@@ -291,19 +291,65 @@ def tensor_names_from_structed(concrete_func, input_names, output_names):
291291 return tensors_to_rename
292292
293293
294+ def _rename_duplicate_keras_model_names (model ):
295+ """
296+ In very rare cases, keras has a bug where it will give multiple outputs the same name.
297+ We must edit the model or the TF trace will fail. Returns old_out_names (or None if no edit was made).
298+ IMPORTANT: model may be edited. Assign model.output_names to old_out_names to restore.
299+ """
300+ old_out_names = None
301+ if model .output_names and len (set (model .output_names )) != len (model .output_names ):
302+ # In very rare cases, keras has a bug where it will give multiple outputs the same name
303+ # We must edit the model or the TF trace will fail
304+ old_out_names = model .output_names
305+ used_names = set ()
306+ new_out_names = []
307+ for name in model .output_names :
308+ new_name = name
309+ i = 0
310+ while new_name in used_names :
311+ i += 1
312+ new_name = name + "_" + str (i )
313+ used_names .add (new_name )
314+ new_out_names .append (new_name )
315+ model .output_names = new_out_names
316+ return old_out_names
317+
318+
319+ def _is_legacy_keras_model (model ):
320+ """Inspects model class to determine if it is from tf or legacy keras"""
321+
322+ logger = logging .getLogger (constants .TF2ONNX_PACKAGE_NAME )
323+ unknown_type_err = "model is not instance of tf.keras.Model or keras.Model"
324+ if isinstance (model , tf .keras .Model ):
325+ return False
326+ try :
327+ import keras # pylint: disable=import-outside-toplevel
328+ if isinstance (model , keras .Model ):
329+ return True
330+ logger .warning (unknown_type_err )
331+ except ImportError :
332+ logger .warning (unknown_type_err )
333+ return False
334+
335+
294336def _from_keras_tf1 (model , input_signature = None , opset = None , custom_ops = None , custom_op_handlers = None ,
295337 custom_rewriter = None , inputs_as_nchw = None , extra_opset = None , shape_override = None ,
296338 target = None , large_model = False , output_path = None ):
297339 """from_keras for tf 1.15"""
298-
299340 input_names = [t .name for t in model .inputs ]
300341 output_names = [t .name for t in model .outputs ]
342+ old_out_names = _rename_duplicate_keras_model_names (model )
301343 tensors_to_rename = dict (zip (input_names , model .input_names ))
302- if len ( set ( model . output_names )) == len ( model .output_names ):
303- # In very rare cases, keras has a bug where it will give multiple outputs the same name
304- tensors_to_rename . update ( zip ( output_names , model .output_names ))
344+ tensors_to_rename . update ( zip ( output_names , model .output_names ))
345+ if old_out_names is not None :
346+ model .output_names = old_out_names
305347
306- sess = tf .keras .backend .get_session (model .outputs )
348+ if _is_legacy_keras_model (model ):
349+ import keras # pylint: disable=import-outside-toplevel
350+ sess = keras .backend .get_session ()
351+ else :
352+ sess = tf .keras .backend .get_session (model .outputs )
307353
308354 with tf .device ("/cpu:0" ):
309355 frozen_graph , initialized_tables = tf_loader .freeze_session (sess , input_names , output_names , get_tables = True )
@@ -351,6 +397,7 @@ def from_keras(model, input_signature=None, opset=None, custom_ops=None, custom_
351397 Returns:
352398 An ONNX model_proto and an external_tensor_storage dict.
353399 """
400+ old_out_names = _rename_duplicate_keras_model_names (model )
354401 if LooseVersion (tf .__version__ ) < "2.0" :
355402 return _from_keras_tf1 (model , input_signature , opset , custom_ops , custom_op_handlers , custom_rewriter ,
356403 inputs_as_nchw , extra_opset , shape_override , target , large_model , output_path )
@@ -370,9 +417,21 @@ def wrap_call(*args, training=False, **kwargs):
370417 return model_call (* args , ** kwargs )
371418 model .call = wrap_call
372419 function = _saving_utils .trace_model_call (model , input_signature )
373- concrete_func = function .get_concrete_function ()
374- # Put it back
375- model .call = model_call
420+ try :
421+ # Legacy keras get make TF erroneously enter eager mode when it should be making symbolic tensors
422+ import tensorflow_core # pylint: disable=import-outside-toplevel
423+ old_get_learning_phase = tensorflow_core .python .keras .backend .learning_phase
424+ tensorflow_core .python .keras .backend .learning_phase = \
425+ tensorflow_core .python .keras .backend .symbolic_learning_phase
426+ except ImportError :
427+ old_get_learning_phase = None
428+ try :
429+ concrete_func = function .get_concrete_function ()
430+ finally :
431+ # Put everything back
432+ model .call = model_call
433+ if old_get_learning_phase is not None :
434+ tensorflow_core .python .keras .backend .learning_phase = old_get_learning_phase
376435
377436 # These inputs will be removed during freezing (includes resources, etc.)
378437 graph_captures = concrete_func .graph ._captures # pylint: disable=protected-access
@@ -392,6 +451,9 @@ def wrap_call(*args, training=False, **kwargs):
392451 # Other models specify output order using the key order of structured_outputs
393452 output_names = [reverse_lookup [out ] for out in concrete_func .structured_outputs .keys ()]
394453
454+ if old_out_names is not None :
455+ model .output_names = old_out_names
456+
395457 with tf .device ("/cpu:0" ):
396458 frozen_graph , initialized_tables = \
397459 tf_loader .from_trackable (model , concrete_func , input_names , output_names , large_model )
0 commit comments