xref: /external/tensorflow/tensorflow/python/keras/engine/base_layer_v1.py

# Copyright 2015 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.
# ==============================================================================
# pylint: disable=protected-access
"""Contains the base Layer class, from which all layers inherit."""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import collections
import functools
import itertools
import threading
import warnings

import numpy as np
import six
from six.moves import zip  # pylint: disable=redefined-builtin

from tensorflow.python.autograph.core import ag_ctx
from tensorflow.python.autograph.impl import api as autograph
from tensorflow.python.distribute import distribution_strategy_context as ds_context
from tensorflow.python.eager import context
from tensorflow.python.framework import dtypes
from tensorflow.python.framework import errors
from tensorflow.python.framework import func_graph
from tensorflow.python.framework import ops
from tensorflow.python.framework import sparse_tensor
from tensorflow.python.framework import tensor_spec
from tensorflow.python.framework import tensor_util
from tensorflow.python.keras import backend
from tensorflow.python.keras import constraints
from tensorflow.python.keras import initializers
from tensorflow.python.keras import regularizers
from tensorflow.python.keras.engine import base_layer
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.mixed_precision import loss_scale_optimizer
from tensorflow.python.keras.mixed_precision import policy
from tensorflow.python.keras.saving.saved_model import layer_serialization
from tensorflow.python.keras.utils import generic_utils
from tensorflow.python.keras.utils import layer_utils
from tensorflow.python.keras.utils import object_identity
from tensorflow.python.keras.utils import tf_inspect
from tensorflow.python.keras.utils import tf_utils
# A module that only depends on `keras.layers` import these from here.
from tensorflow.python.keras.utils.generic_utils import to_snake_case  # pylint: disable=unused-import
from tensorflow.python.keras.utils.tf_utils import is_tensor_or_tensor_list  # pylint: disable=unused-import
from tensorflow.python.module import module
from tensorflow.python.ops import array_ops
from tensorflow.python.ops import math_ops
from tensorflow.python.ops import variables as tf_variables
from tensorflow.python.ops.ragged import ragged_tensor
from tensorflow.python.platform import tf_logging
from tensorflow.python.training.tracking import base as trackable
from tensorflow.python.training.tracking import data_structures
from tensorflow.python.training.tracking import tracking
from tensorflow.python.util import nest
from tensorflow.tools.docs import doc_controls


# pylint: disable=g-classes-have-attributes
class Layer(base_layer.Layer):
  """Base layer class.

  This is the class from which all layers inherit.

  A layer is a class implementing common neural networks operations, such
  as convolution, batch norm, etc. These operations require managing weights,
  losses, updates, and inter-layer connectivity.

  Users will just instantiate a layer and then treat it as a callable.

  We recommend that descendants of `Layer` implement the following methods:

  * `__init__()`: Save configuration in member variables
  * `build()`: Called once from `__call__`, when we know the shapes of inputs
    and `dtype`. Should have the calls to `add_weight()`, and then
    call the super's `build()` (which sets `self.built = True`, which is
    nice in case the user wants to call `build()` manually before the
    first `__call__`).
  * `call()`: Called in `__call__` after making sure `build()` has been called
    once. Should actually perform the logic of applying the layer to the
    input tensors (which should be passed in as the first argument).

  Args:
    trainable: Boolean, whether the layer's variables should be trainable.
    name: String name of the layer.
    dtype: The dtype of the layer's computations and weights (default of
      `None` means use `tf.keras.backend.floatx` in TensorFlow 2, or the type
      of the first input in TensorFlow 1).
    dynamic: Set this to `True` if your layer should only be run eagerly, and
      should not be used to generate a static computation graph.
      This would be the case for a Tree-RNN or a recursive network,
      for example, or generally for any layer that manipulates tensors
      using Python control flow. If `False`, we assume that the layer can
      safely be used to generate a static computation graph.

  Attributes:
    name: The name of the layer (string).
    dtype: The dtype of the layer's computations and weights. If mixed
      precision is used with a `tf.keras.mixed_precision.Policy`, this is
      instead just the dtype of the layer's weights, as the computations are
      done in a different dtype.
    updates: List of update ops of this layer.
    losses: List of losses added by this layer.
    trainable_weights: List of variables to be included in backprop.
    non_trainable_weights: List of variables that should not be
      included in backprop.
    weights: The concatenation of the lists trainable_weights and
      non_trainable_weights (in this order).
    trainable: Whether the layer should be trained (boolean).
    input_spec: Optional (list of) `InputSpec` object(s) specifying the
      constraints on inputs that can be accepted by the layer.

  Each layer has a dtype, which is typically the dtype of the layer's
  computations and variables. A layer's dtype can be queried via the
  `Layer.dtype` property. The dtype is specified with the `dtype` constructor
  argument. In TensorFlow 2, the dtype defaults to `tf.keras.backend.floatx()`
  if no dtype is passed. `floatx()` itself defaults to "float32". Additionally,
  layers will cast their inputs to the layer's dtype in TensorFlow 2. When mixed
  precision is used, layers may have different computation and variable dtypes.
  See `tf.keras.mixed_precision.Policy` for details on layer dtypes.
  """

  # See tf.Module for the usage of this property.
  # The key for _obj_reference_counts_dict is a Trackable, which could be a
  # variable or layer etc. tf.Module._flatten will fail to flatten the key
  # since it is trying to convert Trackable to a string. This attribute can be
  # ignored even after the fix of nest lib, since the trackable object should
  # already been available as individual attributes. _obj_reference_counts_dict
  # just contains a copy of them.
  _TF_MODULE_IGNORED_PROPERTIES = frozenset(itertools.chain(
      ('_obj_reference_counts_dict',),
      module.Module._TF_MODULE_IGNORED_PROPERTIES
  ))

  @trackable.no_automatic_dependency_tracking
  def __init__(self, trainable=True, name=None, dtype=None, dynamic=False,
               **kwargs):
    self._instrument_layer_creation()

    # These properties should be set by the user via keyword arguments.
    # note that 'dtype', 'input_shape' and 'batch_input_shape'
    # are only applicable to input layers: do not pass these keywords
    # to non-input layers.
    allowed_kwargs = {
        'input_dim', 'input_shape', 'batch_input_shape', 'batch_size',
        'weights', 'activity_regularizer', 'autocast', 'implementation'
    }
    # Validate optional keyword arguments.
    generic_utils.validate_kwargs(kwargs, allowed_kwargs)

    # Mutable properties
    # Indicates whether the layer's weights are updated during training
    # and whether the layer's updates are run during training.
    self._trainable = trainable
    # A stateful layer is a layer whose updates are run during inference too,
    # for instance stateful RNNs.
    self._stateful = False
    # Indicates whether `build` needs to be called upon layer call, to create
    # the layer's weights.
    self.built = False
    self._build_input_shape = None
    # Provides information about which inputs are compatible with the layer.
    self._input_spec = None
    self.supports_masking = False

    self._init_set_name(name)
    self._activity_regularizer = regularizers.get(
        kwargs.pop('activity_regularizer', None))
    self._maybe_create_attribute('_trainable_weights', [])
    self._maybe_create_attribute('_non_trainable_weights', [])
    self._updates = []
    # Object to store all thread local layer properties.
    self._thread_local = threading.local()
    # A list of zero-argument lambdas which return Tensors, used for variable
    # regularizers.
    self._callable_losses = []
    # A list of symbolic Tensors containing activity regularizers and losses
    # manually added through `add_loss` in graph-building mode.
    self._losses = []
    # A list of metric instances corresponding to the symbolic metric tensors
    # added using the `add_metric` API.
    self._metrics = []

    # Both graph and subclassed networks have a dtype policy. For graph
    # networks, the policy's compute and variable dtypes are ignored. Such
    # networks only use the policy if it is a PolicyV1, in which case it uses
    # the PolicyV1's loss_scale (Policy does not have a loss_scale). For
    # subclassed networks, the compute and variable dtypes are used as like any
    # ordinary layer.
    self._set_dtype_policy(dtype)
    # Boolean indicating whether the layer automatically casts its inputs to the
    # layer's compute_dtype.
    self._autocast = kwargs.get('autocast',
                                base_layer_utils.v2_dtype_behavior_enabled())

    # Dependencies tracked via attribute assignment.
    # All layers in order of horizontal graph traversal.
    # Entries are unique. For models includes input and output layers.
    self._maybe_create_attribute('_self_tracked_trackables', [])

    # These lists will be filled via successive calls
    # to self._add_inbound_node().
    # Used in symbolic mode only, only in conjunction with graph-networks
    self._inbound_nodes_value = []
    self._outbound_nodes_value = []

    self._init_call_fn_args()

    # Whether the `call` method can be used to build a TF graph without issues.
    # This attribute has no effect if the model is created using the Functional
    # API. Instead, `model.dynamic` is determined based on the internal layers.
    self._dynamic = dynamic

    # Manage input shape information if passed.
    if 'input_dim' in kwargs and 'input_shape' not in kwargs:
      # Backwards compatibility: alias 'input_dim' to 'input_shape'.
      kwargs['input_shape'] = (kwargs['input_dim'],)
    if 'input_shape' in kwargs or 'batch_input_shape' in kwargs:
      # In this case we will later create an input layer
      # to insert before the current layer
      if 'batch_input_shape' in kwargs:
        batch_input_shape = tuple(kwargs['batch_input_shape'])
      elif 'input_shape' in kwargs:
        if 'batch_size' in kwargs:
          batch_size = kwargs['batch_size']
        else:
          batch_size = None
        batch_input_shape = (batch_size,) + tuple(kwargs['input_shape'])
      self._batch_input_shape = batch_input_shape

    # Manage initial weight values if passed.
    self._initial_weights = kwargs.get('weights', None)

    # Whether the layer will track any layers that is set as attribute on itself
    # as sub-layers, the weights from the sub-layers will be included in the
    # parent layer's variables() as well.
    # Default to True, which means auto tracking is turned on. Certain subclass
    # might want to turn it off, like Sequential model.
    self._auto_track_sub_layers = True

    # Mark this layer as having been originally built as a tf1 layer/model
    self._originally_built_as_v1 = True

    # For backwards compat reasons, most built-in layers do not guarantee
    # That they will 100% preserve the structure of input args when saving
    # / loading configs. E.g. they may un-nest an arg that is
    # a list with one element.
    self._preserve_input_structure_in_config = False

  @trackable.no_automatic_dependency_tracking
  @generic_utils.default
  def build(self, input_shape):
    """Creates the variables of the layer (optional, for subclass implementers).

    This is a method that implementers of subclasses of `Layer` or `Model`
    can override if they need a state-creation step in-between
    layer instantiation and layer call.

    This is typically used to create the weights of `Layer` subclasses.

    Args:
      input_shape: Instance of `TensorShape`, or list of instances of
        `TensorShape` if the layer expects a list of inputs
        (one instance per input).
    """
    if not hasattr(self.build, '_is_default'):
      self._build_input_shape = input_shape
    self.built = True

  @doc_controls.for_subclass_implementers
  def call(self, inputs, **kwargs):  # pylint: disable=unused-argument
    """This is where the layer's logic lives.

    Args:
        inputs: Input tensor, or list/tuple of input tensors.
        **kwargs: Additional keyword arguments.

    Returns:
        A tensor or list/tuple of tensors.
    """
    return inputs

  @doc_controls.for_subclass_implementers
  def _add_trackable(self, trackable_object, trainable):
    """Adds a Trackable object to this layer's state.

    Args:
      trackable_object: The tf.tracking.Trackable object to add.
      trainable: Boolean, whether the variable should be part of the layer's
        "trainable_variables" (e.g. variables, biases) or
        "non_trainable_variables" (e.g. BatchNorm mean and variance).

    Returns:
      The TrackableWeightHandler used to track this object.
    """
    handler = base_layer_utils.TrackableWeightHandler(trackable_object)
    if trainable:
      self._trainable_weights.append(handler)
    else:
      self._non_trainable_weights.append(handler)
    return handler

  @doc_controls.for_subclass_implementers
  def add_weight(self,
                 name=None,
                 shape=None,
                 dtype=None,
                 initializer=None,
                 regularizer=None,
                 trainable=None,
                 constraint=None,
                 partitioner=None,
                 use_resource=None,
                 synchronization=tf_variables.VariableSynchronization.AUTO,
                 aggregation=tf_variables.VariableAggregation.NONE,
                 **kwargs):
    """Adds a new variable to the layer.

    Args:
      name: Variable name.
      shape: Variable shape. Defaults to scalar if unspecified.
      dtype: The type of the variable. Defaults to `self.dtype` or `float32`.
      initializer: Initializer instance (callable).
      regularizer: Regularizer instance (callable).
      trainable: Boolean, whether the variable should be part of the layer's
        "trainable_variables" (e.g. variables, biases)
        or "non_trainable_variables" (e.g. BatchNorm mean and variance).
        Note that `trainable` cannot be `True` if `synchronization`
        is set to `ON_READ`.
      constraint: Constraint instance (callable).
      partitioner: Partitioner to be passed to the `Trackable` API.
      use_resource: Whether to use `ResourceVariable`.
      synchronization: Indicates when a distributed a variable will be
        aggregated. Accepted values are constants defined in the class
        `tf.VariableSynchronization`. By default the synchronization is set to
        `AUTO` and the current `DistributionStrategy` chooses
        when to synchronize. If `synchronization` is set to `ON_READ`,
        `trainable` must not be set to `True`.
      aggregation: Indicates how a distributed variable will be aggregated.
        Accepted values are constants defined in the class
        `tf.VariableAggregation`.
      **kwargs: Additional keyword arguments. Accepted values are `getter`,
        `collections`, `experimental_autocast` and `caching_device`.

    Returns:
      The created variable. Usually either a `Variable` or `ResourceVariable`
      instance. If `partitioner` is not `None`, a `PartitionedVariable`
      instance is returned.

    Raises:
      RuntimeError: If called with partitioned variable regularization and
        eager execution is enabled.
      ValueError: When giving unsupported dtype and no initializer or when
        trainable has been set to True with synchronization set as `ON_READ`.
    """
    if shape is None:
      shape = ()
    # Validate optional keyword arguments.
    for kwarg in kwargs:
      if kwarg not in ['getter', 'collections', 'experimental_autocast',
                       'caching_device']:
        raise TypeError('Unknown keyword argument:', kwarg)
    getter = kwargs.pop('getter', base_layer_utils.make_variable)
    collections_arg = kwargs.pop('collections', None)
    # 'experimental_autocast' can be set to False by the caller to indicate an
    # AutoCastVariable should never be created.
    autocast = kwargs.pop('experimental_autocast', True)
    # See the docstring for tf.Variable about the details for caching_device.
    caching_device = kwargs.pop('caching_device', None)

    if dtype is None:
      dtype = self.dtype or backend.floatx()
    dtype = dtypes.as_dtype(dtype)
    if self._dtype_policy.variable_dtype is None:
      # The policy is "_infer", so we infer the policy from the variable dtype.
      self._set_dtype_policy(policy.Policy(dtype.base_dtype.name))
    initializer = initializers.get(initializer)
    regularizer = regularizers.get(regularizer)
    constraint = constraints.get(constraint)

    if synchronization == tf_variables.VariableSynchronization.ON_READ:
      if trainable:
        raise ValueError(
            'Synchronization value can be set to '
            'VariableSynchronization.ON_READ only for non-trainable variables. '
            'You have specified trainable=True and '
            'synchronization=VariableSynchronization.ON_READ.')
      else:
        # Set trainable to be false when variable is to be synced on read.
        trainable = False
    elif trainable is None:
      trainable = True

    # Initialize variable when no initializer provided
    if initializer is None:
      # If dtype is DT_FLOAT, provide a uniform unit scaling initializer
      if dtype.is_floating:
        initializer = initializers.get('glorot_uniform')
      # If dtype is DT_INT/DT_UINT, provide a default value `zero`
      # If dtype is DT_BOOL, provide a default value `FALSE`
      elif dtype.is_integer or dtype.is_unsigned or dtype.is_bool:
        initializer = initializers.zeros()
      # NOTES:Do we need to support for handling DT_STRING and DT_COMPLEX here?
      else:
        raise ValueError('An initializer for variable %s of type %s is required'
                         ' for layer %s' % (name, dtype.base_dtype, self.name))

    if (autocast and
        self._dtype_policy.compute_dtype != self._dtype_policy.variable_dtype
        and dtype.is_floating):
      # Wrap 'getter' with a version that returns an AutoCastVariable.
      old_getter = getter
      def getter(*args, **kwargs):  # pylint: disable=function-redefined
        variable = old_getter(*args, **kwargs)
        return autocast_variable.create_autocast_variable(variable)
      # Also the caching_device does not work with the mixed precision API,
      # disable it if it is specified.
      # TODO(b/142020079): Reenable it once the bug is fixed.
      if caching_device is not None:
        tf_logging.warn('`caching_device` does not work with mixed precision '
                        'API. Ignoring user specified `caching_device`.')
        caching_device = None

    variable = self._add_variable_with_custom_getter(
        name=name,
        shape=shape,
        # TODO(allenl): a `make_variable` equivalent should be added as a
        # `Trackable` method.
        getter=getter,
        # Manage errors in Layer rather than Trackable.
        overwrite=True,
        initializer=initializer,
        dtype=dtype,
        constraint=constraint,
        trainable=trainable,
        partitioner=partitioner,
        use_resource=use_resource,
        collections=collections_arg,
        synchronization=synchronization,
        aggregation=aggregation,
        caching_device=caching_device)
    if regularizer is not None:
      # TODO(fchollet): in the future, this should be handled at the
      # level of variable creation, and weight regularization losses
      # should be variable attributes.
      name_in_scope = variable.name[:variable.name.find(':')]
      self._handle_weight_regularization(name_in_scope,
                                         variable,
                                         regularizer)
    if base_layer_utils.is_split_variable(variable):
      for v in variable:
        backend.track_variable(v)
        if trainable:
          self._trainable_weights.append(v)
        else:
          self._non_trainable_weights.append(v)
    else:
      backend.track_variable(variable)
      if trainable:
        self._trainable_weights.append(variable)
      else:
        self._non_trainable_weights.append(variable)
    return variable

  @generic_utils.default
  def get_config(self):
    """Returns the config of the layer.

    A layer config is a Python dictionary (serializable)
    containing the configuration of a layer.
    The same layer can be reinstantiated later
    (without its trained weights) from this configuration.

    The config of a layer does not include connectivity
    information, nor the layer class name. These are handled
    by `Network` (one layer of abstraction above).

    Returns:
        Python dictionary.
    """
    all_args = tf_inspect.getfullargspec(self.__init__).args
    config = {'name': self.name, 'trainable': self.trainable}
    if hasattr(self, '_batch_input_shape'):
      config['batch_input_shape'] = self._batch_input_shape
    config['dtype'] = policy.serialize(self._dtype_policy)
    if hasattr(self, 'dynamic'):
      # Only include `dynamic` in the `config` if it is `True`
      if self.dynamic:
        config['dynamic'] = self.dynamic
      elif 'dynamic' in all_args:
        all_args.remove('dynamic')
    expected_args = config.keys()
    # Finds all arguments in the `__init__` that are not in the config:
    extra_args = [arg for arg in all_args if arg not in expected_args]
    # Check that either the only argument in the `__init__` is  `self`,
    # or that `get_config` has been overridden:
    if len(extra_args) > 1 and hasattr(self.get_config, '_is_default'):
      raise NotImplementedError('Layers with arguments in `__init__` must '
                                'override `get_config`.')
    return config

  @classmethod
  def from_config(cls, config):
    """Creates a layer from its config.

    This method is the reverse of `get_config`,
    capable of instantiating the same layer from the config
    dictionary. It does not handle layer connectivity
    (handled by Network), nor weights (handled by `set_weights`).

    Args:
        config: A Python dictionary, typically the
            output of get_config.

    Returns:
        A layer instance.
    """
    return cls(**config)

  def compute_output_shape(self, input_shape):
    """Computes the output shape of the layer.

    If the layer has not been built, this method will call `build` on the
    layer. This assumes that the layer will later be used with inputs that
    match the input shape provided here.

    Args:
        input_shape: Shape tuple (tuple of integers)
            or list of shape tuples (one per output tensor of the layer).
            Shape tuples can include None for free dimensions,
            instead of an integer.

    Returns:
        An input shape tuple.
    """
    if context.executing_eagerly():
      # In this case we build the model first in order to do shape inference.
      # This is acceptable because the framework only calls
      # `compute_output_shape` on shape values that the layer would later be
      # built for. It would however cause issues in case a user attempts to
      # use `compute_output_shape` manually with shapes that are incompatible
      # with the shape the Layer will be called on (these users will have to
      # implement `compute_output_shape` themselves).
      self._maybe_build(input_shape)
      with ops.get_default_graph().as_default():
        graph = func_graph.FuncGraph('graph')
        with graph.as_default():
          input_shape = tf_utils.convert_shapes(input_shape, to_tuples=False)
          inputs = nest.map_structure(
              base_layer_utils.generate_placeholders_from_shape, input_shape)
          try:
            outputs = self(inputs, training=False)
          except TypeError as e:
            six.raise_from(
                NotImplementedError(
                    'We could not automatically infer the static shape of the '
                    'layer\'s output. Please implement the '
                    '`compute_output_shape` method on your layer (%s).' %
                    self.__class__.__name__), e)
      return nest.map_structure(lambda t: t.shape, outputs)
    raise NotImplementedError

  @doc_controls.for_subclass_implementers
  def compute_output_signature(self, input_signature):
    """Compute the output tensor signature of the layer based on the inputs.

    Unlike a TensorShape object, a TensorSpec object contains both shape
    and dtype information for a tensor. This method allows layers to provide
    output dtype information if it is different from the input dtype.
    For any layer that doesn't implement this function,
    the framework will fall back to use `compute_output_shape`, and will
    assume that the output dtype matches the input dtype.

    Args:
      input_signature: Single TensorSpec or nested structure of TensorSpec
        objects, describing a candidate input for the layer.

    Returns:
      Single TensorSpec or nested structure of TensorSpec objects, describing
        how the layer would transform the provided input.

    Raises:
      TypeError: If input_signature contains a non-TensorSpec object.
    """
    def check_type_return_shape(s):
      if not isinstance(s, tensor_spec.TensorSpec):
        raise TypeError('Only TensorSpec signature types are supported, '
                        'but saw signature entry: {}.'.format(s))
      return s.shape
    input_shape = nest.map_structure(check_type_return_shape, input_signature)
    output_shape = self.compute_output_shape(input_shape)
    dtype = self._compute_dtype
    if dtype is None:
      input_dtypes = [s.dtype for s in nest.flatten(input_signature)]
      # Default behavior when self.dtype is None, is to use the first input's
      # dtype.
      dtype = input_dtypes[0]
    return nest.map_structure(
        lambda s: tensor_spec.TensorSpec(dtype=dtype, shape=s),
        output_shape)

  @generic_utils.default
  def compute_mask(self, inputs, mask=None):  # pylint: disable=unused-argument
    """Computes an output mask tensor.

    Args:
        inputs: Tensor or list of tensors.
        mask: Tensor or list of tensors.

    Returns:
        None or a tensor (or list of tensors,
            one per output tensor of the layer).
    """
    if not self.supports_masking:
      if any(m is not None for m in nest.flatten(mask)):
        raise TypeError('Layer ' + self.name + ' does not support masking, '
                        'but was passed an input_mask: ' + str(mask))
      # masking not explicitly supported: return None as mask.
      return None
    # if masking is explicitly supported, by default
    # carry over the input mask
    return mask

  def __call__(self, *args, **kwargs):
    """Wraps `call`, applying pre- and post-processing steps.

    Args:
      *args: Positional arguments to be passed to `self.call`.
      **kwargs: Keyword arguments to be passed to `self.call`.

    Returns:
      Output tensor(s).

    Note:
      - The following optional keyword arguments are reserved for specific uses:
        * `training`: Boolean scalar tensor of Python boolean indicating
          whether the `call` is meant for training or inference.
        * `mask`: Boolean input mask.
      - If the layer's `call` method takes a `mask` argument (as some Keras
        layers do), its default value will be set to the mask generated
        for `inputs` by the previous layer (if `input` did come from
        a layer that generated a corresponding mask, i.e. if it came from
        a Keras layer with masking support.

    Raises:
      ValueError: if the layer's `call` method returns None (an invalid value).
      RuntimeError: if `super().__init__()` was not called in the constructor.
    """
    self._assert_built_as_v1()

    if not hasattr(self, '_thread_local'):
      raise RuntimeError(
          'You must call `super().__init__()` in the layer constructor.')

    # Grab the first positional or keyword argument.
    if args:
      inputs = args[0]
      args = args[1:]
    elif self._call_fn_args[0] in kwargs:
      inputs = kwargs.pop(self._call_fn_args[0])
    else:
      raise ValueError(
          'The first argument to `Layer.call` must always be passed.')

    call_context = base_layer_utils.call_context()
    input_list = nest.flatten(inputs)

    # We will attempt to build a TF graph if & only if all inputs are symbolic.
    # This is always the case in graph mode. It can also be the case in eager
    # mode when all inputs can be traced back to `keras.Input()` (when building
    # models using the functional API).
    build_graph = tf_utils.are_all_symbolic_tensors(input_list)

    # Accept NumPy and scalar inputs by converting to Tensors.
    if any(isinstance(x, (np.ndarray, float, int)) for x in input_list):
      def _convert_non_tensor(x):
        # Don't call `ops.convert_to_tensor` on all `inputs` because
        # `SparseTensors` can't be converted to `Tensor`.
        if isinstance(x, (np.ndarray, float, int)):
          return ops.convert_to_tensor_v2_with_dispatch(x)
        return x
      inputs = nest.map_structure(_convert_non_tensor, inputs)
      input_list = nest.flatten(inputs)

    # Handle `mask` propagation from previous layer to current layer. Masks can
    # be propagated explicitly via the `mask` argument, or implicitly via
    # setting the `_keras_mask` attribute on the inputs to a Layer. Masks passed
    # explicitly take priority.
    mask_arg_passed_by_framework = False
    input_masks = self._collect_input_masks(inputs, args, kwargs)
    if (self._expects_mask_arg and input_masks is not None and
        not self._call_arg_was_passed('mask', args, kwargs)):
      mask_arg_passed_by_framework = True
      kwargs['mask'] = input_masks

    # If `training` argument is None or not explicitly passed,
    # propagate `training` value from this layer's calling layer.
    training_value = None
    training_arg_passed_by_framework = False
    # Priority 1: `training` was explicitly passed.
    if self._call_arg_was_passed('training', args, kwargs):
      training_value = self._get_call_arg_value('training', args, kwargs)
      if not self._expects_training_arg:
        kwargs.pop('training')

    if training_value is None:
      # Priority 2: `training` was passed to a parent layer.
      if call_context.training is not None:
        training_value = call_context.training
      # Priority 3a: `learning_phase()` has been set.
      elif backend.global_learning_phase_is_set():
        training_value = backend.learning_phase()
      # Priority 3b: Pass the `learning_phase()` if in the Keras FuncGraph.
      elif build_graph:
        with backend.get_graph().as_default():
          if base_layer_utils.is_in_keras_graph():
            training_value = backend.learning_phase()

      if self._expects_training_arg and training_value is not None:
        # Force the training_value to be bool type which matches to the contract
        # for layer/model call args.
        if tensor_util.is_tf_type(training_value):
          training_value = math_ops.cast(training_value, dtypes.bool)
        else:
          training_value = bool(training_value)
        args, kwargs = self._set_call_arg_value(
            'training', training_value, args, kwargs)
        training_arg_passed_by_framework = True

    # Only create Keras history if at least one tensor originates from a
    # `keras.Input`. Otherwise this Layer may be being used outside the Keras
    # framework.
    if build_graph and base_layer_utils.needs_keras_history(inputs):
      base_layer_utils.create_keras_history(inputs)

    with call_context.enter(self, inputs, build_graph, training_value):
      # Check input assumptions set after layer building, e.g. input shape.
      if build_graph:
        # Symbolic execution on symbolic tensors. We will attempt to build
        # the corresponding TF subgraph inside `backend.get_graph()`
        # TODO(reedwm): We should assert input compatibility after the inputs
        # are casted, not before.
        input_spec.assert_input_compatibility(self.input_spec, inputs,
                                              self.name)
        graph = backend.get_graph()
        with graph.as_default(), backend.name_scope(self._name_scope()):
          # Build layer if applicable (if the `build` method has been
          # overridden).
          self._maybe_build(inputs)
          cast_inputs = self._maybe_cast_inputs(inputs)

          # Wrapping `call` function in autograph to allow for dynamic control
          # flow and control dependencies in call. We are limiting this to
          # subclassed layers as autograph is strictly needed only for
          # subclassed layers and models.
          # tf_convert will respect the value of autograph setting in the
          # enclosing tf.function, if any.
          if (base_layer_utils.is_subclassed(self) and
              not base_layer_utils.from_saved_model(self)):
            call_fn = autograph.tf_convert(
                self.call, ag_ctx.control_status_ctx())
          else:
            call_fn = self.call

          if not self.dynamic:
            try:
              with autocast_variable.enable_auto_cast_variables(
                  self._compute_dtype_object):
                outputs = call_fn(cast_inputs, *args, **kwargs)

            except errors.OperatorNotAllowedInGraphError as e:
              raise TypeError('You are attempting to use Python control '
                              'flow in a layer that was not declared to be '
                              'dynamic. Pass `dynamic=True` to the class '
                              'constructor.\nEncountered error:\n"""\n' +
                              str(e) + '\n"""')
          else:
            # We will use static shape inference to return symbolic tensors
            # matching the specifications of the layer outputs.
            # Since `self.dynamic` is True, we will never attempt to
            # run the underlying TF graph (which is disconnected).
            # TODO(fchollet): consider py_func as an alternative, which
            # would enable us to run the underlying graph if needed.
            outputs = self._symbolic_call(inputs)

          if outputs is None:
            raise ValueError('A layer\'s `call` method should return a '
                             'Tensor or a list of Tensors, not None '
                             '(layer: ' + self.name + ').')
          if base_layer_utils.have_all_keras_metadata(inputs):
            if training_arg_passed_by_framework:
              args, kwargs = self._set_call_arg_value(
                  'training', None, args, kwargs, pop_kwarg_if_none=True)
            if mask_arg_passed_by_framework:
              kwargs.pop('mask')
            outputs = self._set_connectivity_metadata((inputs,) + args, kwargs,
                                                      outputs)
          self._handle_activity_regularization(inputs, outputs)
          self._set_mask_metadata(inputs, outputs, input_masks)
          if hasattr(self, '_set_inputs') and not self.inputs:
            # Subclassed network: explicitly set metadata normally set by
            # a call to self._set_inputs().
            # TODO(b/120997007): This should be done in Eager as well, but
            # causes garbage collection issues because of the placeholders
            # created on the default Keras graph.
            self._set_inputs(inputs, outputs)
      else:
        # Eager execution on data tensors.
        with backend.name_scope(self._name_scope()):
          self._maybe_build(inputs)
          cast_inputs = self._maybe_cast_inputs(inputs)
          with autocast_variable.enable_auto_cast_variables(
              self._compute_dtype_object):
            outputs = self.call(cast_inputs, *args, **kwargs)
          self._handle_activity_regularization(inputs, outputs)
          self._set_mask_metadata(inputs, outputs, input_masks)

    return outputs

  def _assert_built_as_v1(self):
    if not hasattr(self, '_originally_built_as_v1'):
      raise ValueError(
          'Your Layer or Model is in an invalid state. '
          'This can happen for the following cases:\n '
          '1. You might be interleaving estimator/non-estimator models or '
          'interleaving models/layers made in tf.compat.v1.Graph.as_default() '
          'with models/layers created outside of it. '
          'Converting a model to an estimator (via model_to_estimator) '
          'invalidates all models/layers made before the conversion (even '
          'if they were not the model converted to an estimator). '
          'Similarly, making a layer or a model inside a '
          'a tf.compat.v1.Graph invalidates all layers/models you previously '
          'made outside of the graph.\n'
          '2. You might be using a custom keras layer implementation with '
          ' custom __init__ which didn\'t call super().__init__. '
          ' Please check the implementation of %s and its bases.' %
          (type(self),))

  @property
  def dtype(self):
    return self._dtype_policy.variable_dtype

  @property
  def name(self):
    return self._name

  @property
  def dynamic(self):
    return any(layer._dynamic for layer in self._flatten_layers())

  @property
  @doc_controls.do_not_generate_docs
  def stateful(self):
    return any(layer._stateful for layer in self._flatten_layers())

  @stateful.setter
  def stateful(self, value):
    self._stateful = value

  @property
  def trainable(self):
    return self._trainable

  @trainable.setter
  def trainable(self, value):
    self._trainable = value
    for layer in getattr(self, '_self_tracked_trackables', []):
      layer.trainable = value

  @property
  def activity_regularizer(self):
    """Optional regularizer function for the output of this layer."""
    return self._activity_regularizer

  @activity_regularizer.setter
  def activity_regularizer(self, regularizer):
    """Optional regularizer function for the output of this layer."""
    self._activity_regularizer = regularizer

  @property
  def input_spec(self):
    return self._input_spec

  @input_spec.setter
  # Must be decorated to prevent tracking, since the input_spec can be nested
  # InputSpec objects.
  @trackable.no_automatic_dependency_tracking
  def input_spec(self, value):
    for v in nest.flatten(value):
      if v is not None and not isinstance(v, base_layer.InputSpec):
        raise TypeError('Layer input_spec must be an instance of InputSpec. '
                        'Got: {}'.format(v))
    self._input_spec = value

  @property
  def updates(self):
    collected_updates = []
    all_layers = self._flatten_layers()
    with backend.get_graph().as_default():
      for layer in all_layers:
        if not layer.trainable and not layer.stateful:
          continue
        for u in layer._updates:
          if callable(u):
            try:
              u = u()
            except ValueError as e:
              if 'InaccessibleTensorError' in type(e).__name__:
                # For one specific case of error we try to raise
                # a more meaningful error message about the graph if we can.
                # This error is an internal TF symbol that is not
                # publicly exposed, so we check the name directly rather
                # than using a direct import.
                base_layer_utils.check_graph_consistency(
                    method='add_update', force_raise=True)
              raise  # check_graph_consistency may not always raise.
          base_layer_utils.check_graph_consistency(u, method='add_update')
          collected_updates.append(u)
    return collected_updates

  @property
  def losses(self):
    """Losses which are associated with this `Layer`.

    Variable regularization tensors are created when this property is accessed,
    so it is eager safe: accessing `losses` under a `tf.GradientTape` will
    propagate gradients back to the corresponding variables.

    Returns:
      A list of tensors.
    """
    collected_losses = []
    all_layers = self._flatten_layers()
    for layer in all_layers:
      # If any eager losses are present, we assume the model to be part of an
      # eager training loop (either a custom one or the one used when
      # `run_eagerly=True`) and so we always return just the eager losses.
      collected_losses.extend(layer._losses)
      for regularizer in layer._callable_losses:
        loss_tensor = regularizer()
        if loss_tensor is not None:
          collected_losses.append(loss_tensor)
    return collected_losses

  @doc_controls.for_subclass_implementers
  def add_loss(self, losses, inputs=None):
    """Add loss tensor(s), potentially dependent on layer inputs.

    Some losses (for instance, activity regularization losses) may be dependent
    on the inputs passed when calling a layer. Hence, when reusing the same
    layer on different inputs `a` and `b`, some entries in `layer.losses` may
    be dependent on `a` and some on `b`. This method automatically keeps track
    of dependencies.

    This method can be used inside a subclassed layer or model's `call`
    function, in which case `losses` should be a Tensor or list of Tensors.

    Example:

    ```python
    class MyLayer(tf.keras.layers.Layer):
      def call(inputs, self):
        self.add_loss(tf.abs(tf.reduce_mean(inputs)), inputs=True)
        return inputs
    ```

    This method can also be called directly on a Functional Model during
    construction. In this case, any loss Tensors passed to this Model must
    be symbolic and be able to be traced back to the model's `Input`s. These
    losses become part of the model's topology and are tracked in `get_config`.

    Example:

    ```python
    inputs = tf.keras.Input(shape=(10,))
    x = tf.keras.layers.Dense(10)(inputs)
    outputs = tf.keras.layers.Dense(1)(x)
    model = tf.keras.Model(inputs, outputs)
    # Activity regularization.
    model.add_loss(tf.abs(tf.reduce_mean(x)))
    ```

    If this is not the case for your loss (if, for example, your loss references
    a `Variable` of one of the model's layers), you can wrap your loss in a
    zero-argument lambda. These losses are not tracked as part of the model's
    topology since they can't be serialized.

    Example:

    ```python
    inputs = tf.keras.Input(shape=(10,))
    x = tf.keras.layers.Dense(10)(inputs)
    outputs = tf.keras.layers.Dense(1)(x)
    model = tf.keras.Model(inputs, outputs)
    # Weight regularization.
    model.add_loss(lambda: tf.reduce_mean(x.kernel))
    ```

    The `get_losses_for` method allows to retrieve the losses relevant to a
    specific set of inputs.

    Args:
      losses: Loss tensor, or list/tuple of tensors. Rather than tensors, losses
        may also be zero-argument callables which create a loss tensor.
      inputs: Ignored when executing eagerly. If anything other than None is
        passed, it signals the losses are conditional on some of the layer's
        inputs, and thus they should only be run where these inputs are
        available. This is the case for activity regularization losses, for
        instance. If `None` is passed, the losses are assumed
        to be unconditional, and will apply across all dataflows of the layer
        (e.g. weight regularization losses).
    """
    def _tag_unconditional(loss):
      """Process the loss and tag it by setting loss._unconditional_loss."""
      if callable(loss):
        # We run the loss without autocasting, as regularizers are often
        # numerically unstable in float16.
        with autocast_variable.enable_auto_cast_variables(None):
          loss = loss()
      if loss is None:
        return None  # Will be filtered out when computing the .losses property
      if not tensor_util.is_tf_type(loss):
        loss = ops.convert_to_tensor_v2_with_dispatch(
            loss, dtype=backend.floatx())
      loss._unconditional_loss = (inputs is None)  # pylint: disable=protected-access
      return loss

    losses = nest.flatten(losses)

    callable_losses = []
    symbolic_losses = []
    for loss in losses:
      if callable(loss):
        callable_losses.append(functools.partial(_tag_unconditional, loss))
        continue
      if loss is None:
        continue
      if not tensor_util.is_tf_type(loss):
        loss = ops.convert_to_tensor_v2_with_dispatch(
            loss, dtype=backend.floatx())
      # TF Functions should take the eager path.
      if (tf_utils.is_symbolic_tensor(loss) and
          not base_layer_utils.is_in_tf_function()):
        symbolic_losses.append(_tag_unconditional(loss))
        base_layer_utils.check_graph_consistency(loss, method='add_loss')

    self._callable_losses.extend(callable_losses)

    in_call_context = base_layer_utils.call_context().in_call

    if in_call_context:
      for symbolic_loss in symbolic_losses:
        self._losses.append(symbolic_loss)
    else:
      for symbolic_loss in symbolic_losses:
        if getattr(self, '_is_graph_network', False):
          self._graph_network_add_loss(symbolic_loss)
        else:
          # Possible a loss was added in a Layer's `build`.
          self._losses.append(symbolic_loss)

  @property
  def metrics(self):
    collected_metrics = []
    for layer in self._flatten_layers():
      collected_metrics.extend(layer._metrics)
    return collected_metrics

  @doc_controls.for_subclass_implementers
  def add_metric(self, value, aggregation=None, name=None):
    """Adds metric tensor to the layer.

    Args:
      value: Metric tensor.
      aggregation: Sample-wise metric reduction function. If `aggregation=None`,
        it indicates that the metric tensor provided has been aggregated
        already. eg, `bin_acc = BinaryAccuracy(name='acc')` followed by
        `model.add_metric(bin_acc(y_true, y_pred))`. If aggregation='mean', the
        given metric tensor will be sample-wise reduced using `mean` function.
        eg, `model.add_metric(tf.reduce_sum(outputs), name='output_mean',
        aggregation='mean')`.
      name: String metric name.

    Raises:
      ValueError: If `aggregation` is anything other than None or `mean`.
    """
    if aggregation is not None and aggregation != 'mean':
      raise ValueError(
          'We currently support only `mean` sample-wise metric aggregation. '
          'You provided aggregation=`%s`' % aggregation)

    from_metric_obj = hasattr(value, '_metric_obj')
    is_symbolic = tf_utils.is_symbolic_tensor(value)
    in_call_context = base_layer_utils.call_context().in_call

    if name is None and not from_metric_obj:
      # Eg. `self.add_metric(math_ops.reduce_sum(x), aggregation='mean')`
      # In eager mode, we use metric name to lookup a metric. Without a name,
      # a new Mean metric wrapper will be created on every model/layer call.
      # So, we raise an error when no name is provided.
      # We will do the same for symbolic mode for consistency although a name
      # will be generated if no name is provided.

      # We will not raise this error in the foll use case for the sake of
      # consistency as name in provided in the metric constructor.
      # mean = metrics.Mean(name='my_metric')
      # model.add_metric(mean(outputs))
      raise ValueError('Please provide a name for your metric like '
                       '`self.add_metric(tf.reduce_sum(inputs), '
                       'name=\'mean_activation\', aggregation=\'mean\')`')
    elif from_metric_obj:
      name = value._metric_obj.name

    if in_call_context:
      # TF Function path should take the eager path.
      self._symbolic_add_metric(value, aggregation, name)
    else:
      if not is_symbolic:
        raise ValueError('Expected a symbolic Tensor for the metric value, '
                         'received: ' + str(value))

      # Possible a metric was added in a Layer's `build`.
      if not getattr(self, '_is_graph_network', False):
        with backend.get_graph().as_default():
          self._symbolic_add_metric(value, aggregation, name)
        return

      if from_metric_obj:
        raise ValueError('Using the result of calling a `Metric` object '
                         'when calling `add_metric` on a Functional '
                         'Model is not supported. Please pass the '
                         'Tensor to monitor directly.')

      # Insert layers into the Keras Graph Network.
      self._graph_network_add_metric(value, aggregation, name)

  @doc_controls.for_subclass_implementers
  def add_update(self, updates, inputs=None):
    """Add update op(s), potentially dependent on layer inputs.

    Weight updates (for instance, the updates of the moving mean and variance
    in a BatchNormalization layer) may be dependent on the inputs passed
    when calling a layer. Hence, when reusing the same layer on
    different inputs `a` and `b`, some entries in `layer.updates` may be
    dependent on `a` and some on `b`. This method automatically keeps track
    of dependencies.

    The `get_updates_for` method allows to retrieve the updates relevant to a
    specific set of inputs.

    This call is ignored when eager execution is enabled (in that case, variable
    updates are run on the fly and thus do not need to be tracked for later
    execution).

    Args:
      updates: Update op, or list/tuple of update ops, or zero-arg callable
        that returns an update op. A zero-arg callable should be passed in
        order to disable running the updates by setting `trainable=False`
        on this Layer, when executing in Eager mode.
      inputs: Deprecated, will be automatically inferred.
    """
    if inputs is not None:
      tf_logging.warning(
          '`add_update` `inputs` kwarg has been deprecated. You no longer need '
          'to pass a value to `inputs` as it is being automatically inferred.')
    call_context = base_layer_utils.call_context()

    if (ds_context.has_strategy() and
        ds_context.in_cross_replica_context() and
        # When saving the model, the distribution strategy context should be
        # ignored, following the default path for adding updates.
        not call_context.saving):
      # Updates don't need to be run in a cross-replica context.
      return

    updates = generic_utils.to_list(updates)

    if call_context.in_call:
      relevant_inputs = call_context.inputs
    else:
      inbound_nodes = getattr(self, '_inbound_nodes', [])
      relevant_inputs = [node.input_tensors for node in inbound_nodes]

    def process_update(x):
      """Standardize update ops.

      Args:
        x: Tensor, op, or callable.

      Returns:
        An update op.
      """
      if callable(x):
        update = lambda: process_update(x())
        return update()
      elif isinstance(x, ops.Operation):
        update = x
      elif hasattr(x, 'op'):
        update = x.op
      else:
        update = ops.convert_to_tensor_v2_with_dispatch(x)

      reachable = tf_utils.get_reachable_from_inputs(relevant_inputs, [update])
      update._unconditional_update = update not in reachable
      return update

    updates = [process_update(x) for x in updates]
    self._updates.extend(updates)

  def set_weights(self, weights):
    """Sets the weights of the layer, from Numpy arrays.

    The weights of a layer represent the state of the layer. This function
    sets the weight values from numpy arrays. The weight values should be
    passed in the order they are created by the layer. Note that the layer's
    weights must be instantiated before calling this function by calling
    the layer.

    For example, a Dense layer returns a list of two values-- per-output
    weights and the bias value. These can be used to set the weights of another
    Dense layer:

    >>> a = tf.keras.layers.Dense(1,
    ...   kernel_initializer=tf.constant_initializer(1.))
    >>> a_out = a(tf.convert_to_tensor([[1., 2., 3.]]))
    >>> a.get_weights()
    [array([[1.],
           [1.],
           [1.]], dtype=float32), array([0.], dtype=float32)]
    >>> b = tf.keras.layers.Dense(1,
    ...   kernel_initializer=tf.constant_initializer(2.))
    >>> b_out = b(tf.convert_to_tensor([[10., 20., 30.]]))
    >>> b.get_weights()
    [array([[2.],
           [2.],
           [2.]], dtype=float32), array([0.], dtype=float32)]
    >>> b.set_weights(a.get_weights())
    >>> b.get_weights()
    [array([[1.],
           [1.],
           [1.]], dtype=float32), array([0.], dtype=float32)]

    Args:
        weights: a list of Numpy arrays. The number
            of arrays and their shape must match
            number of the dimensions of the weights
            of the layer (i.e. it should match the
            output of `get_weights`).

    Raises:
        ValueError: If the provided weights list does not match the
            layer's specifications.
    """
    params = self.weights

    expected_num_weights = 0
    for param in params:
      if isinstance(param, base_layer_utils.TrackableWeightHandler):
        expected_num_weights += param.num_tensors
      else:
        expected_num_weights += 1

    if expected_num_weights != len(weights):
      raise ValueError(
          'You called `set_weights(weights)` on layer "%s" '
          'with a weight list of length %s, but the layer was '
          'expecting %s weights. Provided weights: %s...' %
          (self.name, len(weights), expected_num_weights, str(weights)[:50]))

    weight_index = 0
    weight_value_tuples = []
    for param in params:
      if isinstance(param, base_layer_utils.TrackableWeightHandler):
        num_tensors = param.num_tensors
        tensors = weights[weight_index:weight_index + num_tensors]
        param.set_weights(tensors)
        weight_index += num_tensors
      else:
        weight = weights[weight_index]
        ref_shape = param.shape
        if not ref_shape.is_compatible_with(weight.shape):
          raise ValueError(
              'Layer weight shape %s not compatible with provided weight '
              'shape %s' % (ref_shape, weight.shape))
        weight_value_tuples.append((param, weight))
        weight_index += 1

    backend.batch_set_value(weight_value_tuples)

  def get_weights(self):
    """Returns the current weights of the layer.

    The weights of a layer represent the state of the layer. This function
    returns both trainable and non-trainable weight values associated with this
    layer as a list of Numpy arrays, which can in turn be used to load state
    into similarly parameterized layers.

    For example, a Dense layer returns a list of two values-- per-output
    weights and the bias value. These can be used to set the weights of another
    Dense layer:

    >>> a = tf.keras.layers.Dense(1,
    ...   kernel_initializer=tf.constant_initializer(1.))
    >>> a_out = a(tf.convert_to_tensor([[1., 2., 3.]]))
    >>> a.get_weights()
    [array([[1.],
           [1.],
           [1.]], dtype=float32), array([0.], dtype=float32)]
    >>> b = tf.keras.layers.Dense(1,
    ...   kernel_initializer=tf.constant_initializer(2.))
    >>> b_out = b(tf.convert_to_tensor([[10., 20., 30.]]))
    >>> b.get_weights()
    [array([[2.],
           [2.],
           [2.]], dtype=float32), array([0.], dtype=float32)]
    >>> b.set_weights(a.get_weights())
    >>> b.get_weights()
    [array([[1.],
           [1.],
           [1.]], dtype=float32), array([0.], dtype=float32)]

    Returns:
        Weights values as a list of numpy arrays.
    """
    weights = self.weights
    output_weights = []
    for weight in weights:
      if isinstance(weight, base_layer_utils.TrackableWeightHandler):
        output_weights.extend(weight.get_tensors())
      else:
        output_weights.append(weight)
    return backend.batch_get_value(output_weights)

  def get_updates_for(self, inputs):
    """Retrieves updates relevant to a specific set of inputs.

    Args:
      inputs: Input tensor or list/tuple of input tensors.

    Returns:
      List of update ops of the layer that depend on `inputs`.
    """
    if inputs is None:
      # Requesting unconditional updates.
      return [u for u in self.updates if u._unconditional_update]

    # Requesting input-conditional updates.
    updates = [u for u in self.updates if not u._unconditional_update]
    inputs = nest.flatten(inputs)
    reachable = tf_utils.get_reachable_from_inputs(inputs, updates)
    return [u for u in updates if u in reachable]

  def get_losses_for(self, inputs):
    """Retrieves losses relevant to a specific set of inputs.

    Args:
      inputs: Input tensor or list/tuple of input tensors.

    Returns:
      List of loss tensors of the layer that depend on `inputs`.
    """
    if inputs is None:
      # Requesting unconditional losses.
      return [l for l in self.losses if l._unconditional_loss]

    # Requesting input-conditional losses.
    losses = [l for l in self.losses if not l._unconditional_loss]
    inputs = nest.flatten(inputs)
    reachable = tf_utils.get_reachable_from_inputs(inputs, losses)
    return [l for l in losses if l in reachable]

  def get_input_mask_at(self, node_index):
    """Retrieves the input mask tensor(s) of a layer at a given node.

    Args:
        node_index: Integer, index of the node
            from which to retrieve the attribute.
            E.g. `node_index=0` will correspond to the
            first time the layer was called.

    Returns:
        A mask tensor
        (or list of tensors if the layer has multiple inputs).
    """
    inputs = self.get_input_at(node_index)
    if isinstance(inputs, list):
      return [getattr(x, '_keras_mask', None) for x in inputs]
    else:
      return getattr(inputs, '_keras_mask', None)

  def get_output_mask_at(self, node_index):
    """Retrieves the output mask tensor(s) of a layer at a given node.

    Args:
        node_index: Integer, index of the node
            from which to retrieve the attribute.
            E.g. `node_index=0` will correspond to the
            first time the layer was called.

    Returns:
        A mask tensor
        (or list of tensors if the layer has multiple outputs).
    """
    output = self.get_output_at(node_index)
    if isinstance(output, list):
      return [getattr(x, '_keras_mask', None) for x in output]
    else:
      return getattr(output, '_keras_mask', None)

  @property
  def input_mask(self):
    """Retrieves the input mask tensor(s) of a layer.

    Only applicable if the layer has exactly one inbound node,
    i.e. if it is connected to one incoming layer.

    Returns:
        Input mask tensor (potentially None) or list of input
        mask tensors.

    Raises:
        AttributeError: if the layer is connected to
        more than one incoming layers.
    """
    inputs = self.input
    if isinstance(inputs, list):
      return [getattr(x, '_keras_mask', None) for x in inputs]
    else:
      return getattr(inputs, '_keras_mask', None)

  @property
  def output_mask(self):
    """Retrieves the output mask tensor(s) of a layer.

    Only applicable if the layer has exactly one inbound node,
    i.e. if it is connected to one incoming layer.

    Returns:
        Output mask tensor (potentially None) or list of output
        mask tensors.

    Raises:
        AttributeError: if the layer is connected to
        more than one incoming layers.
    """
    output = self.output
    if isinstance(output, list):
      return [getattr(x, '_keras_mask', None) for x in output]
    else:
      return getattr(output, '_keras_mask', None)

  def get_input_shape_at(self, node_index):
    """Retrieves the input shape(s) of a layer at a given node.

    Args:
        node_index: Integer, index of the node
            from which to retrieve the attribute.
            E.g. `node_index=0` will correspond to the
            first time the layer was called.

    Returns:
        A shape tuple
        (or list of shape tuples if the layer has multiple inputs).

    Raises:
      RuntimeError: If called in Eager mode.
    """
    return self._get_node_attribute_at_index(node_index, 'input_shapes',
                                             'input shape')

  def get_output_shape_at(self, node_index):
    """Retrieves the output shape(s) of a layer at a given node.

    Args:
        node_index: Integer, index of the node
            from which to retrieve the attribute.
            E.g. `node_index=0` will correspond to the
            first time the layer was called.

    Returns:
        A shape tuple
        (or list of shape tuples if the layer has multiple outputs).

    Raises:
      RuntimeError: If called in Eager mode.
    """
    return self._get_node_attribute_at_index(node_index, 'output_shapes',
                                             'output shape')

  def get_input_at(self, node_index):
    """Retrieves the input tensor(s) of a layer at a given node.

    Args:
        node_index: Integer, index of the node
            from which to retrieve the attribute.
            E.g. `node_index=0` will correspond to the
            first input node of the layer.

    Returns:
        A tensor (or list of tensors if the layer has multiple inputs).

    Raises:
      RuntimeError: If called in Eager mode.
    """
    return self._get_node_attribute_at_index(node_index, 'input_tensors',
                                             'input')

  def get_output_at(self, node_index):
    """Retrieves the output tensor(s) of a layer at a given node.

    Args:
        node_index: Integer, index of the node
            from which to retrieve the attribute.
            E.g. `node_index=0` will correspond to the
            first output node of the layer.

    Returns:
        A tensor (or list of tensors if the layer has multiple outputs).

    Raises:
      RuntimeError: If called in Eager mode.
    """
    return self._get_node_attribute_at_index(node_index, 'output_tensors',
                                             'output')

  @property
  def input(self):
    """Retrieves the input tensor(s) of a layer.

    Only applicable if the layer has exactly one input,
    i.e. if it is connected to one incoming layer.

    Returns:
        Input tensor or list of input tensors.

    Raises:
      RuntimeError: If called in Eager mode.
      AttributeError: If no inbound nodes are found.
    """
    if not self._inbound_nodes:
      raise AttributeError('Layer ' + self.name +
                           ' is not connected, no input to return.')
    return self._get_node_attribute_at_index(0, 'input_tensors', 'input')

  @property
  def output(self):
    """Retrieves the output tensor(s) of a layer.

    Only applicable if the layer has exactly one output,
    i.e. if it is connected to one incoming layer.

    Returns:
      Output tensor or list of output tensors.

    Raises:
      AttributeError: if the layer is connected to more than one incoming
        layers.
      RuntimeError: if called in Eager mode.
    """
    if not self._inbound_nodes:
      raise AttributeError('Layer ' + self.name + ' has no inbound nodes.')
    return self._get_node_attribute_at_index(0, 'output_tensors', 'output')

  @property
  def input_shape(self):
    """Retrieves the input shape(s) of a layer.

    Only applicable if the layer has exactly one input,
    i.e. if it is connected to one incoming layer, or if all inputs
    have the same shape.

    Returns:
        Input shape, as an integer shape tuple
        (or list of shape tuples, one tuple per input tensor).

    Raises:
        AttributeError: if the layer has no defined input_shape.
        RuntimeError: if called in Eager mode.
    """
    if not self._inbound_nodes:
      raise AttributeError('The layer has never been called '
                           'and thus has no defined input shape.')
    all_input_shapes = set(
        [str(node.input_shapes) for node in self._inbound_nodes])
    if len(all_input_shapes) == 1:
      return self._inbound_nodes[0].input_shapes
    else:
      raise AttributeError('The layer "' + str(self.name) +
                           ' has multiple inbound nodes, '
                           'with different input shapes. Hence '
                           'the notion of "input shape" is '
                           'ill-defined for the layer. '
                           'Use `get_input_shape_at(node_index)` '
                           'instead.')

  def count_params(self):
    """Count the total number of scalars composing the weights.

    Returns:
        An integer count.

    Raises:
        ValueError: if the layer isn't yet built
          (in which case its weights aren't yet defined).
    """
    if not self.built:
      if getattr(self, '_is_graph_network', False):
        with tf_utils.maybe_init_scope(self):
          self._maybe_build(self.inputs)
      else:
        raise ValueError('You tried to call `count_params` on ' + self.name +
                         ', but the layer isn\'t built. '
                         'You can build it manually via: `' + self.name +
                         '.build(batch_input_shape)`.')
    return layer_utils.count_params(self.weights)

  @property
  def output_shape(self):
    """Retrieves the output shape(s) of a layer.

    Only applicable if the layer has one output,
    or if all outputs have the same shape.

    Returns:
        Output shape, as an integer shape tuple
        (or list of shape tuples, one tuple per output tensor).

    Raises:
        AttributeError: if the layer has no defined output shape.
        RuntimeError: if called in Eager mode.
    """
    if not self._inbound_nodes:
      raise AttributeError('The layer has never been called '
                           'and thus has no defined output shape.')
    all_output_shapes = set(
        [str(node.output_shapes) for node in self._inbound_nodes])
    if len(all_output_shapes) == 1:
      return self._inbound_nodes[0].output_shapes
    else:
      raise AttributeError('The layer "%s"'
                           ' has multiple inbound nodes, '
                           'with different output shapes. Hence '
                           'the notion of "output shape" is '
                           'ill-defined for the layer. '
                           'Use `get_output_shape_at(node_index)` '
                           'instead.' % self.name)

  @property
  @doc_controls.do_not_doc_inheritable
  def inbound_nodes(self):
    """Deprecated, do NOT use! Only for compatibility with external Keras."""
    return self._inbound_nodes

  @property
  @doc_controls.do_not_doc_inheritable
  def outbound_nodes(self):
    """Deprecated, do NOT use! Only for compatibility with external Keras."""
    return self._outbound_nodes

  ##############################################################################
  # Methods & attributes below are public aliases of other methods.            #
  ##############################################################################

  @doc_controls.do_not_doc_inheritable
  def apply(self, inputs, *args, **kwargs):
    """Deprecated, do NOT use!

    This is an alias of `self.__call__`.

    Args:
      inputs: Input tensor(s).
      *args: additional positional arguments to be passed to `self.call`.
      **kwargs: additional keyword arguments to be passed to `self.call`.

    Returns:
      Output tensor(s).
    """
    warnings.warn('`layer.apply` is deprecated and '
                  'will be removed in a future version. '
                  'Please use `layer.__call__` method instead.')
    return self.__call__(inputs, *args, **kwargs)

  @doc_controls.do_not_doc_inheritable
  def add_variable(self, *args, **kwargs):
    """Deprecated, do NOT use! Alias for `add_weight`."""
    warnings.warn('`layer.add_variable` is deprecated and '
                  'will be removed in a future version. '
                  'Please use `layer.add_weight` method instead.')
    return self.add_weight(*args, **kwargs)

  @property
  def variables(self):
    """Returns the list of all layer variables/weights.

    Alias of `self.weights`.

    Returns:
      A list of variables.
    """
    return self.weights

  @property
  def trainable_variables(self):
    return self.trainable_weights

  @property
  def non_trainable_variables(self):
    return self.non_trainable_weights

  ##############################################################################
  # Methods & attributes below are all private and only used by the framework. #
  ##############################################################################

  @property
  def _inbound_nodes(self):
    return self._inbound_nodes_value

  @_inbound_nodes.setter
  @trackable.no_automatic_dependency_tracking
  def _inbound_nodes(self, value):
    self._inbound_nodes_value = value

  @property
  def _outbound_nodes(self):
    return self._outbound_nodes_value

  @_outbound_nodes.setter
  @trackable.no_automatic_dependency_tracking
  def _outbound_nodes(self, value):
    self._outbound_nodes_value = value

  def _set_dtype_policy(self, dtype):
    """Sets self._dtype_policy."""
    if isinstance(dtype, policy.Policy):
      self._dtype_policy = dtype
    elif isinstance(dtype, dict):
      self._dtype_policy = policy.deserialize(dtype)
    elif dtype:
      self._dtype_policy = policy.Policy(dtypes.as_dtype(dtype).name)
    else:
      self._dtype_policy = policy.global_policy()
    if (self._dtype_policy.name == 'mixed_float16' and
        not loss_scale_optimizer.strategy_supports_loss_scaling()):
      # Although only loss scaling doesn't support certain strategies, to avoid
      # confusion, we disallow the 'mixed_float16' policy with unsupported
      # strategies. This is because 'mixed_float16' requires loss scaling for
      # numeric stability.
      strategy = ds_context.get_strategy()
      raise ValueError('Mixed precision is not supported with the '
                       'tf.distribute.Strategy: %s. Either stop using mixed '
                       'precision by removing the use of the "%s" policy or '
                       'use a different Strategy, e.g. a MirroredStrategy.' %
                       (strategy.__class__.__name__, self._dtype_policy.name))

    # Performance optimization: cache the compute dtype as a Dtype object or
    # None, so that str to Dtype conversion doesn't happen in Layer.__call__.
    if self._dtype_policy.compute_dtype:
      self._compute_dtype_object = dtypes.as_dtype(
          self._dtype_policy.compute_dtype)
    else:
      self._compute_dtype_object = None

  # TODO(reedwm): Expose this property?
  @property
  def _compute_dtype(self):
    """The layer's compute dtype.

    Unless mixed-precision is used, this is the same as `Layer.dtype`.

    If self._autocast is True, layer's will cast floating-point inputs to this.

    Returns:
      The layer's compute dtype.
    """
    return self._dtype_policy.compute_dtype

  def _maybe_cast_inputs(self, inputs):
    """Maybe casts the inputs to the compute dtype.

    If self._compute_dtype is floating-point, and self_autocast is True,
    floating-point inputs are casted to self._compute_dtype.

    Args:
      inputs: Input tensor, or structure of input tensors.

    Returns:
      `inputs`, but tensors may have been casted to self._compute_dtype
    """
    compute_dtype = self._compute_dtype
    if (self._autocast and compute_dtype and
        dtypes.as_dtype(compute_dtype).is_floating):
      def f(x):
        """Cast a single Tensor or TensorSpec to the compute dtype."""
        cast_types = (ops.Tensor, sparse_tensor.SparseTensor,
                      ragged_tensor.RaggedTensor)
        if (isinstance(x, cast_types) and x.dtype.is_floating and
            x.dtype.base_dtype.name != compute_dtype):
          return math_ops.cast(x, compute_dtype)
        elif isinstance(x, tensor_spec.TensorSpec) and x.dtype.is_floating:
          # Inputs may be TensorSpecs when this function is called from
          # model._set_inputs.
          return tensor_spec.TensorSpec(x.shape, compute_dtype, x.name)
        else:
          return x
      return nest.map_structure(f, inputs)
    else:
      return inputs

  # _dtype used to be an attribute set in the constructor. We still expose it
  # because some clients still use it.
  # TODO(reedwm): Deprecate, then remove the _dtype property.
  @property
  def _dtype(self):
    # This is equivalent to returning self.dtype . We do not return self.dtype
    # as it would cause infinite recursion in a few subclasses, which override
    # "dtype" to return self._dtype.
    return self._dtype_policy.variable_dtype

  @_dtype.setter
  def _dtype(self, value):
    value = dtypes.as_dtype(value).name
    self._set_dtype_policy(policy.Policy(value))

  def _name_scope(self):
    return self.name

  def _init_set_name(self, name, zero_based=True):
    if not name:
      self._name = backend.unique_object_name(
          generic_utils.to_snake_case(self.__class__.__name__),
          zero_based=zero_based)
    else:
      self._name = name

  def _get_existing_metric(self, name=None):
    match = [m for m in self._metrics if m.name == name]
    if not match:
      return
    if len(match) > 1:
      raise ValueError(
          'Please provide different names for the metrics you have added. '
          'We found {} metrics with the name: "{}"'.format(len(match), name))
    return match[0]

  def _symbolic_add_metric(self, value, aggregation=None, name=None):
    base_layer_utils.check_graph_consistency(value, method='add_metric')
    match = self._get_existing_metric(name)
    if aggregation is None:
      # Iterate over the metrics and check if the given metric exists already.
      # This can happen when a metric instance is created in subclassed model
      # layer `__init__` and we have tracked that instance already in
      # model.__setattr__.
      if match:
        result_tensor = value
        metric_obj = match
      elif hasattr(value, '_metric_obj'):
        # We track the instance using the metadata on the result tensor.
        result_tensor = value
        metric_obj = result_tensor._metric_obj
        self._metrics.append(metric_obj)
      else:
        raise ValueError(
            'We do not support adding an aggregated metric result tensor that '
            'is not the output of a `tf.keras.metrics.Metric` metric instance. '
            'Without having access to the metric instance we cannot reset the '
            'state of a metric after every epoch during training. You can '
            'create a `tf.keras.metrics.Metric` instance and pass the result '
            'here or pass an un-aggregated result with `aggregation` parameter '
            'set as `mean`. For example: `self.add_metric(tf.reduce_sum(inputs)'
            ', name=\'mean_activation\', aggregation=\'mean\')`')
    else:
      # If a non-aggregated tensor is given as input (ie. `aggregation` is
      # explicitly set to `mean`), we wrap the tensor in `Mean` metric.
      if match:
        result_tensor = match(value)
        metric_obj = match
      else:
        metric_obj, result_tensor = base_layer_utils.create_mean_metric(
            value, name)
        self._metrics.append(metric_obj)

  def _handle_weight_regularization(self, name, variable, regularizer):
    """Create lambdas which compute regularization losses."""

    def _loss_for_variable(v):
      """Creates a regularization loss `Tensor` for variable `v`."""
      with backend.name_scope(name + '/Regularizer'):
        regularization = regularizer(v)
      return regularization

    if base_layer_utils.is_split_variable(variable):
      for v in variable:
        self.add_loss(functools.partial(_loss_for_variable, v))
    else:
      self.add_loss(functools.partial(_loss_for_variable, variable))

  def _handle_activity_regularization(self, inputs, outputs):
    # Apply activity regularization.
    # Note that it should be applied every time the layer creates a new
    # output, since it is output-specific.
    if self._activity_regularizer:
      output_list = nest.flatten(outputs)
      with backend.name_scope('ActivityRegularizer'):
        for output in output_list:
          activity_loss = self._activity_regularizer(output)
          batch_size = math_ops.cast(
              array_ops.shape(output)[0], activity_loss.dtype)
          # Make activity regularization strength batch-agnostic.
          mean_activity_loss = activity_loss / batch_size
          base_layer_utils.check_graph_consistency(
              mean_activity_loss, method='activity_regularizer')
          self.add_loss(mean_activity_loss, inputs=inputs)

  def _set_mask_metadata(self, inputs, outputs, previous_mask):
    flat_outputs = nest.flatten(outputs)

    mask_already_computed = (
        getattr(self, '_compute_output_and_mask_jointly', False) or
        all(getattr(x, '_keras_mask', None) is not None for x in flat_outputs))

    # Only compute the mask if the Layer explicitly supports masking or has
    # overridden `compute_mask`.
    should_compute_mask = (
        hasattr(self, 'compute_mask') and
        (self.supports_masking or
         not getattr(self.compute_mask, '_is_default', False)))

    if mask_already_computed:
      flat_masks = [getattr(x, '_keras_mask', None) for x in flat_outputs]
    elif not should_compute_mask:
      flat_masks = [None for _ in flat_outputs]
    else:
      output_masks = self.compute_mask(inputs, previous_mask)
      # `compute_mask` can return a single `None` even when a Layer
      # has multiple outputs.
      if output_masks is None:
        flat_masks = [None for _ in flat_outputs]
      else:
        flat_masks = nest.flatten(output_masks)

    for output, mask in zip(flat_outputs, flat_masks):
      try:
        output._keras_mask = mask
      except AttributeError:
        # C Type such as np.ndarray.
        pass

    if tf_utils.are_all_symbolic_tensors(flat_outputs):
      for output in flat_outputs:
        if getattr(output, '_keras_mask', None) is not None:
          # Do not track masks for `TensorFlowOpLayer` construction.
          output._keras_mask._keras_history_checked = True

  def _collect_input_masks(self, inputs, args, kwargs):
    """Checks if `mask` argument was passed, else gathers mask from inputs."""
    if self._call_arg_was_passed('mask', args, kwargs):
      return self._get_call_arg_value('mask', args, kwargs)

    if not self._should_compute_mask:
      return None

    input_masks = nest.map_structure(lambda t: getattr(t, '_keras_mask', None),
                                     inputs)
    if generic_utils.is_all_none(input_masks):
      return None
    return input_masks

  def _call_arg_was_passed(self, arg_name, args, kwargs, inputs_in_args=False):
    if arg_name in kwargs:
      return True
    call_fn_args = self._call_fn_args
    if not inputs_in_args:
      # Ignore `inputs` arg.
      call_fn_args = call_fn_args[1:]
    if arg_name in dict(zip(call_fn_args, args)):
      return True
    return False

  def _get_call_arg_value(self, arg_name, args, kwargs, inputs_in_args=False):
    if arg_name in kwargs:
      return kwargs[arg_name]
    call_fn_args = self._call_fn_args
    if not inputs_in_args:
      # Ignore `inputs` arg.
      call_fn_args = call_fn_args[1:]
    args_dict = dict(zip(call_fn_args, args))
    return args_dict[arg_name]

  def _set_call_arg_value(
      self, arg_name, new_value, args,
      kwargs, inputs_in_args=False, pop_kwarg_if_none=False):
    arg_pos = self._call_fn_arg_positions.get(arg_name, None)
    if arg_pos is not None:
      if not inputs_in_args:
        # Ignore `inputs` arg.
        arg_pos = arg_pos - 1
      if len(args) > arg_pos:
        args = list(args)
        args[arg_pos] = new_value
        return args, kwargs
    if new_value is None and pop_kwarg_if_none:
      kwargs.pop(arg_name, None)
    else:
      kwargs[arg_name] = new_value
    return args, kwargs

  def _get_node_attribute_at_index(self, node_index, attr, attr_name):
    """Private utility to retrieves an attribute (e.g. inputs) from a node.

    This is used to implement the methods:
        - get_input_shape_at
        - get_output_shape_at
        - get_input_at
        etc...

    Args:
        node_index: Integer index of the node from which
            to retrieve the attribute.
        attr: Exact node attribute name.
        attr_name: Human-readable attribute name, for error messages.

    Returns:
        The layer's attribute `attr` at the node of index `node_index`.

    Raises:
        RuntimeError: If the layer has no inbound nodes, or if called in Eager
        mode.
        ValueError: If the index provided does not match any node.
    """
    if not self._inbound_nodes:
      raise RuntimeError('The layer has never been called '
                         'and thus has no defined ' + attr_name + '.')
    if not len(self._inbound_nodes) > node_index:
      raise ValueError('Asked to get ' + attr_name + ' at node ' +
                       str(node_index) + ', but the layer has only ' +
                       str(len(self._inbound_nodes)) + ' inbound nodes.')
    values = getattr(self._inbound_nodes[node_index], attr)
    if isinstance(values, list) and len(values) == 1:
      return values[0]
    else:
      return values

  def _maybe_build(self, inputs):
    # Check input assumptions set before layer building, e.g. input rank.
    if not self.built:
      input_spec.assert_input_compatibility(
          self.input_spec, inputs, self.name)
      input_list = nest.flatten(inputs)
      if input_list and self._dtype_policy.compute_dtype is None:
        try:
          dtype = input_list[0].dtype.base_dtype.name
        except AttributeError:
          pass
        else:
          self._set_dtype_policy(policy.Policy(dtype))
      input_shapes = None
      if all(hasattr(x, 'shape') for x in input_list):
        input_shapes = nest.map_structure(lambda x: x.shape, inputs)
      # Only call `build` if the user has manually overridden the build method.
      if not hasattr(self.build, '_is_default'):
        # Any setup work performed only once should happen in an `init_scope`
        # to avoid creating symbolic Tensors that will later pollute any eager
        # operations.
        with tf_utils.maybe_init_scope(self):
          self.build(input_shapes)
      # We must set also ensure that the layer is marked as built, and the build
      # shape is stored since user defined build functions may not be calling
      # `super.build()`
      Layer.build(self, input_shapes)

    # Optionally load weight values specified at layer instantiation.
    if self._initial_weights is not None:
      self.set_weights(self._initial_weights)
      self._initial_weights = None

  def _symbolic_call(self, inputs):
    input_shapes = nest.map_structure(lambda x: x.shape, inputs)
    output_shapes = self.compute_output_shape(input_shapes)

    def _make_placeholder_like(shape):
      ph = backend.placeholder(shape=shape, dtype=self.dtype)
      ph._keras_mask = None
      return ph

    return nest.map_structure(_make_placeholder_like, output_shapes)

  def _get_trainable_state(self):
    """Get the `trainable` state of each sublayer.

    Returns:
      A dict mapping all sublayers to their `trainable` value.
    """
    layers = self._flatten_layers(include_self=False, recursive=False)
    trainable_state = {self: self.trainable}
    for l in layers:
      trainable_state.update(l._get_trainable_state())
    return trainable_state

  def _set_trainable_state(self, trainable_state):
    """Set `trainable` state for each sublayer."""
    if self in trainable_state:
      self.trainable = trainable_state[self]
    layers = self._flatten_layers(include_self=False, recursive=False)
    for l in layers:
      if l in trainable_state:
        l._set_trainable_state(trainable_state)

  @property
  def _obj_reference_counts(self):
    """A dictionary counting the number of attributes referencing an object."""
    self._maybe_create_attribute('_obj_reference_counts_dict',
                                 object_identity.ObjectIdentityDictionary())
    return self._obj_reference_counts_dict

  @trackable.no_automatic_dependency_tracking
  def _maybe_create_attribute(self, name, default_value):
    """Create the attribute with the default value if it hasn't been created.

    This is useful for fields that is used for tracking purpose,
    _trainable_weights, or _layers. Note that user could create a layer subclass
    and assign an internal field before invoking the Layer.__init__(), the
    __setattr__() need to create the tracking fields and __init__() need to not
    override them.

    Args:
      name: String, the name of the attribute.
      default_value: Object, the default value of the attribute.
    """
    if not hasattr(self, name):
      self.__setattr__(name, default_value)

  def __delattr__(self, name):
    # For any super.__delattr__() call, we will directly use the implementation
    # in Trackable and skip the behavior in AutoTrackable. The Layer was
    # originally use Trackable as base class, the change of using Module as base
    # class forced us to have AutoTrackable in the class hierarchy. Skipping
    # the __delattr__ and __setattr__ in AutoTrackable will keep the status quo.
    existing_value = getattr(self, name, None)

    # If this value is replacing an existing object assigned to an attribute, we
    # should clean it out to avoid leaking memory. First we check if there are
    # other attributes referencing it.
    reference_counts = self._obj_reference_counts
    if existing_value not in reference_counts:
      super(tracking.AutoTrackable, self).__delattr__(name)
      return

    reference_count = reference_counts[existing_value]
    if reference_count > 1:
      # There are other remaining references. We can't remove this object from
      # _layers etc.
      reference_counts[existing_value] = reference_count - 1
      super(tracking.AutoTrackable, self).__delattr__(name)
      return
    else:
      # This is the last remaining reference.
      del reference_counts[existing_value]

    super(tracking.AutoTrackable, self).__delattr__(name)

    if (isinstance(existing_value, Layer)
        or base_layer_utils.has_weights(existing_value)):
      super(tracking.AutoTrackable, self).__setattr__(
          '_self_tracked_trackables',
          [l for l in self._self_tracked_trackables if l is not existing_value])
    if isinstance(existing_value, tf_variables.Variable):
      super(tracking.AutoTrackable, self).__setattr__(
          '_trainable_weights',
          [w for w in self._trainable_weights if w is not existing_value])
      super(tracking.AutoTrackable, self).__setattr__(
          '_non_trainable_weights',
          [w for w in self._non_trainable_weights if w is not existing_value])

  def __setattr__(self, name, value):
    if (name == '_self_setattr_tracking' or
        not getattr(self, '_self_setattr_tracking', True) or
        # Exclude @property.setters from tracking
        hasattr(self.__class__, name)):
      try:
        super(tracking.AutoTrackable, self).__setattr__(name, value)
      except AttributeError:
        raise AttributeError(
            ('Can\'t set the attribute "{}", likely because it conflicts with '
             'an existing read-only @property of the object. Please choose a '
             'different name.').format(name))
      return

    # Keep track of trackable objects, for the needs of `Network.save_weights`.
    value = data_structures.sticky_attribute_assignment(
        trackable=self, value=value, name=name)

    reference_counts = self._obj_reference_counts
    reference_counts[value] = reference_counts.get(value, 0) + 1

    # Clean out the old attribute, which clears _layers and _trainable_weights
    # if necessary.
    try:
      self.__delattr__(name)
    except AttributeError:
      pass

    # Keep track of metric instance created in subclassed layer.
    from tensorflow.python.keras import metrics as metrics_module  # pylint: disable=g-import-not-at-top
    for val in nest.flatten(value):
      if isinstance(val, metrics_module.Metric) and hasattr(self, '_metrics'):
        self._metrics.append(val)

    # TODO(scottzhu): Need to track Module object as well for weight tracking.
    # Be careful about metric if it becomes a Module in future.
    # Append value to self._layers if relevant
    if (getattr(self, '_auto_track_sub_layers', True) and
        (isinstance(value, Layer) or base_layer_utils.has_weights(value))):
      self._maybe_create_attribute('_self_tracked_trackables', [])
      # We need to check object identity to avoid de-duplicating empty
      # container types which compare equal.
      if not any((layer is value for layer in self._self_tracked_trackables)):
        self._self_tracked_trackables.append(value)
        if hasattr(value, '_use_resource_variables'):
          # Legacy layers (V1 tf.layers) must always use
          # resource variables.
          value._use_resource_variables = True

    # Append value to list of trainable / non-trainable weights if relevant
    # TODO(b/125122625): This won't pick up on any variables added to a
    # list/dict after creation.
    for val in nest.flatten(value):
      if not isinstance(val, tf_variables.Variable):
        continue

      # Users may add extra weights/variables
      # simply by assigning them to attributes (invalid for graph networks)
      self._maybe_create_attribute('_trainable_weights', [])
      self._maybe_create_attribute('_non_trainable_weights', [])
      if val.trainable:
        if any(val is w for w in self._trainable_weights):
          continue
        self._trainable_weights.append(val)
      else:
        if any(val is w for w in self._non_trainable_weights):
          continue
        self._non_trainable_weights.append(val)

      backend.track_variable(val)

    # Skip the auto trackable from tf.Module to keep status quo. See the comment
    # at __delattr__.
    super(tracking.AutoTrackable, self).__setattr__(name, value)

  # This is a hack so that the is_layer (within
  # training/trackable/layer_utils.py) check doesn't get the weights attr.
  # TODO(b/110718070): Remove when fixed.
  def _is_layer(self):
    return True

  def _init_call_fn_args(self):
    # Clear cached call function arguments.
    self.__class__._call_full_argspec.fget.cache.pop(self, None)
    self.__class__._call_fn_args.fget.cache.pop(self, None)
    self.__class__._call_accepts_kwargs.fget.cache.pop(self, None)

    call_fn_args = self._call_fn_args
    self._expects_training_arg = ('training' in call_fn_args or
                                  self._call_accepts_kwargs)
    self._expects_mask_arg = ('mask' in call_fn_args or
                              self._call_accepts_kwargs)

  @property
  @layer_utils.cached_per_instance
  def _call_full_argspec(self):
    # Argspec inspection is expensive and the call spec is used often, so it
    # makes sense to cache the result.
    return tf_inspect.getfullargspec(self.call)

  @property
  @layer_utils.cached_per_instance
  def _call_fn_args(self):
    all_args = self._call_full_argspec.args
    # Scrub `self` that appears if a decorator was applied.
    if all_args and all_args[0] == 'self':
      return all_args[1:]
    return all_args

  @property
  @layer_utils.cached_per_instance
  def _call_fn_arg_positions(self):
    call_fn_arg_positions = dict()
    for pos, arg in enumerate(self._call_fn_args):
      call_fn_arg_positions[arg] = pos
    return call_fn_arg_positions

  @property
  @layer_utils.cached_per_instance
  def _call_accepts_kwargs(self):
    return self._call_full_argspec.varkw is not None

  @property
  @layer_utils.cached_per_instance
  def _should_compute_mask(self):
    return ('mask' in self._call_fn_args or
            getattr(self, 'compute_mask', None) is not None)

  def _dedup_weights(self, weights):
    """Dedupe weights while maintaining order as much as possible."""
    output, seen_ids = [], set()
    for w in weights:
      if id(w) not in seen_ids:
        output.append(w)
        # Track the Variable's identity to avoid __eq__ issues.
        seen_ids.add(id(w))

    return output

  # SavedModel properties. Please see keras/saving/saved_model for details.

  @property
  def _trackable_saved_model_saver(self):
    return layer_serialization.LayerSavedModelSaver(self)

  @property
  def _object_identifier(self):
    return self._trackable_saved_model_saver.object_identifier

  @property
  def _tracking_metadata(self):
    return self._trackable_saved_model_saver.tracking_metadata

  def _list_extra_dependencies_for_serialization(self, serialization_cache):
    return (self._trackable_saved_model_saver
            .list_extra_dependencies_for_serialization(serialization_cache))

  def _list_functions_for_serialization(self, serialization_cache):
    return (self._trackable_saved_model_saver
            .list_functions_for_serialization(serialization_cache))

  def __getstate__(self):
    # Override to support `copy.deepcopy` and pickling.
    # Thread-local objects cannot be copied in Python 3, so pop these.
    # Thread-local objects are used to cache losses in MirroredStrategy, and
    # so shouldn't be copied.
    state = self.__dict__.copy()
    state.pop('_thread_local', None)
    return state

  def __setstate__(self, state):
    state['_thread_local'] = threading.local()
    # Bypass Trackable logic as `__dict__` already contains this info.
    object.__setattr__(self, '__dict__', state)


class KerasHistory(
    collections.namedtuple('KerasHistory',
                           ['layer', 'node_index', 'tensor_index'])):
  """Tracks the Layer call that created a Tensor, for Keras Graph Networks.

  During construction of Keras Graph Networks, this metadata is added to
  each Tensor produced as the output of a Layer, starting with an
  `InputLayer`. This allows Keras to track how each Tensor was produced, and
  this information is later retraced by the `keras.engine.Network` class to
  reconstruct the Keras Graph Network.

  Attributes:
    layer: The Layer that produced the Tensor.
    node_index: The specific call to the Layer that produced this Tensor. Layers
      can be called multiple times in order to share weights. A new node is
      created every time a Tensor is called.
    tensor_index: The output index for this Tensor. Always zero if the Layer
      that produced this Tensor only has one output. Nested structures of
      Tensors are deterministically assigned an index via `nest.flatten`.
  """
  # Added to maintain memory and performance characteristics of `namedtuple`
  # while subclassing.
  __slots__ = ()


# Avoid breaking users who directly import this symbol from this file.
# TODO(fchollet): remove this.
InputSpec = input_spec.InputSpec  # pylint:disable=invalid-name