tensorflow/python/keras/saving/saved_model/save.py

# 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."""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import functools
import os
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.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.io_utils import ask_to_proceed_with_overwrite
from tensorflow.python.platform import tf_logging as logging
from tensorflow.python.saved_model import save as save_lib
from tensorflow.python.training.tracking import base as trackable
from tensorflow.python.training.tracking import data_structures
from tensorflow.python.training.tracking import layer_utils as trackable_layer_utils
from tensorflow.python.util import nest
from tensorflow.python.util import tf_decorator
from tensorflow.python.util import tf_inspect
from tensorflow.python.util.lazy_loader import LazyLoader

# 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")
training_lib = LazyLoader(
    "training_lib", globals(),
    "tensorflow.python.keras.engine.training")
# pylint:enable=g-inconsistent-quotes


def save(model, filepath, overwrite, include_optimizer, signatures=None):
  """Saves a model as a SavedModel to the filepath.

  Args:
    model: Keras model instance to be saved.
    filepath: String path to save the model.
    overwrite: whether to overwrite the existing filepath.
    include_optimizer: If True, save the model's optimizer state.
    signatures: Signatures to save with the SavedModel. Applicable to the 'tf'
      format only. Please see the `signatures` argument in `tf.saved_model.save`
      for details.

  Raises:
    ValueError: if the model's inputs have not been defined.
  """
  # If file exists and should not be overwritten.
  if not overwrite and os.path.exists(filepath):
    proceed = ask_to_proceed_with_overwrite(filepath)
    if not proceed:
      return

  if _should_skip_serialization(model):
    saving_utils.raise_model_input_error(model)

  if not include_optimizer:
    orig_optimizer = model.optimizer
    model.optimizer = None

  # Trace all functions and signatures with `training=0` instead of using the
  # default learning phase placeholder.
  with K.learning_phase_scope(0):
    save_lib.save(model, filepath, signatures)

  if not include_optimizer:
    model.optimizer = orig_optimizer


def serialize_all_attributes(layer, serialization_cache):
  """Serialize all attributes in the layer."""
  save_model_default_signature = False
  if constants.KERAS_CACHE_KEY not in serialization_cache:
    keras_cache = serialization_cache[constants.KERAS_CACHE_KEY] = {}
    if isinstance(layer, training_lib.Model):
      # Only trace default signature if the root object is a Model. Since the
      # keras cache key is only created in this method, we know that the object
      # is root if the key does not yet exist in the cache.
      save_model_default_signature = True
  else:
    keras_cache = serialization_cache[constants.KERAS_CACHE_KEY]

  if layer in keras_cache:
    return keras_cache[layer]
  serialized_attr = keras_cache[layer] = (
      serialized_attributes.SerializedAttributes.new(layer))

  if _should_skip_serialization(layer):
    return serialized_attr

  function_dict = {}
  if save_model_default_signature:
    # For compatibility with the tf.Lite Converter, the default save signature
    # should be traced without nested calls to other wrapped functions.
    # TODO(kathywu): Investigate why having nested calls results in a stateful
    # function. Perhaps something to do with losses, which are traced in nested
    # calls but not in the flat call.
    function_dict['_default_save_signature'] = _default_save_signature(layer)
  else:
    function_dict['_default_save_signature'] = None

  object_dict = _wrap_layer_objects(layer, serialization_cache)
  try:
    function_dict.update(_wrap_layer_functions(layer, serialization_cache))
  except (ValueError, TypeError) as e:
    logging.warning('Skipping full serialization of object {}, because an '
                    'error occurred while tracing layer functions. Error '
                    'message: {}'.format(layer, e))
  else:
    # Add checkpointable objects and functions to the SerializedAttribute object
    # only if all functions are successfully traced.
    # The `set_and_validate_*` function ensures that all required attributes are
    # exported with the correct type.
    serialized_attr.set_and_validate_objects(object_dict)
    serialized_attr.set_and_validate_functions(function_dict)
  return serialized_attr


def _should_skip_serialization(layer):
  """Skip serializing extra objects and functions if layer inputs aren't set."""
  if isinstance(layer, training_lib.Model):
    try:
      # pylint:disable=pointless-statement
      layer.inputs
      layer.input_names
      # pylint:enable=pointless-statement
    except AttributeError:
      # If the model does not have inputs set, because it was not called or its
      # input shapes were not recorded, we won't have a signature so can't trace
      # a function. But the user may still save an object with this Model
      # attached; we won't fail the whole tf.saved_model.save.
      logging.warning('Skipping full serialization of Keras model {}, because '
                      'its inputs are not defined.'.format(layer))
      return True
    else:
      return False
  else:
    if not layer.built:
      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 _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
  return dict(
      variables=data_structures.ListWrapper(layer.variables),
      trainable_variables=data_structures.ListWrapper(
          layer.trainable_variables),
      non_trainable_variables=data_structures.ListWrapper(
          layer.non_trainable_variables),
      layers=data_structures.ListWrapper(_list_all_layers(layer)),
      metrics=data_structures.ListWrapper(layer.metrics),
      regularization_losses=data_structures.ListWrapper(
          wrapped_loss_functions),
      layer_regularization_losses=data_structures.ListWrapper(
          wrapped_layer_losses))


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, keras_load.RevivedSequential)):
    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))
  call_fn = call_collection.add_function(
      _extract_outputs_from_fn(layer, call_fn_with_losses),
      '{}_layer_call_fn'.format(layer.name))

  fns = {'call_and_return_conditional_losses': call_fn_with_losses,
         '__call__': call_fn}

  if layer.activity_regularizer is not None:
    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)
            ))
  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 base_layer_utils.call_context().enter(layer, None, True, None):
    for fn in fns.values():
      if fn is not None and fn.input_signature is not None:
        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 _list_all_layers(obj):
  if isinstance(obj, training_lib.Model):
    return obj.layers
  else:
    return trackable_layer_utils.filter_empty_layer_containers(obj._layers)  # pylint: disable=protected-access


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 = {}
  for child_layer in _list_all_layers(layer):
    if child_layer not in serialization_cache[constants.KERAS_CACHE_KEY]:
      layer_fns = (serialize_all_attributes(child_layer, serialization_cache)
                   .functions)
    else:
      layer_fns = (
          serialization_cache[constants.KERAS_CACHE_KEY][child_layer].functions)
    if not layer_fns:
      # 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
    original_fns[child_layer] = {
        'call': child_layer.call,
        'activity_regularizer': child_layer.activity_regularizer
    }
    with trackable.no_automatic_dependency_tracking_scope(child_layer):
      try:
        child_layer.activity_regularizer = layer_fns.get(
            'activity_regularizer_fn')
      except AttributeError:
        # Some layers have an unsettable activity regularizer.
        pass
      child_layer.call = utils.use_wrapped_call(
          child_layer, layer_fns['call_and_return_conditional_losses'],
          default_training_value=False)
  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 trackable.no_automatic_dependency_tracking_scope(child_layer):
      child_layer.call = fns['call']
      try:
        child_layer.activity_regularizer = fns['activity_regularizer']
      except AttributeError:
        pass


# 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 _list_all_layers(parent_layer) + [parent_layer]:
    losses_dict[layer] = {'losses': layer._losses[:],
                          'eager_losses': layer._eager_losses[:]}
    with trackable.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 trackable.no_automatic_dependency_tracking_scope(layer):
      layer._losses = losses_dict[layer]['losses']
      layer._eager_losses = losses_dict[layer]['eager_losses']
# pylint: enable=protected-access


def layer_uses_training_bool(layer):
  """Returns whether this layer or any of its children uses the training arg."""
  if layer._expects_training_arg:  # pylint: disable=protected-access
    return True
  visited = {layer}
  to_visit = _list_all_layers(layer)
  while to_visit:
    layer = to_visit.pop()
    if layer in visited:
      continue
    if layer._expects_training_arg:  # pylint: disable=protected-access
      return True
    visited.add(layer)
    to_visit.extend(_list_all_layers(layer))
  return False


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._expects_training_arg = layer_uses_training_bool(layer)
    self._training_arg_index = utils.get_training_arg_index(layer.call)

    self._input_signature = self._generate_input_signature(layer)
    self._functions = weakref.WeakValueDictionary()
    # Bool indicating whether this object is currently tracing the layer call
    # functions.
    self.tracing = False

  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
    else:
      if isinstance(layer, training_lib.Model):
        return saving_utils.model_input_signature(layer)
      elif layer.input_spec is not None:

        def to_tensor_spec_or_none(x):
          spec = input_spec.to_tensor_spec(x, layer.dtype)
          # 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()
    self.tracing = True
    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)
          with K.learning_phase_scope(value):
            fn.get_concrete_function(*args, **kwargs)

        trace_with_training(True)
        trace_with_training(False)
      else:
        fn.get_concrete_function(*args, **kwargs)
    self.tracing = False

  @property
  def fn_input_signature(self):
    """Returns input signature for the wrapped layer call function."""
    if self._expects_training_arg:
      # The training arg is left as a python boolean, so the call functions
      # will not have an input signature (input signatures may only describe
      # tensor arguments).
      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 _maybe_wrap_with_training_arg(self, call_fn):
    """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):
        # 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):
    """Adds a layer call function to the collection."""
    self._functions[name] = fn = LayerCall(
        self, self._maybe_wrap_with_training_arg(call_fn), name,
        input_signature=self.fn_input_signature)

    if (None not in nest.flatten(self._input_signature) and
        self._expects_training_arg):
      # Manually add traces for layers that expect a training argument and have
      # a fully defined input signature.
      self.add_trace(*self._input_signature)
    return fn


def layer_call_wrapper(call_collection, method):
  """Ensures layer losses are kept the same, and runs method in call context."""
  def wrapper(*args, **kwargs):
    """Calls method within call context."""
    layer = call_collection.layer
    training = None
    inputs = None
    # pylint: disable=protected-access
    if (args or kwargs) and call_collection.training_arg_was_passed(
        args, kwargs):
      inputs = args[0]
      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):
      ret = method(*args, **kwargs)
    _restore_layer_losses(original_losses)
    return ret
  return tf_decorator.make_decorator(target=method, decorator_func=wrapper)


class LayerCall(def_function.Function):
  """Function that triggers traces of other functions in the same collection."""

  def __init__(self, call_collection, python_function, *args, **kwargs):
    self.call_collection = call_collection
    self.original_call = call_collection.layer.call
    python_function = layer_call_wrapper(call_collection, python_function)
    super(LayerCall, self).__init__(python_function, *args, **kwargs)

  def __call__(self, *args, **kwargs):
    if not self.call_collection.tracing:
      self.call_collection.add_trace(*args, **kwargs)
    return super(LayerCall, self).__call__(*args, **kwargs)

  def get_concrete_function(self, *args, **kwargs):
    if not self.call_collection.tracing:
      self.call_collection.add_trace(*args, **kwargs)
    return super(LayerCall, self).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 = layer.call
  def call_and_return_conditional_losses(inputs, *args, **kwargs):
    return layer_call(inputs, *args, **kwargs), layer.get_losses_for(inputs)
  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):
  fn, arg_spec = utils.maybe_add_training_arg(
      layer.call, wrapped_call, layer._expects_training_arg,  # pylint: disable=protected-access
      default_training_value=False)
  return tf_decorator.make_decorator(
      target=layer.call,
      decorator_func=fn,
      decorator_argspec=arg_spec)


def _wrap_unconditional_loss(loss_fn, index):
  """Wraps callable/unconditonal 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."""
  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.dtype or K.floatx())])