ambiser
/
human
mirror of https://github.com/vladmandic/human
9
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

cleanup blazepose code

pull/91/head
Vladimir Mandic 2021-03-05 08:03:00 -05:00
parent 02439d2c53
commit 32247c1b06
1 changed files with 13 additions and 11 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

24
src/blazepose/blazepose.ts
View File

@ -8,6 +8,7 @@ let model;
export async function load(config) { export async function load(config) {
if (!model) { if (!model) {
model = await tf.loadGraphModel(config.body.modelPath); model = await tf.loadGraphModel(config.body.modelPath);
// blazepose inputSize is 256x256px, but we can find that out dynamically
model.width = parseInt(model.signature.inputs['input_1:0'].tensorShape.dim[2].size); model.width = parseInt(model.signature.inputs['input_1:0'].tensorShape.dim[2].size);
model.height = parseInt(model.signature.inputs['input_1:0'].tensorShape.dim[1].size); model.height = parseInt(model.signature.inputs['input_1:0'].tensorShape.dim[1].size);
if (config.debug) log(`load model: ${config.body.modelPath.match(/\/(.*)\./)[1]}`); if (config.debug) log(`load model: ${config.body.modelPath.match(/\/(.*)\./)[1]}`);
@ -25,31 +26,29 @@ export async function predict(image, config) {
// let segmentation; // not used right now since we have keypoints and don't need to go through matrix using strides // let segmentation; // not used right now since we have keypoints and don't need to go through matrix using strides
// let poseflag; // irrelevant // let poseflag; // irrelevant
let points; let points;
if (!config.profile) { if (!config.profile) { // run through profiler or just execute
const resT = await model.predict(normalize); const resT = await model.predict(normalize);
// segmentation = resT[0].dataSync(); // segmentation = resT[0].dataSync();
// poseflag = resT[1].dataSync(); // poseflag = resT[1].dataSync();
points = resT.find((t) => (t.size === 195 || t.size === 155)).dataSync(); points = resT.find((t) => (t.size === 195 || t.size === 155)).dataSync(); // order of output tensors may change between models, full has 195 and upper has 155 items
resT.forEach((t) => t.dispose()); resT.forEach((t) => t.dispose());
} else { } else {
const profileData = await tf.profile(() => model.predict(normalize)); const profileData = await tf.profile(() => model.predict(normalize));
// segmentation = profileData.result[0].dataSync(); points = profileData.result.find((t) => (t.size === 195 || t.size === 155)).dataSync();
// poseflag = profileData.result[1].dataSync();
points = profileData.result.find((t) => t.size === 195).dataSync(); // find a tensor with 195 items which is 39 points with 5 properties
profileData.result.forEach((t) => t.dispose()); profileData.result.forEach((t) => t.dispose());
profile.run('blazepose', profileData); profile.run('blazepose', profileData);
} }
normalize.dispose(); normalize.dispose();
const keypoints: Array<{ id, part, position: { x, y, z }, score, presence }> = []; const keypoints: Array<{ id, part, position: { x, y, z }, score, presence }> = [];
const labels = points.length === 195 ? annotations.full : annotations.upper; const labels = points.length === 195 ? annotations.full : annotations.upper; // full model has 39 keypoints, upper has 31 keypoints
const depth = 5; const depth = 5; // each points has x,y,z,visibility,presence
for (let i = 0; i < points.length / depth; i++) { for (let i = 0; i < points.length / depth; i++) {
keypoints.push({ keypoints.push({
id: i, id: i,
part: labels[i], part: labels[i],
position: { position: {
x: Math.trunc(imgSize.width * points[depth * i + 0] / 255), x: Math.trunc(imgSize.width * points[depth * i + 0] / 255), // return normalized x value istead of 0..255
y: Math.trunc(imgSize.height * points[depth * i + 1] / 255), y: Math.trunc(imgSize.height * points[depth * i + 1] / 255), // return normalized y value istead of 0..255
z: Math.trunc(points[depth * i + 2]) + 0, // fix negative zero z: Math.trunc(points[depth * i + 2]) + 0, // fix negative zero
}, },
score: (100 - Math.trunc(100 / (1 + Math.exp(points[depth * i + 3])))) / 100, // reverse sigmoid value score: (100 - Math.trunc(100 / (1 + Math.exp(points[depth * i + 3])))) / 100, // reverse sigmoid value
@ -61,6 +60,9 @@ export async function predict(image, config) {
} }
/* /*
Model card: https://drive.google.com/file/d/10IU-DRP2ioSNjKFdiGbmmQX81xAYj88s/view Model card:
Download: https://github.com/PINTO0309/PINTO_model_zoo/tree/main/058_BlazePose_Full_Keypoints - https://drive.google.com/file/d/10IU-DRP2ioSNjKFdiGbmmQX81xAYj88s/view
Download:
- https://github.com/PINTO0309/PINTO_model_zoo/tree/main/058_BlazePose_Full_Keypoints/10_new_256x256/saved_model/tfjs_model_float16
- https://github.com/PINTO0309/PINTO_model_zoo/tree/main/053_BlazePose/20_new_256x256/saved_model/tfjs_model_float16
*/ */