# Copyright 2018 The TensorFlow Authors. All Rights Reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # ============================================================================== """Keras SavedModel serialization. TODO (kathywu): Move to layer_serialization.py. Some model-specific logic should go to model_serialization.py. """ import functools import threading import weakref from tensorflow.python.eager import def_function from tensorflow.python.framework import tensor_shape from tensorflow.python.framework import tensor_spec from tensorflow.python.keras import backend as K from tensorflow.python.keras.engine import base_layer_utils from tensorflow.python.keras.engine import input_spec from tensorflow.python.keras.mixed_precision import autocast_variable from tensorflow.python.keras.saving import saving_utils from tensorflow.python.keras.saving.saved_model import constants from tensorflow.python.keras.saving.saved_model import load as keras_load from tensorflow.python.keras.saving.saved_model import serialized_attributes from tensorflow.python.keras.saving.saved_model import utils from tensorflow.python.keras.utils import tf_contextlib from tensorflow.python.keras.utils import tf_inspect from tensorflow.python.keras.utils import tf_utils from tensorflow.python.keras.utils import version_utils from tensorflow.python.keras.utils.generic_utils import LazyLoader from tensorflow.python.platform import tf_logging as logging from tensorflow.python.trackable import data_structures from tensorflow.python.util import nest from tensorflow.python.util import tf_decorator # To avoid circular dependencies between keras/engine and keras/saving, # code in keras/saving must delay imports. # TODO(b/134426265): Switch back to single-quotes to match the rest of the file # once the issue with copybara is fixed. # pylint:disable=g-inconsistent-quotes base_layer = LazyLoader( "base_layer", globals(), "tensorflow.python.keras.engine.base_layer") metrics = LazyLoader("metrics", globals(), "tensorflow.python.keras.metrics") input_layer = LazyLoader( "input_layer", globals(), "tensorflow.python.keras.engine.input_layer") training_lib = LazyLoader( "training_lib", globals(), "tensorflow.python.keras.engine.training") sequential_lib = LazyLoader( "sequential_lib", globals(), "tensorflow.python.keras.engine.sequential") # pylint:enable=g-inconsistent-quotes def should_skip_serialization(layer): """Skip serializing extra objects and functions if layer inputs aren't set.""" saved_model_input_spec_set = (isinstance(layer, training_lib.Model) and layer._saved_model_inputs_spec is not None) # pylint: disable=protected-access if not layer.built and not saved_model_input_spec_set: logging.warning('Skipping full serialization of Keras layer {}, because ' 'it is not built.'.format(layer)) return True return False def wrap_layer_objects(layer, serialization_cache): """Returns extra trackable objects to attach to the serialized layer. Args: layer: Keras Layer object. serialization_cache: Dictionary shared between all objects during serialization. Returns: A dictionary containing all checkpointable objects from a SerializedAttributes object. See LayerAttributes and ModelAttributes for entire list of objects """ # Wrap all regularization losses as tf.functions. # First, generate list of all regularization losses in this layer and # sublayers. all_losses = layer._callable_losses[:] # pylint: disable=protected-access for child_layer in utils.list_all_layers(layer): all_losses.extend(child_layer._callable_losses) # pylint: disable=protected-access # Next, wrap all loss functions as tf.functions. Use the serialization cache # to store already-wrapped functions. keras_loss_cache = serialization_cache.setdefault('keras_losses', {}) wrapped_loss_functions = [] for loss_fn in all_losses: if loss_fn in keras_loss_cache: wrapped_loss_functions.append(keras_loss_cache[loss_fn]) else: wrapped_loss = _wrap_unconditional_loss(loss_fn, len(keras_loss_cache)) keras_loss_cache[loss_fn] = wrapped_loss wrapped_loss_functions.append(wrapped_loss) wrapped_layer_losses = [keras_loss_cache[fn] for fn in layer._callable_losses[:]] # pylint: disable=protected-access layer_metrics = data_structures.wrap_or_unwrap( {m.name: m for m in layer._metrics}) # pylint: disable=protected-access return dict( variables=data_structures.wrap_or_unwrap(layer.variables), trainable_variables=data_structures.wrap_or_unwrap( layer.trainable_variables), non_trainable_variables=data_structures.wrap_or_unwrap( layer.non_trainable_variables), layers=data_structures.wrap_or_unwrap(utils.list_all_layers(layer)), metrics=data_structures.wrap_or_unwrap(layer.metrics), regularization_losses=data_structures.wrap_or_unwrap( wrapped_loss_functions), layer_regularization_losses=data_structures.wrap_or_unwrap( wrapped_layer_losses), layer_metrics=layer_metrics) # pylint: disable=protected-access def wrap_layer_functions(layer, serialization_cache): """Returns dict of wrapped layer call function and losses in tf.functions. Args: layer: Keras Layer object. serialization_cache: Dictionary shared between all objects during serialization. Returns: A dictionary containing all keras tf.functions to serialize. See LayerAttributes and ModelAttributes for the list of all attributes. """ # Since Sequential models may be modified in place using model.add() or # model.pop(), don't use saved functions. if (isinstance(layer, keras_load.RevivedLayer) and not isinstance(layer, sequential_lib.Sequential)): return {fn_name: getattr(layer.keras_api, fn_name, None) for fn_name in serialized_attributes.LayerAttributes.all_functions} # Reset the losses of the layer and its children. The call function in each # child layer is replaced with tf.functions. original_fns = _replace_child_layer_functions(layer, serialization_cache) original_losses = _reset_layer_losses(layer) # Wrap all the layer call and activity regularizer functions. # Use LayerCallCollection to ensure that all layer call functions (__call__, # call with losses) are traced with the same inputs. call_collection = LayerCallCollection(layer) call_fn_with_losses = call_collection.add_function( _wrap_call_and_conditional_losses(layer), '{}_layer_call_and_return_conditional_losses'.format(layer.name), # If any of this layer's child layers use the training arg, the traced # call functions of this layer will have a training keyword argument. If # the original layer does not expect the training arg, then it will have # to be removed (by setting `match_layer_training_arg`). match_layer_training_arg=True) call_fn = call_collection.add_function( _extract_outputs_from_fn(layer, call_fn_with_losses), '{}_layer_call_fn'.format(layer.name), # Since `call_fn` wraps call_fn_with_losses and not the original call # function, `match_layer_training_arg` should be set to False. match_layer_training_arg=False) fns = {'call_and_return_conditional_losses': call_fn_with_losses, '__call__': call_fn} if layer._activity_regularizer is not None: # pylint: disable=protected-access fns['activity_regularizer_fn'] = _wrap_activity_regularizer(layer) fns['call_and_return_all_conditional_losses'] = ( call_collection.add_function( _append_activity_regularizer_loss( layer, call_fn_with_losses, fns['activity_regularizer_fn']), '{}_layer_call_and_return_all_conditional_losses'.format( layer.name), match_layer_training_arg=False)) else: fns['activity_regularizer_fn'] = None fns['call_and_return_all_conditional_losses'] = call_fn_with_losses # Manually trigger traces before restoring the overwritten functions. The # functions are traced within the layer call context to ensure that layer # functions (e.g. add_loss) behave as though running in graph mode. with tracing_scope(): call_collection.trace_with_input_signature() with base_layer_utils.call_context().enter( layer, inputs=None, build_graph=True, training=None, saving=True): for fn in fns.values(): if fn is not None and fn.input_signature is not None: if isinstance(fn, LayerCall): fn = fn.wrapped_call fn.get_concrete_function() # Restore overwritten functions and losses _restore_child_layer_functions(original_fns) _restore_layer_losses(original_losses) return fns def default_save_signature(layer): original_losses = _reset_layer_losses(layer) fn = saving_utils.trace_model_call(layer) fn.get_concrete_function() _restore_layer_losses(original_losses) return fn def _replace_child_layer_functions(layer, serialization_cache): """Replaces functions in the children layers with wrapped tf.functions. This step allows functions from parent layers to reference the wrapped functions from their children layers instead of retracing the ops. This function also resets all losses stored in the layer. These are stored in the returned dictionary. Use `_restore_child_layer_functions` to restore the original attributes. Args: layer: Keras Layer object. serialization_cache: Dictionary shared between all objects during serialization. Returns: Dictionary mapping layer objects -> original functions and losses: { Child layer 1: { 'losses': Original losses, 'call': Original call function '_activity_regularizer': Original activity regularizer}, Child layer 2: ... } """ # pylint: disable=protected-access original_fns = {} def replace_layer_functions(child_layer, serialized_fns): """Replaces layer call and activity regularizer with wrapped functions.""" original_fns[child_layer] = { 'call': child_layer.call, '_activity_regularizer': child_layer._activity_regularizer } with utils.no_automatic_dependency_tracking_scope(child_layer): try: child_layer._activity_regularizer = serialized_fns.get( 'activity_regularizer_fn') except AttributeError: # Some layers have an unsettable activity regularizer. pass child_layer.call = utils.use_wrapped_call( child_layer, serialized_fns['call_and_return_conditional_losses'], default_training_value=False) def replace_metric_functions(child_layer, serialized_fns): """Replaces metric functions with wrapped functions.""" original_fns[child_layer] = { '__call__': child_layer.__call__, 'result': child_layer.result, 'update_state': child_layer.update_state } with utils.no_automatic_dependency_tracking_scope(child_layer): child_layer.__call__ = serialized_fns['__call__'] child_layer.result = serialized_fns['result'] child_layer.update_state = serialized_fns['update_state'] for child_layer in utils.list_all_layers(layer): if isinstance(child_layer, input_layer.InputLayer): continue if child_layer not in serialization_cache[constants.KERAS_CACHE_KEY]: serialized_functions = ( child_layer._trackable_saved_model_saver._get_serialized_attributes( serialization_cache).functions) else: serialized_functions = ( serialization_cache[constants.KERAS_CACHE_KEY][child_layer].functions) if not serialized_functions: # This indicates either: # - circular dependency, which means the current layer's functions # should be wrapped first. # - Child layer's inputs are not defined, so its functions have not been # wrapped. In this case, no replacement is necessary so move on to the # next child. continue if isinstance(child_layer, metrics.Metric): replace_metric_functions(child_layer, serialized_functions) else: replace_layer_functions(child_layer, serialized_functions) return original_fns # pylint: enable=protected-access def _restore_child_layer_functions(original_fns): """Restores attributes replaced with `_replace_child_layer_functions`.""" for child_layer, fns in original_fns.items(): with utils.no_automatic_dependency_tracking_scope(child_layer): for fn_name, fn in fns.items(): try: setattr(child_layer, fn_name, fn) # pylint: disable=protected-access except AttributeError: pass # In the case of _activity_regularizer, setting the attribute # may be disallowed. # pylint: disable=protected-access def _reset_layer_losses(parent_layer): """Resets losses of layer and its sublayers, and returns original losses.""" losses_dict = {} for layer in utils.list_all_layers_and_sublayers(parent_layer): losses_dict[layer] = {'losses': layer._losses[:], 'eager_losses': layer._eager_losses[:]} with utils.no_automatic_dependency_tracking_scope(layer): layer._losses = [] layer._eager_losses = [] return losses_dict def _restore_layer_losses(losses_dict): for layer in losses_dict: with utils.no_automatic_dependency_tracking_scope(layer): layer._losses = losses_dict[layer]['losses'] layer._eager_losses = losses_dict[layer]['eager_losses'] # pylint: enable=protected-access class LayerTracingContext(threading.local): def __init__(self): super(LayerTracingContext, self).__init__() self.enable_call_tracing = False self.trace_queue = [] _thread_local_data = LayerTracingContext() @tf_contextlib.contextmanager def tracing_scope(): """Enables tracing scope.""" # This enables the LayerCallCollection's tracing mechanism to trace all call # functions in the collection. previous_value = _thread_local_data.enable_call_tracing previous_queue = _thread_local_data.trace_queue try: _thread_local_data.enable_call_tracing = True _thread_local_data.trace_queue = [] yield finally: # Run traces from the queue. while _thread_local_data.trace_queue: fn, args, kwargs, training = _thread_local_data.trace_queue.pop() if training is not None: with K.deprecated_internal_learning_phase_scope(training): fn.get_concrete_function(*args, **kwargs) else: fn.get_concrete_function(*args, **kwargs) _thread_local_data.trace_queue = previous_queue _thread_local_data.enable_call_tracing = previous_value def add_trace_to_queue(fn, args, kwargs, training=None): if tracing_enabled(): _thread_local_data.trace_queue.append( (fn, args[:], kwargs.copy(), training)) def tracing_enabled(): """Whether to add extra traces to the queue.""" return _thread_local_data.enable_call_tracing class LayerCallCollection(object): """Groups wrapped layer call functions. This is used to ensure that all layer call functions are traced with the same inputs- - call - call_and_return_conditional_losses - call_and_return_all_conditional_losses """ def __init__(self, layer): self.layer = layer self.layer_call_method = _get_layer_call_method(layer) self._expects_training_arg = utils.layer_uses_training_bool(layer) self._training_arg_index = utils.get_training_arg_index( self.layer_call_method) # If the layer call function has kwargs, then the traced function cannot # have an input signature. arg_spec = tf_inspect.getfullargspec(self.layer_call_method) self._has_kwargs = bool(self._expects_training_arg or arg_spec.defaults or arg_spec.kwonlyargs or arg_spec.varkw) self._input_signature = self._generate_input_signature(layer) self._functions = weakref.WeakValueDictionary() # Get the input argument name from the args. args = arg_spec.args if tf_inspect.ismethod(self.layer_call_method): args = args[1:] self._input_arg_name = args[0] if args else 'inputs' def _generate_input_signature(self, layer): """Inspects layer object and returns the inferred input signature. Args: layer: Layer object. Returns: List of possibly nested TensorSpecs of the layer call function inputs. The list does not contain the `training` argument. """ if (isinstance(layer.call, def_function.Function) and layer.call.input_signature is not None): return layer.call.input_signature elif isinstance(layer, training_lib.Model): return saving_utils.model_input_signature(layer) elif (layer.input_spec is not None and layer._use_input_spec_as_call_signature): # pylint: disable=protected-access def to_tensor_spec_or_none(x): spec = input_spec.to_tensor_spec(x, layer._compute_dtype) # pylint: disable=protected-access # If the shape is too general (e.g. multiple dimensions are allowed), # return None so that separate functions can be generated for each # inferred input signature. # TODO(b/134962016): currently partial signatures are not supported. if spec.shape == tensor_shape.TensorShape(None): return None return spec input_signature = [nest.map_structure( to_tensor_spec_or_none, layer.input_spec)] return input_signature else: return None def add_trace(self, *args, **kwargs): """Traces all functions with the same args and kwargs. Args: *args: Positional args passed to the original function. **kwargs: Keyword args passed to the original function. """ args = list(args) kwargs = kwargs.copy() for fn in self._functions.values(): # TODO(kathywu): Replace arguments with broader shapes defined in the # input signature. if self._expects_training_arg: def trace_with_training(value, fn=fn): utils.set_training_arg(value, self._training_arg_index, args, kwargs) add_trace_to_queue(fn, args, kwargs, value) trace_with_training(True) trace_with_training(False) else: add_trace_to_queue(fn, args, kwargs) @property def fn_input_signature(self): """Returns input signature for the wrapped layer call function.""" if self._has_kwargs: # Input signatures may only describe tensor arguments and kwargs are not # supported. return None if None in nest.flatten(self._input_signature): # TODO(b/134962016): If input signature cannot be partially defined. return None return self._input_signature def training_arg_was_passed(self, args, kwargs): if not self.layer._expects_training_arg and self._expects_training_arg: # pylint: disable=protected-access return (utils.get_training_arg(self._training_arg_index, args, kwargs) is not None) else: return self.layer._call_arg_was_passed( # pylint: disable=protected-access 'training', args, kwargs, inputs_in_args=True) def get_training_arg_value(self, args, kwargs): if not self.layer._expects_training_arg and self._expects_training_arg: # pylint: disable=protected-access return utils.get_training_arg(self._training_arg_index, args, kwargs) else: return self.layer._get_call_arg_value( # pylint: disable=protected-access 'training', args, kwargs, inputs_in_args=True) def get_input_arg_value(self, args, kwargs): return self.layer._get_call_arg_value( # pylint: disable=protected-access self._input_arg_name, args, kwargs, inputs_in_args=True) def _maybe_wrap_with_training_arg(self, call_fn, match_layer_training_arg): """Wraps call function with added training argument if necessary.""" if not self.layer._expects_training_arg and self._expects_training_arg: # pylint: disable=protected-access # Add training arg to wrapper function. arg_spec = tf_inspect.getfullargspec(call_fn) args = arg_spec.args + ['training'] defaults = list(arg_spec.defaults or []) defaults.append(False) new_arg_spec = tf_inspect.FullArgSpec( args=args, varargs=arg_spec.varargs, varkw=arg_spec.varkw, defaults=defaults, kwonlyargs=arg_spec.kwonlyargs, kwonlydefaults=arg_spec.kwonlydefaults, annotations=arg_spec.annotations) # Set new training arg index self._training_arg_index = len(args) - 1 if tf_inspect.ismethod(call_fn): self._training_arg_index -= 1 def wrap_with_training_arg(*args, **kwargs): if match_layer_training_arg: # Remove the training value, since the original call_fn does not # expect a training arg. Instead, the training value will be # propagated using the call context created in LayerCall. args = list(args) kwargs = kwargs.copy() utils.remove_training_arg(self._training_arg_index, args, kwargs) return call_fn(*args, **kwargs) return tf_decorator.make_decorator( target=call_fn, decorator_func=wrap_with_training_arg, decorator_argspec=new_arg_spec) return call_fn def add_function(self, call_fn, name, match_layer_training_arg): """Adds a layer call function to the collection. Args: call_fn: a python function name: Name of call function match_layer_training_arg: If True, removes the `training` from the function arguments when calling `call_fn`. Returns: LayerCall (tf.function) """ fn = LayerCall( self, self._maybe_wrap_with_training_arg(call_fn, match_layer_training_arg), name, input_signature=self.fn_input_signature) self._functions[name] = fn.wrapped_call return fn def trace_with_input_signature(self): """Trace with the layer/models inferred input signature if possible.""" if (None not in nest.flatten(self._input_signature) and self._has_kwargs): # Manually add traces for layers that have keyword arguments and have # a fully defined input signature. self.add_trace(*self._input_signature) def _filtered_inputs(inputs): return list(filter(tf_utils.is_tensor_or_variable, nest.flatten(inputs))) def layer_call_wrapper(call_collection, method, name): """Ensures layer losses are kept the same, and runs method in call context.""" # Create wrapper that deals with losses and call context. def wrapper(*args, **kwargs): """Calls method within call context.""" layer = call_collection.layer training = None inputs = _filtered_inputs([args, kwargs]) # pylint: disable=protected-access if (args or kwargs) and call_collection.training_arg_was_passed( args, kwargs): training = call_collection.get_training_arg_value(args, kwargs) # pylint: enable=protected-access original_losses = _reset_layer_losses(layer) with base_layer_utils.call_context().enter( layer, inputs=inputs, build_graph=False, training=training, saving=True): with autocast_variable.enable_auto_cast_variables( layer._compute_dtype_object): # pylint: disable=protected-access ret = method(*args, **kwargs) _restore_layer_losses(original_losses) return ret # Rename to `name`, since tf.function doesn't have a name argument. Without # this, all functions returned by this method will be named "call", which # would be a nightmare to debug. fn = tf_decorator.make_decorator(target=method, decorator_func=wrapper) fn.__name__ = name return fn class LayerCall(object): """Function that triggers traces of other functions in the same collection.""" def __init__(self, call_collection, call_fn, name, input_signature): """Initializes a LayerCall object. Args: call_collection: a LayerCallCollection, which contains the other layer call functions (e.g. call_with_conditional_losses, call). These functions should be traced with the same arguments. call_fn: A call function. name: Name of the call function. input_signature: Input signature of call_fn (can be None). """ self.call_collection = call_collection self.input_signature = input_signature self.wrapped_call = def_function.function( layer_call_wrapper(call_collection, call_fn, name), input_signature=input_signature) self.original_layer_call = call_collection.layer_call_method def _maybe_trace(self, args, kwargs): # Trigger traces of other call functions + extra training-arg traces. if tracing_enabled(): self.call_collection.add_trace(*args, **kwargs) def __call__(self, *args, **kwargs): self._maybe_trace(args, kwargs) return self.wrapped_call(*args, **kwargs) def get_concrete_function(self, *args, **kwargs): self._maybe_trace(args, kwargs) return self.wrapped_call.get_concrete_function(*args, **kwargs) def _wrap_call_and_conditional_losses(layer): """Wraps call function that returns a tuple of (outputs, losses). The losses returned are conditional on the inputs passed to the call function. Unconditional losses (e.g. weight regularizeration) are wrapped separately. Args: layer: a Keras layer object Returns: python call function that returns outputs and conditional losses -- excludes activity regularizer """ # Create function that generates both outputs and losses layer_call = _get_layer_call_method(layer) def call_and_return_conditional_losses(*args, **kwargs): """Returns layer (call_output, conditional losses) tuple.""" call_output = layer_call(*args, **kwargs) if version_utils.is_v1_layer_or_model(layer): conditional_losses = layer.get_losses_for( _filtered_inputs([args, kwargs])) else: conditional_losses = [ l for l in layer.losses if not hasattr(l, '_unconditional_loss') ] return call_output, conditional_losses return _create_call_fn_decorator(layer, call_and_return_conditional_losses) def _extract_outputs_from_fn(layer, call_and_return_conditional_losses): """Returns a function that returns only call function outputs.""" if isinstance(layer, keras_load.RevivedLayer): return layer.keras_api.__call__ # pylint: disable=protected-access def call(inputs, *args, **kwargs): return call_and_return_conditional_losses(inputs, *args, **kwargs)[0] return _create_call_fn_decorator(layer, call) def _append_activity_regularizer_loss( layer, call_fn_with_losses, activity_regularizer_fn): """Appends activity regularizer loss to losses returned by the wrapped fn.""" def fn(inputs, *args, **kwargs): outputs, losses = call_fn_with_losses(inputs, *args, **kwargs) losses.append(activity_regularizer_fn(outputs)) return outputs, losses return _create_call_fn_decorator(layer, fn) def _create_call_fn_decorator(layer, wrapped_call): call_fn = _get_layer_call_method(layer) fn, arg_spec = utils.maybe_add_training_arg( call_fn, wrapped_call, layer._expects_training_arg, # pylint: disable=protected-access default_training_value=False) return tf_decorator.make_decorator( target=call_fn, decorator_func=fn, decorator_argspec=arg_spec) def _wrap_unconditional_loss(loss_fn, index): """Wraps callable/unconditional loss, returning a serializable function.""" # Extract original loss function from partial function fn = loss_fn.args[0] if isinstance(loss_fn, functools.partial) else loss_fn if isinstance(fn, def_function.Function): return fn else: return def_function.Function( fn, 'loss_fn_{}'.format(index), input_signature=[]) def _wrap_activity_regularizer(layer): """Wraps the activity regularizer.""" # pylint: disable=protected-access if isinstance(layer._activity_regularizer, def_function.Function): return layer._activity_regularizer return def_function.Function( layer._activity_regularizer, '{}_activity_regularizer'.format(layer.name), input_signature=[ tensor_spec.TensorSpec(None, layer._compute_dtype or K.floatx()) ]) # pylint: enable=protected-access def _get_layer_call_method(layer): if isinstance(layer.call, (def_function.Function)): return layer.call.python_function return layer.call