Coverage for /pythoncovmergedfiles/medio/medio/usr/local/lib/python3.8/site-packages/keras/src/models/cloning.py: 14%
281 statements
« prev ^ index » next coverage.py v7.4.0, created at 2024-01-03 07:57 +0000
1# Copyright 2015 The TensorFlow Authors. All Rights Reserved.
2#
3# Licensed under the Apache License, Version 2.0 (the "License");
4# you may not use this file except in compliance with the License.
5# You may obtain a copy of the License at
6#
7# http://www.apache.org/licenses/LICENSE-2.0
8#
9# Unless required by applicable law or agreed to in writing, software
10# distributed under the License is distributed on an "AS IS" BASIS,
11# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12# See the License for the specific language governing permissions and
13# limitations under the License.
14# ==============================================================================
16"""Code for model cloning, plus model-related API entries."""
18import tensorflow.compat.v2 as tf
20from keras.src import backend
21from keras.src import metrics as metrics_module
22from keras.src.engine import functional
23from keras.src.engine import sequential
24from keras.src.engine import training
25from keras.src.engine import training_v1
26from keras.src.engine.base_layer import AddMetric
27from keras.src.engine.base_layer import Layer
28from keras.src.engine.input_layer import Input
29from keras.src.engine.input_layer import InputLayer
30from keras.src.optimizers import optimizer_v1
31from keras.src.saving.legacy import serialization
32from keras.src.saving.legacy.saved_model.utils import keras_option_scope
33from keras.src.saving.object_registration import CustomObjectScope
34from keras.src.utils import generic_utils
35from keras.src.utils import version_utils
37# isort: off
38from tensorflow.python.platform import tf_logging as logging
39from tensorflow.python.util.tf_export import keras_export
41# API entries importable from `keras.models`:
42Model = training.Model
43Sequential = sequential.Sequential
46# Callable used to clone a layer with weights preserved.
47def share_weights(layer):
48 return layer
51def _clone_layer(layer):
52 return layer.__class__.from_config(layer.get_config())
55def _insert_ancillary_layers(model, ancillary_layers, metrics_names, new_nodes):
56 """Inserts ancillary layers into the model with the proper order."""
57 # Sort `AddMetric` layers so they agree with metrics_names.
58 metric_layers = [
59 layer for layer in ancillary_layers if isinstance(layer, AddMetric)
60 ]
61 metric_layers.sort(key=lambda layer: metrics_names.index(layer.metric_name))
62 ancillary_layers = [
63 layer for layer in ancillary_layers if not isinstance(layer, AddMetric)
64 ] + metric_layers
65 model._insert_layers(ancillary_layers, relevant_nodes=list(new_nodes))
68def _make_new_nodes(nodes_by_depth, layer_fn, layer_map, tensor_map):
69 """Make new nodes with the layers in `layer_map` based on `nodes_by_depth`.
71 Args:
72 nodes_by_depth: Provides structure information to create new nodes.
73 layer_fn: Function to clone layers.
74 layer_map: Map from layers in `model` to new layers.
75 tensor_map: Map from tensors in `model` to newly compute tensors.
77 Returns:
78 A set of new nodes. `layer_map` and `tensor_map` are updated.
79 """
80 # Iterated over every node in the reference model, in depth order.
81 new_nodes = set()
82 depth_keys = list(nodes_by_depth.keys())
83 depth_keys.sort(reverse=True)
84 for depth in depth_keys:
85 nodes = nodes_by_depth[depth]
86 for node in nodes:
87 # Recover the corresponding layer.
88 layer = node.outbound_layer
90 # Get or create layer.
91 if layer not in layer_map:
92 new_layer = layer_fn(layer)
93 layer_map[layer] = new_layer
94 layer = new_layer
95 else:
96 # Reuse previously cloned layer.
97 layer = layer_map[layer]
98 # Don't call InputLayer multiple times.
99 if isinstance(layer, InputLayer):
100 continue
102 # If all previous input tensors are available in tensor_map,
103 # then call node.inbound_layer on them.
104 if all(
105 tensor in tensor_map
106 for tensor in tf.nest.flatten(node.input_tensors)
107 ):
108 # Call layer.
109 args = tf.nest.map_structure(
110 lambda t: tensor_map.get(t, t), node.call_args
111 )
112 kwargs = tf.nest.map_structure(
113 lambda t: tensor_map.get(t, t), node.call_kwargs
114 )
115 output_tensors = layer(*args, **kwargs)
117 # Thread-safe way to keep track of what node was created.
118 first_output_tensor = tf.nest.flatten(output_tensors)[0]
119 new_nodes.add(
120 layer._inbound_nodes[
121 first_output_tensor._keras_history.node_index
122 ]
123 )
125 for x, y in zip(
126 tf.nest.flatten(node.output_tensors),
127 tf.nest.flatten(output_tensors),
128 ):
129 tensor_map[x] = y
130 return new_nodes
133def _clone_functional_model(model, input_tensors=None, layer_fn=_clone_layer):
134 """Clone a functional `Model` instance.
136 Model cloning is similar to calling a model on new inputs,
137 except that it creates new layers (and thus new weights) instead
138 of sharing the weights of the existing layers.
140 Input layers are always cloned.
142 Args:
143 model: Instance of `Model`.
144 input_tensors: optional list of input tensors
145 to build the model upon. If not provided,
146 placeholders will be created.
147 layer_fn: callable to be applied on non-input layers in the model. By
148 default it clones the layer. Another example is to preserve the
149 layer to share the weights. This is required when we create a
150 per-replica copy of the model with distribution strategy; we want
151 the weights to be shared but still feed inputs separately so we
152 create new input layers.
154 Returns:
155 An instance of `Model` reproducing the behavior
156 of the original model, on top of new inputs tensors,
157 using newly instantiated weights.
159 Raises:
160 ValueError: in case of invalid `model` argument value or `layer_fn`
161 argument value.
162 """
163 if layer_fn is None:
164 layer_fn = _clone_layer
166 if not isinstance(model, Model):
167 raise ValueError(
168 "Expected `model` argument "
169 f"to be a `Model` instance. Received: model={model}"
170 )
171 if isinstance(model, Sequential):
172 raise ValueError(
173 "Expected `model` argument "
174 "to be a functional `Model` instance, "
175 f"got a `Sequential` instance instead: {model}"
176 )
177 if not model._is_graph_network:
178 raise ValueError(
179 "Expected `model` argument "
180 "to be a functional `Model` instance, "
181 f"but got a subclassed model instead: {model}"
182 )
184 new_input_layers = {} # Cache for created layers.
185 if input_tensors is not None:
186 # Make sure that all input tensors come from a Keras layer.
187 input_tensors = tf.nest.flatten(input_tensors)
188 for i, input_tensor in enumerate(input_tensors):
189 original_input_layer = model._input_layers[i]
191 # Cache input layer. Create a new layer if the tensor is originally
192 # not from a Keras layer.
193 if not backend.is_keras_tensor(input_tensor):
194 name = original_input_layer.name
195 input_tensor = Input(
196 tensor=input_tensor, name="input_wrapper_for_" + name
197 )
198 newly_created_input_layer = input_tensor._keras_history.layer
199 new_input_layers[
200 original_input_layer
201 ] = newly_created_input_layer
202 else:
203 new_input_layers[
204 original_input_layer
205 ] = input_tensor._keras_history.layer
207 if not callable(layer_fn):
208 raise ValueError(
209 "Expected `layer_fn` argument to be a callable. "
210 f"Received: layer_fn={layer_fn}"
211 )
213 # For affected g3 users who need to default to old serialization in cloning
214 if getattr(model, "use_legacy_config", False):
215 with keras_option_scope(
216 save_traces=False, in_tf_saved_model_scope=True
217 ):
218 model_configs, created_layers = _clone_layers_and_model_config(
219 model, new_input_layers, layer_fn
220 )
221 else:
222 model_configs, created_layers = _clone_layers_and_model_config(
223 model, new_input_layers, layer_fn
224 )
225 # Reconstruct model from the config, using the cloned layers.
226 (
227 input_tensors,
228 output_tensors,
229 created_layers,
230 ) = functional.reconstruct_from_config(
231 model_configs, created_layers=created_layers
232 )
233 metrics_names = model.metrics_names
234 if functional.has_functional_like_constructor(model.__class__):
235 new_model = model.__class__(
236 input_tensors, output_tensors, name=model.name
237 )
238 else:
239 # This may be incorrect: the new model will end up having a different
240 # class than the original. However various existing models rely
241 # on this behavior, so we keep it.
242 new_model = Model(input_tensors, output_tensors, name=model.name)
244 # Layers not directly tied to outputs of the Model, such as loss layers
245 # created in `add_loss` and `add_metric`.
246 ancillary_layers = [
247 layer
248 for layer in created_layers.values()
249 if layer not in new_model.layers
250 ]
251 # TODO(b/162887610): This may need to adjust the inbound node index if the
252 # created layers had already been used to define other models.
253 if ancillary_layers:
254 new_nodes = tf.nest.flatten(
255 [
256 layer.inbound_nodes[1:]
257 if functional._should_skip_first_node(layer)
258 else layer.inbound_nodes
259 for layer in created_layers.values()
260 ]
261 )
262 _insert_ancillary_layers(
263 new_model, ancillary_layers, metrics_names, new_nodes
264 )
265 return new_model
268def _clone_layers_and_model_config(model, input_layers, layer_fn):
269 """Clones all layers; returns the model config without serializing layers.
271 This function ensures that only the node graph is retrieved when getting the
272 model config. The `layer_fn` used to clone layers might not rely on
273 `layer.get_config()`, so some custom layers do not define `get_config`.
274 Trying to retrieve the config results in errors.
276 Args:
277 model: A Functional model.
278 input_layers: Dictionary mapping input layers in `model` to new input
279 layers.
280 layer_fn: Function used to clone all non-input layers.
282 Returns:
283 Model config object, and a dictionary of newly created layers.
284 """
285 created_layers = {}
287 def _copy_layer(layer):
288 # Whenever the network config attempts to get the layer serialization,
289 # return a dummy dictionary.
290 if layer in input_layers:
291 created_layers[layer.name] = input_layers[layer]
292 elif layer in model._input_layers:
293 created_layers[layer.name] = InputLayer(**layer.get_config())
294 else:
295 created_layers[layer.name] = layer_fn(layer)
296 return {}
298 config = functional.get_network_config(
299 model, serialize_layer_fn=_copy_layer
300 )
301 return config, created_layers
304def _remove_ancillary_layers(model, layer_map, layers):
305 """Removes and returns any ancillary layers from `layers` based on `model`.
307 Ancillary layers are part of the model topology but not used to compute the
308 model outputs, e.g., layers from `add_loss` and `add_metric`.
310 Args:
311 model: A Keras Model.
312 layer_map: A map to from layers in the `model` to those in `layers`.
313 layers: A list of all layers.
315 Returns:
316 Two lists of layers: (1) `layers` with the ancillary layers removed, and
317 (2) the ancillary layers.
318 """
319 ancillary_layers = [] # Additional layers for computing losses and metrics.
320 if not model._is_graph_network:
321 return layers, ancillary_layers
323 # Ancillary layers are those with depth < 0.
324 depths = [depth for depth in model._nodes_by_depth.keys() if depth < 0]
325 depths.sort(reverse=True) # Order topologically from inputs to outputs.
326 for depth in depths:
327 for node in model._nodes_by_depth[depth]:
328 ancillary_layers.append(layer_map[node.outbound_layer])
330 return [l for l in layers if l not in ancillary_layers], ancillary_layers
333def _clone_sequential_model(model, input_tensors=None, layer_fn=_clone_layer):
334 """Clone a `Sequential` model instance.
336 Model cloning is similar to calling a model on new inputs,
337 except that it creates new layers (and thus new weights) instead
338 of sharing the weights of the existing layers.
340 Args:
341 model: Instance of `Sequential`.
342 input_tensors: optional list of input tensors
343 to build the model upon. If not provided,
344 placeholders will be created.
345 layer_fn: callable to be applied on non-input layers in the model. By
346 default it clones the layer. Another example is to preserve the
347 layer to share the weights. This is required when we create a
348 per-replica copy of the model with distribution strategy; we want
349 the weights to be shared but still feed inputs separately so we
350 create new input layers.
352 Returns:
353 An instance of `Sequential` reproducing the behavior
354 of the original model, on top of new inputs tensors,
355 using newly instantiated weights.
357 Raises:
358 ValueError: in case of invalid `model` argument value or `layer_fn`
359 argument value.
360 """
361 if layer_fn is None:
362 layer_fn = _clone_layer
364 if not isinstance(model, Sequential):
365 raise ValueError(
366 "Expected `model` argument "
367 "to be a `Sequential` model instance. "
368 f"Received: model={model}"
369 )
371 if not callable(layer_fn):
372 raise ValueError(
373 "Expected `layer_fn` argument to be a callable. "
374 f"Received: layer_fn={layer_fn}"
375 )
377 layers = [] # Layers needed to compute the model's outputs.
378 layer_map = {}
379 # Ensure that all layers are cloned. The model's layers
380 # property will exclude the initial InputLayer (if it exists) in the model,
381 # resulting in a different Sequential model structure.
382 for layer in model._flatten_layers(include_self=False, recursive=False):
383 if isinstance(layer, InputLayer) and input_tensors is not None:
384 # If input tensors are provided, the original model's InputLayer is
385 # overwritten with a different InputLayer.
386 continue
387 cloned_layer = (
388 _clone_layer(layer)
389 if isinstance(layer, InputLayer)
390 else layer_fn(layer)
391 )
392 layers.append(cloned_layer)
393 layer_map[layer] = cloned_layer
394 layers, ancillary_layers = _remove_ancillary_layers(
395 model, layer_map, layers
396 )
398 if input_tensors is None:
399 cloned_model = Sequential(layers=layers, name=model.name)
400 elif len(generic_utils.to_list(input_tensors)) != 1:
401 raise ValueError(
402 "To clone a `Sequential` model, we expect at most one tensor as "
403 f"part of `input_tensors`. Received: input_tensors={input_tensors}"
404 )
405 else:
406 # Overwrite the original model's input layer.
407 if isinstance(input_tensors, tuple):
408 input_tensors = list(input_tensors)
409 x = generic_utils.to_list(input_tensors)[0]
410 if backend.is_keras_tensor(x):
411 origin_layer = x._keras_history.layer
412 if isinstance(origin_layer, InputLayer):
413 cloned_model = Sequential(
414 layers=[origin_layer] + layers, name=model.name
415 )
416 else:
417 raise ValueError(
418 "Cannot clone a `Sequential` model on top "
419 "of a tensor that comes from a Keras layer "
420 "other than an `InputLayer`. "
421 "Use the Functional API instead. "
422 f"Received: input_tensors={input_tensors}"
423 )
424 else:
425 input_tensor = Input(
426 tensor=x, name="input_wrapper_for_" + str(x.name)
427 )
428 input_layer = input_tensor._keras_history.layer
429 cloned_model = Sequential(
430 layers=[input_layer] + layers, name=model.name
431 )
433 if not ancillary_layers:
434 return cloned_model
436 tensor_map = {} # Maps tensors from `model` to those in `cloned_model`.
437 for depth, cloned_nodes in cloned_model._nodes_by_depth.items():
438 nodes = model._nodes_by_depth[depth]
439 # This should be safe in a Sequential model. In an arbitrary network,
440 # you need to sort using the outbound layer of the node as a key.
441 for cloned_node, node in zip(cloned_nodes, nodes):
442 if isinstance(cloned_node.output_tensors, list):
443 for j, output_tensor in enumerate(cloned_node.output_tensors):
444 tensor_map[node.output_tensors[j]] = output_tensor
445 else:
446 tensor_map[node.output_tensors] = cloned_node.output_tensors
447 # Ancillary nodes have negative depth.
448 new_nodes = _make_new_nodes(
449 {
450 depth: nodes
451 for depth, nodes in model._nodes_by_depth.items()
452 if depth < 0
453 },
454 layer_fn,
455 layer_map,
456 tensor_map,
457 )
458 _insert_ancillary_layers(
459 cloned_model, ancillary_layers, model.metrics_names, new_nodes
460 )
461 return cloned_model
464@keras_export("keras.models.clone_model")
465def clone_model(model, input_tensors=None, clone_function=None):
466 """Clone a Functional or Sequential `Model` instance.
468 Model cloning is similar to calling a model on new inputs,
469 except that it creates new layers (and thus new weights) instead
470 of sharing the weights of the existing layers.
472 Note that
473 `clone_model` will not preserve the uniqueness of shared objects within the
474 model (e.g. a single variable attached to two distinct layers will be
475 restored as two separate variables).
477 Args:
478 model: Instance of `Model`
479 (could be a Functional model or a Sequential model).
480 input_tensors: optional list of input tensors or InputLayer objects
481 to build the model upon. If not provided,
482 new `Input` objects will be created.
483 clone_function: Callable to be used to clone each layer in the target
484 model (except `InputLayer` instances). It takes as argument the
485 layer instance to be cloned, and returns the corresponding layer
486 instance to be used in the model copy. If unspecified, this callable
487 defaults to the following serialization/deserialization function:
488 `lambda layer: layer.__class__.from_config(layer.get_config())`.
489 By passing a custom callable, you can customize your copy of the
490 model, e.g. by wrapping certain layers of interest (you might want
491 to replace all `LSTM` instances with equivalent
492 `Bidirectional(LSTM(...))` instances, for example).
494 Returns:
495 An instance of `Model` reproducing the behavior
496 of the original model, on top of new inputs tensors,
497 using newly instantiated weights. The cloned model may behave
498 differently from the original model if a custom `clone_function`
499 modifies the layer.
501 Example:
503 ```python
504 # Create a test Sequential model.
505 model = keras.Sequential([
506 keras.Input(shape=(728,)),
507 keras.layers.Dense(32, activation='relu'),
508 keras.layers.Dense(1, activation='sigmoid'),
509 ])
510 # Create a copy of the test model (with freshly initialized weights).
511 new_model = clone_model(model)
512 ```
514 Note that subclassed models cannot be cloned, since their internal
515 layer structure is not known. To achieve equivalent functionality
516 as `clone_model` in the case of a subclassed model, simply make sure
517 that the model class implements `get_config()`
518 (and optionally `from_config()`), and call:
520 ```python
521 new_model = model.__class__.from_config(model.get_config())
522 ```
523 """
524 with serialization.DisableSharedObjectScope():
525 if isinstance(model, Sequential):
526 return _clone_sequential_model(
527 model, input_tensors=input_tensors, layer_fn=clone_function
528 )
529 if isinstance(model, functional.Functional):
530 # If the get_config() method is the same as a regular Functional
531 # model, we're safe to use _clone_functional_model (which relies
532 # on a Functional constructor). In the case where the get_config
533 # is custom, this may not necessarily work, but if clone_function
534 # or input_tensors are passed, we attempt it anyway
535 # in order to preserve backwards compatibility.
536 if generic_utils.is_default(model.get_config) or (
537 clone_function or input_tensors
538 ):
539 return _clone_functional_model(
540 model, input_tensors=input_tensors, layer_fn=clone_function
541 )
543 # Case of a custom model class
544 if clone_function or input_tensors:
545 raise ValueError(
546 "Arguments clone_function and input_tensors "
547 "are only supported for Sequential models "
548 "or Functional models. Received model of "
549 f"type '{model.__class__.__name__}', with "
550 f"clone_function={clone_function} and "
551 f"input_tensors={input_tensors}"
552 )
553 # Note that a custom object scope may be required in this case.
554 return model.__class__.from_config(model.get_config())
557# "Clone" a subclassed model by resetting all of the attributes.
558def _in_place_subclassed_model_reset(model):
559 """Substitute for model cloning that works for subclassed models.
561 Subclassed models cannot be cloned because their topology is not
562 serializable. To "instantiate" an identical model in a new TF graph, we
563 reuse the original model object, but we clear its state.
565 After calling this function on a model instance, you can use the model
566 instance as if it were a model clone (in particular you can use it in a new
567 graph).
569 This method clears the state of the input model. It is thus destructive.
570 However the original state can be restored fully by calling
571 `_in_place_subclassed_model_state_restoration`.
573 Args:
574 model: Instance of a Keras model created via subclassing.
576 Raises:
577 ValueError: In case the model uses a subclassed model as inner layer.
578 """
579 assert (
580 not model._is_graph_network
581 ) # Only makes sense for subclassed networks
582 # Select correct base class for new Model.
583 version_utils.swap_class(
584 model.__class__,
585 training.Model,
586 training_v1.Model,
587 tf.compat.v1.executing_eagerly_outside_functions(),
588 )
589 # Retrieve all layers tracked by the model as well as their attribute names
590 attributes_cache = {}
591 for name in dir(model):
592 # Skip attrs that track other trackables.
593 if name == "submodules" or name == "_self_tracked_trackables":
594 continue
596 try:
597 value = getattr(model, name)
598 except (AttributeError, ValueError, TypeError):
599 continue
600 if isinstance(value, Layer):
601 attributes_cache[name] = value
602 assert value in model.layers
603 if hasattr(value, "layers") and value.layers:
604 raise ValueError(
605 "We do not support the use of nested layers "
606 "in `model_to_estimator` at this time. Found nested "
607 f"layer: {value}"
608 )
609 elif isinstance(value, (list, tuple)) and name not in (
610 "layers",
611 "_layers",
612 "metrics",
613 "_compile_metric_functions",
614 "_output_loss_metrics",
615 ):
616 # Handle case: list/tuple of layers (also tracked by the Network
617 # API).
618 if value and all(isinstance(val, Layer) for val in value):
619 raise ValueError(
620 "We do not support the use of list-of-layers "
621 "attributes in subclassed models used with "
622 "`model_to_estimator` at this time. Found list "
623 f"model: {name}"
624 )
626 # Replace layers on the model with fresh layers
627 layers_to_names = {value: key for key, value in attributes_cache.items()}
628 original_layers = list(
629 model._flatten_layers(include_self=False, recursive=False)
630 )
631 setattr_tracking = model._setattr_tracking
632 model._setattr_tracking = False
633 model._self_tracked_trackables = []
634 for layer in original_layers: # We preserve layer order.
635 config = layer.get_config()
636 # This will not work for nested subclassed models used as layers.
637 # This would be theoretically possible to support, but would add
638 # complexity. Only do it if users complain.
639 if isinstance(layer, training.Model) and not layer._is_graph_network:
640 raise ValueError(
641 "We do not support the use of nested subclassed models "
642 "in `model_to_estimator` at this time. Found nested "
643 f"model: {layer}"
644 )
645 fresh_layer = layer.__class__.from_config(config)
646 name = layers_to_names[layer]
647 setattr(model, name, fresh_layer)
648 model._self_tracked_trackables.append(fresh_layer)
650 # Cache original model build attributes (in addition to layers)
651 if (
652 not hasattr(model, "_original_attributes_cache")
653 or model._original_attributes_cache is None
654 ):
655 if model.built:
656 attributes_to_cache = [
657 "inputs",
658 "outputs",
659 "total_loss",
660 "optimizer",
661 "train_function",
662 "test_function",
663 "predict_function",
664 "_training_endpoints",
665 "_collected_trainable_weights",
666 "_feed_inputs",
667 "_feed_input_names",
668 "_feed_input_shapes",
669 ]
670 for name in attributes_to_cache:
671 attributes_cache[name] = getattr(model, name)
672 model._original_attributes_cache = attributes_cache
673 _reset_build_compile_trackers(model)
674 model._setattr_tracking = setattr_tracking
677def _reset_build_compile_trackers(model):
678 """Reset state trackers for model.
680 Note that we do not actually zero out attributes such as optimizer,
681 but instead rely on the expectation that all of the attrs will be
682 over-written on calling build/compile/etc. This is somewhat fragile,
683 insofar as we check elsewhere for the presence of these attributes as
684 evidence of having been built/compiled/etc. Pending a better way to do this,
685 we reset key attributes here to allow building and compiling.
687 Args:
688 model: the model that is being reset
689 """
690 # Reset build state
691 model.built = False
692 model.inputs = None
693 model.outputs = None
694 # Reset compile state
695 model._is_compiled = False
696 if not tf.compat.v1.executing_eagerly_outside_functions():
697 model._v1_compile_was_called = False
698 model.optimizer = None
701@keras_export(
702 "keras.__internal__.models.in_place_subclassed_model_state_restoration",
703 v1=[],
704)
705def in_place_subclassed_model_state_restoration(model):
706 """Restores the original state of a model after it was "reset".
708 This undoes this action of `_in_place_subclassed_model_reset`, which is
709 called in `clone_and_build_model` if `in_place_reset` is set to True.
711 Args:
712 model: Instance of a Keras model created via subclassing, on which
713 `_in_place_subclassed_model_reset` was previously called.
714 """
715 assert not model._is_graph_network
716 # Restore layers and build attributes
717 if (
718 hasattr(model, "_original_attributes_cache")
719 and model._original_attributes_cache is not None
720 ):
721 # Models have sticky attribute assignment, so we want to be careful to
722 # add back the previous attributes and track Layers by their original
723 # names without adding dependencies on "utility" attributes which Models
724 # exempt when they're constructed.
725 setattr_tracking = model._setattr_tracking
726 model._setattr_tracking = False
727 model._self_tracked_trackables = []
728 for name, value in model._original_attributes_cache.items():
729 setattr(model, name, value)
730 if isinstance(value, Layer):
731 model._self_tracked_trackables.append(value)
732 model._original_attributes_cache = None
733 model._setattr_tracking = setattr_tracking
734 else:
735 # Restore to the state of a never-called model.
736 _reset_build_compile_trackers(model)
739@keras_export("keras.__internal__.models.clone_and_build_model", v1=[])
740def clone_and_build_model(
741 model,
742 input_tensors=None,
743 target_tensors=None,
744 custom_objects=None,
745 compile_clone=True,
746 in_place_reset=False,
747 optimizer_iterations=None,
748 optimizer_config=None,
749):
750 """Clone a `Model` and build/compile it with the same settings used before.
752 This function can be run in the same graph or in a separate graph from the
753 model. When using a separate graph, `in_place_reset` must be `False`.
755 Note that, currently, the clone produced from this function may not work
756 with TPU DistributionStrategy. Try at your own risk.
758 Args:
759 model: `tf.keras.Model` object. Can be Functional, Sequential, or
760 sub-classed.
761 input_tensors: Optional list or dictionary of input tensors to build the
762 model upon. If not provided, placeholders will be created.
763 target_tensors: Optional list of target tensors for compiling the model.
764 If not provided, placeholders will be created.
765 custom_objects: Optional dictionary mapping string names to custom classes
766 or functions.
767 compile_clone: Boolean, whether to compile model clone (default `True`).
768 in_place_reset: Boolean, whether to reset the model in place. Only used if
769 the model is a subclassed model. In the case of a subclassed model,
770 this argument must be set to `True` (default `False`). To restore the
771 original model, use the function
772 `in_place_subclassed_model_state_restoration(model)`.
773 optimizer_iterations: An iterations variable that will be incremented by
774 the optimizer if the clone is compiled. This argument is used when a
775 Keras model is cloned into an Estimator model function, because
776 Estimators create their own global step variable.
777 optimizer_config: Optimizer config dictionary or list of dictionary
778 returned from `get_config()`. This argument should be defined if
779 `clone_and_build_model` is called in a different graph or session from
780 the original model, and the optimizer is an instance of `OptimizerV2`.
782 Returns:
783 Clone of the model.
785 Raises:
786 ValueError: Cloning fails in the following cases
787 - cloning a subclassed model with `in_place_reset` set to False.
788 - compiling the clone when the original model has not been compiled.
789 """
790 # Grab optimizer now, as we reset-in-place for subclassed models, but
791 # want to maintain access to the original optimizer.
792 orig_optimizer = model.optimizer
793 if compile_clone and not orig_optimizer:
794 raise ValueError(
795 "Error when cloning model: `compile_clone` was set to True, but "
796 f"the original model has not been compiled. Received: model={model}"
797 )
799 if compile_clone:
800 compile_args = model._get_compile_args()
801 # Allows this method to be robust to switching graph and eager classes.
802 model._get_compile_args = lambda: compile_args
804 with CustomObjectScope(custom_objects or {}):
805 if model._is_graph_network:
806 clone = clone_model(model, input_tensors=input_tensors)
807 elif isinstance(model, Sequential):
808 clone = clone_model(model, input_tensors=input_tensors)
809 if (
810 not clone._is_graph_network
811 and model._build_input_shape is not None
812 ):
813 if tf.compat.v1.executing_eagerly_outside_functions():
814 clone.build(model._build_input_shape)
815 else:
816 clone._set_inputs(
817 backend.placeholder(
818 model._build_input_shape,
819 dtype=model.inputs[0].dtype,
820 )
821 )
822 else:
823 try:
824 # Prefer cloning the model if serial/deserial logic is
825 # implemented for subclassed model.
826 clone = model.__class__.from_config(model.get_config())
827 except NotImplementedError:
828 logging.warning(
829 "This model is a subclassed model. Please implement "
830 "`get_config` and `from_config` to better support "
831 "cloning the model."
832 )
833 if not in_place_reset:
834 raise ValueError(
835 f"This model ({model}) is a subclassed model. "
836 "Such a model cannot be cloned, but there is a "
837 "workaround where the model is reset in-place. "
838 "To use this, please set the "
839 "argument `in_place_reset` to `True`. This will reset "
840 "the attributes in the original model. "
841 "To restore the attributes, call "
842 "`in_place_subclassed_model_state_restoration(model)`."
843 )
844 clone = model
845 _in_place_subclassed_model_reset(clone)
846 if input_tensors is not None:
847 if (
848 isinstance(input_tensors, (list, tuple))
849 and len(input_tensors) == 1
850 ):
851 input_tensors = input_tensors[0]
852 clone._set_inputs(input_tensors)
854 if compile_clone:
855 if isinstance(orig_optimizer, optimizer_v1.TFOptimizer):
856 optimizer = optimizer_v1.TFOptimizer(
857 orig_optimizer.optimizer, optimizer_iterations
858 )
859 backend.track_tf_optimizer(optimizer)
860 else:
861 if not isinstance(orig_optimizer, (tuple, list)):
862 orig_optimizer = [orig_optimizer]
863 if optimizer_config is None:
864 optimizer = [
865 opt.__class__.from_config(opt.get_config())
866 for opt in orig_optimizer
867 ]
868 elif isinstance(optimizer_config, dict):
869 optimizer = [
870 orig_optimizer[0].__class__.from_config(optimizer_config)
871 ]
872 else:
873 # optimizer config is list of dict, same order as
874 # orig_optimizer.
875 optimizer = [
876 opt.__class__.from_config(opt_config)
877 for (opt, opt_config) in zip(
878 orig_optimizer, optimizer_config
879 )
880 ]
881 if optimizer_iterations is not None:
882 for opt in optimizer:
883 opt.iterations = optimizer_iterations
885 if len(optimizer) == 1:
886 optimizer = optimizer[0]
888 compile_args["optimizer"] = optimizer
889 if target_tensors is not None:
890 compile_args["target_tensors"] = target_tensors
891 # Ensure Metric objects in new model are separate from existing model.
892 compile_args["metrics"] = metrics_module.clone_metrics(
893 compile_args["metrics"]
894 )
895 compile_args["weighted_metrics"] = metrics_module.clone_metrics(
896 compile_args["weighted_metrics"]
897 )
898 clone.compile(**compile_args)
900 return clone