ2O1p89NRji9J+cei1xMc3wXi+KZj1j1dTTaqmor06WjvWW+ddcu8XK8BZmAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAMMmWmKu952UZ9XFZmuP3revlDTtzWnmvO8q3XGmfHb9ZanV3yxtWeWn7y4es
|
|
vPNtDqZJ6Ts5mppvdl/XXRMyfGvSNlu/RVvtOzLfoipLT1VTKbSpvfogRkvtDVyZOhkyvQcA4Dzz
|
|
XV6yvTvTHMfvK+c9U3rkW+zvA/D21urr789cdZ8vi9KDb45rejl8Rry6iJ/FV1HP4vXbBTJEfYt1
|
|
+UpiHM295bXsqrO9l8QkZ0lZEqqLeyBZHZLGvZkhIndADKJ3TMoqWQMZ6pjsxll2jsCLSrmU2lFY
|
|
36gieyu0LJk3jbsga0wdqzK20QpyztQGprL/AFMrOE05NLkt6qdVWZxNrSe5o9vWBLiUjnzXn0vL
|
|
q555dHt8HOwV928/1z/LpzXxbYccRvzTB+jucOwxh0dI22mY3ltIrHLWIjyjZKyoAAAAACJiJjaY
|
|
3iXleM8InR5J1GniZw2n3oj7s/8Ao9Wi9a3rNbRE1mNpifNFnVs65XhcWTdt47bnFuF24dm8TFEz
|
|
p7T0/pn0a+HJux1OOrOux08d1ndqY7tillVkzExLOk7yd4YxGwluViJhE45raL0na0dtlWO0+bZr
|
|
1TKi+2zptZGTamT3b/tLacvJjiY3XaTWdYxZZ6/dtPm1zrv1z78fPcbwC7EAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAABhkyV
|
|
xUm152iAZWtFazNp2iGhm1Vss8uP3aevnKrNntqLdelI7VRHRnrX/HRjx/tZREVjZXeybW6KbWZt
|
|
pCZ6S08tN7Nmbb7zCrJtyoS5145bSx5mWafelr3tsKmS/o08uXyhlly7RPV2+AcBnPNdZrK+53pS
|
|
fP4ytnPVda4y4BwHxOXV6uvu96Unz+MvVxG0bQRG0bR2G0nHLb2gCUDX12LxtFmpHeazt82wT1gH
|
|
mMN4tWs+rcr2aEV8DU5sM/cvO3yb+O0csLUTSdrLphRE8tlkZI7Atr2ZMazDJVKTYSCawi7Ksq7z
|
|
1QERvLK3ZGPrKbyCrbdnMcsbeaa18/RhvvM7oGEwTG0JmYYTIML22a2e28xELM19oURPNO4lOem+
|
|
n3ZY5+prVnMc2GYU4/L4A0a15cNf6rz/AC6fC6+NxCPOuOu/5tHJTbHj+F5/l1+BYumXJMd9o3/d
|
|
MRXYASgAAAAAAABhlxUz4rY8lYtS0bTEvH8R4ffhmo6bzhtPu29Pg9mq1Gnx6rDbFmrzVsizq2df
|
|
zXkMWTeIbNL7tbXaHLwzUctvexWn3bmPL8WFnHVL326VZ91MfFVjvvVlz79kLrcf2m7j7bNHH3bl
|
|
J2SirLQoy4t1++7G0dBC/RanxI8PJPv18/WG241+alovSdrV6w6mDNGfFF4/OPSW2b1zeTPL1aAs
|
|
zAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAVZ9RXBTe3WZ7R6iZOpzZq4ac1p+UermZMl89+a/byj0Ra9815ted59PQ32hlrXXRjH
|
|
DpCLX6ML5NlNsm/ZRqstfdXzbsZt06sLZNvNB1Za8RDWyZdo7q8udq5Mu/mIMt4md2lmy7JzZuWJ
|
|
dHgfBL8RvGo1MTXTxPSPx/8AstJ1XWpIs4BwSdbeNVqq/URPu0n73/s9hEREbRG0QUpWlYrWIisR
|
|
tER5JbSccur2gCUAAAAPM8Sry8Uyz67fwuxbzVPGsE49XGbvF42V4M0TEL33ERnktsxpk3sumK2j
|
|
admFdPFZ33VS2Mdui2J3UU6LYlFSsN2O5NkCyJ6K7T1TEsbAsxdpReerKkTFGMxvYEz0rsqtbbpC
|
|
b2VT1QEzuwtbaGUxspuJU3neWdKoiu8rq12gCI92YatLcublnzbEz1aOptyZqTuDHLfxN6R0+t5X
|
|
qdJhjBp6UiPLeXl9NSMnEKxHa1+bb8nrlvxUAAAAAAAAAAABTqtNj1eC2LLXeto/R43VabJw/VTh
|
|
ydY+7b1h7ho8V4dXiGlmvbJXrS3xRZ1fGv5rzeHN02bEW3cys3xZJx5ImtqztMS3MeTeGFjqlb2O
|
|
8btql3NpbZtYsnSBLeiWfdTjtutid+ghherHS5p0+f3vsX6T8Fkw181d4lMvEWdnHaGnw/UeNh5L
|
|
T7+PpPxbjdyWcvAAQAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAo1Oprgr63ntAmTqdRqK4K9etp7Q5d7Wy2m953lNrWyWm953mVd77R0
|
|
Za1104xxlN9lV8qnJl2a9s3xUXX2ybsJyRDWtl3YWydEC+2VRkzeW6q+T4tbJm+KRdfK1cmWZnlr
|
|
vNp7RC/R6HU8SycmCk7ed57Q9ZwvgOn4fEXtHi5/O9o7fJaZ6z1uRyOEezVstq6jiEbV71xevzer
|
|
rWtKxWsRFY6REeSRrJxz22gCUAAAAAANbX6aNVpL0npMRvWfSXlKamsRMVvXm+EvZXjmpaPWHzfL
|
|
oNRjzXicfWJ8phfPxFejx72x7xMzK+sXiNoiXlq+Pi6fWV/VfTNqfLJl/WTg9Pji8R70LqvMV1Gq
|
|
j/zcv6yz+lanzzZP1lWpelTET6S81Gp1P/Gyf90s412rjtnyfqql6asREdWM9+jz9eJ6yP8Az7uh
|
|
odZqMt458tpB1JvEViI3/RhzRt13/R1MNaziiZiJn5K9ZNceKZiIiQcu/WekT+iYrWI3lzdTrs+8
|
|
8uW0fJzcur1Np/zsn6g79phVaIeetqNR/wAXJ/3SwnUaj/i5P+6UD0ldonum161h5mNRqP8Ai5P1
|
|
lNtRqJjacuT9Qd22WN5aGeZyZd/KHJy59RHbLf8AVq31Gp/4uT9ZEvS8Lr/vSs2npzRtL1z53wK+
|
|
oza/HW2XJNd99pmX0Rb8VAAAAAAAAAAAAAAcHj/C5yV+l4I9+v24jzj1cLFk8nu5jeNpeW41wmdL
|
|
knU6ev1Vp96sfdn/ANFdTrXG+eq1q5F2LLtbZoY8m8d11bbSydErsYsm+zZrO/zcnBm226uhiyRK
|
|
EtrvCrJDOJTeu8A1MWX6Lqq5N/dnpb5O5ExMbx2cPNTeJb/DM/iYPDtPvY+nzhri/jDy5/W6AuwA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAa2p1UYo5adbz+xbxMlvqJ1OqjDHLXree0ejmzNrWm953tPmTPWbWneZ7yoy5YhjrXXTjH8s75N
|
|
mtkyxt0VZM2/m175N1V03yTKubMLXVXybeYLLX2VXy7eam+b0bOg4VquJW+rry4/O9uyZOq3UjVm
|
|
9r25axMzPaIdvhns1kzbZddM0p5Y47z8/R2+HcF03Doi1a8+Xzvbv+TotJnjDXkt+K8ODHp8cY8N
|
|
IpSO0RCwF2YAAAAAAAAACvUZYw6fJkntWN3k8dfHz2vLucdz8mkjFE9bz1+UOZosX1UzPm0nqI/W
|
|
MYo9FlcPNklfFGeH/NshLGun+Cz6PtHZtVZWlRLS+jxPkRpIn7rdoupHTdA5s6SI+7H6Mfo+32Y2
|
|
+To3neSIiZ7A0IjPXpXLePlMotGW3272t85datKzHZjbTVnsDj+FG/2Y/RlGP4R+jo20u7H6N1Ql
|
|
o+H8I/REY957R+jpfReiK6eOYHLtj2tttH6KrY/6Y/R2c+kjeJiFVtLG24hxpw7/AHY/RRkw9O37
|
|
O99Hrt1YX0tfOBLjcGp4XF8c+u8fs9c4dcVcGemSI61nd3IneN1orQAAAAAAAAAAAAABFqxes1tE
|
|
TE9JiUgPKcX4RbRXnNgiZwWnrH4XPi28PdXpW9JraImsxtMS8pxXhF9DecuGJtgmf+1TWW2N/la1
|
|
L7N7T5e3Vy6W3hsYcvLbqzbO9jvvCzvDR0+XeO7crO6FmGSvRThy/RtVXJ92elvk2rRvDUzU7pl4
|
|
izsd2J3jeBpcNz+Lg5LT7+Pp+Xk3W7js5eAAgAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAADs0NTrN96Yp6edkW8Wzm6+LNTq4pvTHO9vOfRoWtt
|
|
1mes95YWvs1s2fZldddOczLPLn2ju0MmebT3YZc2/mpm3qqllN1drsbZIhr3yzvtHf4AsvlYYseb
|
|
V5Yx4KTe0+UQ6nDvZ3UazbJqd8OKeu33peq0eh0+hxcmnxxWPOfOfm0mP+steT/ji8N9mKY9suum
|
|
L37+HHaPm9DSlaVitKxWsdohI0Y22gAgAAAAAAAAAABXnyRhw3yT92Nwef4xm8bVzET0rPJH5d12
|
|
CvLhho3rN9RWs9Z23n5y6O21YhrVYbdGOCfrrLPJRpv863zVS6FS09SvZj3lVZZRdPSqmnSWdrIE
|
|
ebOkK4ldTsgW1WKqd1oMZhEVZyRAImOjGI6rJ7IiATNd46qL02bHkiaxaoNGY2n4ImPgtyV2n0Vo
|
|
Gvlx7x2beiyTk08RPevSVUxux00+Fn2n7N+n5rRFb4AAAAAAAAAAAAAAACLVres1tETWekxKQHlu
|
|
L8InR2nPp43wz3j8P/s5dLveWrFqzW0bxPeJeV4xwmdFec+CJnDM9Y/CrY1xv8qvTZ+WYdbDk5oh
|
|
5zHk283U0eo3jaZZ2N5XYjrCnLSJhOK+8d1kxvCqzSwZvousrb7k9LfJ3nB1OLeJdLhufx9LEWn3
|
|
6e7LXN9Ofy5/W4AuxAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAETaKxMzO0Qi9646Ta07RDmZ9VbPbaOlI7Qi3i+c3TPUaqcu9adKfy0722ZXvFa9
|
|
XO1OrjrESxt66ZJmcjPUanlidmhkzTZVfLN5VWvsC2b7R3U3yqrZZtO1esz2h2+F+zWTUcuXXTNM
|
|
feKR3n5+iZLVbqRzNJo9TxHLyaekz62ntD1fDOA6fQbZL7Zc/wCKY6R8odLBgxabFGPDSKUjyiFj
|
|
SZkYa3aALKAAAAAAAAAAAAAADQ4pl2pTFH3p3n5Q33E12Tn1eSfKscsLZ+orS00eJqbW+Lfnu1tF
|
|
XaJnZsz3WpCfsyp00fWSvmPdVYOmSUDd8kR3InoQosy7JmUX7MdwZ17ro7KKT1XRPRAsrO0rYndr
|
|
79V1ZBaQiJ6JgCSIJASwrO07MpV2nqBlrv1a1o2bf2qtfLXaQUTO0sb05o3jv3ZXhjS20xEphW5h
|
|
yeJjjf7UdJWNKLziyRePsz0lux1SgAQAAAAAAAAAAAAAADG9K5KTS8Rato2mJZAPIcU4ZbQZuekT
|
|
OC3afT4NXFkmlntc2GmoxWx5K71tG0vHa/RX0GpmlutJ61t6wrY2xr8dXS5uesN+tt4ef0eaa223
|
|
2dnHk3juyreM81OaFGiy/RtZET9jJ7s/2bdutd2jqKeic3iNTsd8a2h1H0jTVtP2o6W+bZbOO+gA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAABje9cdJt
|
|
adohGTLXFTmvO0fy52bJfU23t0pHaqLeL5xdK9Rnvqb+cUjtCi94xxvK3JetKuHrdZvaa1ljb10y
|
|
cnIs1Wt3naJc++TmVWvMz1YWybfMGdsm3eWek0mo4jm8PT0mfW3lDf4V7P5tdMZdRviwfvZ6/TaX
|
|
DpMMYsFIpWPTzXmf+steT8jn8L4Dp+HxF77Zc/4pjpHydYGjC3oAAAAAAAAAAAAAAAAADG9opS1p
|
|
7RG7zszN6WtPe0zLua+3Joss/wBOzhzG2OsL5+IrY09dsSyYRijbHEMvOChb7KjF0yS2LQ169Mso
|
|
S24noyrPVXWejNVKbTuw3T3REdQWU6LYlVvsyiUDPfqupPRr79VuOQX1lZEqoZxIMksd0gT2VT0l
|
|
bPZVbuCaW8i8bwr32WxbcGnkjaZa9p2ndv5qbw5+aNugLItF6TEtvTX5sMb969HMpfazc0d9stqe
|
|
vVZDdAQAAAAAAAAAAAAAAAADV1+iprtPOO/2u9bektoB4TJTJpNRbHkja1Z6uto8viVht+0HDvpG
|
|
H6Tjj6zHHvbecONw7Ltfkmeqmo6Ma69DXbbZTkr1mGWO3RneOaGbZRoM30fVzSelMnT83aef1FZ7
|
|
x3h1tBqfpGnjmn369LNc3sc3kzy9bQCzIAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAa+q1dNNXr7157VhGp1Xh70x+9f9ocy283m1p5rz3mVbrjXHjt91lz
|
|
5c9+fJ1nyjyhdM8lZlOOIiqrUXikd+kMreunnI5XEdX4dZiZcG+XmtNl/F83PeeWWHDOGanieSKY
|
|
q+5H2rz2hMzWd1Iqx1yajJXHhrNrW6REeb1nCPZumn2z62Ivl7xTyr/6uhwzhGn4Zj2xxzZJ+1kn
|
|
vLoNJnjHW7TbbsAszAAAAAAAAAAAAAAAAAAAAaPFrbaSK/itEOXt0rDf4xb/ACa/GZacRvaF58Q2
|
|
IjasQnzPIhCU92tMbZGzHmotG10C6nZkwpPRmipIllEbMIZIE7solgmJBnCyk9VMM6z1BtVllEqK
|
|
z0WRILYlluriWcSDJVbusV27gwInaSWM9ECyZ3hqamnSWxFmOSOaqRx725bNnSZNs9J+OynVY+WZ
|
|
YYr7TE+nVaIr0Ais81Yn1hKAAAAAAAAAAAAAAAAAABExvG09peU4nov9n66L0j6q/WPg9Y1OJaON
|
|
ZpL0+9HWs/EWzeVz9PbmrEtnyc3h9reHy26TWdnSr2YX6657ijLXpLX0+onSamL/AHJ6W+Tbv2aW
|
|
ekTv16JzeI1Ox6KJiYiY7Slz+E6jxdN4dp3vj6fl5Og2clnKACAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAACZ2jeQRMxEbzO0Q08uqtkma4ulfO3r8lefUePMxWf
|
|
cjy9WvlzVxV6T1Z61/x0Y8f7Wc7Ur1lqVy+LqOWJ2hp6rXddon5rOF1tfmz5OkT0qzb8dWbxjp1c
|
|
biuuilJ5Z6r+IcQrixzEy8zl1E6rNt1tMztFY81sztU1eRucN4ffi2p5esRM72n0h7rS6XFo8FcO
|
|
CkVpX082nwXh3+z9FWLxHi36328vg6TZyW9ABAAAAAAAAAAAAAAAAAAAAAADj8Unm1tK/hqppHvw
|
|
y1k8/EMk+m0GOPeafiFpCZYwolnXspvHvLa9mF46gmnZmwozRUiUCBKYYsoBLOFbKAX0llEqqyzi
|
|
QXRLOJVRLOOwLIljZMEgrlhKyYYTAK5nZPN0RZjugUanHzVlz6xtLq361c+9eXItPpXX0dubTU+E
|
|
bL2lw2++O1fSW6m/VYAISAAAAAAAAAAAAAAAAAp1GbwcfTreelYEydcuMcRrM/L9nnlsV6wqpi2r
|
|
tv133mfWVkRyRtEdGFva7MzkYZNoamWN4bV4mYa9qztKIujhVppxGI8r1mJegeZpknBqKZY+7L0t
|
|
LRekWrO8TG8Ns/HJ5ZypAWZAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAADS12fp4VJ6z9qVuq1HgUiI+3bpDl589cOKZmevqprXPTbx477rDJlrhr1nq4+s182tMRP
|
|
RqaziXiZJrWekNG17ZbxWJ336M5LXRbI3dLTJrs07RMY6fan1dHLrowY+X7MVjt6N3R6Kul0EbWm
|
|
s7bz8Z+LnabQX43r7Y53php/mXj+Dnv0f1JO1x/8ZxbUzj02O15mfLtD13AvZqnDds+pmMmo26el
|
|
XX0Wh0/D8EYtNjilY7+s/NstpOOTW7QBKgAAAAAAAAAAAAAAAAAAAAAADG88tLW9I3BwJtz6nNf1
|
|
vK/DHVqYJ3pzT5y3MPZeojOWMQylEKpTVjZnDCwkqzYQyRRICATCITAJZQxhMAshnEq4ZQC2srKq
|
|
qrIBZCWNZZgwswmFloVyCu0dFcx1WyrtCBhv5NTPHXds2U5o3hIz4ffbPt+KHUcTSW5c9Jme0u2v
|
|
VYAKpAAAAAAAAAAAAAAAAYZctcVOa35R6tLrltN795/YvknNqrfhpPLH92V5isd9mWq6fHjk6rn0
|
|
ZxG8KK5Jm/wbVZiYZtqrmkqL023bkxvCiY3lJHNyRG81mHS4Rn5sNsNp64+3yaWaNrzOzHBl+i6q
|
|
mT7s9J+S+ay8mex6EIneN47SNXKAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAImYiJme0JafEs3h6fkidrZOn5eaLeJk7eOdm1Hi2vmtPTry/CHmOJcUvmvOPF1n09Pm
|
|
6HF9ZGm01qxO3R5vSY7XwzmzTy47zzTEd7en5Mfvt2/PURWdo3tvPrPlKymbktFqTtMTvHzbOLDG
|
|
f63JXbFX7FdnoODcDprZpq9TjiMMTvSn4vj8l5fxnrk91saPSa7i2hpOfbTVt5x1m0fLydzR6PDo
|
|
dPGHBXasd585n1lsRERG0dIF5OOe6tAEqgAAAAAAAAAAAAAAAAAAAAAAADX11+TRZrf0y2Gjxe22
|
|
gtH4piP3TPpXKwxtjhuYo9xq442iIblI2pC1RET2ILd9kxCqRjZmwlCSEohIJAQAAJZISDKGUd2M
|
|
MoBnVbVVCyAWVWeSuqyOwIlXZZKue4MJV2WWYT2QKbKL9YlfdRdIo35b7/Hd3KTzUrPrDh27uxpb
|
|
c2mpPwX/ABX9XAKpAAAAAAAAAAAAAACekTIp1eTwtJmv+GkyJn1oafeazbfpMzLR4jq/o8b823zX
|
|
6XNF8ERCvTcNpxLV5LauvPhx9Irv3lhztdtv8TtaWLicXrt03jzjzb2k1nid56ty3s/w+a7Uwzjn
|
|
1raejlarhmbhl/FpbxMO/fzj5p/ixSeXOvTtRfeI280ZI26tfDm3pWe63LaZx7qtGvniJ6tPLvOK
|
|
fOa9WzbJvTbza02jl3n5SSljscK1MajSxWZ96nSW88xw/VfQ9XMT9nfa3yemid43jtLeXsce88qQ
|
|
EqAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAADia3UTm1l4j7OP3Y/u
|
|
7Vp2rM+kPJW1PhYcmS0+9MzKm/jbwz31weMzbV8UppazPL9q0/BF4rk1GLDSNqxPWPhCnHmnNrtT
|
|
qPKteWPm6U6OdHaZvO+SaRNvhv12Ub/q3FhtrNVj0uKOt56z6R5y9zix1w4qY6RtWsREOJ7L6OKa
|
|
S2rvX6zNM7T6Vh3mmZyOfya7eACzIAAAAAAAAAAAAAAAAAAAAAAAAAAczjVvqMVfW/8AZ03I41bf
|
|
Lp6/OVs/UVrY47NyOzUxd4bUJpEbb3Z7IiOrKIVSjZhMLJYyhKIgmGUQSDESIEbJEgQmCITEAmGU
|
|
IiGUAyhZVhDOoM4Wx2VQtqBKuyyWEgqlhKyyuyBVaGtkbNmvk7A15l1eH2300R6TMORPSXT4ZO+O
|
|
8fFefEX63gEAAAAAAAAAAAAAAAq1WPxdLlp+Kkx+y1Fvsz8gjhaDauGK8sx07y3OE3m1tT6RaP4c
|
|
vU6yMNKUx73zT0ilY3l2eF6a+m0kRl/zbzz3+Ez5M8z26fJruW6wzYq5sV8d43raNpZjRzPPaTmx
|
|
5b6bJ9rHO3zb2WJ8GWPEscY9bgzxH2t62n19GWW0eHOzHU5XbjXZ1x8WTnz2iZ7S2M1IjH2+LX0V
|
|
KTqs8zO9ot0j8nUthi1J3UaOFMTfLFo6xMbS9BwHWTqdHOO8+/hnln5eTjYMFo1WTH5VnePzXcIm
|
|
2k4zlpPSmXy/hfF5eMfJns69OA2cgAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAADG/2LfJ874rW845mubliY7bPoto5qzHrDz0+yePNF41OotaJ7RWNtpV1OtfHqZ715fhu
|
|
j8adNpcVfeyzE2/vLuanhOu1nEctIxTTFa/+ZPbZ3eHcF0vDbTfFE2yzG03t32+DokynXl9+leDB
|
|
TTYKYccbUpWIhYCzEAAAAAAAAAAAAAAAAAAAAAAAAAAAAcXjE/4zDH9M/wAu04XF5/3jj/0f3Wz9
|
|
RUYmzDWxS2I7FSyjuzY1ZKpRKEygEwiWUIkGIk2QJNhKQhMIhkCYZQxhlAMoZwwZwgWQshVCyATL
|
|
CWc9ldpBhZXLOVdpQK7NfJPRdaWvknoDVvPvOnwuel4+TlXn3nS4VPvXj4QtEV0wAAAAAAAAAAAA
|
|
AAAAAVV02CmTxK4qRf8AFFeq0AAAanEsfPpZmO9Ji0NDLfkwdOsulrumiyzHlVzJrz4Ovoy26vB8
|
|
cTBa9NffLtMY77Rv8Yegx5ImkKdJoY1HC81Y+3OSbVn0mGGkmbY45u6tnrrTOu2xGO0RxCd+nNVj
|
|
qKxTV1vH2pjaGtnyzXXYdo96ZmGXEMk15b7/AGZiVerWPTYckZcNbx5wzc7hGbnxXxzPWk7x8pdF
|
|
0S9jh1OXgAlUAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAcPjEf4/FP9H93ccXjMf4vDP9Mx+62fqKrx+S+GvibEFSsqyYwlVK
|
|
ZYsmIMoRKYJQIPIEiQ2ATCUQygCGUIhMAyhnDCGUIFkLIV1ZxIMpVWWSrsCuyqyyyq09ECq8tfJK
|
|
66jJ2Bp5J6upwn7dv9Lk5J951uE/av8AJaIrqAAAAAAAAAAAAAAAAAAAAAAq1Mc2myxPnWf4cmtu
|
|
XT9fR0tffk0WSe28bfq5Wbamm3326MtunwfK6PCv/AxPraZ/dz9PO97/AOqf5dHhdZrw7Dv3mOb9
|
|
XOxRFM+avpe38mvkPHf/AFWlrKba7Tzt99ZxKkfR7euyNXMTrtPHfa0z+zPiM/UR8Zj+Wbdu8HpN
|
|
M2bfzrV13M4dO2pyR61dNvj44/J/oAWZgAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAADj8bj63BPzdhyeNx0wz8ZWz9RWri7Nmv
|
|
VrYu0NmqaRZHZlDGGSiwxZSgCEkCBCQSCQBMJRCYgEsoYx3Z17AlMIhlCBnDOGEM4AlhZZKq4KrK
|
|
7LLKrIFN2vdfZReAaObu6/CO9vk5OePR1uEd7fJeIrqAIAAAAAAAAAAAAAAAAAAAAGtxCk5NFliI
|
|
3mI32+XVyNTyZOHTee946PQKPoeDffw4777eW/yVs60xv+ZxOnr4Okx1t05KRv8Ao41Z5q3yed5m
|
|
XY1szXRZ5jvFJ/hxItP0aOSN9q7yrtr4f2tHFM5+KT16Yq/vK/iGSbXw4vO14UcPx5MGfNbPG18m
|
|
1oj4THRsTw7VanPXVYpi3gzMcnrvCnG11JOupwuN8+a3pEQ6jT4divjxWnJExa09pbjbM5HHu90A
|
|
JUAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAHM41H1GOf6nTc/jEf4Ws+lls/UX45uGekNujTwdm5RNIthKIZKLDFlsiQIShIC
|
|
EgCUJ7AmGTGO7IDzZQhMSDJMMYZQgZwzhhDOATuqssmVdgVWVWWyqtCBTeVF19lF+wNLNG7q8I+9
|
|
8nLyupwnt+S8RXUAQAAAAAAAAAAAAAAAAAAAAAAItWL1mto3iY2lyrcLyUxzix2ia2nvPeK+jrCL
|
|
OrTVnxpanhuPPemSs8l6RtE7dJj0ldpNP9GwRSZ3neZmV4cR/Vs4AJQAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAANHi1d9H
|
|
M+kt5ra+vPoskfDdOfqK4mn7Q3aNHBPZu0W0RdDOGFWcKLCJZeTGQQlCQSgASBsCYZQxhlAJTAmA
|
|
TsmAgGcM4YQyjsgRLC3VnaVcgwsrt3Z2V2QK7tbJ1bN5a9waeWO7p8Knt8nNyebpcK8vkvlFdQBA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAK9RXmwZI+ErEWjesx6wQeZwejeo0cccuW8
|
|
elpblJaaRGxVnCuss4ZrMvJEgCAASISCQIBlCYYpieoM0wx8k7gzIRueYM4Z79FcSy3QEsLJmWFp
|
|
BjaVVpZWlXMoGNmvkXXlr3kGtknu6XCf7OXkl1OEdl8orqgIAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAHmskcmtzV/rls0U62OXiWX4zErcc9GmkRfWVkSqqziWayxCPIANwBIhIJSxS
|
|
CRG6dwZwlhEs4BluMdzfqgZxLLdXuy3AmVdpZTKuZBjaVVpWWV2QlhZRdfZRcGpl7urwfrzfJy8r
|
|
rcH61vPyWitdMAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAHA4nHLxKZ9awnH2ZcY
|
|
jbW459aq8fZpfiI2IZwrqzhmsz3Ebm4JN0AMhCQSIASndiAziWUSriWcAyRujc80DM3RCfIETLCW
|
|
UsZEsJYSslXZAwlTddPZTkBp5e7r8Gj6rJPxhx8k9Xa4PG2C8/FaK10QAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAcfjcbZMFvnDWx9m5x2PqcNvS+zSxT7sNPxH62YZQwqzhRZO6UCB
|
|
KUAJTux3SDIRuAncQAmJZRLBMSgZ7iIAZRKd2DICUSlAljLCYWMLIFVukNfI2bNbIDTyT7zu8Ijb
|
|
Sz/qcG/2nf4T/wCE/wD2WnxWt4AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAHL9oL
|
|
+Hw2cm28VvEuPptfgyVj6yIn0no7/FtJfW8NzYMe3PaPd39d3iMug1WktNc2C9dvPbeP1aZ9xF+v
|
|
T471tHu2iflK2HkqWmvaZj5Surqc9Ps5bx+alTHqYHm68S1Vf/NmfnC2vGNTXvyT84Ql6A3cSvHM
|
|
sfaxVn5Ssrxyv3sM/lKB1xza8bwT3pePyWV4tpZ+/MfOEjfGrXiGlt2zV/PotrqcN/s5aT/+wLRj
|
|
FontMSlAlKEgndO6IAZQljDIEgeQljLCzOVdkCu/SGrkbF56NPNeKxMzMRHxENe0+89DwuNtHHzl
|
|
5PJr8NcnLW3Pbf7r1nCZm2gpae8zMrz4i/W6AgAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAETETG0xukB4HVaeMHEtRi26RedvkyjBSfX9W77QYvC4xz7dMlYlrU7M929dWJLFc6aPK0q
|
|
7YLxPS0S22FlP6q38Zac0yR92s/KVc3tHfFf8tpbcsLRvB/dR/8ALLVnU0r9uL1+dZI1mnmdvGpv
|
|
6TOy6ym+Oto2tWJ+cJ/tW+KLK5KW+zes/KU7tG+h01p64qx8Y6NXNo6Y+uPJlp8rLf0rfG7MXtHa
|
|
0x8pZxqs9e2a8f8A7Oj7HaTHn0+f6RWM23LETfr6vRW4PoL99NT8ui7F4+vEdXXtnt+fVbXjGsr/
|
|
AOZE/OsPS29nuH27YrV+VpeV9pdPXhOtw49NG9Mld55+vXcTPd42I47qo7xSfyWV9oM8d8VJ/VxM
|
|
d8l46xWF9cV7en6o/qLfxp2I9ob+eCv/AHMo9op89P8A/wBORGmyT5R+qfo2X8P7n9Q/jTsx7RR5
|
|
6ef+4/8AuHftg/8A6cWcOSO9J/WEbWr3pY7Efzp2Lcfv5YK/9zWy8d1E/ZpSv5Oba1/+Hb9lc+LP
|
|
bFt87I7E/wAabWbiurvEx4nL/pjZzc2bJkn372t85ZXx55/BX85lucC0vPxnTxlnnjm32mOiZqUu
|
|
LJ2p4TwnVavNWaYbRTfre0bQ99pcH0bT0xb78vmtiIiNojaErMwAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAHnfarF7umzRHaZrLjYrdIen9ocPi8JyTt1xzF4eUw23rCm3R4r6bMy
|
|
wt6kdTaWLdjswmNoZontsCm0K5XWjopnuDC0dGpqG5bs08/daKV672MjbSaif6oh6Z5f2LtvptRX
|
|
0tEvUN3Jfo8f7cYve0eX4zV7B5z20xc/C8eSPuZIRficfXlcPaG7ino08HWIbePpLF2NuiyOyrHK
|
|
3fZFSwuovHVfaVF4QK5YWTM9UT0EKry6Ps1Tn4zjn8NZn9nOtLseydObiWW34cf918fWfk+PYANn
|
|
KAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAq1WKM+ly4p+/WYeBxTNd6zG0xO0
|
|
vobw3FcP0bi2em20Tbmj5Srr418V9sa2Z7qKyzi07MXUylhaU7yjqhLCeiq3ddaFNxFYW7NLNG8t
|
|
zya+WO6Va9J7FW66mvwidnrXiPY3Ny8RyUn71Jj9Ht3RPjk19HK9pMHj8D1ER3rHN+jqqtTjjNps
|
|
uOe16zAifXzfTz7kNyndpYazS9qT0mszDdoxrsi6m8LazMq6zDOsq1ZEyrt1WWlXaUCqyq0rbKbi
|
|
Fdp6PReyFd8uqv8ACsfy83aXrPZHHto89/xX2/SP/dpj6y8vx6EBq5gAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAB5n2q03LfDqqx39y39npmlxbS/TOG5se29tuavzgWzeV4mtui2
|
|
O3RRSY2hdVhqO2MvI36iu9lUsrSrvDHn6spnmSiq5jooyV6tq1VV69RC32byTh43h8otMx+r6I+Z
|
|
aK/g8TwX7bXh9Mid4iW+fjl8n1ICWb57xLBOm4zqse20Tbmj8+qKdnS9q8PhcTw5tumSm0/OHMxz
|
|
0Za+uzx3sX1t0Zxurr1ZxvspWiZYWZbsbT0QK7KLrZVZJFaqt5vbezNOTg9J/FaZeJns93wCvLwb
|
|
T/GJn92uGHldIBowAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAADuAPA67F9H4l
|
|
qMW20VvO3yRWW97T4fC4rXJHSMtI/WGhVlue3b473K2KzMML4+62tujG9pnozXaOSOVFMnVbmq1t
|
|
trJRW5E7wwvUxTvCyY6CHOt7moxz6Wh9PxTzYaT61h8x1MbZK/OH0zTf+Fxf6I/htj45vL9WgLMn
|
|
mvbPFvocGWO9L7fq85p5maw9d7VYvE4JkmPu2if3eW0+PasdFNOnxfF1Y2hlykRsmY+LJ0MZjZXa
|
|
eq2eyi8oQTO0KLdZWzPRjWu6VaqtHR73g0bcI0sf0Q8Nkq93wqNuFaWP+XDTDDytwBowAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAef9q8HNpcGaI60vtPyl56k9Iew49j8ThGe
|
|
PwxFv0l4zH2U26fDfTYiyJljvsjf4sm6vJ1hrXjq2MkqLdZEVbgbMx0auGdmzNt6iHN1Ub5af6of
|
|
TdPG2nxx6Vj+HzaaTm1+nx/iyVj930ysbViPRrj45vL9SAuyc7j1efguqj+jd4/T33rD3HEcPj8O
|
|
1GP8WOY/Z4TTT7sKadHhbcsZnaCJ3TPZk6VdrKbTutmP0U2nqgrGOsr8deiuI2X09EqKM1dt3uuG
|
|
f/jdN/06/wAPE546S9rwud+Gaaf+XH8NMMPK2wGjAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAABrcRp4nDtRWPPHP8PCYusPoWSvNjtX1iYfPuWaXtX8MzCuvjfw32siu8ptXoxi
|
|
0wy5t4YulReqmazu2skbquURWFInddM7VYRGyL291KFnCcfj8e0le/Lbmn8n0N4b2Ur4nHLWmPsY
|
|
5e5a5+OXyXugBZmiY3iY9Xz7NjnTa3Ph/BeYj5PoTxftFg8Hjk2iOmWkW/Psrr418V5WrWd2faFc
|
|
V2jdnEMXWxntupmN7NiYU27iWML6dVMVnddjgVqMsdHr+CW5uE6f4Rt+7yuSsTDv+zWXn0WTHP3L
|
|
/tK+GHl+O0A1c4AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA8Dn93W56/wDM
|
|
t/L3z59qp24jn+OS38lnpr4r7ZxHQ2TEstt3PXUrt27K57rr1VT0BjKnJPRbMqMs7QlV2fYvHvrd
|
|
VknyrEfu9m8f7FZI8fVU85iJewbT45NfQBKo817W4eulzxHaZrL0rje09ItwqbfhtBVs3leai8RD
|
|
KLw1sduesL606dWFdsZT1jdhNeq6K9DlhCVUU6s4jZnt1YzAhnM71dH2bycmszY/K1d/0c6OzY4R
|
|
fwuK4p8rTstn6z8k7HrwGzkAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAHz3
|
|
Vxvr80/8y38voTwGpj/F5/8AqT/JfjTx/WVeyY6FPspc9dZPVXaOq2WEwIUTVRmjo2rNfLHRI3vZ
|
|
DJycXtX8dZh7t879nsnhcbwz23tt+r6I2nxyb+gCVBzuPY/E4PqI9K7ui19fTxNBnp60n+Aj5/pJ
|
|
3jZu1aOnnltMNussdfXbm+l3ZM9URHREdZVXTuT1Nk7boQiOkJw28PU47/htEp5eivJPLMTCZ9Vv
|
|
x7mJ3iJ9UqNHk8XR4b+tIXuhxAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAD
|
|
weqjbWZ4/wCZP8vePCaz/wDIaiP+Zb+UX408f0r9lOxWOifJhXWjfyYWllPRXYQxnrCrJHRd3YZI
|
|
6A1NJecHEsN/S0T+76bE7xE+r5dk93LW3pL6ZpMni6PDf8VIn9m2fjm8s9rgFmQxvHNS0esbMiew
|
|
PnHLyai9fS0w2aNfUTtrs3+uf5bGPqy068fF227KtSsdFlKqNGMV6myyY6sbdIQI8tlOWOi6Jhhk
|
|
j3RD0vA8nicMx9etZmHRcT2Zyb6XNT8N9/2dt0T449T2AJVAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAHhdfG3E9TH9cvdPEcXjk4zqI/q3L8aeP6xr2TsxpLOekMK6mFo6qpXSrm
|
|
OqBixvHSVmzC4OfqK7S9/wAByeLwbTW9K7fo8Fqo6Paeyl+fglI/Da0NcMPK7QC7AAB8313TiOf/
|
|
AKk/y2MHWrX4jG3E9R/1Lfyv0/aFNOrHxuU7LI7MMayGTVlHWUXhNe6Z6wIUsb9d1m20q7dkDpez
|
|
N9tRqKT5xEvRvKez9+Xis1/FSYerb5+OTyf6AFlAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAB43j9eXjN/jWJ/Z7J5L2mry8Upb8VIF8f6aGOey2eynHvOy7bowrrYSxZSwQJ2YXZ
|
|
92N4BoanrEvVexmTm4blr+HJ/aHltRHSXofYm/1Wrp5RaJaYY+X49WA0c4AD51xONuKan/qW/lbp
|
|
+0MOLRtxbU/9SU4J7KadWPjep2WQrr2WRPRk1TvsndXMpiRCb9FNu0rbTuqvKBscCjfi9PhWZeue
|
|
V9n434rafTHL1TfPxy+T/QAszAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAHmv
|
|
avHtfTZfnV6VxPajHzcNrf8ABeJFs/XnMcr4no18c+6vr2YadkY2YM57sEDLyY37Mo7MMnYGlqO0
|
|
vQ+xNfqNVb1tEfs87qZ2rL0/sVX/AHdnt65P7Q0wx8vx6UBo5wAHz/jUbcX1PT78qtO2vaCnJxjP
|
|
8Zif2amnnspp04+OjWejKJ6MKdmcMmyJn4m5ZHzEVPMwtJv0VZLbQDqezcb8RzT6Y/7vUPM+ytZt
|
|
n1OTyiIh6Ztn45N/6AFlAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAABocbxeLw
|
|
nUR5xXm/Rvq8+OMuDJjntaswEeBxT0bNZ6NatZpNqz3rO0rqsdO3PxlaWEMpY+aqWXkryT0ZT2V3
|
|
7A0dVPuy9f7G124NM/iyT/Z4zWT7sw957MYfB4Fp4/FE2/WWmGHldcBowAAeM9qKcvFeb8VIly9P
|
|
0nq7ntbTbVYL+tJj93CwT76unR4/jo0nozhhTsy3Y1sWljM9Ce7HyQIm3RRlttVbaWrnt0Sh6n2U
|
|
x8vD8mSfv3/h3XN4Bi8Lg2nj8Uc36y6TeOPXugCUAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAPD8RxeBxXUU26Tbmj8+quro+02Lw+I4ssdslNvzhzazvDPbq8d7GW7Dfqz2VzG
|
|
0s2qd+iu/Zn5Ksk9BVztX1mI8930zh2LwOHabH+HHWP2fNYp4+vwYvxXiP3fUqxtWIjyjZtj45/L
|
|
faQFmQADzftfj3w6fJ6WmHmsP23rvaqnNwqLfhvEvIYZ+sV038bo0noy36MK9oZQxrdMyrlnMbMZ
|
|
QKrS1M07zEestq/RRjr4utwY/wAV4j91p9V18fQdJj8LR4ccfdpEfsuREbREJbuMAAAAAAAAAAAA
|
|
BAJAAAAEAJEAJQAJQAJEAJQAJQAJEACUJAQlAJEAJQAJQJAAAEAJEAJBAAAJAABAJEJAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAABwvanDzaPFmjvjv8A
|
|
tLztJ3h7HjGHx+FainnFeaPnHV4vFbeIU038VbHeGF+kso7Mb9mTdhKnLK3dRm7SIrHhGPxeP6Sv
|
|
9cT/AHfSnz72Zx+J7Q45/BWZ/Z9BbZ+OXyfQBZQABzeP4/E4NqI9Ii36S8Ng/wAx9C4jTxOH6ivr
|
|
jn+Hz3B/mQi/GvjdCnWNlsdI2V07LIlg6USrt2ZzZXMoFV+zPhGLxeOaavpbm/RVltEN72Yx+Jxm
|
|
b7dKUmf7L5+s9/HtRA2cqRACRACRACRACUAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAACQAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAQCQQCRACRACRCQBCQBCQB
|
|
ACRACRACRACRACL1i9LVntMbPATTwdRkxT3pea/u+gPE8Xx+DxrPHlaYt+qNfGvjvtXXsi0dOrKk
|
|
dEXjZg6VMtbP2bMtXUdpEV0/Y2nNxbNf8OP+727xvsXH+N1U/wBEfy9k3nxyb+gCVQAGOWvNivX1
|
|
rMPnGGOXNNfOJ2fSZ6w+dZKeHxDPX8N7R+6L8a+L63KdoZ7q6zvEMpnowdKJ6ywmWUyqvIKM0vQ+
|
|
x+D6rU55+9aKx+TzWa36vbezmDwODYenW+95/Nphj5L6dQBo5wAAAAAAAAAAAAAAAAAAAAAAAAAA
|
|
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAEgAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAACAEiAAAEoA
|
|
AAAAAAAAAAAAAEAkEAkRuAkQbgkQAkQAkQAkQAl5T2nx8nEMOT8dNv0l6pwfarHvpcGWPu32/WCr
|
|
YvK4mOem6b9mGKd4Z3idmFdka0y1c892zfpMtLPaNpEV6D2Kj/Eauf6YeweQ9ieuTVz8K/3evbT4
|
|
5NfQBKoAA8FxCvJxrUx/XMvevD8Zry8fz/Haf2RfjTx/6RSOnRMyypHu9kXjowrqVSrvPRnZVl6V
|
|
kK0775MsUjvadn0nT4ow6bFijtSsVfPuFYvpPGtNTy54mfy6vorXDm8l9pEC7JIgBIgBIgBIgBIg
|
|
BIgBIhIAgBIhIAgBIgBIIBIAAhIAhIAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAJAAAAAAAAAAAAAAA
|
|
AAAAAAAAABAJQkAEAAAAAAAAAAjc3BIjdG4Mkbo5kcwMjdhzHMDPc3V8xzAs3N1fMjmBZubq+Y5g
|
|
Wbm6vmOYFm5ur5jmBZubq+Y5gWbm6vmOYFm5ur5jmBZubq+Y5gWbm6vmTzAz3N2HMnmBlu5ftFTx
|
|
OEZJ/DMW/d0t2rxKni8N1FPWkiZ9eS08e7Cy8dGGn6UhZaJljXZGnmc3UT3dPP2cnUT78xCIV6j2
|
|
H/8A9c/6f7vXPI+w8bU1U+vL/d63du5NfUiDcVSIAS8b7RV5eOb/AIqRL2TyXtNX/e2KfXH/AHlF
|
|
+NPH/pr4+2xcxx0hFpY11K7R16KM32ZWz3UaidqSgrc9kcPicWyZJjfw6T+727y3sXh2xarN+K0V
|
|
h6lvPjj3e0ASqAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAJQAAAAAkQAkQAkAAAAAAAAAAAAAAA
|
|
EgAAAAAAAAAAAAAAAAAAAAAgAAABKDcAN0bgkY8xzAyRux5kcwM9zdXNkTcFm6OZXzMeYFvMibKu
|
|
ZHMC2bo51U2RuC2bom6rc3BZzom6sBZzI52ADPnOdggFnMc6skFnMc6rc3BbznOp3RzAv50c6nml
|
|
HMC/nOf4qOY5wX85zqOc5wbHOc7X5znBsc6edr85zg2ec52vzpi4NjmY5bROG+/bllVzsNTk5dLl
|
|
n0pP8BHmMHWNmzt0aum8obm08vVjfrtnxztR0mXHzTvaZdjVRMTLkZo6yiFen9iZ2pqY/wBP93rN
|
|
3kPY+/LfPX1rE/u9XzN3HfqzdO6vmTuIZ7m7Hc3Bnu8t7TR/vHBP9E/y9Pu837SV31umn+if5Rfi
|
|
/j/01MMb1hjkrtKzBG0bMsmOZY11tOYamr6Und0LUc7XT7u3rJPqL8er9lcPhcFpbzyWm39v7O00
|
|
+FYvA4Zpsc94xxu227jv1IAgAAAAAAAAABKAAAASgASgBIgBIgBIgBIhIAAAAAAAAAAAAAAAAAAC
|
|
UACUJAAAAAAAAAAAABIAAAAAAAAAAAAAAAAAAAAg3AEbomQZbo3YzLGbAz3RNlc3YzcFs2YzdVN2
|
|
M2Bdzom6nmNwW86JurTAMuY3REJ2BB1ZRVMVBhsbSsiqeUFXLucq3lTygp5TlXcpygp5TlXcpygp
|
|
5TlXcqOUFXKjlXcrGYBXysdlswiYBVMdUTCyY6sZBWxlnMMZgGLGZZSwkDdHMiWO4MuY5mEyjcFn
|
|
N1OdVzHMC3nTzqeY5gX85zqOZPMC+Lqdbk20eb/RKOZr8QybaK/XvtH7iZ9aGlp2luzT3fg19NHS
|
|
OjbmPcYX67XH1XSZ9XIzRvMuzrK7zLkZYmYnciunb9lZ5dTk+OP+71cXeP8AZnJ/ip2nf3J/l6iL
|
|
/Fu5L9bMWZczXi6YuIbEWTzKIuyiwLt3nuO25uI4a/hx7/rLuczg8TicvFLbfdpEK6+NPH/phhjo
|
|
stLGkctUWnoxrrU3j1cnWTzZq1jzl1clo5Zcu8c+txR63iP3Tn6pv4+g4o5cVI9IiGe7CJ2iE7t3
|
|
GyN2O6dwSINwSISAlAAlACRAAlAAlACRACRCQAAAAAAAAAASgASISAAAAAAAAAAAAACQAAAAAAAA
|
|
AAAAAASAAAAAAAAAAAAAAAAIAAAQCAJljuljsCJlhMs9mOwMJYys5TkBVsjZdyHICrZPKt5E8oK4
|
|
qmKrOVOwMIqyirPY2Bjyp2ZbAI2NmSARsbMgEbI2ZAMdjZICNkbMkSCNmOzJEgx2YyzljMAwlhKy
|
|
WEwCuWErJhhMArlhLOWEgxljMpljIImWMyTKJA3N0IBO5vux3NwZbnMx3NwZczT4jf3MdPW27a3a
|
|
fJOq1XNP2KdIRfi+J2trSYfcjeF+Wm1OicVeWIiN9kai8xjY12ORqultnI1Ecsujq79XP1FovWYI
|
|
rTgeq+j8QrWZ+3Mx+r2UXeC0WG2Ti2kiN5mL807eUREvbzbaejefHJv62Iv8WUXa0WTFhVtRdlF2
|
|
rz9WUXBtc7jR9dqc2T1ttHyhvZMvJitb0jdq6XHNcNenWVN3028U99WRj6Kb02be3Tq18/SN2Lpc
|
|
3UdN9nOmZrqKX/DaJ/d0svvTLRzV3jomK6+Pd1vvWJj0ZczT0mXxNJht60hfFnQ4qu3N1cWTEgs3
|
|
Tur5k7gz3N2O5uDM3Y7m4MtxBuCQASIASIASAAAAAAACRCQAAAAAAAAEoSAAAAAAAAAAAlAAlCQA
|
|
AAAAAAAAAAASAAAAAAAAAAAAIASgAAAEJAQJQCNkbMgGOyOVnsAw5TlZ7GwMOVPKy2NgY7GzIBGx
|
|
skA2AAAAAAAAAAQkBAEghEskAxYzDPZGwK5hjMLJhjMAqmGEwumrCagomFcw2JqqtUFEsLLrV82F
|
|
o7gqljKyYYTGwMZRKUSCAQAboJnaN5Bjkneu0d5W4ccViIiOzHFWbTzNumP1Zarr8eeRMbxDW1Mx
|
|
NO67NbkhzNVnmInqzaOZrL93JyZeV0M1++7S02jvxDWxhxx033tPpC8Z6rrezWjmZyazJG2/u03h
|
|
2vFibTHoqvamiwVwY+nLGzV0+SZ1Mx8G0/45tOhzJ5lXMc3UVXRdlF1HP+iYsDPLPPy49/tz1+Te
|
|
pSIr0ho6ak5Ms5J8o2q6NImOrHV7XX488ypzTtHXo0s9t6zG7c1G1qz6ubeZiZ3UatXJG3yauSO7
|
|
cvMTEx5tPLb3prPRMVr0HB8vicNxf0+7+kt+LOJwTJyY/Bnz3tH93X36N58cWvq6LSyiyndMSlC7
|
|
mZcymLJiwLosmJVRLKLAtiU7q4lMSCzc3YxJuDMRuAlKAEgAAAlAkAAAAAABKAEgAAAAAJAAAAAA
|
|
AAAAAAAEgAAAAAAAAAAAAAkAAAAAAAAEAAAAAAAAAAAAAAAAAAAAAhIAAACAAAASgAAAAAAEAAAA
|
|
hGzJAImGMwzQDDZjNVuyNgUTVhNGxysZqDVmiu1G5NN2M4waM0+DCaN2cbGcQNGaMZq3JxMJxA1J
|
|
qx2bU4kU09slorWNwa20z02RXHbJbl26QvtFovbHWkxEdJt5y2MOHlr2U1W3jx+1hiw8vSO63lmI
|
|
XRTaEWmtY6snRHO1VpmJ+DjavpSZl2s8b7y4HFcnh0n0gha5ebJN55KRM2mdoiPN6fh+kpwXh0Wy
|
|
RHj5Otp/s5Ps1p62y31+em9aTMYt/OfVfxTiPjZ52naI7fBrI5t66xz5+a1rW7yx0eSL6iZjtEOX
|
|
qNbSletom3lENjh2fbHzbbWt3iVozruc+5ztWubf4M4ybpQ2Oboyrva0Vjza8WdDR4OkXt3n9ldX
|
|
kaePP9VtYqctYhdvt5oivTeCZ2YOxXk6ubqMfV0b9mrljfqlFcq88k7z2U5axeItDa1OPessuC8P
|
|
ya7XRWYnwqdbT/ZMilvIu4dpslNdixXja8Y5tt85djZdbDWnGOesRtXFtuw6T27No5Kx2OrKYQlC
|
|
ExKJgBnEpiyvdlEgsizKLKollFgWxLKJVRLKJBbEp3VxLKJBnuMWQJEbpBIAAAJAAAABIAAAAAAA
|
|
lAJAAAAAAAAAAAAAASAAAAAAAAAAAAAJAAAABAJABAlAAAAAAAAAAAAAAAAAAAAAAAAIAAAAAAAA
|
|
AAABAJQAAAAgAABAAI2EoBGyJhkgGPKxmqxAKpownHC+YRMdN5BrTj67R3bOn01o7p01Iv71u89o
|
|
b9a7LfBTfS1vWI2jf12VfQPSW8KX2mas+NC2iv6xMNfJpMnLtEbuuxtMRCtzF55NR5rPps1N/ctP
|
|
y6uHreE6nXZ4pak48X3rT06fB7fNeI33cbX6mI32R/MWu7XF116aDSRhxbRERs8f499bkyZeeKae
|
|
kzE2mdon81/tfxDLGOunwbzlzbx08oaHBvZHJlx48mrvaa94pu04y617576rNGLRRM0397JEd/lu
|
|
9Dw/S3x4qxffo6mm4NjwUiKY4iI9Ib1dHFY6QIaNabbrYrLfrpJtaK1rMzPZb/s+05IpP59OyLeJ
|
|
k7eNfRaOc1ue32I7fGXYpi5Y77M8OGMeOKxHSFsU3Y29deZMzirl6dlVvhLatCjJHeYQv1rXnps1
|
|
8k9/VsW6qLVmZIi1rzitlvFKRvaZ2h6TSaenC9FFY+3brM+sqeG8Prp4+kZ+lvuxPkr1mqm95nfp
|
|
DXM459676a2q1dsV7XietvNno78+CJn1cjX6mOeIm0bR33dfRU5NJjidt9t5afjG/V6JZ7I2QMNh
|
|
nyo2BhsMuVG3wAhMSbbQRAMolnE+iuGUSCyJZRKuGUSCyJZK4llEgyZMYTuCUsYSCQASISAAAlCQ
|
|
AAAAAAEoASCASAAAAAAAAAAAAlACRACQAAAAAAAAAEgCEoASCAAAAAAAAAAAAAAAAAAAAAAABAAA
|
|
AAAAAAAISAIAAAAAAQAAACASgAAAQJAQAAhIDHZhln3do7z0WS18mWsajHjmes7pg3dNi5aRMNqO
|
|
yvDHTpPRaigHZhN4hHRlaVN59JY3zRENLUavaO+yq0iNVlitJ6vNcR1MVi0zO0era1/Ea0rPvbz5
|
|
PM5MWp45qvo2GZrhmfrsnpHpHzTCseEcM/2vrr8Q1Eb4qzy44nziPN63HpYiIiI7LNHoqabBTFii
|
|
IpSNohuVxrKtWMEejPwY9G1FFmHB4mWJn7MdfnIM9JpIx15to5pbUaas/a6rqViI7MxPxqX0UT1r
|
|
O3wVzpbR2hviP5i03Y5s6a879FNtHljydhExCv8AMTPJXBnRZbz0iG5ptFjwe/l96zctMVamTJtE
|
|
yTMibu1VrdTzRMR0j0ed4lr64MVpm0RERvMz5NvX62uOJ69XhOKX1HH9bHDtFvNYnfJeOy0Z2ojX
|
|
6jjnEq6fRUmccTvN/J9H0eKcOnx45neaxEbubwHgOHg+milI3vP2resu3Wu0JQmITsmISDHZHKz2
|
|
JgFc1RMLJhGwK9iIZ7MZgEdgmAEwyiWCdwWRLKJVxKYsC2JTuriWUSDNlEsIlMAySx3SCRCQSIAS
|
|
AAACRACQAAAAAAASIASAAAAAAAAAAAAAAACRACRACQASIAAAAAAAAAAAAAAAAAAAAAAAAQCUAAAA
|
|
AAAAAAIAAAAAAAAQAAAAAACBICBICAAEJAQJQCJcLjuS2ny6fPG/LWdpd1o8T0X07SXx/e7wCdJx
|
|
Wa0jmneHQpxPDMdZmJfNtZm49weZrh0/j4o7VtSZ2+Uw0/8A7o49k92vBLc/ntFohFW9PqGXimOI
|
|
6Tu1L8T3eCx6r2t1O3JwvHjifO99v7t/Bwf2l1PXU6rS6eJ8qUm8x+so5TsekzcSjbvs4mt4rzW5
|
|
K2mbT0itesy2cHsvbvqtbmyz5xERWP2jd1tJwrTaONsOKtZ8585+cnDrzmn4Rq+IZObUROHD32n7
|
|
Vv8A0ej0uhxaXFGPFSK1j0bkY4jyZRVZVXFGUVWbGwKsk8mObekNrSW3pWf1a2aYjHbm7bNnQ1id
|
|
PW0TvuDdhJEbQABMsLW2R0ZTMQrvfbz2YWzVhpanUxEd0dWkW5c8R5uXxDX1w4pnfr5Q19XxKuOJ
|
|
2neXltVqtVxbV/RdJ715+1bypANfiOu1HENV9C0MTfNeesx2rD1PAeBYuE6aKx72W3W9/WVnBuB4
|
|
eF4dqRzZbdb5J72l160WVK02ZxCYhOwI23TsnY2BGxsnYBjsiYZsZBjMMZZSgGEolMsQDdG6NwZ7
|
|
piVe6YkFsSziVMWZRILolMSriWUSCyJTuwhMSDMRCQSI3SAlACRCQAAEoAEoASAAAAAAAAACUACR
|
|
ACQAAAAAAAAAAAAASAAAAAAAAAAAAAAAAAAACAAAAAAAAAAAAAABAAAAAAAAAAAAACBKAAAAAAAQ
|
|
JQAAAhICEbJAYTWJ7wx8KvpC0BV4ceieWGewDHlNmWwCNjZICNhIDmcZredBecdpiY69FXCOLW+i
|
|
UiZidukulmxxlx2paN4mNng+K4+I8Hy2yaTfl37TXetoCPfRxfp1qi3F48ofKMvtvxak8s6LDv61
|
|
rZji9rPaLUf5PC+bfttS0q8q3p9W/wBrRMdpUZuKdN99nzvFqPbTVz7nD8OKs+do2/mW3h4D7Xaq
|
|
ZnPrtNpqz35aRaYOHY9Zk4pNt9rR+rl6zi+OnS+WN57Rv1lXp/YrNaYtruL6zNPnGO3hxP6O5w/2
|
|
f0HDuun09Yv55Le9afznqcOvO4tBreMTHu30unnva0bWt8on+70nDuE4OHYYx4Kbesz3tPrMuhGO
|
|
IjpDOKrK9YVpsyiGUQnYGOyUgI2SlAIEmwMWMs9kTAMJYzDOYRMArmGErZhhMArlHmzmGMwDE3Ts
|
|
bAbs4swj5pgFkSziVcM4BZEsolXDKAZwyhjCYBkACQhIAAAAAAAJAAAAAAAAAAAAAAAAAAAShIAA
|
|
AAAAAAJAAAAAAAAAAAAAABAJEAAAAAAAAAAAAAAAIEoBKAAAAAAAAAAAAAAABAlAAAAAAAIAAAAA
|
|
BAkBAkBAkBAlACEgMZjdjbFW8bWrEx8YWANb6Fp+bfwab+vLDKMFK9qxH5L0bAr8OPRPKz2AY7J2
|
|
SbAjYZAI2E7AIEgIEgIEgMdkSy2NgY7MdlmyNoBXsxmFuyNgVTVjNV3KjlBRNTlXTVHKCrlIqt5T
|
|
lBhEMohlFerLlBjEMohMVTEARDKCITsAk2AEgAAAkAAAAAAAAAAAAAAAAAAAAAAAASAAAAAAAAD/
|
|
2Q==`;async function G3e(e){let t=(r,s="application/octet-stream")=>fetch(`data:${s};base64,${r}`).then(i=>i.blob()),a,n;switch(e.config.warmup){case"face":a=await t(nf);break;case"body":case"full":a=await t(rf);break;default:a=null}if(a){let r=await createImageBitmap(a);n=await e.detect(r,e.config),r.close()}return n}async function H3e(e){return new Promise(t=>{let a;switch(e.config.warmup){case"face":a="data:image/jpeg;base64,"+nf;break;case"full":case"body":a="data:image/jpeg;base64,"+rf;break;default:a=""}let n;if(typeof Image!="undefined")n=new Image;else if(ne.Image)n=new ne.Image;else{t(void 0);return}n.onload=async()=>{let r=Fn(n.naturalWidth,n.naturalHeight);if(!r)K("Warmup: Canvas not found"),t(void 0);else{let s=r.getContext("2d");s&&s.drawImage(n,0,0);let i=await e.image(r,!0),o=i.tensor?await e.detect(i.tensor,e.config):void 0;t(o)}},a?n.src=a:t(void 0)})}async function j3e(e){let t=r=>Buffer.from(r,"base64"),a;e.config.warmup==="face"?a=t(nf):a=t(rf);let n;if("node"in je&&ha()==="tensorflow"){let r=(void 0).decodeJpeg(a),s=Ht(r,0);e.tf.dispose(r),n=await e.detect(s,e.config),e.tf.dispose(s)}else e.config.debug&&K("Warmup tfjs-node not loaded");return n}async function q3e(e){let t;return typeof createImageBitmap=="function"?t=await G3e(e):typeof Image!="undefined"||ne.Canvas!==void 0?t=await H3e(e):t=await j3e(e),t}async function X3e(e){var o,l,u,p;if(!W().flagRegistry.ENGINE_COMPILE_ONLY)return;let t=ha(),a=nr();if(t!=="webgl"&&t!=="humangl"||!(a!=null&&a.checkCompileCompletion))return;W().set("ENGINE_COMPILE_ONLY",!0);let n=It().state.numTensors,r=[];for(let[c,d]of Object.entries(e.models.models)){if(!d)continue;let h=d!=null&&d.modelSignature&&((l=(o=d==null?void 0:d.inputs)==null?void 0:o[0])!=null&&l.shape)?[...d.inputs[0].shape]:[1,64,64,3],f=d!=null&&d.modelSignature&&((p=(u=d==null?void 0:d.inputs)==null?void 0:u[0])!=null&&p.dtype)?d.inputs[0].dtype:"float32";for(let g=0;g<h.length;g++)h[g]===-1&&(h[g]=g===0?1:64);let m=gn(h,f);try{let g=d.execute(m);r.push(c),Array.isArray(g)?g.forEach(y=>J(y)):J(g)}catch(g){e.config.debug&&K("compile fail model:",c)}J(m)}let s=await a.checkCompileCompletionAsync();a.getUniformLocations(),e.config.debug&&K("compile pass:",{models:r,kernels:s.length}),W().set("ENGINE_COMPILE_ONLY",!1);let i=It().state.numTensors;i-n>0&&K("tensor leak:",i-n)}async function gS(e,t){await Jp(e,!1);let a=ae();return e.state="warmup",t&&(e.config=Et(e.config,t)),!e.config.warmup||e.config.warmup.length===0||e.config.warmup==="none"?pr():new Promise(async n=>{await e.models.load(),await X3e(e);let r=await q3e(e),s=ae();e.config.debug&&K("warmup",e.config.warmup,Math.round(s-a),"ms"),e.emit("warmup"),n(r)})}var rd,oc,lc,sf,Ns,h5=class{constructor(t){he(this,"version");he(this,"config");he(this,"result");he(this,"state");he(this,"process");he(this,"tf");he(this,"env",ne);he(this,"draw",v0);he(this,"match",K0);he(this,"models");he(this,"events");he(this,"faceTriangulation");he(this,"faceUVMap");he(this,"performance");qn(this,rd,void 0);qn(this,oc,void 0);qn(this,lc,void 0);he(this,"analyze",(...t)=>{if(!Xa(this,oc))return;let a=this.tf.engine().state.numTensors,n=Xa(this,rd);xr(this,rd,a);let r=a-n;r!==0&&K(...t,r)});qn(this,sf,t=>{if(!Xa(this,lc))return null;if(!t)return"input is not defined";if(this.env.node&&!(t instanceof gt))return"input must be a tensor";try{this.tf.getBackend()}catch(a){return"backend not loaded"}return null});he(this,"webcam",new f0);he(this,"emit",t=>{var a;(a=this.events)!=null&&a.dispatchEvent&&this.events.dispatchEvent(new Event(t))});qn(this,Ns,{});let a=(qp.tfjs||j3).replace(/-(.*)/,"");Qo.wasmPath=`https://cdn.jsdelivr.net/npm/@tensorflow/tfjs-backend-wasm@${a}/dist/`,Qo.modelBasePath=ne.browser?"../models/":"file://models/",this.version=Vg,Object.defineProperty(this,"version",{value:Vg}),this.config=JSON.parse(JSON.stringify(Qo)),Object.seal(this.config),this.config.cacheModels=typeof indexedDB!="undefined",t&&(this.config=Et(this.config,t)),Tk(this.config),this.tf=je,this.state="idle",xr(this,rd,0),xr(this,oc,!1),xr(this,lc,!1),this.performance={},this.events=typeof EventTarget!="undefined"?new EventTarget:void 0,this.models=new ic(this),Kg(),this.result=pr(),this.process={tensor:null,canvas:null},this.faceTriangulation=b9,this.faceUVMap=v9,af(this,null,""),this.emit("create"),(this.config.debug||this.env.browser)&&K(`version: ${this.version}`),this.config.debug&&K(`tfjs version: ${this.tf.version["tfjs-core"]}`);let n=JSON.parse(JSON.stringify(this.env));delete n.kernels,delete n.initial,delete n.perfadd,this.config.debug&&K("environment:",n)}reset(){let t=this.config.backend;this.config=JSON.parse(JSON.stringify(Qo)),this.config.backend=t,Wg(),ne.initial=!0}validate(t){let a=Og(Qo,t||this.config);return a.length===0&&(this.config=Et(this.config,t)),a}now(){return ae()}image(t,a=!1){return c0(t,this.config,a)}async segmentation(t,a){var s,i,o;if(a&&(this.config=Et(this.config,a)),!this.config.segmentation.enabled)return null;let n=await c0(t,this.config);if(!n.tensor)return null;let r=null;return(s=this.config.segmentation.modelPath)!=null&&s.includes("rvm")&&(r=await pS(n.tensor,this.config)),(i=this.config.segmentation.modelPath)!=null&&i.includes("meet")&&(r=await UI(n.tensor,this.config)),(o=this.config.segmentation.modelPath)!=null&&o.includes("selfie")&&(r=await hS(n.tensor,this.config)),J(n.tensor),r}compare(t,a){return Ck(this.config,t,a)}async init(){await Jp(this,!0),await this.tf.ready(),Wg()}async load(t){this.state="load";let a=ae(),n=Object.values(this.models.models).filter(i=>i).length;t&&(this.config=Et(this.config,t)),this.env.initial&&(await Jp(this,!1)||K("error: backend check failed"),await Np(),this.env.browser&&(this.config.debug&&K("configuration:",this.config),this.config.debug&&K("tf flags:",this.tf.ENV.flags))),await this.models.load(this),this.env.initial&&this.config.debug&&K("tf engine state:",this.tf.engine().state.numBytes,"bytes",this.tf.engine().state.numTensors,"tensors"),this.env.initial=!1,Object.values(this.models.models).filter(i=>i).length!==n&&(this.models.validate(),this.emit("load"));let s=Math.trunc(ae()-a);s>(this.performance.loadModels||0)&&(this.performance.loadModels=this.env.perfadd?(this.performance.loadModels||0)+s:s)}next(t=this.result){return VI(t,this.config)}async warmup(t){let a=ae(),n=await gS(this,t),r=ae();return this.performance.warmup=Math.trunc(r-a),n}async profile(t,a){let n=await this.tf.profile(()=>this.detect(t,a)),r={},s=0;for(let o of n.kernels){let l=Number(o.kernelTimeMs)||0;r[o.name]?r[o.name]+=l:r[o.name]=l,s+=l}let i=[];Object.entries(r).forEach(o=>i.push({kernel:o[0],time:o[1],perc:0}));for(let o of i)o.perc=Math.round(1e3*o.time/s)/1e3,o.time=Math.round(1e3*o.time)/1e3;return i.sort((o,l)=>l.time-o.time),i.length=20,i}async detect(t,a){return this.state="detect",new Promise(async n=>{var g,y,x,A,b,w,C,T,N,M,F,E,I,_,O,B,$,U,G,q,H;this.state="config";let r;this.config=Et(this.config,a),this.state="check";let s=Xa(this,sf).call(this,t);s&&(K(s,t),this.emit("error"),n(pr(s)));let i=ae();await this.load(),r=ae(),this.state="image";let o=await c0(t,this.config);if(this.process=o,this.performance.inputProcess=this.env.perfadd?(this.performance.inputProcess||0)+Math.trunc(ae()-r):Math.trunc(ae()-r),this.analyze("Get Image:"),!o.tensor){this.config.debug&&K("could not convert input to tensor"),this.emit("error"),n(pr("could not convert input to tensor"));return}this.emit("image"),r=ae(),this.config.skipAllowed=await Sk(this.config,o.tensor),this.config.filter.autoBrightness=(this.config.filter.autoBrightness||!1)&&this.config.skipAllowed,this.performance.totalFrames||(this.performance.totalFrames=0),this.performance.cachedFrames||(this.performance.cachedFrames=0),this.performance.totalFrames++,this.config.skipAllowed&&this.performance.cachedFrames++,this.performance.cacheCheck=this.env.perfadd?(this.performance.cacheCheck||0)+Math.trunc(ae()-r):Math.trunc(ae()-r),this.analyze("Check Changed:");let l=[],u=[],p=[],c=[];this.state="detect:face",this.config.async?(l=this.config.face.enabled?Oy(this,o.tensor):[],this.performance.face&&delete this.performance.face):(r=ae(),l=this.config.face.enabled?await Oy(this,o.tensor):[],this.performance.face=this.env.perfadd?(this.performance.face||0)+Math.trunc(ae()-r):Math.trunc(ae()-r)),this.config.async&&(this.config.body.maxDetected===-1||this.config.hand.maxDetected===-1)&&(l=await l),this.analyze("Start Body:"),this.state="detect:body";let d=this.config.body.maxDetected===-1?Et(this.config,{body:{maxDetected:this.config.face.enabled?1*l.length:1}}):this.config;this.config.async?((g=this.config.body.modelPath)!=null&&g.includes("posenet")?u=this.config.body.enabled?o5(o.tensor,d):[]:(y=this.config.body.modelPath)!=null&&y.includes("blazepose")?u=this.config.body.enabled?ey(o.tensor,d):[]:(x=this.config.body.modelPath)!=null&&x.includes("efficientpose")?u=this.config.body.enabled?oy(o.tensor,d):[]:(A=this.config.body.modelPath)!=null&&A.includes("movenet")&&(u=this.config.body.enabled?e5(o.tensor,d):[]),this.performance.body&&delete this.performance.body):(r=ae(),(b=this.config.body.modelPath)!=null&&b.includes("posenet")?u=this.config.body.enabled?await o5(o.tensor,d):[]:(w=this.config.body.modelPath)!=null&&w.includes("blazepose")?u=this.config.body.enabled?await ey(o.tensor,d):[]:(C=this.config.body.modelPath)!=null&&C.includes("efficientpose")?u=this.config.body.enabled?await oy(o.tensor,d):[]:(T=this.config.body.modelPath)!=null&&T.includes("movenet")&&(u=this.config.body.enabled?await e5(o.tensor,d):[]),this.performance.body=this.env.perfadd?(this.performance.body||0)+Math.trunc(ae()-r):Math.trunc(ae()-r)),this.analyze("End Body:"),this.analyze("Start Hand:"),this.state="detect:hand";let h=this.config.hand.maxDetected===-1?Et(this.config,{hand:{maxDetected:this.config.face.enabled?2*l.length:1}}):this.config;this.config.async?((M=(N=this.config.hand.detector)==null?void 0:N.modelPath)!=null&&M.includes("handdetect")?p=this.config.hand.enabled?Vy(o.tensor,h):[]:(E=(F=this.config.hand.detector)==null?void 0:F.modelPath)!=null&&E.includes("handtrack")&&(p=this.config.hand.enabled?Hy(o.tensor,h):[]),this.performance.hand&&delete this.performance.hand):(r=ae(),(_=(I=this.config.hand.detector)==null?void 0:I.modelPath)!=null&&_.includes("handdetect")?p=this.config.hand.enabled?await Vy(o.tensor,h):[]:(B=(O=this.config.hand.detector)==null?void 0:O.modelPath)!=null&&B.includes("handtrack")&&(p=this.config.hand.enabled?await Hy(o.tensor,h):[]),this.performance.hand=this.env.perfadd?(this.performance.hand||0)+Math.trunc(ae()-r):Math.trunc(ae()-r)),this.analyze("End Hand:"),this.analyze("Start Object:"),this.state="detect:object",this.config.async?(($=this.config.object.modelPath)!=null&&$.includes("nanodet")?c=this.config.object.enabled?a5(o.tensor,this.config):[]:(U=this.config.object.modelPath)!=null&&U.includes("centernet")&&(c=this.config.object.enabled?ny(o.tensor,this.config):[]),this.performance.object&&delete this.performance.object):(r=ae(),(G=this.config.object.modelPath)!=null&&G.includes("nanodet")?c=this.config.object.enabled?await a5(o.tensor,this.config):[]:(q=this.config.object.modelPath)!=null&&q.includes("centernet")&&(c=this.config.object.enabled?await ny(o.tensor,this.config):[]),this.performance.object=this.env.perfadd?(this.performance.object||0)+Math.trunc(ae()-r):Math.trunc(ae()-r)),this.analyze("End Object:"),this.state="detect:await",this.config.async&&([l,u,p,c]=await Promise.all([l,u,p,c])),this.state="detect:gesture";let f=[];this.config.gesture.enabled&&(r=ae(),f=[...xI(l),...yI(u),...bI(p),...AI(l)],this.config.async?this.performance.gesture&&delete this.performance.gesture:this.performance.gesture=this.env.perfadd?(this.performance.gesture||0)+Math.trunc(ae()-r):Math.trunc(ae()-r)),this.performance.total=this.env.perfadd?(this.performance.total||0)+Math.trunc(ae()-i):Math.trunc(ae()-i);let m=((H=this.process.tensor)==null?void 0:H.shape)||[0,0,0,0];this.result={face:l,body:u,hand:p,gesture:f,object:c,performance:this.performance,canvas:this.process.canvas,timestamp:Date.now(),error:null,width:m[2],height:m[1],get persons(){return mS(l,u,p,f,m)}},J(o.tensor),this.emit("detect"),this.state="idle",n(this.result)})}async sleep(t){return new Promise(a=>{setTimeout(a,t)})}async video(t,a=!0,n=0){a?(Xa(this,Ns)[t.id]||(this.config.debug&&K("video start",t.id),Xa(this,Ns)[t.id]=!0),!t.paused&&Xa(this,Ns)[t.id]&&t.readyState>=2&&await this.detect(t),n>0&&await this.sleep(n),Xa(this,Ns)[t.id]&&requestAnimationFrame(()=>this.video(t,a,n))):(this.config.debug&&K("video stop",t.id),Xa(this,Ns)[t.id]=!1)}};rd=new WeakMap,oc=new WeakMap,lc=new WeakMap,sf=new WeakMap,Ns=new WeakMap;return tC(Y3e);})();
|