update

src/tfjs-to-saved.py

@@ -1,36 +0,0 @@
-import os
-import sys
-import glob
-import tensorflow as tf
-import tfjs_graph_converter.api as tfjs
-
-graphDir = 'models/'
-savedDir = 'saved/'
-tfliteDir = 'tflite/'
-
-def main() -> None:
-  for f in glob.glob(os.path.join(graphDir, '*.json')):
-    modelName = os.path.basename(f).split('.')[0]
-    print('graph model: ' + modelName + ' path: ' + f)
-    savedModel = os.path.join(savedDir, modelName)
-    try:
-      tfjs.graph_model_to_saved_model(f, savedModel)  # type: ignore
-    except:
-      print('saved convert failed')
-    else:
-      converter = tf.lite.TFLiteConverter.from_saved_model(savedModel)
-      converter.optimizations = [tf.lite.Optimize.DEFAULT]
-      converter.target_spec.supported_ops = [ tf.lite.OpsSet.TFLITE_BUILTINS, tf.lite.OpsSet.SELECT_TF_OPS ]
-      converter.target_spec.supported_types = [tf.float16]
-      tfliteModel = os.path.join(tfliteDir, modelName)
-      try:
-        tflite_model = converter.convert()
-      except:
-        print('tflite convert failed')
-      else:
-        with open(tfliteModel, 'wb') as f:
-          f.write(tflite_model)
-        print('saved:' + savedModel + ' tflite: ' + tfliteModel)
-
-if __name__ == '__main__':
-  main()

src/tfjs2tflite.py

@@ -0,0 +1,57 @@
+import os
+import glob
+import tensorflow as tf
+import tfjs_graph_converter.api as tfjs
+
+graphDir = 'models/'
+savedDir = 'saved/'
+tfliteDir = 'tflite/'
+
+def saved2tflite(savedModelDir, tfliteModelName):
+  if (os.path.isfile(os.path.join(savedModelDir, 'saved_model.pb'))):
+    converter = tf.lite.TFLiteConverter.from_saved_model(savedModelDir)
+    converter.optimizations = [tf.lite.Optimize.DEFAULT] # type: ignore
+    converter.target_spec.supported_ops = [ tf.lite.OpsSet.TFLITE_BUILTINS, tf.lite.OpsSet.SELECT_TF_OPS ]
+    # converter.target_spec.experimental_supported_backends = ["CPU", "GPU"]
+    converter.target_spec.supported_types = [tf.float16]
+    converter.allow_custom_ops = True
+    converter.exclude_conversion_metadata = True
+    try:
+      tfliteModel = converter.convert()
+      # tf.lite.experimental.Analyzer.analyze(model_content = tfliteModel)
+    except:
+      print('  tflite convert failed')
+    else:
+      with open(tfliteModelName, 'wb') as f:
+        f.write(tfliteModel)
+      print('  tflite model', tfliteModelName)
+  else:
+    print('  tf saved model missing:', savedModelDir)
+
+
+def tfjs2saved(graphJsonFile, savedModelDir):
+  if (not os.path.exists(savedModelDir)):
+    try:
+      tfjs.graph_model_to_saved_model(graphJsonFile, savedModelDir) # type: ignore
+    except:
+      print('  tf saved convert failed:', graphJsonFile)
+    else:
+      print('  tf saved model:', savedModelDir)
+  else:
+    print('  tf saved model exists:', savedModelDir)
+
+
+def main():
+  tf.compat.v1.enable_control_flow_v2()
+  for graphJsonFile in glob.glob(os.path.join(graphDir, '*.json')):
+    modelName = os.path.basename(graphJsonFile).split('.')[0]
+    print('model:', modelName)
+    print('  tfjs graph model:', graphJsonFile)
+    savedModelDir = os.path.join(savedDir, modelName)
+    tfliteModelFile = os.path.join(tfliteDir, modelName) + '.tflite'
+    tfjs2saved(graphJsonFile, savedModelDir)
+    saved2tflite(savedModelDir, tfliteModelFile)
+
+
+if __name__ == '__main__':
+  main()