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# ============================================================================== 15# pylint: disable=protected-access 16"""Code for model cloning, plus model-related API entries.""" 17 18from tensorflow.python.framework import ops 19from tensorflow.python.keras import backend 20from tensorflow.python.keras import metrics as metrics_module 21from tensorflow.python.keras import optimizer_v1 22from tensorflow.python.keras.engine import functional 23from tensorflow.python.keras.engine import sequential 24from tensorflow.python.keras.engine import training 25from tensorflow.python.keras.engine import training_v1 26from tensorflow.python.keras.engine.base_layer import AddMetric 27from tensorflow.python.keras.engine.base_layer import Layer 28from tensorflow.python.keras.engine.input_layer import Input 29from tensorflow.python.keras.engine.input_layer import InputLayer 30from tensorflow.python.keras.saving import model_config 31from tensorflow.python.keras.saving import save 32from tensorflow.python.keras.utils import generic_utils 33from tensorflow.python.keras.utils import version_utils 34from tensorflow.python.keras.utils.generic_utils import CustomObjectScope 35from tensorflow.python.platform import tf_logging as logging 36from tensorflow.python.util import nest 37from tensorflow.python.util.tf_export import keras_export 38 39 40# API entries importable from `keras.models`: 41Model = training.Model # pylint: disable=invalid-name 42Sequential = sequential.Sequential # pylint: disable=invalid-name 43Functional = functional.Functional # pylint: disable=invalid-name 44save_model = save.save_model 45load_model = save.load_model 46model_from_config = model_config.model_from_config 47model_from_yaml = model_config.model_from_yaml 48model_from_json = model_config.model_from_json 49 50 51# Callable used to clone a layer with weights preserved. 52def share_weights(layer): 53 return layer 54 55 56def _clone_layer(layer): 57 return layer.__class__.from_config(layer.get_config()) 58 59 60def _insert_ancillary_layers(model, ancillary_layers, metrics_names, new_nodes): 61 """Inserts ancillary layers into the model with the proper order.""" 62 # Sort `AddMetric` layers so they agree with metrics_names. 63 metric_layers = [ 64 layer for layer in ancillary_layers if isinstance(layer, AddMetric) 65 ] 66 metric_layers.sort(key=lambda layer: metrics_names.index(layer.metric_name)) 67 ancillary_layers = [ 68 layer for layer in ancillary_layers if not isinstance(layer, AddMetric) 69 ] + metric_layers 70 model._insert_layers(ancillary_layers, relevant_nodes=list(new_nodes)) 71 72 73def _make_new_nodes(nodes_by_depth, layer_fn, layer_map, tensor_map): 74 """Uses the layers in `layer_map` to make new nodes based on `nodes_by_depth`. 75 76 Args: 77 nodes_by_depth: Provides structure information to create new nodes. 78 layer_fn: Function to clone layers. 79 layer_map: Map from layers in `model` to new layers. 80 tensor_map: Map from tensors in `model` to newly compute tensors. 81 82 Returns: 83 A set of new nodes. `layer_map` and `tensor_map` are updated. 84 """ 85 # Iterated over every node in the reference model, in depth order. 86 new_nodes = set() 87 depth_keys = list(nodes_by_depth.keys()) 88 depth_keys.sort(reverse=True) 89 for depth in depth_keys: 90 nodes = nodes_by_depth[depth] 91 for node in nodes: 92 # Recover the corresponding layer. 93 layer = node.outbound_layer 94 95 # Get or create layer. 96 if layer not in layer_map: 97 new_layer = layer_fn(layer) 98 layer_map[layer] = new_layer 99 layer = new_layer 100 else: 101 # Reuse previously cloned layer. 102 layer = layer_map[layer] 103 # Don't call InputLayer multiple times. 104 if isinstance(layer, InputLayer): 105 continue 106 107 # If all previous input tensors are available in tensor_map, 108 # then call node.inbound_layer on them. 109 if all( 110 tensor in tensor_map for tensor in nest.flatten(node.input_tensors)): 111 # Call layer. 112 args = nest.map_structure(lambda t: tensor_map.get(t, t), 113 node.call_args) 114 kwargs = nest.map_structure(lambda t: tensor_map.get(t, t), 115 node.call_kwargs) 116 output_tensors = layer(*args, **kwargs) 117 118 # Thread-safe way to keep track of what node was created. 119 first_output_tensor = nest.flatten(output_tensors)[0] 120 new_nodes.add( 121 layer._inbound_nodes[first_output_tensor._keras_history.node_index]) 122 123 for x, y in zip( 124 nest.flatten(node.output_tensors), nest.flatten(output_tensors)): 125 tensor_map[x] = y 126 return new_nodes 127 128 129def _clone_functional_model(model, input_tensors=None, layer_fn=_clone_layer): 130 """Clone a functional `Model` instance. 131 132 Model cloning is similar to calling a model on new inputs, 133 except that it creates new layers (and thus new weights) instead 134 of sharing the weights of the existing layers. 135 136 Input layers are always cloned. 137 138 Args: 139 model: Instance of `Model`. 140 input_tensors: optional list of input tensors 141 to build the model upon. If not provided, 142 placeholders will be created. 143 layer_fn: callable to be applied on non-input layers in the model. By 144 default it clones the layer. Another example is to preserve the layer 145 to share the weights. This is required when we create a per-replica 146 copy of the model with distribution strategy; we want the weights to 147 be shared but still feed inputs separately so we create new input 148 layers. 149 150 Returns: 151 An instance of `Model` reproducing the behavior 152 of the original model, on top of new inputs tensors, 153 using newly instantiated weights. 154 155 Raises: 156 ValueError: in case of invalid `model` argument value or `layer_fn` 157 argument value. 158 """ 159 if not isinstance(model, Model): 160 raise ValueError('Expected `model` argument ' 161 'to be a `Model` instance, got ', model) 162 if isinstance(model, Sequential): 163 raise ValueError('Expected `model` argument ' 164 'to be a functional `Model` instance, ' 165 'got a `Sequential` instance instead:', model) 166 if not model._is_graph_network: 167 raise ValueError('Expected `model` argument ' 168 'to be a functional `Model` instance, ' 169 'but got a subclass model instead.') 170 171 new_input_layers = {} # Cache for created layers. 172 if input_tensors is not None: 173 # Make sure that all input tensors come from a Keras layer. 174 input_tensors = nest.flatten(input_tensors) 175 for i, input_tensor in enumerate(input_tensors): 176 original_input_layer = model._input_layers[i] 177 178 # Cache input layer. Create a new layer if the tensor is originally not 179 # from a Keras layer. 180 if not backend.is_keras_tensor(input_tensor): 181 name = original_input_layer.name 182 input_tensor = Input(tensor=input_tensor, 183 name='input_wrapper_for_' + name) 184 newly_created_input_layer = input_tensor._keras_history.layer 185 new_input_layers[original_input_layer] = newly_created_input_layer 186 else: 187 new_input_layers[original_input_layer] = original_input_layer 188 189 if not callable(layer_fn): 190 raise ValueError('Expected `layer_fn` argument to be a callable.') 191 192 model_configs, created_layers = _clone_layers_and_model_config( 193 model, new_input_layers, layer_fn) 194 # Reconstruct model from the config, using the cloned layers. 195 input_tensors, output_tensors, created_layers = ( 196 functional.reconstruct_from_config(model_configs, 197 created_layers=created_layers)) 198 metrics_names = model.metrics_names 199 model = Model(input_tensors, output_tensors, name=model.name) 200 # Layers not directly tied to outputs of the Model, such as loss layers 201 # created in `add_loss` and `add_metric`. 202 ancillary_layers = [ 203 layer for layer in created_layers.values() if layer not in model.layers 204 ] 205 # TODO(b/162887610): This may need to adjust the inbound node index if the 206 # created layers had already been used to define other models. 207 if ancillary_layers: 208 new_nodes = nest.flatten([ 209 layer.inbound_nodes[1:] 210 if functional._should_skip_first_node(layer) 211 else layer.inbound_nodes for layer in created_layers.values() 212 ]) 213 _insert_ancillary_layers(model, ancillary_layers, metrics_names, new_nodes) 214 return model 215 216 217def _clone_layers_and_model_config(model, input_layers, layer_fn): 218 """Clones all layers, and returns the model config without serializing layers. 219 220 This function ensures that only the node graph is retrieved when getting the 221 model config. The `layer_fn` used to clone layers might not rely on 222 `layer.get_config()`, so some custom layers do not define `get_config`. 223 Trying to retrieve the config results in errors. 224 225 Args: 226 model: A Functional model. 227 input_layers: Dictionary mapping input layers in `model` to new input layers 228 layer_fn: Function used to clone all non-input layers. 229 230 Returns: 231 Model config object, and a dictionary of newly created layers. 232 """ 233 created_layers = {} 234 def _copy_layer(layer): 235 # Whenever the network config attempts to get the layer serialization, 236 # return a dummy dictionary. 237 if layer in input_layers: 238 created_layers[layer.name] = input_layers[layer] 239 elif layer in model._input_layers: 240 created_layers[layer.name] = InputLayer(**layer.get_config()) 241 else: 242 created_layers[layer.name] = layer_fn(layer) 243 return {} 244 245 config = functional.get_network_config( 246 model, serialize_layer_fn=_copy_layer) 247 return config, created_layers 248 249 250def _remove_ancillary_layers(model, layer_map, layers): 251 """Removes and returns any ancillary layers from `layers` based on `model`. 252 253 Ancillary layers are part of the model topology but not used to compute the 254 model outputs, e.g., layers from `add_loss` and `add_metric`. 255 256 Args: 257 model: A Keras Model. 258 layer_map: A map to from layers in the `model` to those in `layers`. 259 layers: A list of all layers. 260 261 Returns: 262 Two lists of layers: (1) `layers` with the ancillary layers removed, and (2) 263 the ancillary layers. 264 """ 265 ancillary_layers = [] # Additional layers for computing losses and metrics. 266 if not model._is_graph_network: 267 return layers, ancillary_layers 268 269 # Ancillary layers are those with depth < 0. 270 depths = [depth for depth in model._nodes_by_depth.keys() if depth < 0] 271 depths.sort(reverse=True) # Order topologically from inputs to outputs. 272 for depth in depths: 273 for node in model._nodes_by_depth[depth]: 274 ancillary_layers.append(layer_map[node.outbound_layer]) 275 276 return [l for l in layers if l not in ancillary_layers], ancillary_layers 277 278 279def _clone_sequential_model(model, input_tensors=None, layer_fn=_clone_layer): 280 """Clone a `Sequential` model instance. 281 282 Model cloning is similar to calling a model on new inputs, 283 except that it creates new layers (and thus new weights) instead 284 of sharing the weights of the existing layers. 285 286 Args: 287 model: Instance of `Sequential`. 288 input_tensors: optional list of input tensors 289 to build the model upon. If not provided, 290 placeholders will be created. 291 layer_fn: callable to be applied on non-input layers in the model. By 292 default it clones the layer. Another example is to preserve the layer 293 to share the weights. This is required when we create a per-replica 294 copy of the model with distribution strategy; we want the weights to 295 be shared but still feed inputs separately so we create new input 296 layers. 297 298 Returns: 299 An instance of `Sequential` reproducing the behavior 300 of the original model, on top of new inputs tensors, 301 using newly instantiated weights. 302 303 Raises: 304 ValueError: in case of invalid `model` argument value or `layer_fn` 305 argument value. 306 """ 307 if not isinstance(model, Sequential): 308 raise ValueError('Expected `model` argument ' 309 'to be a `Sequential` model instance, ' 310 'but got:', model) 311 312 if not callable(layer_fn): 313 raise ValueError('Expected `layer_fn` argument to be a callable.') 314 315 layers = [] # Layers needed to compute the model's outputs. 316 layer_map = {} 317 # Ensure that all layers are cloned. The model's layers 318 # property will exclude the initial InputLayer (if it exists) in the model, 319 # resulting in a different Sequential model structure. 320 for layer in model._flatten_layers(include_self=False, recursive=False): 321 if isinstance(layer, InputLayer) and input_tensors is not None: 322 # If input tensors are provided, the original model's InputLayer is 323 # overwritten with a different InputLayer. 324 continue 325 cloned_layer = ( 326 _clone_layer(layer) 327 if isinstance(layer, InputLayer) else layer_fn(layer)) 328 layers.append(cloned_layer) 329 layer_map[layer] = cloned_layer 330 layers, ancillary_layers = _remove_ancillary_layers(model, layer_map, layers) 331 332 if input_tensors is None: 333 cloned_model = Sequential(layers=layers, name=model.name) 334 elif len(generic_utils.to_list(input_tensors)) != 1: 335 raise ValueError('To clone a `Sequential` model, we expect ' 336 ' at most one tensor ' 337 'as part of `input_tensors`.') 338 else: 339 # Overwrite the original model's input layer. 340 if isinstance(input_tensors, tuple): 341 input_tensors = list(input_tensors) 342 x = generic_utils.to_list(input_tensors)[0] 343 if backend.is_keras_tensor(x): 344 origin_layer = x._keras_history.layer 345 if isinstance(origin_layer, InputLayer): 346 cloned_model = Sequential( 347 layers=[origin_layer] + layers, name=model.name) 348 else: 349 raise ValueError('Cannot clone a `Sequential` model on top ' 350 'of a tensor that comes from a Keras layer ' 351 'other than an `InputLayer`. ' 352 'Use the functional API instead.') 353 else: 354 input_tensor = Input(tensor=x, name='input_wrapper_for_' + str(x.name)) 355 input_layer = input_tensor._keras_history.layer 356 cloned_model = Sequential(layers=[input_layer] + layers, name=model.name) 357 358 if not ancillary_layers: 359 return cloned_model 360 361 tensor_map = {} # Maps tensors from `model` to those in `cloned_model`. 362 for depth, cloned_nodes in cloned_model._nodes_by_depth.items(): 363 nodes = model._nodes_by_depth[depth] 364 # This should be safe in a Sequential model. In an arbitrary network, you 365 # need to sort using the outbound layer of the node as a key. 366 for cloned_node, node in zip(cloned_nodes, nodes): 367 if isinstance(cloned_node.output_tensors, list): 368 for j, output_tensor in enumerate(cloned_node.output_tensors): 369 tensor_map[node.output_tensors[j]] = output_tensor 370 else: 371 tensor_map[node.output_tensors] = cloned_node.output_tensors 372 # Ancillary nodes have negative depth. 373 new_nodes = _make_new_nodes( 374 { 375 depth: nodes 376 for depth, nodes in model._nodes_by_depth.items() 377 if depth < 0 378 }, layer_fn, layer_map, tensor_map) 379 _insert_ancillary_layers(cloned_model, ancillary_layers, model.metrics_names, 380 new_nodes) 381 return cloned_model 382 383 384@keras_export('keras.models.clone_model') 385def clone_model(model, input_tensors=None, clone_function=None): 386 """Clone a Functional or Sequential `Model` instance. 387 388 Model cloning is similar to calling a model on new inputs, 389 except that it creates new layers (and thus new weights) instead 390 of sharing the weights of the existing layers. 391 392 Note that 393 `clone_model` will not preserve the uniqueness of shared objects within the 394 model (e.g. a single variable attached to two distinct layers will be 395 restored as two separate variables). 396 397 Args: 398 model: Instance of `Model` 399 (could be a Functional model or a Sequential model). 400 input_tensors: optional list of input tensors or InputLayer objects 401 to build the model upon. If not provided, 402 new `Input` objects will be created. 403 clone_function: Callable to be used to clone each layer in the target 404 model (except `InputLayer` instances). It takes as argument the layer 405 instance to be cloned, and returns the corresponding layer instance to 406 be used in the model copy. If unspecified, this callable defaults to 407 the following serialization/deserialization function: 408 `lambda layer: layer.__class__.from_config(layer.get_config())`. 409 By passing a custom callable, you can customize your copy of the 410 model, e.g. by wrapping certain layers of interest (you might want to 411 replace all `LSTM` instances with equivalent 412 `Bidirectional(LSTM(...))` instances, for example). 413 414 Returns: 415 An instance of `Model` reproducing the behavior 416 of the original model, on top of new inputs tensors, 417 using newly instantiated weights. The cloned model may behave 418 differently from the original model if a custom `clone_function` 419 modifies the layer. 420 421 Example: 422 423 ```python 424 # Create a test Sequential model. 425 model = keras.Sequential([ 426 keras.Input(shape=(728,)), 427 keras.layers.Dense(32, activation='relu'), 428 keras.layers.Dense(1, activation='sigmoid'), 429 ]) 430 # Create a copy of the test model (with freshly initialized weights). 431 new_model = clone_model(model) 432 ``` 433 434 Note that subclassed models cannot be cloned, since their internal 435 layer structure is not known. To achieve equivalent functionality 436 as `clone_model` in the case of a subclassed model, simply make sure 437 that the model class implements `get_config()` 438 (and optionally `from_config()`), and call: 439 440 ```python 441 new_model = model.__class__.from_config(model.get_config()) 442 ``` 443 """ 444 with generic_utils.DisableSharedObjectScope(): 445 if clone_function is None: 446 clone_function = _clone_layer 447 448 if isinstance(model, Sequential): 449 return _clone_sequential_model( 450 model, input_tensors=input_tensors, layer_fn=clone_function) 451 else: 452 return _clone_functional_model( 453 model, input_tensors=input_tensors, layer_fn=clone_function) 454 455 456# "Clone" a subclassed model by reseting all of the attributes. 457def _in_place_subclassed_model_reset(model): 458 """Substitute for model cloning that works for subclassed models. 459 460 Subclassed models cannot be cloned because their topology is not serializable. 461 To "instantiate" an identical model in a new TF graph, we reuse the original 462 model object, but we clear its state. 463 464 After calling this function on a model instance, you can use the model 465 instance as if it were a model clone (in particular you can use it in a new 466 graph). 467 468 This method clears the state of the input model. It is thus destructive. 469 However the original state can be restored fully by calling 470 `_in_place_subclassed_model_state_restoration`. 471 472 Args: 473 model: Instance of a Keras model created via subclassing. 474 475 Raises: 476 ValueError: In case the model uses a subclassed model as inner layer. 477 """ 478 assert not model._is_graph_network # Only makes sense for subclassed networks 479 # Select correct base class for new Model. 480 version_utils.swap_class(model.__class__, training.Model, training_v1.Model, 481 ops.executing_eagerly_outside_functions()) 482 # Retrieve all layers tracked by the model as well as their attribute names 483 attributes_cache = {} 484 for name in dir(model): 485 # Skip attrs that track other trackables. 486 if name == 'submodules' or name == '_self_tracked_trackables': 487 continue 488 489 try: 490 value = getattr(model, name) 491 except (AttributeError, ValueError, TypeError): 492 continue 493 if isinstance(value, Layer): 494 attributes_cache[name] = value 495 assert value in model.layers 496 if hasattr(value, 'layers') and value.layers: 497 raise ValueError('We do not support the use of nested layers ' 498 'in `model_to_estimator` at this time. Found nested ' 499 'layer: %s' % value) 500 elif isinstance( 501 value, (list, tuple)) and name not in ('layers', '_layers', 'metrics', 502 '_compile_metric_functions', 503 '_output_loss_metrics'): 504 # Handle case: list/tuple of layers (also tracked by the Network API). 505 if value and all(isinstance(val, Layer) for val in value): 506 raise ValueError('We do not support the use of list-of-layers ' 507 'attributes in subclassed models used with ' 508 '`model_to_estimator` at this time. Found list ' 509 'model: %s' % name) 510 511 # Replace layers on the model with fresh layers 512 layers_to_names = {value: key for key, value in attributes_cache.items()} 513 original_layers = list( 514 model._flatten_layers(include_self=False, recursive=False)) 515 setattr_tracking = model._setattr_tracking 516 model._setattr_tracking = False 517 model._self_tracked_trackables = [] 518 for layer in original_layers: # We preserve layer order. 519 config = layer.get_config() 520 # This will not work for nested subclassed models used as layers. 521 # This would be theoretically possible to support, but would add complexity. 522 # Only do it if users complain. 523 if isinstance(layer, training.Model) and not layer._is_graph_network: 524 raise ValueError('We do not support the use of nested subclassed models ' 525 'in `model_to_estimator` at this time. Found nested ' 526 'model: %s' % layer) 527 fresh_layer = layer.__class__.from_config(config) 528 name = layers_to_names[layer] 529 setattr(model, name, fresh_layer) 530 model._self_tracked_trackables.append(fresh_layer) 531 532 # Cache original model build attributes (in addition to layers) 533 if (not hasattr(model, '_original_attributes_cache') or 534 model._original_attributes_cache is None): 535 if model.built: 536 attributes_to_cache = [ 537 'inputs', 538 'outputs', 539 'total_loss', 540 'optimizer', 541 'train_function', 542 'test_function', 543 'predict_function', 544 '_training_endpoints', 545 '_collected_trainable_weights', 546 '_feed_inputs', 547 '_feed_input_names', 548 '_feed_input_shapes', 549 ] 550 for name in attributes_to_cache: 551 attributes_cache[name] = getattr(model, name) 552 model._original_attributes_cache = attributes_cache 553 _reset_build_compile_trackers(model) 554 model._setattr_tracking = setattr_tracking 555 556 557def _reset_build_compile_trackers(model): 558 """Reset state trackers for model. 559 560 Note that we do not actually zero out attributes such as optimizer, 561 but instead rely on the expectation that all of the attrs will be 562 over-written on calling build/compile/etc. This is somewhat fragile, 563 insofar as we check elsewhere for the presence of these attributes as 564 evidence of having been built/compiled/etc. Pending a better way to do this, 565 we reset key attributes here to allow building and compiling. 566 567 Args: 568 model: the model that is being reset 569 """ 570 # Reset build state 571 model.built = False 572 model.inputs = None 573 model.outputs = None 574 # Reset compile state 575 model._is_compiled = False # pylint:disable=protected-access 576 if not ops.executing_eagerly_outside_functions(): 577 model._v1_compile_was_called = False 578 model.optimizer = None 579 580 581@keras_export( 582 'keras.__internal__.models.in_place_subclassed_model_state_restoration', 583 v1=[]) 584def in_place_subclassed_model_state_restoration(model): 585 """Restores the original state of a model after it was "reset". 586 587 This undoes this action of `_in_place_subclassed_model_reset`, which is called 588 in `clone_and_build_model` if `in_place_reset` is set to True. 589 590 Args: 591 model: Instance of a Keras model created via subclassing, on which 592 `_in_place_subclassed_model_reset` was previously called. 593 """ 594 assert not model._is_graph_network 595 # Restore layers and build attributes 596 if (hasattr(model, '_original_attributes_cache') and 597 model._original_attributes_cache is not None): 598 # Models have sticky attribute assignment, so we want to be careful to add 599 # back the previous attributes and track Layers by their original names 600 # without adding dependencies on "utility" attributes which Models exempt 601 # when they're constructed. 602 setattr_tracking = model._setattr_tracking 603 model._setattr_tracking = False 604 model._self_tracked_trackables = [] 605 for name, value in model._original_attributes_cache.items(): 606 setattr(model, name, value) 607 if isinstance(value, Layer): 608 model._self_tracked_trackables.append(value) 609 model._original_attributes_cache = None 610 model._setattr_tracking = setattr_tracking 611 else: 612 # Restore to the state of a never-called model. 613 _reset_build_compile_trackers(model) 614 615 616@keras_export('keras.__internal__.models.clone_and_build_model', v1=[]) 617def clone_and_build_model( 618 model, input_tensors=None, target_tensors=None, custom_objects=None, 619 compile_clone=True, in_place_reset=False, optimizer_iterations=None, 620 optimizer_config=None): 621 """Clone a `Model` and build/compile it with the same settings used before. 622 623 This function can be run in the same graph or in a separate graph from the 624 model. When using a separate graph, `in_place_reset` must be `False`. 625 626 Note that, currently, the clone produced from this function may not work with 627 TPU DistributionStrategy. Try at your own risk. 628 629 Args: 630 model: `tf.keras.Model` object. Can be Functional, Sequential, or 631 sub-classed. 632 input_tensors: Optional list or dictionary of input tensors to build the 633 model upon. If not provided, placeholders will be created. 634 target_tensors: Optional list of target tensors for compiling the model. If 635 not provided, placeholders will be created. 636 custom_objects: Optional dictionary mapping string names to custom classes 637 or functions. 638 compile_clone: Boolean, whether to compile model clone (default `True`). 639 in_place_reset: Boolean, whether to reset the model in place. Only used if 640 the model is a subclassed model. In the case of a subclassed model, 641 this argument must be set to `True` (default `False`). To restore the 642 original model, use the function 643 `in_place_subclassed_model_state_restoration(model)`. 644 optimizer_iterations: An iterations variable that will be incremented by the 645 optimizer if the clone is compiled. This argument is used when a Keras 646 model is cloned into an Estimator model function, because Estimators 647 create their own global step variable. 648 optimizer_config: Optimizer config dictionary or list of dictionary 649 returned from `get_config()`. This argument should be defined if 650 `clone_and_build_model` is called in a different graph or session from 651 the original model, and the optimizer is an instance of `OptimizerV2`. 652 653 Returns: 654 Clone of the model. 655 656 Raises: 657 ValueError: Cloning fails in the following cases 658 - cloning a subclassed model with `in_place_reset` set to False. 659 - compiling the clone when the original model has not been compiled. 660 """ 661 # Grab optimizer now, as we reset-in-place for subclassed models, but 662 # want to maintain access to the original optimizer. 663 orig_optimizer = model.optimizer 664 if compile_clone and not orig_optimizer: 665 raise ValueError( 666 'Error when cloning model: compile_clone was set to True, but the ' 667 'original model has not been compiled.') 668 669 if compile_clone: 670 compile_args = model._get_compile_args() # pylint: disable=protected-access 671 # Allows this method to be robust to switching graph and eager classes. 672 model._get_compile_args = lambda: compile_args 673 674 with CustomObjectScope(custom_objects or {}): 675 if model._is_graph_network: 676 clone = clone_model(model, input_tensors=input_tensors) 677 elif isinstance(model, Sequential): 678 clone = clone_model(model, input_tensors=input_tensors) 679 if (not clone._is_graph_network and model._build_input_shape is not None): 680 if ops.executing_eagerly_outside_functions(): 681 clone.build(model._build_input_shape) 682 else: 683 clone._set_inputs( 684 backend.placeholder( 685 model._build_input_shape, dtype=model.inputs[0].dtype)) 686 else: 687 try: 688 # Prefer cloning the model if serial/deserial logic is implemented for 689 # subclassed model. 690 clone = model.__class__.from_config(model.get_config()) 691 except NotImplementedError: 692 logging.warning('This model is a subclassed model. Please implement ' 693 '`get_config` and `from_config` to better support ' 694 'cloning the model.') 695 if not in_place_reset: 696 raise ValueError( 697 'This model is a subclassed model. ' 698 'Such a model cannot be cloned, but there is a workaround where ' 699 'the model is reset in-place. To use this, please set the ' 700 'argument `in_place_reset` to `True`. This will reset the ' 701 'attributes in the original model. To restore the attributes, ' 702 'call `in_place_subclassed_model_state_restoration(model)`.') 703 clone = model 704 _in_place_subclassed_model_reset(clone) 705 if input_tensors is not None: 706 if isinstance(input_tensors, (list, tuple)) and len(input_tensors) == 1: 707 input_tensors = input_tensors[0] 708 clone._set_inputs(input_tensors) 709 710 if compile_clone: 711 if isinstance(orig_optimizer, optimizer_v1.TFOptimizer): 712 optimizer = optimizer_v1.TFOptimizer( 713 orig_optimizer.optimizer, optimizer_iterations) 714 backend.track_tf_optimizer(optimizer) 715 else: 716 if not isinstance(orig_optimizer, (tuple, list)): 717 orig_optimizer = [orig_optimizer] 718 if optimizer_config is None: 719 optimizer = [ 720 opt.__class__.from_config(opt.get_config()) 721 for opt in orig_optimizer 722 ] 723 elif isinstance(optimizer_config, dict): 724 optimizer = [orig_optimizer[0].__class__.from_config(optimizer_config)] 725 else: 726 # optimizer config is list of dict, same order as orig_optimizer. 727 optimizer = [ 728 opt.__class__.from_config(opt_config) 729 for (opt, opt_config) in zip(orig_optimizer, optimizer_config) 730 ] 731 if optimizer_iterations is not None: 732 for opt in optimizer: 733 opt.iterations = optimizer_iterations 734 735 if len(optimizer) == 1: 736 optimizer = optimizer[0] 737 738 compile_args['optimizer'] = optimizer 739 if target_tensors is not None: 740 compile_args['target_tensors'] = target_tensors 741 # Ensure Metric objects in new model are separate from existing model. 742 compile_args['metrics'] = metrics_module.clone_metrics( 743 compile_args['metrics']) 744 compile_args['weighted_metrics'] = metrics_module.clone_metrics( 745 compile_args['weighted_metrics']) 746 clone.compile(**compile_args) 747 748 return clone 749