Remove training=True from inbound nodes during conversion, fixes #284

keras_export/convert_model.py

@@ -6,6 +6,7 @@
 import datetime
 import hashlib
 import json
+import os
 import sys
 
 import numpy as np
@@ -596,10 +597,6 @@ def get_all_weights(model, prefix):
     layers = model.layers
     assert K.image_data_format() == 'channels_last'
     for layer in layers:
-        for node in layer.inbound_nodes:
-            if "training" in node.call_kwargs:
-                assert node.call_kwargs["training"] is not True, \
-                    "training=true is not supported, see https://github.com/Dobiasd/frugally-deep/issues/284"
         layer_type = type(layer).__name__
         name = prefix + layer.name
         assert is_ascii(name)
@@ -750,6 +747,32 @@ def singleton_list_to_value(value_or_values):
     return value_or_values
 
 
+def remove_training_flags(model):
+    """
+    In case a layer has set training=True, this flag is removed.
+    """
+    layers = model.layers
+    for layer in layers:
+        for node in layer.inbound_nodes:
+            if "training" in node.call_kwargs:
+                print(f"Removing training=True from inbound node to layer named {layer.name}.")
+                del node.call_kwargs["training"]
+        layer_type = type(layer).__name__
+        if layer_type in ['Model', 'Sequential', 'Functional']:
+            remove_training_flags(layer)
+
+
+def remove_training_flags_and_reload(model, temp_model_filename):
+    """
+    The reload is needed to propagate the change in the model.
+    """
+    remove_training_flags(model)
+    model.save(temp_model_filename, include_optimizer=False)
+    model = load_model(temp_model_filename)
+    os.remove(temp_model_filename)
+    return model
+
+
 def model_to_fdeep_json(model, no_tests=False):
     """Convert any Keras model to the frugally-deep model format."""
 
@@ -788,6 +811,7 @@ def convert(in_path, out_path, no_tests=False):
 
     print('loading {}'.format(in_path))
     model = load_model(in_path)
+    model = remove_training_flags_and_reload(model, out_path + ".fdeep_temp.h5")
     json_output = model_to_fdeep_json(model, no_tests)
     print('writing {}'.format(out_path))
     write_text_file(out_path, json.dumps(