STT-tensorflow/tensorflow/python/keras/layers/recurrent.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
"""Recurrent layers and their base classes.
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import numpy as np

from tensorflow.python.distribute import distribution_strategy_context as ds_context
from tensorflow.python.eager import context
from tensorflow.python.framework import ops
from tensorflow.python.framework import tensor_shape
from tensorflow.python.keras import activations
from tensorflow.python.keras import backend as K
from tensorflow.python.keras import constraints
from tensorflow.python.keras import initializers
from tensorflow.python.keras import regularizers
from tensorflow.python.keras.engine.base_layer import Layer
from tensorflow.python.keras.engine.input_spec import InputSpec
from tensorflow.python.keras.saving.saved_model import layer_serialization
from tensorflow.python.keras.utils import generic_utils
from tensorflow.python.keras.utils import tf_utils
from tensorflow.python.ops import array_ops
from tensorflow.python.ops import control_flow_ops
from tensorflow.python.ops import control_flow_util
from tensorflow.python.ops import math_ops
from tensorflow.python.ops import state_ops
from tensorflow.python.platform import tf_logging as logging
from tensorflow.python.training.tracking import base as trackable
from tensorflow.python.training.tracking import data_structures
from tensorflow.python.util import nest
from tensorflow.python.util.tf_export import keras_export
from tensorflow.tools.docs import doc_controls

try:
  from collections import abc as collections_abc  # pylint: disable=g-import-not-at-top
except ImportError:  # For Python 2
  import collections as collections_abc  # pylint: disable=g-import-not-at-top


RECURRENT_DROPOUT_WARNING_MSG = (
    'RNN `implementation=2` is not supported when `recurrent_dropout` is set. '
    'Using `implementation=1`.')


@keras_export('keras.layers.StackedRNNCells')
class StackedRNNCells(Layer):
  """Wrapper allowing a stack of RNN cells to behave as a single cell.

  Used to implement efficient stacked RNNs.

  Arguments:
    cells: List of RNN cell instances.

  Examples:

  ```python
  batch_size = 3
  sentence_max_length = 5
  n_features = 2
  new_shape = (batch_size, sentence_max_length, n_features)
  x = tf.constant(np.reshape(np.arange(30), new_shape), dtype = tf.float32)

  rnn_cells = [tf.keras.layers.LSTMCell(128) for _ in range(2)]
  stacked_lstm = tf.keras.layers.StackedRNNCells(rnn_cells)
  lstm_layer = tf.keras.layers.RNN(stacked_lstm)

  result = lstm_layer(x)
  ```
  """

  def __init__(self, cells, **kwargs):
    for cell in cells:
      if not 'call' in dir(cell):
        raise ValueError('All cells must have a `call` method. '
                         'received cells:', cells)
      if not 'state_size' in dir(cell):
        raise ValueError('All cells must have a '
                         '`state_size` attribute. '
                         'received cells:', cells)
    self.cells = cells
    # reverse_state_order determines whether the state size will be in a reverse
    # order of the cells' state. User might want to set this to True to keep the
    # existing behavior. This is only useful when use RNN(return_state=True)
    # since the state will be returned as the same order of state_size.
    self.reverse_state_order = kwargs.pop('reverse_state_order', False)
    if self.reverse_state_order:
      logging.warning('reverse_state_order=True in StackedRNNCells will soon '
                      'be deprecated. Please update the code to work with the '
                      'natural order of states if you rely on the RNN states, '
                      'eg RNN(return_state=True).')
    super(StackedRNNCells, self).__init__(**kwargs)

  @property
  def state_size(self):
    return tuple(c.state_size for c in
                 (self.cells[::-1] if self.reverse_state_order else self.cells))

  @property
  def output_size(self):
    if getattr(self.cells[-1], 'output_size', None) is not None:
      return self.cells[-1].output_size
    elif _is_multiple_state(self.cells[-1].state_size):
      return self.cells[-1].state_size[0]
    else:
      return self.cells[-1].state_size

  def get_initial_state(self, inputs=None, batch_size=None, dtype=None):
    initial_states = []
    for cell in self.cells[::-1] if self.reverse_state_order else self.cells:
      get_initial_state_fn = getattr(cell, 'get_initial_state', None)
      if get_initial_state_fn:
        initial_states.append(get_initial_state_fn(
            inputs=inputs, batch_size=batch_size, dtype=dtype))
      else:
        initial_states.append(_generate_zero_filled_state_for_cell(
            cell, inputs, batch_size, dtype))

    return tuple(initial_states)

  def call(self, inputs, states, constants=None, training=None, **kwargs):
    # Recover per-cell states.
    state_size = (self.state_size[::-1]
                  if self.reverse_state_order else self.state_size)
    nested_states = nest.pack_sequence_as(state_size, nest.flatten(states))

    # Call the cells in order and store the returned states.
    new_nested_states = []
    for cell, states in zip(self.cells, nested_states):
      states = states if nest.is_sequence(states) else [states]
      # TF cell does not wrap the state into list when there is only one state.
      is_tf_rnn_cell = getattr(cell, '_is_tf_rnn_cell', None) is not None
      states = states[0] if len(states) == 1 and is_tf_rnn_cell else states
      if generic_utils.has_arg(cell.call, 'training'):
        kwargs['training'] = training
      else:
        kwargs.pop('training', None)
      # Use the __call__ function for callable objects, eg layers, so that it
      # will have the proper name scopes for the ops, etc.
      cell_call_fn = cell.__call__ if callable(cell) else cell.call
      if generic_utils.has_arg(cell.call, 'constants'):
        inputs, states = cell_call_fn(inputs, states,
                                      constants=constants, **kwargs)
      else:
        inputs, states = cell_call_fn(inputs, states, **kwargs)
      new_nested_states.append(states)

    return inputs, nest.pack_sequence_as(state_size,
                                         nest.flatten(new_nested_states))

  @tf_utils.shape_type_conversion
  def build(self, input_shape):
    if isinstance(input_shape, list):
      input_shape = input_shape[0]
    for cell in self.cells:
      if isinstance(cell, Layer) and not cell.built:
        with K.name_scope(cell.name):
          cell.build(input_shape)
          cell.built = True
      if getattr(cell, 'output_size', None) is not None:
        output_dim = cell.output_size
      elif _is_multiple_state(cell.state_size):
        output_dim = cell.state_size[0]
      else:
        output_dim = cell.state_size
      input_shape = tuple([input_shape[0]] +
                          tensor_shape.as_shape(output_dim).as_list())
    self.built = True

  def get_config(self):
    cells = []
    for cell in self.cells:
      cells.append({
          'class_name': cell.__class__.__name__,
          'config': cell.get_config()
      })
    config = {'cells': cells}
    base_config = super(StackedRNNCells, self).get_config()
    return dict(list(base_config.items()) + list(config.items()))

  @classmethod
  def from_config(cls, config, custom_objects=None):
    from tensorflow.python.keras.layers import deserialize as deserialize_layer  # pylint: disable=g-import-not-at-top
    cells = []
    for cell_config in config.pop('cells'):
      cells.append(
          deserialize_layer(cell_config, custom_objects=custom_objects))
    return cls(cells, **config)


@keras_export('keras.layers.RNN')
class RNN(Layer):
  """Base class for recurrent layers.

  See [the Keras RNN API guide](https://www.tensorflow.org/guide/keras/rnn)
  for details about the usage of RNN API.

  Arguments:
    cell: A RNN cell instance or a list of RNN cell instances.
      A RNN cell is a class that has:
      - A `call(input_at_t, states_at_t)` method, returning
        `(output_at_t, states_at_t_plus_1)`. The call method of the
        cell can also take the optional argument `constants`, see
        section "Note on passing external constants" below.
      - A `state_size` attribute. This can be a single integer
        (single state) in which case it is the size of the recurrent
        state. This can also be a list/tuple of integers (one size per state).
        The `state_size` can also be TensorShape or tuple/list of
        TensorShape, to represent high dimension state.
      - A `output_size` attribute. This can be a single integer or a
        TensorShape, which represent the shape of the output. For backward
        compatible reason, if this attribute is not available for the
        cell, the value will be inferred by the first element of the
        `state_size`.
      - A `get_initial_state(inputs=None, batch_size=None, dtype=None)`
        method that creates a tensor meant to be fed to `call()` as the
        initial state, if the user didn't specify any initial state via other
        means. The returned initial state should have a shape of
        [batch_size, cell.state_size]. The cell might choose to create a
        tensor full of zeros, or full of other values based on the cell's
        implementation.
        `inputs` is the input tensor to the RNN layer, which should
        contain the batch size as its shape[0], and also dtype. Note that
        the shape[0] might be `None` during the graph construction. Either
        the `inputs` or the pair of `batch_size` and `dtype` are provided.
        `batch_size` is a scalar tensor that represents the batch size
        of the inputs. `dtype` is `tf.DType` that represents the dtype of
        the inputs.
        For backward compatible reason, if this method is not implemented
        by the cell, the RNN layer will create a zero filled tensor with the
        size of [batch_size, cell.state_size].
      In the case that `cell` is a list of RNN cell instances, the cells
      will be stacked on top of each other in the RNN, resulting in an
      efficient stacked RNN.
    return_sequences: Boolean (default `False`). Whether to return the last
      output in the output sequence, or the full sequence.
    return_state: Boolean (default `False`). Whether to return the last state
      in addition to the output.
    go_backwards: Boolean (default `False`).
      If True, process the input sequence backwards and return the
      reversed sequence.
    stateful: Boolean (default `False`). If True, the last state
      for each sample at index i in a batch will be used as initial
      state for the sample of index i in the following batch.
    unroll: Boolean (default `False`).
      If True, the network will be unrolled, else a symbolic loop will be used.
      Unrolling can speed-up a RNN, although it tends to be more
      memory-intensive. Unrolling is only suitable for short sequences.
    time_major: The shape format of the `inputs` and `outputs` tensors.
      If True, the inputs and outputs will be in shape
      `(timesteps, batch, ...)`, whereas in the False case, it will be
      `(batch, timesteps, ...)`. Using `time_major = True` is a bit more
      efficient because it avoids transposes at the beginning and end of the
      RNN calculation. However, most TensorFlow data is batch-major, so by
      default this function accepts input and emits output in batch-major
      form.
    zero_output_for_mask: Boolean (default `False`).
      Whether the output should use zeros for the masked timesteps. Note that
      this field is only used when `return_sequences` is True and mask is
      provided. It can useful if you want to reuse the raw output sequence of
      the RNN without interference from the masked timesteps, eg, merging
      bidirectional RNNs.

  Call arguments:
    inputs: Input tensor.
    mask: Binary tensor of shape `[batch_size, timesteps]` indicating whether
      a given timestep should be masked.
    training: Python boolean indicating whether the layer should behave in
      training mode or in inference mode. This argument is passed to the cell
      when calling it. This is for use with cells that use dropout.
    initial_state: List of initial state tensors to be passed to the first
      call of the cell.
    constants: List of constant tensors to be passed to the cell at each
      timestep.

  Input shape:
    N-D tensor with shape `[batch_size, timesteps, ...]` or
    `[timesteps, batch_size, ...]` when time_major is True.

  Output shape:
    - If `return_state`: a list of tensors. The first tensor is
      the output. The remaining tensors are the last states,
      each with shape `[batch_size, state_size]`, where `state_size` could
      be a high dimension tensor shape.
    - If `return_sequences`: N-D tensor with shape
      `[batch_size, timesteps, output_size]`, where `output_size` could
      be a high dimension tensor shape, or
      `[timesteps, batch_size, output_size]` when `time_major` is True.
    - Else, N-D tensor with shape `[batch_size, output_size]`, where
      `output_size` could be a high dimension tensor shape.

  Masking:
    This layer supports masking for input data with a variable number
    of timesteps. To introduce masks to your data,
    use an [tf.keras.layers.Embedding] layer with the `mask_zero` parameter
    set to `True`.

  Note on using statefulness in RNNs:
    You can set RNN layers to be 'stateful', which means that the states
    computed for the samples in one batch will be reused as initial states
    for the samples in the next batch. This assumes a one-to-one mapping
    between samples in different successive batches.

    To enable statefulness:
      - Specify `stateful=True` in the layer constructor.
      - Specify a fixed batch size for your model, by passing
        If sequential model:
          `batch_input_shape=(...)` to the first layer in your model.
        Else for functional model with 1 or more Input layers:
          `batch_shape=(...)` to all the first layers in your model.
        This is the expected shape of your inputs
        *including the batch size*.
        It should be a tuple of integers, e.g. `(32, 10, 100)`.
      - Specify `shuffle=False` when calling fit().

    To reset the states of your model, call `.reset_states()` on either
    a specific layer, or on your entire model.

  Note on specifying the initial state of RNNs:
    You can specify the initial state of RNN layers symbolically by
    calling them with the keyword argument `initial_state`. The value of
    `initial_state` should be a tensor or list of tensors representing
    the initial state of the RNN layer.

    You can specify the initial state of RNN layers numerically by
    calling `reset_states` with the keyword argument `states`. The value of
    `states` should be a numpy array or list of numpy arrays representing
    the initial state of the RNN layer.

  Note on passing external constants to RNNs:
    You can pass "external" constants to the cell using the `constants`
    keyword argument of `RNN.__call__` (as well as `RNN.call`) method. This
    requires that the `cell.call` method accepts the same keyword argument
    `constants`. Such constants can be used to condition the cell
    transformation on additional static inputs (not changing over time),
    a.k.a. an attention mechanism.

  Examples:

  ```python
  # First, let's define a RNN Cell, as a layer subclass.

  class MinimalRNNCell(keras.layers.Layer):

      def __init__(self, units, **kwargs):
          self.units = units
          self.state_size = units
          super(MinimalRNNCell, self).__init__(**kwargs)

      def build(self, input_shape):
          self.kernel = self.add_weight(shape=(input_shape[-1], self.units),
                                        initializer='uniform',
                                        name='kernel')
          self.recurrent_kernel = self.add_weight(
              shape=(self.units, self.units),
              initializer='uniform',
              name='recurrent_kernel')
          self.built = True

      def call(self, inputs, states):
          prev_output = states[0]
          h = K.dot(inputs, self.kernel)
          output = h + K.dot(prev_output, self.recurrent_kernel)
          return output, [output]

  # Let's use this cell in a RNN layer:

  cell = MinimalRNNCell(32)
  x = keras.Input((None, 5))
  layer = RNN(cell)
  y = layer(x)

  # Here's how to use the cell to build a stacked RNN:

  cells = [MinimalRNNCell(32), MinimalRNNCell(64)]
  x = keras.Input((None, 5))
  layer = RNN(cells)
  y = layer(x)
  ```
  """

  def __init__(self,
               cell,
               return_sequences=False,
               return_state=False,
               go_backwards=False,
               stateful=False,
               unroll=False,
               time_major=False,
               **kwargs):
    if isinstance(cell, (list, tuple)):
      cell = StackedRNNCells(cell)
    if not 'call' in dir(cell):
      raise ValueError('`cell` should have a `call` method. '
                       'The RNN was passed:', cell)
    if not 'state_size' in dir(cell):
      raise ValueError('The RNN cell should have '
                       'an attribute `state_size` '
                       '(tuple of integers, '
                       'one integer per RNN state).')
    # If True, the output for masked timestep will be zeros, whereas in the
    # False case, output from previous timestep is returned for masked timestep.
    self.zero_output_for_mask = kwargs.pop('zero_output_for_mask', False)

    if 'input_shape' not in kwargs and (
        'input_dim' in kwargs or 'input_length' in kwargs):
      input_shape = (kwargs.pop('input_length', None),
                     kwargs.pop('input_dim', None))
      kwargs['input_shape'] = input_shape

    super(RNN, self).__init__(**kwargs)
    self.cell = cell
    self.return_sequences = return_sequences
    self.return_state = return_state
    self.go_backwards = go_backwards
    self.stateful = stateful
    self.unroll = unroll
    self.time_major = time_major

    self.supports_masking = True
    # The input shape is unknown yet, it could have nested tensor inputs, and
    # the input spec will be the list of specs for nested inputs, the structure
    # of the input_spec will be the same as the input.
    self.input_spec = None
    self.state_spec = None
    self._states = None
    self.constants_spec = None
    self._num_constants = 0

    if stateful:
      if ds_context.has_strategy():
        raise ValueError('RNNs with stateful=True not yet supported with '
                         'tf.distribute.Strategy.')

  @property
  def states(self):
    if self._states is None:
      state = nest.map_structure(lambda _: None, self.cell.state_size)
      return state if nest.is_sequence(self.cell.state_size) else [state]
    return self._states

  @states.setter
  # Automatic tracking catches "self._states" which adds an extra weight and
  # breaks HDF5 checkpoints.
  @trackable.no_automatic_dependency_tracking
  def states(self, states):
    self._states = states

  def compute_output_shape(self, input_shape):
    if isinstance(input_shape, list):
      input_shape = input_shape[0]
    # Check whether the input shape contains any nested shapes. It could be
    # (tensor_shape(1, 2), tensor_shape(3, 4)) or (1, 2, 3) which is from numpy
    # inputs.
    try:
      input_shape = tensor_shape.as_shape(input_shape)
    except (ValueError, TypeError):
      # A nested tensor input
      input_shape = nest.flatten(input_shape)[0]

    batch = input_shape[0]
    time_step = input_shape[1]
    if self.time_major:
      batch, time_step = time_step, batch

    if _is_multiple_state(self.cell.state_size):
      state_size = self.cell.state_size
    else:
      state_size = [self.cell.state_size]

    def _get_output_shape(flat_output_size):
      output_dim = tensor_shape.as_shape(flat_output_size).as_list()
      if self.return_sequences:
        if self.time_major:
          output_shape = tensor_shape.as_shape([time_step, batch] + output_dim)
        else:
          output_shape = tensor_shape.as_shape([batch, time_step] + output_dim)
      else:
        output_shape = tensor_shape.as_shape([batch] + output_dim)
      return output_shape

    if getattr(self.cell, 'output_size', None) is not None:
      # cell.output_size could be nested structure.
      output_shape = nest.flatten(nest.map_structure(
          _get_output_shape, self.cell.output_size))
      output_shape = output_shape[0] if len(output_shape) == 1 else output_shape
    else:
      # Note that state_size[0] could be a tensor_shape or int.
      output_shape = _get_output_shape(state_size[0])

    if self.return_state:
      def _get_state_shape(flat_state):
        state_shape = [batch] + tensor_shape.as_shape(flat_state).as_list()
        return tensor_shape.as_shape(state_shape)
      state_shape = nest.map_structure(_get_state_shape, state_size)
      return generic_utils.to_list(output_shape) + nest.flatten(state_shape)
    else:
      return output_shape

  def compute_mask(self, inputs, mask):
    # Time step masks must be the same for each input.
    # This is because the mask for an RNN is of size [batch, time_steps, 1],
    # and specifies which time steps should be skipped, and a time step
    # must be skipped for all inputs.
    # TODO(scottzhu): Should we accept multiple different masks?
    mask = nest.flatten(mask)[0]
    output_mask = mask if self.return_sequences else None
    if self.return_state:
      state_mask = [None for _ in self.states]
      return [output_mask] + state_mask
    else:
      return output_mask

  def build(self, input_shape):
    if isinstance(input_shape, list):
      input_shape = input_shape[0]
      # The input_shape here could be a nest structure.

    # do the tensor_shape to shapes here. The input could be single tensor, or a
    # nested structure of tensors.
    def get_input_spec(shape):
      """Convert input shape to InputSpec."""
      if isinstance(shape, tensor_shape.TensorShape):
        input_spec_shape = shape.as_list()
      else:
        input_spec_shape = list(shape)
      batch_index, time_step_index = (1, 0) if self.time_major else (0, 1)
      if not self.stateful:
        input_spec_shape[batch_index] = None
      input_spec_shape[time_step_index] = None
      return InputSpec(shape=tuple(input_spec_shape))

    def get_step_input_shape(shape):
      if isinstance(shape, tensor_shape.TensorShape):
        shape = tuple(shape.as_list())
      # remove the timestep from the input_shape
      return shape[1:] if self.time_major else (shape[0],) + shape[2:]

    # Check whether the input shape contains any nested shapes. It could be
    # (tensor_shape(1, 2), tensor_shape(3, 4)) or (1, 2, 3) which is from numpy
    # inputs.
    try:
      input_shape = tensor_shape.as_shape(input_shape)
    except (ValueError, TypeError):
      # A nested tensor input
      pass

    if not nest.is_sequence(input_shape):
      # This indicates the there is only one input.
      if self.input_spec is not None:
        self.input_spec[0] = get_input_spec(input_shape)
      else:
        self.input_spec = [get_input_spec(input_shape)]
      step_input_shape = get_step_input_shape(input_shape)
    else:
      if self.input_spec is not None:
        self.input_spec[0] = nest.map_structure(get_input_spec, input_shape)
      else:
        self.input_spec = generic_utils.to_list(
            nest.map_structure(get_input_spec, input_shape))
      step_input_shape = nest.map_structure(get_step_input_shape, input_shape)

    # allow cell (if layer) to build before we set or validate state_spec.
    if isinstance(self.cell, Layer) and not self.cell.built:
      with K.name_scope(self.cell.name):
        self.cell.build(step_input_shape)
        self.cell.built = True

    # set or validate state_spec
    if _is_multiple_state(self.cell.state_size):
      state_size = list(self.cell.state_size)
    else:
      state_size = [self.cell.state_size]

    if self.state_spec is not None:
      # initial_state was passed in call, check compatibility
      self._validate_state_spec(state_size, self.state_spec)
    else:
      self.state_spec = [
          InputSpec(shape=[None] + tensor_shape.as_shape(dim).as_list())
          for dim in state_size
      ]
    if self.stateful:
      self.reset_states()
    self.built = True

  @staticmethod
  def _validate_state_spec(cell_state_sizes, init_state_specs):
    """Validate the state spec between the initial_state and the state_size.

    Args:
      cell_state_sizes: list, the `state_size` attribute from the cell.
      init_state_specs: list, the `state_spec` from the initial_state that is
        passed in `call()`.

    Raises:
      ValueError: When initial state spec is not compatible with the state size.
    """
    validation_error = ValueError(
        'An `initial_state` was passed that is not compatible with '
        '`cell.state_size`. Received `state_spec`={}; '
        'however `cell.state_size` is '
        '{}'.format(init_state_specs, cell_state_sizes))
    flat_cell_state_sizes = nest.flatten(cell_state_sizes)
    flat_state_specs = nest.flatten(init_state_specs)

    if len(flat_cell_state_sizes) != len(flat_state_specs):
      raise validation_error
    for cell_state_spec, cell_state_size in zip(flat_state_specs,
                                                flat_cell_state_sizes):
      if not tensor_shape.TensorShape(
          # Ignore the first axis for init_state which is for batch
          cell_state_spec.shape[1:]).is_compatible_with(
              tensor_shape.TensorShape(cell_state_size)):
        raise validation_error

  @doc_controls.do_not_doc_inheritable
  def get_initial_state(self, inputs):
    get_initial_state_fn = getattr(self.cell, 'get_initial_state', None)

    if nest.is_sequence(inputs):
      # The input are nested sequences. Use the first element in the seq to get
      # batch size and dtype.
      inputs = nest.flatten(inputs)[0]

    input_shape = array_ops.shape(inputs)
    batch_size = input_shape[1] if self.time_major else input_shape[0]
    dtype = inputs.dtype
    if get_initial_state_fn:
      init_state = get_initial_state_fn(
          inputs=None, batch_size=batch_size, dtype=dtype)
    else:
      init_state = _generate_zero_filled_state(batch_size, self.cell.state_size,
                                               dtype)
    # Keras RNN expect the states in a list, even if it's a single state tensor.
    if not nest.is_sequence(init_state):
      init_state = [init_state]
    # Force the state to be a list in case it is a namedtuple eg LSTMStateTuple.
    return list(init_state)

  def __call__(self, inputs, initial_state=None, constants=None, **kwargs):
    inputs, initial_state, constants = _standardize_args(inputs,
                                                         initial_state,
                                                         constants,
                                                         self._num_constants)

    if initial_state is None and constants is None:
      return super(RNN, self).__call__(inputs, **kwargs)

    # If any of `initial_state` or `constants` are specified and are Keras
    # tensors, then add them to the inputs and temporarily modify the
    # input_spec to include them.

    additional_inputs = []
    additional_specs = []
    if initial_state is not None:
      additional_inputs += initial_state
      self.state_spec = nest.map_structure(
          lambda s: InputSpec(shape=K.int_shape(s)), initial_state)
      additional_specs += self.state_spec
    if constants is not None:
      additional_inputs += constants
      self.constants_spec = [
          InputSpec(shape=K.int_shape(constant)) for constant in constants
      ]
      self._num_constants = len(constants)
      additional_specs += self.constants_spec
    # additional_inputs can be empty if initial_state or constants are provided
    # but empty (e.g. the cell is stateless).
    flat_additional_inputs = nest.flatten(additional_inputs)
    is_keras_tensor = K.is_keras_tensor(
        flat_additional_inputs[0]) if flat_additional_inputs else True
    for tensor in flat_additional_inputs:
      if K.is_keras_tensor(tensor) != is_keras_tensor:
        raise ValueError('The initial state or constants of an RNN'
                         ' layer cannot be specified with a mix of'
                         ' Keras tensors and non-Keras tensors'
                         ' (a "Keras tensor" is a tensor that was'
                         ' returned by a Keras layer, or by `Input`)')

    if is_keras_tensor:
      # Compute the full input spec, including state and constants
      full_input = [inputs] + additional_inputs
      if self.built:
        # Keep the input_spec since it has been populated in build() method.
        full_input_spec = self.input_spec + additional_specs
      else:
        # The original input_spec is None since there could be a nested tensor
        # input. Update the input_spec to match the inputs.
        full_input_spec = generic_utils.to_list(
            nest.map_structure(lambda _: None, inputs)) + additional_specs
      # Perform the call with temporarily replaced input_spec
      self.input_spec = full_input_spec
      output = super(RNN, self).__call__(full_input, **kwargs)
      # Remove the additional_specs from input spec and keep the rest. It is
      # important to keep since the input spec was populated by build(), and
      # will be reused in the stateful=True.
      self.input_spec = self.input_spec[:-len(additional_specs)]
      return output
    else:
      if initial_state is not None:
        kwargs['initial_state'] = initial_state
      if constants is not None:
        kwargs['constants'] = constants
      return super(RNN, self).__call__(inputs, **kwargs)

  def call(self,
           inputs,
           mask=None,
           training=None,
           initial_state=None,
           constants=None):
    # The input should be dense, padded with zeros. If a ragged input is fed
    # into the layer, it is padded and the row lengths are used for masking.
    inputs, row_lengths = K.convert_inputs_if_ragged(inputs)
    is_ragged_input = (row_lengths is not None)
    self._validate_args_if_ragged(is_ragged_input, mask)

    inputs, initial_state, constants = self._process_inputs(
        inputs, initial_state, constants)

    self._maybe_reset_cell_dropout_mask(self.cell)
    if isinstance(self.cell, StackedRNNCells):
      for cell in self.cell.cells:
        self._maybe_reset_cell_dropout_mask(cell)

    if mask is not None:
      # Time step masks must be the same for each input.
      # TODO(scottzhu): Should we accept multiple different masks?
      mask = nest.flatten(mask)[0]

    if nest.is_sequence(inputs):
      # In the case of nested input, use the first element for shape check.
      input_shape = K.int_shape(nest.flatten(inputs)[0])
    else:
      input_shape = K.int_shape(inputs)
    timesteps = input_shape[0] if self.time_major else input_shape[1]
    if self.unroll and timesteps is None:
      raise ValueError('Cannot unroll a RNN if the '
                       'time dimension is undefined. \n'
                       '- If using a Sequential model, '
                       'specify the time dimension by passing '
                       'an `input_shape` or `batch_input_shape` '
                       'argument to your first layer. If your '
                       'first layer is an Embedding, you can '
                       'also use the `input_length` argument.\n'
                       '- If using the functional API, specify '
                       'the time dimension by passing a `shape` '
                       'or `batch_shape` argument to your Input layer.')

    kwargs = {}
    if generic_utils.has_arg(self.cell.call, 'training'):
      kwargs['training'] = training

    # TF RNN cells expect single tensor as state instead of list wrapped tensor.
    is_tf_rnn_cell = getattr(self.cell, '_is_tf_rnn_cell', None) is not None
    # Use the __call__ function for callable objects, eg layers, so that it
    # will have the proper name scopes for the ops, etc.
    cell_call_fn = self.cell.__call__ if callable(self.cell) else self.cell.call
    if constants:
      if not generic_utils.has_arg(self.cell.call, 'constants'):
        raise ValueError('RNN cell does not support constants')

      def step(inputs, states):
        constants = states[-self._num_constants:]  # pylint: disable=invalid-unary-operand-type
        states = states[:-self._num_constants]  # pylint: disable=invalid-unary-operand-type

        states = states[0] if len(states) == 1 and is_tf_rnn_cell else states
        output, new_states = cell_call_fn(
            inputs, states, constants=constants, **kwargs)
        if not nest.is_sequence(new_states):
          new_states = [new_states]
        return output, new_states
    else:

      def step(inputs, states):
        states = states[0] if len(states) == 1 and is_tf_rnn_cell else states
        output, new_states = cell_call_fn(inputs, states, **kwargs)
        if not nest.is_sequence(new_states):
          new_states = [new_states]
        return output, new_states
    last_output, outputs, states = K.rnn(
        step,
        inputs,
        initial_state,
        constants=constants,
        go_backwards=self.go_backwards,
        mask=mask,
        unroll=self.unroll,
        input_length=row_lengths if row_lengths is not None else timesteps,
        time_major=self.time_major,
        zero_output_for_mask=self.zero_output_for_mask)

    if self.stateful:
      updates = [
          state_ops.assign(self_state, state) for self_state, state in zip(
              nest.flatten(self.states), nest.flatten(states))
      ]
      self.add_update(updates)

    if self.return_sequences:
      output = K.maybe_convert_to_ragged(is_ragged_input, outputs, row_lengths)
    else:
      output = last_output

    if self.return_state:
      if not isinstance(states, (list, tuple)):
        states = [states]
      else:
        states = list(states)
      return generic_utils.to_list(output) + states
    else:
      return output

  def _process_inputs(self, inputs, initial_state, constants):
    # input shape: `(samples, time (padded with zeros), input_dim)`
    # note that the .build() method of subclasses MUST define
    # self.input_spec and self.state_spec with complete input shapes.
    if (isinstance(inputs, collections_abc.Sequence)
        and not isinstance(inputs, tuple)):
      # get initial_state from full input spec
      # as they could be copied to multiple GPU.
      if not self._num_constants:
        initial_state = inputs[1:]
      else:
        initial_state = inputs[1:-self._num_constants]
        constants = inputs[-self._num_constants:]
      if len(initial_state) == 0:
        initial_state = None
      inputs = inputs[0]

    if self.stateful:
      if initial_state is not None:
        # When layer is stateful and initial_state is provided, check if the
        # recorded state is same as the default value (zeros). Use the recorded
        # state if it is not same as the default.
        non_zero_count = math_ops.add_n([math_ops.count_nonzero_v2(s)
                                         for s in nest.flatten(self.states)])
        # Set strict = True to keep the original structure of the state.
        initial_state = control_flow_ops.cond(non_zero_count > 0,
                                              true_fn=lambda: self.states,
                                              false_fn=lambda: initial_state,
                                              strict=True)
      else:
        initial_state = self.states
    elif initial_state is None:
      initial_state = self.get_initial_state(inputs)

    if len(initial_state) != len(self.states):
      raise ValueError('Layer has ' + str(len(self.states)) +
                       ' states but was passed ' + str(len(initial_state)) +
                       ' initial states.')
    return inputs, initial_state, constants

  def _validate_args_if_ragged(self, is_ragged_input, mask):
    if not is_ragged_input:
      return

    if mask is not None:
      raise ValueError('The mask that was passed in was ' + str(mask) +
                       ' and cannot be applied to RaggedTensor inputs. Please '
                       'make sure that there is no mask passed in by upstream '
                       'layers.')
    if self.unroll:
      raise ValueError('The input received contains RaggedTensors and does '
                       'not support unrolling. Disable unrolling by passing '
                       '`unroll=False` in the RNN Layer constructor.')

  def _maybe_reset_cell_dropout_mask(self, cell):
    if isinstance(cell, DropoutRNNCellMixin):
      cell.reset_dropout_mask()
      cell.reset_recurrent_dropout_mask()

  def reset_states(self, states=None):
    """Reset the recorded states for the stateful RNN layer.

    Can only be used when RNN layer is constructed with `stateful` = `True`.
    Args:
      states: Numpy arrays that contains the value for the initial state, which
        will be feed to cell at the first time step. When the value is None,
        zero filled numpy array will be created based on the cell state size.

    Raises:
      AttributeError: When the RNN layer is not stateful.
      ValueError: When the batch size of the RNN layer is unknown.
      ValueError: When the input numpy array is not compatible with the RNN
        layer state, either size wise or dtype wise.
    """
    if not self.stateful:
      raise AttributeError('Layer must be stateful.')
    spec_shape = None
    if self.input_spec is not None:
      spec_shape = nest.flatten(self.input_spec[0])[0].shape
    if spec_shape is None:
      # It is possible to have spec shape to be None, eg when construct a RNN
      # with a custom cell, or standard RNN layers (LSTM/GRU) which we only know
      # it has 3 dim input, but not its full shape spec before build().
      batch_size = None
    else:
      batch_size = spec_shape[1] if self.time_major else spec_shape[0]
    if not batch_size:
      raise ValueError('If a RNN is stateful, it needs to know '
                       'its batch size. Specify the batch size '
                       'of your input tensors: \n'
                       '- If using a Sequential model, '
                       'specify the batch size by passing '
                       'a `batch_input_shape` '
                       'argument to your first layer.\n'
                       '- If using the functional API, specify '
                       'the batch size by passing a '
                       '`batch_shape` argument to your Input layer.')
    # initialize state if None
    if nest.flatten(self.states)[0] is None:
      def create_state_variable(state):
        return K.zeros([batch_size] + tensor_shape.as_shape(state).as_list())
      self.states = nest.map_structure(
          create_state_variable, self.cell.state_size)
      if not nest.is_sequence(self.states):
        self.states = [self.states]
    elif states is None:
      for state, size in zip(nest.flatten(self.states),
                             nest.flatten(self.cell.state_size)):
        K.set_value(state, np.zeros([batch_size] +
                                    tensor_shape.as_shape(size).as_list()))
    else:
      flat_states = nest.flatten(self.states)
      flat_input_states = nest.flatten(states)
      if len(flat_input_states) != len(flat_states):
        raise ValueError('Layer ' + self.name + ' expects ' +
                         str(len(flat_states)) + ' states, '
                         'but it received ' + str(len(flat_input_states)) +
                         ' state values. Input received: ' + str(states))
      set_value_tuples = []
      for i, (value, state) in enumerate(zip(flat_input_states,
                                             flat_states)):
        if value.shape != state.shape:
          raise ValueError(
              'State ' + str(i) + ' is incompatible with layer ' +
              self.name + ': expected shape=' + str(
                  (batch_size, state)) + ', found shape=' + str(value.shape))
        set_value_tuples.append((state, value))
      K.batch_set_value(set_value_tuples)

  def get_config(self):
    config = {
        'return_sequences': self.return_sequences,
        'return_state': self.return_state,
        'go_backwards': self.go_backwards,
        'stateful': self.stateful,
        'unroll': self.unroll,
        'time_major': self.time_major
    }
    if self._num_constants:
      config['num_constants'] = self._num_constants
    if self.zero_output_for_mask:
      config['zero_output_for_mask'] = self.zero_output_for_mask

    cell_config = self.cell.get_config()
    config['cell'] = {
        'class_name': self.cell.__class__.__name__,
        'config': cell_config
    }
    base_config = super(RNN, self).get_config()
    return dict(list(base_config.items()) + list(config.items()))

  @classmethod
  def from_config(cls, config, custom_objects=None):
    from tensorflow.python.keras.layers import deserialize as deserialize_layer  # pylint: disable=g-import-not-at-top
    cell = deserialize_layer(config.pop('cell'), custom_objects=custom_objects)
    num_constants = config.pop('num_constants', 0)
    layer = cls(cell, **config)
    layer._num_constants = num_constants
    return layer

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


@keras_export('keras.layers.AbstractRNNCell')
class AbstractRNNCell(Layer):
  """Abstract object representing an RNN cell.

  See [the Keras RNN API guide](https://www.tensorflow.org/guide/keras/rnn)
  for details about the usage of RNN API.

  This is the base class for implementing RNN cells with custom behavior.

  Every `RNNCell` must have the properties below and implement `call` with
  the signature `(output, next_state) = call(input, state)`.

  Examples:

  ```python
    class MinimalRNNCell(AbstractRNNCell):

      def __init__(self, units, **kwargs):
        self.units = units
        super(MinimalRNNCell, self).__init__(**kwargs)

      @property
      def state_size(self):
        return self.units

      def build(self, input_shape):
        self.kernel = self.add_weight(shape=(input_shape[-1], self.units),
                                      initializer='uniform',
                                      name='kernel')
        self.recurrent_kernel = self.add_weight(
            shape=(self.units, self.units),
            initializer='uniform',
            name='recurrent_kernel')
        self.built = True

      def call(self, inputs, states):
        prev_output = states[0]
        h = K.dot(inputs, self.kernel)
        output = h + K.dot(prev_output, self.recurrent_kernel)
        return output, output
  ```

  This definition of cell differs from the definition used in the literature.
  In the literature, 'cell' refers to an object with a single scalar output.
  This definition refers to a horizontal array of such units.

  An RNN cell, in the most abstract setting, is anything that has
  a state and performs some operation that takes a matrix of inputs.
  This operation results in an output matrix with `self.output_size` columns.
  If `self.state_size` is an integer, this operation also results in a new
  state matrix with `self.state_size` columns.  If `self.state_size` is a
  (possibly nested tuple of) TensorShape object(s), then it should return a
  matching structure of Tensors having shape `[batch_size].concatenate(s)`
  for each `s` in `self.batch_size`.
  """

  def call(self, inputs, states):
    """The function that contains the logic for one RNN step calculation.

    Args:
      inputs: the input tensor, which is a slide from the overall RNN input by
        the time dimension (usually the second dimension).
      states: the state tensor from previous step, which has the same shape
        as `(batch, state_size)`. In the case of timestep 0, it will be the
        initial state user specified, or zero filled tensor otherwise.

    Returns:
      A tuple of two tensors:
        1. output tensor for the current timestep, with size `output_size`.
        2. state tensor for next step, which has the shape of `state_size`.
    """
    raise NotImplementedError('Abstract method')

  @property
  def state_size(self):
    """size(s) of state(s) used by this cell.

    It can be represented by an Integer, a TensorShape or a tuple of Integers
    or TensorShapes.
    """
    raise NotImplementedError('Abstract method')

  @property
  def output_size(self):
    """Integer or TensorShape: size of outputs produced by this cell."""
    raise NotImplementedError('Abstract method')

  def get_initial_state(self, inputs=None, batch_size=None, dtype=None):
    return _generate_zero_filled_state_for_cell(self, inputs, batch_size, dtype)


@doc_controls.do_not_generate_docs
class DropoutRNNCellMixin(object):
  """Object that hold dropout related fields for RNN Cell.

  This class is not a standalone RNN cell. It suppose to be used with a RNN cell
  by multiple inheritance. Any cell that mix with class should have following
  fields:
    dropout: a float number within range [0, 1). The ratio that the input
      tensor need to dropout.
    recurrent_dropout: a float number within range [0, 1). The ratio that the
      recurrent state weights need to dropout.
  This object will create and cache created dropout masks, and reuse them for
  the incoming data, so that the same mask is used for every batch input.
  """

  def __init__(self, *args, **kwargs):
    self._create_non_trackable_mask_cache()
    super(DropoutRNNCellMixin, self).__init__(*args, **kwargs)

  @trackable.no_automatic_dependency_tracking
  def _create_non_trackable_mask_cache(self):
    """Create the cache for dropout and recurrent dropout mask.

    Note that the following two masks will be used in "graph function" mode,
    e.g. these masks are symbolic tensors. In eager mode, the `eager_*_mask`
    tensors will be generated differently than in the "graph function" case,
    and they will be cached.

    Also note that in graph mode, we still cache those masks only because the
    RNN could be created with `unroll=True`. In that case, the `cell.call()`
    function will be invoked multiple times, and we want to ensure same mask
    is used every time.

    Also the caches are created without tracking. Since they are not picklable
    by python when deepcopy, we don't want layer._obj_reference_counts_dict
    to track it by default.
    """
    self._dropout_mask_cache = K.ContextValueCache(self._create_dropout_mask)
    self._recurrent_dropout_mask_cache = K.ContextValueCache(
        self._create_recurrent_dropout_mask)

  def reset_dropout_mask(self):
    """Reset the cached dropout masks if any.

    This is important for the RNN layer to invoke this in it call() method so
    that the cached mask is cleared before calling the cell.call(). The mask
    should be cached across the timestep within the same batch, but shouldn't
    be cached between batches. Otherwise it will introduce unreasonable bias
    against certain index of data within the batch.
    """
    self._dropout_mask_cache.clear()

  def reset_recurrent_dropout_mask(self):
    """Reset the cached recurrent dropout masks if any.

    This is important for the RNN layer to invoke this in it call() method so
    that the cached mask is cleared before calling the cell.call(). The mask
    should be cached across the timestep within the same batch, but shouldn't
    be cached between batches. Otherwise it will introduce unreasonable bias
    against certain index of data within the batch.
    """
    self._recurrent_dropout_mask_cache.clear()

  def _create_dropout_mask(self, inputs, training, count=1):
    return _generate_dropout_mask(
        array_ops.ones_like(inputs),
        self.dropout,
        training=training,
        count=count)

  def _create_recurrent_dropout_mask(self, inputs, training, count=1):
    return _generate_dropout_mask(
        array_ops.ones_like(inputs),
        self.recurrent_dropout,
        training=training,
        count=count)

  def get_dropout_mask_for_cell(self, inputs, training, count=1):
    """Get the dropout mask for RNN cell's input.

    It will create mask based on context if there isn't any existing cached
    mask. If a new mask is generated, it will update the cache in the cell.

    Args:
      inputs: The input tensor whose shape will be used to generate dropout
        mask.
      training: Boolean tensor, whether its in training mode, dropout will be
        ignored in non-training mode.
      count: Int, how many dropout mask will be generated. It is useful for cell
        that has internal weights fused together.
    Returns:
      List of mask tensor, generated or cached mask based on context.
    """
    if self.dropout == 0:
      return None
    init_kwargs = dict(inputs=inputs, training=training, count=count)
    return self._dropout_mask_cache.setdefault(kwargs=init_kwargs)

  def get_recurrent_dropout_mask_for_cell(self, inputs, training, count=1):
    """Get the recurrent dropout mask for RNN cell.

    It will create mask based on context if there isn't any existing cached
    mask. If a new mask is generated, it will update the cache in the cell.

    Args:
      inputs: The input tensor whose shape will be used to generate dropout
        mask.
      training: Boolean tensor, whether its in training mode, dropout will be
        ignored in non-training mode.
      count: Int, how many dropout mask will be generated. It is useful for cell
        that has internal weights fused together.
    Returns:
      List of mask tensor, generated or cached mask based on context.
    """
    if self.recurrent_dropout == 0:
      return None
    init_kwargs = dict(inputs=inputs, training=training, count=count)
    return self._recurrent_dropout_mask_cache.setdefault(kwargs=init_kwargs)

  def __getstate__(self):
    # Used for deepcopy. The caching can't be pickled by python, since it will
    # contain tensor and graph.
    state = super(DropoutRNNCellMixin, self).__getstate__()
    state.pop('_dropout_mask_cache', None)
    state.pop('_recurrent_dropout_mask_cache', None)
    return state

  def __setstate__(self, state):
    state['_dropout_mask_cache'] = K.ContextValueCache(
        self._create_dropout_mask)
    state['_recurrent_dropout_mask_cache'] = K.ContextValueCache(
        self._create_recurrent_dropout_mask)
    super(DropoutRNNCellMixin, self).__setstate__(state)


@keras_export('keras.layers.SimpleRNNCell')
class SimpleRNNCell(DropoutRNNCellMixin, Layer):
  """Cell class for SimpleRNN.

  See [the Keras RNN API guide](https://www.tensorflow.org/guide/keras/rnn)
  for details about the usage of RNN API.

  This class processes one step within the whole time sequence input, whereas
  `tf.keras.layer.SimpleRNN` processes the whole sequence.

  Arguments:
    units: Positive integer, dimensionality of the output space.
    activation: Activation function to use.
      Default: hyperbolic tangent (`tanh`).
      If you pass `None`, no activation is applied
      (ie. "linear" activation: `a(x) = x`).
    use_bias: Boolean, (default `True`), whether the layer uses a bias vector.
    kernel_initializer: Initializer for the `kernel` weights matrix,
      used for the linear transformation of the inputs. Default:
      `glorot_uniform`.
    recurrent_initializer: Initializer for the `recurrent_kernel`
      weights matrix, used for the linear transformation of the recurrent state.
      Default: `orthogonal`.
    bias_initializer: Initializer for the bias vector. Default: `zeros`.
    kernel_regularizer: Regularizer function applied to the `kernel` weights
      matrix. Default: `None`.
    recurrent_regularizer: Regularizer function applied to the
      `recurrent_kernel` weights matrix. Default: `None`.
    bias_regularizer: Regularizer function applied to the bias vector. Default:
      `None`.
    kernel_constraint: Constraint function applied to the `kernel` weights
      matrix. Default: `None`.
    recurrent_constraint: Constraint function applied to the `recurrent_kernel`
      weights matrix. Default: `None`.
    bias_constraint: Constraint function applied to the bias vector. Default:
      `None`.
    dropout: Float between 0 and 1. Fraction of the units to drop for the linear
      transformation of the inputs. Default: 0.
    recurrent_dropout: Float between 0 and 1. Fraction of the units to drop for
      the linear transformation of the recurrent state. Default: 0.

  Call arguments:
    inputs: A 2D tensor, with shape of `[batch, feature]`.
    states: A 2D tensor with shape of `[batch, units]`, which is the state from
      the previous time step. For timestep 0, the initial state provided by user
      will be feed to cell.
    training: Python boolean indicating whether the layer should behave in
      training mode or in inference mode. Only relevant when `dropout` or
      `recurrent_dropout` is used.

  Examples:

  ```python
  inputs = np.random.random([32, 10, 8]).astype(np.float32)
  rnn = tf.keras.layers.RNN(tf.keras.layers.SimpleRNNCell(4))

  output = rnn(inputs)  # The output has shape `[32, 4]`.

  rnn = tf.keras.layers.RNN(
      tf.keras.layers.SimpleRNNCell(4),
      return_sequences=True,
      return_state=True)

  # whole_sequence_output has shape `[32, 10, 4]`.
  # final_state has shape `[32, 4]`.
  whole_sequence_output, final_state = rnn(inputs)
  ```
  """

  def __init__(self,
               units,
               activation='tanh',
               use_bias=True,
               kernel_initializer='glorot_uniform',
               recurrent_initializer='orthogonal',
               bias_initializer='zeros',
               kernel_regularizer=None,
               recurrent_regularizer=None,
               bias_regularizer=None,
               kernel_constraint=None,
               recurrent_constraint=None,
               bias_constraint=None,
               dropout=0.,
               recurrent_dropout=0.,
               **kwargs):
    # By default use cached variable under v2 mode, see b/143699808.
    if ops.executing_eagerly_outside_functions():
      self._enable_caching_device = kwargs.pop('enable_caching_device', True)
    else:
      self._enable_caching_device = kwargs.pop('enable_caching_device', False)
    super(SimpleRNNCell, self).__init__(**kwargs)
    self.units = units
    self.activation = activations.get(activation)
    self.use_bias = use_bias

    self.kernel_initializer = initializers.get(kernel_initializer)
    self.recurrent_initializer = initializers.get(recurrent_initializer)
    self.bias_initializer = initializers.get(bias_initializer)

    self.kernel_regularizer = regularizers.get(kernel_regularizer)
    self.recurrent_regularizer = regularizers.get(recurrent_regularizer)
    self.bias_regularizer = regularizers.get(bias_regularizer)

    self.kernel_constraint = constraints.get(kernel_constraint)
    self.recurrent_constraint = constraints.get(recurrent_constraint)
    self.bias_constraint = constraints.get(bias_constraint)

    self.dropout = min(1., max(0., dropout))
    self.recurrent_dropout = min(1., max(0., recurrent_dropout))
    self.state_size = self.units
    self.output_size = self.units

  @tf_utils.shape_type_conversion
  def build(self, input_shape):
    default_caching_device = _caching_device(self)
    self.kernel = self.add_weight(
        shape=(input_shape[-1], self.units),
        name='kernel',
        initializer=self.kernel_initializer,
        regularizer=self.kernel_regularizer,
        constraint=self.kernel_constraint,
        caching_device=default_caching_device)
    self.recurrent_kernel = self.add_weight(
        shape=(self.units, self.units),
        name='recurrent_kernel',
        initializer=self.recurrent_initializer,
        regularizer=self.recurrent_regularizer,
        constraint=self.recurrent_constraint,
        caching_device=default_caching_device)
    if self.use_bias:
      self.bias = self.add_weight(
          shape=(self.units,),
          name='bias',
          initializer=self.bias_initializer,
          regularizer=self.bias_regularizer,
          constraint=self.bias_constraint,
          caching_device=default_caching_device)
    else:
      self.bias = None
    self.built = True

  def call(self, inputs, states, training=None):
    prev_output = states[0] if nest.is_sequence(states) else states
    dp_mask = self.get_dropout_mask_for_cell(inputs, training)
    rec_dp_mask = self.get_recurrent_dropout_mask_for_cell(
        prev_output, training)

    if dp_mask is not None:
      h = K.dot(inputs * dp_mask, self.kernel)
    else:
      h = K.dot(inputs, self.kernel)
    if self.bias is not None:
      h = K.bias_add(h, self.bias)

    if rec_dp_mask is not None:
      prev_output = prev_output * rec_dp_mask
    output = h + K.dot(prev_output, self.recurrent_kernel)
    if self.activation is not None:
      output = self.activation(output)

    new_state = [output] if nest.is_sequence(states) else output
    return output, new_state

  def get_initial_state(self, inputs=None, batch_size=None, dtype=None):
    return _generate_zero_filled_state_for_cell(self, inputs, batch_size, dtype)

  def get_config(self):
    config = {
        'units':
            self.units,
        'activation':
            activations.serialize(self.activation),
        'use_bias':
            self.use_bias,
        'kernel_initializer':
            initializers.serialize(self.kernel_initializer),
        'recurrent_initializer':
            initializers.serialize(self.recurrent_initializer),
        'bias_initializer':
            initializers.serialize(self.bias_initializer),
        'kernel_regularizer':
            regularizers.serialize(self.kernel_regularizer),
        'recurrent_regularizer':
            regularizers.serialize(self.recurrent_regularizer),
        'bias_regularizer':
            regularizers.serialize(self.bias_regularizer),
        'kernel_constraint':
            constraints.serialize(self.kernel_constraint),
        'recurrent_constraint':
            constraints.serialize(self.recurrent_constraint),
        'bias_constraint':
            constraints.serialize(self.bias_constraint),
        'dropout':
            self.dropout,
        'recurrent_dropout':
            self.recurrent_dropout
    }
    config.update(_config_for_enable_caching_device(self))
    base_config = super(SimpleRNNCell, self).get_config()
    return dict(list(base_config.items()) + list(config.items()))


@keras_export('keras.layers.SimpleRNN')
class SimpleRNN(RNN):
  """Fully-connected RNN where the output is to be fed back to input.

  See [the Keras RNN API guide](https://www.tensorflow.org/guide/keras/rnn)
  for details about the usage of RNN API.

  Arguments:
    units: Positive integer, dimensionality of the output space.
    activation: Activation function to use.
      Default: hyperbolic tangent (`tanh`).
      If you pass None, no activation is applied
      (ie. "linear" activation: `a(x) = x`).
    use_bias: Boolean, (default `True`), whether the layer uses a bias vector.
    kernel_initializer: Initializer for the `kernel` weights matrix,
      used for the linear transformation of the inputs. Default:
      `glorot_uniform`.
    recurrent_initializer: Initializer for the `recurrent_kernel`
      weights matrix, used for the linear transformation of the recurrent state.
      Default: `orthogonal`.
    bias_initializer: Initializer for the bias vector. Default: `zeros`.
    kernel_regularizer: Regularizer function applied to the `kernel` weights
      matrix. Default: `None`.
    recurrent_regularizer: Regularizer function applied to the
      `recurrent_kernel` weights matrix. Default: `None`.
    bias_regularizer: Regularizer function applied to the bias vector. Default:
      `None`.
    activity_regularizer: Regularizer function applied to the output of the
      layer (its "activation"). Default: `None`.
    kernel_constraint: Constraint function applied to the `kernel` weights
      matrix. Default: `None`.
    recurrent_constraint: Constraint function applied to the `recurrent_kernel`
      weights matrix.  Default: `None`.
    bias_constraint: Constraint function applied to the bias vector. Default:
      `None`.
    dropout: Float between 0 and 1.
      Fraction of the units to drop for the linear transformation of the inputs.
      Default: 0.
    recurrent_dropout: Float between 0 and 1.
      Fraction of the units to drop for the linear transformation of the
      recurrent state. Default: 0.
    return_sequences: Boolean. Whether to return the last output
      in the output sequence, or the full sequence. Default: `False`.
    return_state: Boolean. Whether to return the last state
      in addition to the output. Default: `False`
    go_backwards: Boolean (default False).
      If True, process the input sequence backwards and return the
      reversed sequence.
    stateful: Boolean (default False). If True, the last state
      for each sample at index i in a batch will be used as initial
      state for the sample of index i in the following batch.
    unroll: Boolean (default False).
      If True, the network will be unrolled,
      else a symbolic loop will be used.
      Unrolling can speed-up a RNN,
      although it tends to be more memory-intensive.
      Unrolling is only suitable for short sequences.

  Call arguments:
    inputs: A 3D tensor, with shape `[batch, timesteps, feature]`.
    mask: Binary tensor of shape `[batch, timesteps]` indicating whether
      a given timestep should be masked.
    training: Python boolean indicating whether the layer should behave in
      training mode or in inference mode. This argument is passed to the cell
      when calling it. This is only relevant if `dropout` or
      `recurrent_dropout` is used.
    initial_state: List of initial state tensors to be passed to the first
      call of the cell.

  Examples:

  ```python
  inputs = np.random.random([32, 10, 8]).astype(np.float32)
  simple_rnn = tf.keras.layers.SimpleRNN(4)

  output = simple_rnn(inputs)  # The output has shape `[32, 4]`.

  simple_rnn = tf.keras.layers.SimpleRNN(
      4, return_sequences=True, return_state=True)

  # whole_sequence_output has shape `[32, 10, 4]`.
  # final_state has shape `[32, 4]`.
  whole_sequence_output, final_state = simple_rnn(inputs)
  ```
  """

  def __init__(self,
               units,
               activation='tanh',
               use_bias=True,
               kernel_initializer='glorot_uniform',
               recurrent_initializer='orthogonal',
               bias_initializer='zeros',
               kernel_regularizer=None,
               recurrent_regularizer=None,
               bias_regularizer=None,
               activity_regularizer=None,
               kernel_constraint=None,
               recurrent_constraint=None,
               bias_constraint=None,
               dropout=0.,
               recurrent_dropout=0.,
               return_sequences=False,
               return_state=False,
               go_backwards=False,
               stateful=False,
               unroll=False,
               **kwargs):
    if 'implementation' in kwargs:
      kwargs.pop('implementation')
      logging.warning('The `implementation` argument '
                      'in `SimpleRNN` has been deprecated. '
                      'Please remove it from your layer call.')
    if 'enable_caching_device' in kwargs:
      cell_kwargs = {'enable_caching_device':
                     kwargs.pop('enable_caching_device')}
    else:
      cell_kwargs = {}
    cell = SimpleRNNCell(
        units,
        activation=activation,
        use_bias=use_bias,
        kernel_initializer=kernel_initializer,
        recurrent_initializer=recurrent_initializer,
        bias_initializer=bias_initializer,
        kernel_regularizer=kernel_regularizer,
        recurrent_regularizer=recurrent_regularizer,
        bias_regularizer=bias_regularizer,
        kernel_constraint=kernel_constraint,
        recurrent_constraint=recurrent_constraint,
        bias_constraint=bias_constraint,
        dropout=dropout,
        recurrent_dropout=recurrent_dropout,
        dtype=kwargs.get('dtype'),
        trainable=kwargs.get('trainable', True),
        **cell_kwargs)
    super(SimpleRNN, self).__init__(
        cell,
        return_sequences=return_sequences,
        return_state=return_state,
        go_backwards=go_backwards,
        stateful=stateful,
        unroll=unroll,
        **kwargs)
    self.activity_regularizer = regularizers.get(activity_regularizer)
    self.input_spec = [InputSpec(ndim=3)]

  def call(self, inputs, mask=None, training=None, initial_state=None):
    self._maybe_reset_cell_dropout_mask(self.cell)
    return super(SimpleRNN, self).call(
        inputs, mask=mask, training=training, initial_state=initial_state)

  @property
  def units(self):
    return self.cell.units

  @property
  def activation(self):
    return self.cell.activation

  @property
  def use_bias(self):
    return self.cell.use_bias

  @property
  def kernel_initializer(self):
    return self.cell.kernel_initializer

  @property
  def recurrent_initializer(self):
    return self.cell.recurrent_initializer

  @property
  def bias_initializer(self):
    return self.cell.bias_initializer

  @property
  def kernel_regularizer(self):
    return self.cell.kernel_regularizer

  @property
  def recurrent_regularizer(self):
    return self.cell.recurrent_regularizer

  @property
  def bias_regularizer(self):
    return self.cell.bias_regularizer

  @property
  def kernel_constraint(self):
    return self.cell.kernel_constraint

  @property
  def recurrent_constraint(self):
    return self.cell.recurrent_constraint

  @property
  def bias_constraint(self):
    return self.cell.bias_constraint

  @property
  def dropout(self):
    return self.cell.dropout

  @property
  def recurrent_dropout(self):
    return self.cell.recurrent_dropout

  def get_config(self):
    config = {
        'units':
            self.units,
        'activation':
            activations.serialize(self.activation),
        'use_bias':
            self.use_bias,
        'kernel_initializer':
            initializers.serialize(self.kernel_initializer),
        'recurrent_initializer':
            initializers.serialize(self.recurrent_initializer),
        'bias_initializer':
            initializers.serialize(self.bias_initializer),
        'kernel_regularizer':
            regularizers.serialize(self.kernel_regularizer),
        'recurrent_regularizer':
            regularizers.serialize(self.recurrent_regularizer),
        'bias_regularizer':
            regularizers.serialize(self.bias_regularizer),
        'activity_regularizer':
            regularizers.serialize(self.activity_regularizer),
        'kernel_constraint':
            constraints.serialize(self.kernel_constraint),
        'recurrent_constraint':
            constraints.serialize(self.recurrent_constraint),
        'bias_constraint':
            constraints.serialize(self.bias_constraint),
        'dropout':
            self.dropout,
        'recurrent_dropout':
            self.recurrent_dropout
    }
    base_config = super(SimpleRNN, self).get_config()
    config.update(_config_for_enable_caching_device(self.cell))
    del base_config['cell']
    return dict(list(base_config.items()) + list(config.items()))

  @classmethod
  def from_config(cls, config):
    if 'implementation' in config:
      config.pop('implementation')
    return cls(**config)


@keras_export(v1=['keras.layers.GRUCell'])
class GRUCell(DropoutRNNCellMixin, Layer):
  """Cell class for the GRU layer.

  Arguments:
    units: Positive integer, dimensionality of the output space.
    activation: Activation function to use.
      Default: hyperbolic tangent (`tanh`).
      If you pass None, no activation is applied
      (ie. "linear" activation: `a(x) = x`).
    recurrent_activation: Activation function to use
      for the recurrent step.
      Default: hard sigmoid (`hard_sigmoid`).
      If you pass `None`, no activation is applied
      (ie. "linear" activation: `a(x) = x`).
    use_bias: Boolean, whether the layer uses a bias vector.
    kernel_initializer: Initializer for the `kernel` weights matrix,
      used for the linear transformation of the inputs.
    recurrent_initializer: Initializer for the `recurrent_kernel`
      weights matrix,
      used for the linear transformation of the recurrent state.
    bias_initializer: Initializer for the bias vector.
    kernel_regularizer: Regularizer function applied to
      the `kernel` weights matrix.
    recurrent_regularizer: Regularizer function applied to
      the `recurrent_kernel` weights matrix.
    bias_regularizer: Regularizer function applied to the bias vector.
    kernel_constraint: Constraint function applied to
      the `kernel` weights matrix.
    recurrent_constraint: Constraint function applied to
      the `recurrent_kernel` weights matrix.
    bias_constraint: Constraint function applied to the bias vector.
    dropout: Float between 0 and 1.
      Fraction of the units to drop for the linear transformation of the inputs.
    recurrent_dropout: Float between 0 and 1.
      Fraction of the units to drop for
      the linear transformation of the recurrent state.
    implementation: Implementation mode, either 1 or 2.
      Mode 1 will structure its operations as a larger number of
      smaller dot products and additions, whereas mode 2 will
      batch them into fewer, larger operations. These modes will
      have different performance profiles on different hardware and
      for different applications.
    reset_after: GRU convention (whether to apply reset gate after or
      before matrix multiplication). False = "before" (default),
      True = "after" (CuDNN compatible).

  Call arguments:
    inputs: A 2D tensor.
    states: List of state tensors corresponding to the previous timestep.
    training: Python boolean indicating whether the layer should behave in
      training mode or in inference mode. Only relevant when `dropout` or
      `recurrent_dropout` is used.
  """

  def __init__(self,
               units,
               activation='tanh',
               recurrent_activation='hard_sigmoid',
               use_bias=True,
               kernel_initializer='glorot_uniform',
               recurrent_initializer='orthogonal',
               bias_initializer='zeros',
               kernel_regularizer=None,
               recurrent_regularizer=None,
               bias_regularizer=None,
               kernel_constraint=None,
               recurrent_constraint=None,
               bias_constraint=None,
               dropout=0.,
               recurrent_dropout=0.,
               implementation=1,
               reset_after=False,
               **kwargs):
    # By default use cached variable under v2 mode, see b/143699808.
    if ops.executing_eagerly_outside_functions():
      self._enable_caching_device = kwargs.pop('enable_caching_device', True)
    else:
      self._enable_caching_device = kwargs.pop('enable_caching_device', False)
    super(GRUCell, self).__init__(**kwargs)
    self.units = units
    self.activation = activations.get(activation)
    self.recurrent_activation = activations.get(recurrent_activation)
    self.use_bias = use_bias

    self.kernel_initializer = initializers.get(kernel_initializer)
    self.recurrent_initializer = initializers.get(recurrent_initializer)
    self.bias_initializer = initializers.get(bias_initializer)

    self.kernel_regularizer = regularizers.get(kernel_regularizer)
    self.recurrent_regularizer = regularizers.get(recurrent_regularizer)
    self.bias_regularizer = regularizers.get(bias_regularizer)

    self.kernel_constraint = constraints.get(kernel_constraint)
    self.recurrent_constraint = constraints.get(recurrent_constraint)
    self.bias_constraint = constraints.get(bias_constraint)

    self.dropout = min(1., max(0., dropout))
    self.recurrent_dropout = min(1., max(0., recurrent_dropout))
    if self.recurrent_dropout != 0 and implementation != 1:
      logging.debug(RECURRENT_DROPOUT_WARNING_MSG)
      self.implementation = 1
    else:
      self.implementation = implementation
    self.reset_after = reset_after
    self.state_size = self.units
    self.output_size = self.units

  @tf_utils.shape_type_conversion
  def build(self, input_shape):
    input_dim = input_shape[-1]
    default_caching_device = _caching_device(self)
    self.kernel = self.add_weight(
        shape=(input_dim, self.units * 3),
        name='kernel',
        initializer=self.kernel_initializer,
        regularizer=self.kernel_regularizer,
        constraint=self.kernel_constraint,
        caching_device=default_caching_device)
    self.recurrent_kernel = self.add_weight(
        shape=(self.units, self.units * 3),
        name='recurrent_kernel',
        initializer=self.recurrent_initializer,
        regularizer=self.recurrent_regularizer,
        constraint=self.recurrent_constraint,
        caching_device=default_caching_device)

    if self.use_bias:
      if not self.reset_after:
        bias_shape = (3 * self.units,)
      else:
        # separate biases for input and recurrent kernels
        # Note: the shape is intentionally different from CuDNNGRU biases
        # `(2 * 3 * self.units,)`, so that we can distinguish the classes
        # when loading and converting saved weights.
        bias_shape = (2, 3 * self.units)
      self.bias = self.add_weight(shape=bias_shape,
                                  name='bias',
                                  initializer=self.bias_initializer,
                                  regularizer=self.bias_regularizer,
                                  constraint=self.bias_constraint,
                                  caching_device=default_caching_device)
    else:
      self.bias = None
    self.built = True

  def call(self, inputs, states, training=None):
    h_tm1 = states[0] if nest.is_sequence(states) else states  # previous memory

    dp_mask = self.get_dropout_mask_for_cell(inputs, training, count=3)
    rec_dp_mask = self.get_recurrent_dropout_mask_for_cell(
        h_tm1, training, count=3)

    if self.use_bias:
      if not self.reset_after:
        input_bias, recurrent_bias = self.bias, None
      else:
        input_bias, recurrent_bias = array_ops.unstack(self.bias)

    if self.implementation == 1:
      if 0. < self.dropout < 1.:
        inputs_z = inputs * dp_mask[0]
        inputs_r = inputs * dp_mask[1]
        inputs_h = inputs * dp_mask[2]
      else:
        inputs_z = inputs
        inputs_r = inputs
        inputs_h = inputs

      x_z = K.dot(inputs_z, self.kernel[:, :self.units])
      x_r = K.dot(inputs_r, self.kernel[:, self.units:self.units * 2])
      x_h = K.dot(inputs_h, self.kernel[:, self.units * 2:])

      if self.use_bias:
        x_z = K.bias_add(x_z, input_bias[:self.units])
        x_r = K.bias_add(x_r, input_bias[self.units: self.units * 2])
        x_h = K.bias_add(x_h, input_bias[self.units * 2:])

      if 0. < self.recurrent_dropout < 1.:
        h_tm1_z = h_tm1 * rec_dp_mask[0]
        h_tm1_r = h_tm1 * rec_dp_mask[1]
        h_tm1_h = h_tm1 * rec_dp_mask[2]
      else:
        h_tm1_z = h_tm1
        h_tm1_r = h_tm1
        h_tm1_h = h_tm1

      recurrent_z = K.dot(h_tm1_z, self.recurrent_kernel[:, :self.units])
      recurrent_r = K.dot(h_tm1_r,
                          self.recurrent_kernel[:, self.units:self.units * 2])
      if self.reset_after and self.use_bias:
        recurrent_z = K.bias_add(recurrent_z, recurrent_bias[:self.units])
        recurrent_r = K.bias_add(recurrent_r,
                                 recurrent_bias[self.units:self.units * 2])

      z = self.recurrent_activation(x_z + recurrent_z)
      r = self.recurrent_activation(x_r + recurrent_r)

      # reset gate applied after/before matrix multiplication
      if self.reset_after:
        recurrent_h = K.dot(h_tm1_h, self.recurrent_kernel[:, self.units * 2:])
        if self.use_bias:
          recurrent_h = K.bias_add(recurrent_h, recurrent_bias[self.units * 2:])
        recurrent_h = r * recurrent_h
      else:
        recurrent_h = K.dot(r * h_tm1_h,
                            self.recurrent_kernel[:, self.units * 2:])

      hh = self.activation(x_h + recurrent_h)
    else:
      if 0. < self.dropout < 1.:
        inputs = inputs * dp_mask[0]

      # inputs projected by all gate matrices at once
      matrix_x = K.dot(inputs, self.kernel)
      if self.use_bias:
        # biases: bias_z_i, bias_r_i, bias_h_i
        matrix_x = K.bias_add(matrix_x, input_bias)

      x_z, x_r, x_h = array_ops.split(matrix_x, 3, axis=-1)

      if self.reset_after:
        # hidden state projected by all gate matrices at once
        matrix_inner = K.dot(h_tm1, self.recurrent_kernel)
        if self.use_bias:
          matrix_inner = K.bias_add(matrix_inner, recurrent_bias)
      else:
        # hidden state projected separately for update/reset and new
        matrix_inner = K.dot(h_tm1, self.recurrent_kernel[:, :2 * self.units])

      recurrent_z, recurrent_r, recurrent_h = array_ops.split(
          matrix_inner, [self.units, self.units, -1], axis=-1)

      z = self.recurrent_activation(x_z + recurrent_z)
      r = self.recurrent_activation(x_r + recurrent_r)

      if self.reset_after:
        recurrent_h = r * recurrent_h
      else:
        recurrent_h = K.dot(r * h_tm1,
                            self.recurrent_kernel[:, 2 * self.units:])

      hh = self.activation(x_h + recurrent_h)
    # previous and candidate state mixed by update gate
    h = z * h_tm1 + (1 - z) * hh
    new_state = [h] if nest.is_sequence(states) else h
    return h, new_state

  def get_config(self):
    config = {
        'units': self.units,
        'activation': activations.serialize(self.activation),
        'recurrent_activation':
            activations.serialize(self.recurrent_activation),
        'use_bias': self.use_bias,
        'kernel_initializer': initializers.serialize(self.kernel_initializer),
        'recurrent_initializer':
            initializers.serialize(self.recurrent_initializer),
        'bias_initializer': initializers.serialize(self.bias_initializer),
        'kernel_regularizer': regularizers.serialize(self.kernel_regularizer),
        'recurrent_regularizer':
            regularizers.serialize(self.recurrent_regularizer),
        'bias_regularizer': regularizers.serialize(self.bias_regularizer),
        'kernel_constraint': constraints.serialize(self.kernel_constraint),
        'recurrent_constraint':
            constraints.serialize(self.recurrent_constraint),
        'bias_constraint': constraints.serialize(self.bias_constraint),
        'dropout': self.dropout,
        'recurrent_dropout': self.recurrent_dropout,
        'implementation': self.implementation,
        'reset_after': self.reset_after
    }
    config.update(_config_for_enable_caching_device(self))
    base_config = super(GRUCell, self).get_config()
    return dict(list(base_config.items()) + list(config.items()))

  def get_initial_state(self, inputs=None, batch_size=None, dtype=None):
    return _generate_zero_filled_state_for_cell(self, inputs, batch_size, dtype)


@keras_export(v1=['keras.layers.GRU'])
class GRU(RNN):
  """Gated Recurrent Unit - Cho et al. 2014.

  There are two variants. The default one is based on 1406.1078v3 and
  has reset gate applied to hidden state before matrix multiplication. The
  other one is based on original 1406.1078v1 and has the order reversed.

  The second variant is compatible with CuDNNGRU (GPU-only) and allows
  inference on CPU. Thus it has separate biases for `kernel` and
  `recurrent_kernel`. Use `'reset_after'=True` and
  `recurrent_activation='sigmoid'`.

  Arguments:
    units: Positive integer, dimensionality of the output space.
    activation: Activation function to use.
      Default: hyperbolic tangent (`tanh`).
      If you pass `None`, no activation is applied
      (ie. "linear" activation: `a(x) = x`).
    recurrent_activation: Activation function to use
      for the recurrent step.
      Default: hard sigmoid (`hard_sigmoid`).
      If you pass `None`, no activation is applied
      (ie. "linear" activation: `a(x) = x`).
    use_bias: Boolean, whether the layer uses a bias vector.
    kernel_initializer: Initializer for the `kernel` weights matrix,
      used for the linear transformation of the inputs.
    recurrent_initializer: Initializer for the `recurrent_kernel`
      weights matrix, used for the linear transformation of the recurrent state.
    bias_initializer: Initializer for the bias vector.
    kernel_regularizer: Regularizer function applied to
      the `kernel` weights matrix.
    recurrent_regularizer: Regularizer function applied to
      the `recurrent_kernel` weights matrix.
    bias_regularizer: Regularizer function applied to the bias vector.
    activity_regularizer: Regularizer function applied to
      the output of the layer (its "activation")..
    kernel_constraint: Constraint function applied to
      the `kernel` weights matrix.
    recurrent_constraint: Constraint function applied to
      the `recurrent_kernel` weights matrix.
    bias_constraint: Constraint function applied to the bias vector.
    dropout: Float between 0 and 1.
      Fraction of the units to drop for
      the linear transformation of the inputs.
    recurrent_dropout: Float between 0 and 1.
      Fraction of the units to drop for
      the linear transformation of the recurrent state.
    implementation: Implementation mode, either 1 or 2.
      Mode 1 will structure its operations as a larger number of
      smaller dot products and additions, whereas mode 2 will
      batch them into fewer, larger operations. These modes will
      have different performance profiles on different hardware and
      for different applications.
    return_sequences: Boolean. Whether to return the last output
      in the output sequence, or the full sequence.
    return_state: Boolean. Whether to return the last state
      in addition to the output.
    go_backwards: Boolean (default False).
      If True, process the input sequence backwards and return the
      reversed sequence.
    stateful: Boolean (default False). If True, the last state
      for each sample at index i in a batch will be used as initial
      state for the sample of index i in the following batch.
    unroll: Boolean (default False).
      If True, the network will be unrolled,
      else a symbolic loop will be used.
      Unrolling can speed-up a RNN,
      although it tends to be more memory-intensive.
      Unrolling is only suitable for short sequences.
    time_major: The shape format of the `inputs` and `outputs` tensors.
      If True, the inputs and outputs will be in shape
      `(timesteps, batch, ...)`, whereas in the False case, it will be
      `(batch, timesteps, ...)`. Using `time_major = True` is a bit more
      efficient because it avoids transposes at the beginning and end of the
      RNN calculation. However, most TensorFlow data is batch-major, so by
      default this function accepts input and emits output in batch-major
      form.
    reset_after: GRU convention (whether to apply reset gate after or
      before matrix multiplication). False = "before" (default),
      True = "after" (CuDNN compatible).

  Call arguments:
    inputs: A 3D tensor.
    mask: Binary tensor of shape `(samples, timesteps)` indicating whether
      a given timestep should be masked.
    training: Python boolean indicating whether the layer should behave in
      training mode or in inference mode. This argument is passed to the cell
      when calling it. This is only relevant if `dropout` or
      `recurrent_dropout` is used.
    initial_state: List of initial state tensors to be passed to the first
      call of the cell.
  """

  def __init__(self,
               units,
               activation='tanh',
               recurrent_activation='hard_sigmoid',
               use_bias=True,
               kernel_initializer='glorot_uniform',
               recurrent_initializer='orthogonal',
               bias_initializer='zeros',
               kernel_regularizer=None,
               recurrent_regularizer=None,
               bias_regularizer=None,
               activity_regularizer=None,
               kernel_constraint=None,
               recurrent_constraint=None,
               bias_constraint=None,
               dropout=0.,
               recurrent_dropout=0.,
               implementation=1,
               return_sequences=False,
               return_state=False,
               go_backwards=False,
               stateful=False,
               unroll=False,
               reset_after=False,
               **kwargs):
    if implementation == 0:
      logging.warning('`implementation=0` has been deprecated, '
                      'and now defaults to `implementation=1`.'
                      'Please update your layer call.')
    if 'enable_caching_device' in kwargs:
      cell_kwargs = {'enable_caching_device':
                     kwargs.pop('enable_caching_device')}
    else:
      cell_kwargs = {}
    cell = GRUCell(
        units,
        activation=activation,
        recurrent_activation=recurrent_activation,
        use_bias=use_bias,
        kernel_initializer=kernel_initializer,
        recurrent_initializer=recurrent_initializer,
        bias_initializer=bias_initializer,
        kernel_regularizer=kernel_regularizer,
        recurrent_regularizer=recurrent_regularizer,
        bias_regularizer=bias_regularizer,
        kernel_constraint=kernel_constraint,
        recurrent_constraint=recurrent_constraint,
        bias_constraint=bias_constraint,
        dropout=dropout,
        recurrent_dropout=recurrent_dropout,
        implementation=implementation,
        reset_after=reset_after,
        dtype=kwargs.get('dtype'),
        trainable=kwargs.get('trainable', True),
        **cell_kwargs)
    super(GRU, self).__init__(
        cell,
        return_sequences=return_sequences,
        return_state=return_state,
        go_backwards=go_backwards,
        stateful=stateful,
        unroll=unroll,
        **kwargs)
    self.activity_regularizer = regularizers.get(activity_regularizer)
    self.input_spec = [InputSpec(ndim=3)]

  def call(self, inputs, mask=None, training=None, initial_state=None):
    self._maybe_reset_cell_dropout_mask(self.cell)
    return super(GRU, self).call(
        inputs, mask=mask, training=training, initial_state=initial_state)

  @property
  def units(self):
    return self.cell.units

  @property
  def activation(self):
    return self.cell.activation

  @property
  def recurrent_activation(self):
    return self.cell.recurrent_activation

  @property
  def use_bias(self):
    return self.cell.use_bias

  @property
  def kernel_initializer(self):
    return self.cell.kernel_initializer

  @property
  def recurrent_initializer(self):
    return self.cell.recurrent_initializer

  @property
  def bias_initializer(self):
    return self.cell.bias_initializer

  @property
  def kernel_regularizer(self):
    return self.cell.kernel_regularizer

  @property
  def recurrent_regularizer(self):
    return self.cell.recurrent_regularizer

  @property
  def bias_regularizer(self):
    return self.cell.bias_regularizer

  @property
  def kernel_constraint(self):
    return self.cell.kernel_constraint

  @property
  def recurrent_constraint(self):
    return self.cell.recurrent_constraint

  @property
  def bias_constraint(self):
    return self.cell.bias_constraint

  @property
  def dropout(self):
    return self.cell.dropout

  @property
  def recurrent_dropout(self):
    return self.cell.recurrent_dropout

  @property
  def implementation(self):
    return self.cell.implementation

  @property
  def reset_after(self):
    return self.cell.reset_after

  def get_config(self):
    config = {
        'units':
            self.units,
        'activation':
            activations.serialize(self.activation),
        'recurrent_activation':
            activations.serialize(self.recurrent_activation),
        'use_bias':
            self.use_bias,
        'kernel_initializer':
            initializers.serialize(self.kernel_initializer),
        'recurrent_initializer':
            initializers.serialize(self.recurrent_initializer),
        'bias_initializer':
            initializers.serialize(self.bias_initializer),
        'kernel_regularizer':
            regularizers.serialize(self.kernel_regularizer),
        'recurrent_regularizer':
            regularizers.serialize(self.recurrent_regularizer),
        'bias_regularizer':
            regularizers.serialize(self.bias_regularizer),
        'activity_regularizer':
            regularizers.serialize(self.activity_regularizer),
        'kernel_constraint':
            constraints.serialize(self.kernel_constraint),
        'recurrent_constraint':
            constraints.serialize(self.recurrent_constraint),
        'bias_constraint':
            constraints.serialize(self.bias_constraint),
        'dropout':
            self.dropout,
        'recurrent_dropout':
            self.recurrent_dropout,
        'implementation':
            self.implementation,
        'reset_after':
            self.reset_after
    }
    config.update(_config_for_enable_caching_device(self.cell))
    base_config = super(GRU, self).get_config()
    del base_config['cell']
    return dict(list(base_config.items()) + list(config.items()))

  @classmethod
  def from_config(cls, config):
    if 'implementation' in config and config['implementation'] == 0:
      config['implementation'] = 1
    return cls(**config)


@keras_export(v1=['keras.layers.LSTMCell'])
class LSTMCell(DropoutRNNCellMixin, Layer):
  """Cell class for the LSTM layer.

  Arguments:
    units: Positive integer, dimensionality of the output space.
    activation: Activation function to use.
      Default: hyperbolic tangent (`tanh`).
      If you pass `None`, no activation is applied
      (ie. "linear" activation: `a(x) = x`).
    recurrent_activation: Activation function to use
      for the recurrent step.
      Default: hard sigmoid (`hard_sigmoid`).
      If you pass `None`, no activation is applied
      (ie. "linear" activation: `a(x) = x`).
    use_bias: Boolean, whether the layer uses a bias vector.
    kernel_initializer: Initializer for the `kernel` weights matrix,
      used for the linear transformation of the inputs.
    recurrent_initializer: Initializer for the `recurrent_kernel`
      weights matrix,
      used for the linear transformation of the recurrent state.
    bias_initializer: Initializer for the bias vector.
    unit_forget_bias: Boolean.
      If True, add 1 to the bias of the forget gate at initialization.
      Setting it to true will also force `bias_initializer="zeros"`.
      This is recommended in [Jozefowicz et al., 2015](
        http://www.jmlr.org/proceedings/papers/v37/jozefowicz15.pdf)
    kernel_regularizer: Regularizer function applied to
      the `kernel` weights matrix.
    recurrent_regularizer: Regularizer function applied to
      the `recurrent_kernel` weights matrix.
    bias_regularizer: Regularizer function applied to the bias vector.
    kernel_constraint: Constraint function applied to
      the `kernel` weights matrix.
    recurrent_constraint: Constraint function applied to
      the `recurrent_kernel` weights matrix.
    bias_constraint: Constraint function applied to the bias vector.
    dropout: Float between 0 and 1.
      Fraction of the units to drop for
      the linear transformation of the inputs.
    recurrent_dropout: Float between 0 and 1.
      Fraction of the units to drop for
      the linear transformation of the recurrent state.
    implementation: Implementation mode, either 1 or 2.
      Mode 1 will structure its operations as a larger number of
      smaller dot products and additions, whereas mode 2 will
      batch them into fewer, larger operations. These modes will
      have different performance profiles on different hardware and
      for different applications.

  Call arguments:
    inputs: A 2D tensor.
    states: List of state tensors corresponding to the previous timestep.
    training: Python boolean indicating whether the layer should behave in
      training mode or in inference mode. Only relevant when `dropout` or
      `recurrent_dropout` is used.
  """

  def __init__(self,
               units,
               activation='tanh',
               recurrent_activation='hard_sigmoid',
               use_bias=True,
               kernel_initializer='glorot_uniform',
               recurrent_initializer='orthogonal',
               bias_initializer='zeros',
               unit_forget_bias=True,
               kernel_regularizer=None,
               recurrent_regularizer=None,
               bias_regularizer=None,
               kernel_constraint=None,
               recurrent_constraint=None,
               bias_constraint=None,
               dropout=0.,
               recurrent_dropout=0.,
               implementation=1,
               **kwargs):
    # By default use cached variable under v2 mode, see b/143699808.
    if ops.executing_eagerly_outside_functions():
      self._enable_caching_device = kwargs.pop('enable_caching_device', True)
    else:
      self._enable_caching_device = kwargs.pop('enable_caching_device', False)
    super(LSTMCell, self).__init__(**kwargs)
    self.units = units
    self.activation = activations.get(activation)
    self.recurrent_activation = activations.get(recurrent_activation)
    self.use_bias = use_bias

    self.kernel_initializer = initializers.get(kernel_initializer)
    self.recurrent_initializer = initializers.get(recurrent_initializer)
    self.bias_initializer = initializers.get(bias_initializer)
    self.unit_forget_bias = unit_forget_bias

    self.kernel_regularizer = regularizers.get(kernel_regularizer)
    self.recurrent_regularizer = regularizers.get(recurrent_regularizer)
    self.bias_regularizer = regularizers.get(bias_regularizer)

    self.kernel_constraint = constraints.get(kernel_constraint)
    self.recurrent_constraint = constraints.get(recurrent_constraint)
    self.bias_constraint = constraints.get(bias_constraint)

    self.dropout = min(1., max(0., dropout))
    self.recurrent_dropout = min(1., max(0., recurrent_dropout))
    if self.recurrent_dropout != 0 and implementation != 1:
      logging.debug(RECURRENT_DROPOUT_WARNING_MSG)
      self.implementation = 1
    else:
      self.implementation = implementation
    # tuple(_ListWrapper) was silently dropping list content in at least 2.7.10,
    # and fixed after 2.7.16. Converting the state_size to wrapper around
    # NoDependency(), so that the base_layer.__setattr__ will not convert it to
    # ListWrapper. Down the stream, self.states will be a list since it is
    # generated from nest.map_structure with list, and tuple(list) will work
    # properly.
    self.state_size = data_structures.NoDependency([self.units, self.units])
    self.output_size = self.units

  @tf_utils.shape_type_conversion
  def build(self, input_shape):
    default_caching_device = _caching_device(self)
    input_dim = input_shape[-1]
    self.kernel = self.add_weight(
        shape=(input_dim, self.units * 4),
        name='kernel',
        initializer=self.kernel_initializer,
        regularizer=self.kernel_regularizer,
        constraint=self.kernel_constraint,
        caching_device=default_caching_device)
    self.recurrent_kernel = self.add_weight(
        shape=(self.units, self.units * 4),
        name='recurrent_kernel',
        initializer=self.recurrent_initializer,
        regularizer=self.recurrent_regularizer,
        constraint=self.recurrent_constraint,
        caching_device=default_caching_device)

    if self.use_bias:
      if self.unit_forget_bias:

        def bias_initializer(_, *args, **kwargs):
          return K.concatenate([
              self.bias_initializer((self.units,), *args, **kwargs),
              initializers.get('ones')((self.units,), *args, **kwargs),
              self.bias_initializer((self.units * 2,), *args, **kwargs),
          ])
      else:
        bias_initializer = self.bias_initializer
      self.bias = self.add_weight(
          shape=(self.units * 4,),
          name='bias',
          initializer=bias_initializer,
          regularizer=self.bias_regularizer,
          constraint=self.bias_constraint,
          caching_device=default_caching_device)
    else:
      self.bias = None
    self.built = True

  def _compute_carry_and_output(self, x, h_tm1, c_tm1):
    """Computes carry and output using split kernels."""
    x_i, x_f, x_c, x_o = x
    h_tm1_i, h_tm1_f, h_tm1_c, h_tm1_o = h_tm1
    i = self.recurrent_activation(
        x_i + K.dot(h_tm1_i, self.recurrent_kernel[:, :self.units]))
    f = self.recurrent_activation(x_f + K.dot(
        h_tm1_f, self.recurrent_kernel[:, self.units:self.units * 2]))
    c = f * c_tm1 + i * self.activation(x_c + K.dot(
        h_tm1_c, self.recurrent_kernel[:, self.units * 2:self.units * 3]))
    o = self.recurrent_activation(
        x_o + K.dot(h_tm1_o, self.recurrent_kernel[:, self.units * 3:]))
    return c, o

  def _compute_carry_and_output_fused(self, z, c_tm1):
    """Computes carry and output using fused kernels."""
    z0, z1, z2, z3 = z
    i = self.recurrent_activation(z0)
    f = self.recurrent_activation(z1)
    c = f * c_tm1 + i * self.activation(z2)
    o = self.recurrent_activation(z3)
    return c, o

  def call(self, inputs, states, training=None):
    h_tm1 = states[0]  # previous memory state
    c_tm1 = states[1]  # previous carry state

    dp_mask = self.get_dropout_mask_for_cell(inputs, training, count=4)
    rec_dp_mask = self.get_recurrent_dropout_mask_for_cell(
        h_tm1, training, count=4)

    if self.implementation == 1:
      if 0 < self.dropout < 1.:
        inputs_i = inputs * dp_mask[0]
        inputs_f = inputs * dp_mask[1]
        inputs_c = inputs * dp_mask[2]
        inputs_o = inputs * dp_mask[3]
      else:
        inputs_i = inputs
        inputs_f = inputs
        inputs_c = inputs
        inputs_o = inputs
      k_i, k_f, k_c, k_o = array_ops.split(
          self.kernel, num_or_size_splits=4, axis=1)
      x_i = K.dot(inputs_i, k_i)
      x_f = K.dot(inputs_f, k_f)
      x_c = K.dot(inputs_c, k_c)
      x_o = K.dot(inputs_o, k_o)
      if self.use_bias:
        b_i, b_f, b_c, b_o = array_ops.split(
            self.bias, num_or_size_splits=4, axis=0)
        x_i = K.bias_add(x_i, b_i)
        x_f = K.bias_add(x_f, b_f)
        x_c = K.bias_add(x_c, b_c)
        x_o = K.bias_add(x_o, b_o)

      if 0 < self.recurrent_dropout < 1.:
        h_tm1_i = h_tm1 * rec_dp_mask[0]
        h_tm1_f = h_tm1 * rec_dp_mask[1]
        h_tm1_c = h_tm1 * rec_dp_mask[2]
        h_tm1_o = h_tm1 * rec_dp_mask[3]
      else:
        h_tm1_i = h_tm1
        h_tm1_f = h_tm1
        h_tm1_c = h_tm1
        h_tm1_o = h_tm1
      x = (x_i, x_f, x_c, x_o)
      h_tm1 = (h_tm1_i, h_tm1_f, h_tm1_c, h_tm1_o)
      c, o = self._compute_carry_and_output(x, h_tm1, c_tm1)
    else:
      if 0. < self.dropout < 1.:
        inputs = inputs * dp_mask[0]
      z = K.dot(inputs, self.kernel)
      z += K.dot(h_tm1, self.recurrent_kernel)
      if self.use_bias:
        z = K.bias_add(z, self.bias)

      z = array_ops.split(z, num_or_size_splits=4, axis=1)
      c, o = self._compute_carry_and_output_fused(z, c_tm1)

    h = o * self.activation(c)
    return h, [h, c]

  def get_config(self):
    config = {
        'units':
            self.units,
        'activation':
            activations.serialize(self.activation),
        'recurrent_activation':
            activations.serialize(self.recurrent_activation),
        'use_bias':
            self.use_bias,
        'kernel_initializer':
            initializers.serialize(self.kernel_initializer),
        'recurrent_initializer':
            initializers.serialize(self.recurrent_initializer),
        'bias_initializer':
            initializers.serialize(self.bias_initializer),
        'unit_forget_bias':
            self.unit_forget_bias,
        'kernel_regularizer':
            regularizers.serialize(self.kernel_regularizer),
        'recurrent_regularizer':
            regularizers.serialize(self.recurrent_regularizer),
        'bias_regularizer':
            regularizers.serialize(self.bias_regularizer),
        'kernel_constraint':
            constraints.serialize(self.kernel_constraint),
        'recurrent_constraint':
            constraints.serialize(self.recurrent_constraint),
        'bias_constraint':
            constraints.serialize(self.bias_constraint),
        'dropout':
            self.dropout,
        'recurrent_dropout':
            self.recurrent_dropout,
        'implementation':
            self.implementation
    }
    config.update(_config_for_enable_caching_device(self))
    base_config = super(LSTMCell, self).get_config()
    return dict(list(base_config.items()) + list(config.items()))

  def get_initial_state(self, inputs=None, batch_size=None, dtype=None):
    return list(_generate_zero_filled_state_for_cell(
        self, inputs, batch_size, dtype))


@keras_export('keras.experimental.PeepholeLSTMCell')
class PeepholeLSTMCell(LSTMCell):
  """Equivalent to LSTMCell class but adds peephole connections.

  Peephole connections allow the gates to utilize the previous internal state as
  well as the previous hidden state (which is what LSTMCell is limited to).
  This allows PeepholeLSTMCell to better learn precise timings over LSTMCell.

  From [Gers et al., 2002](
    http://www.jmlr.org/papers/volume3/gers02a/gers02a.pdf):

  "We find that LSTM augmented by 'peephole connections' from its internal
  cells to its multiplicative gates can learn the fine distinction between
  sequences of spikes spaced either 50 or 49 time steps apart without the help
  of any short training exemplars."

  The peephole implementation is based on:

  [Sak et al., 2014](https://research.google.com/pubs/archive/43905.pdf)

  Example:

  ```python
  # Create 2 PeepholeLSTMCells
  peephole_lstm_cells = [PeepholeLSTMCell(size) for size in [128, 256]]
  # Create a layer composed sequentially of the peephole LSTM cells.
  layer = RNN(peephole_lstm_cells)
  input = keras.Input((timesteps, input_dim))
  output = layer(input)
  ```
  """

  def build(self, input_shape):
    super(PeepholeLSTMCell, self).build(input_shape)
    # The following are the weight matrices for the peephole connections. These
    # are multiplied with the previous internal state during the computation of
    # carry and output.
    self.input_gate_peephole_weights = self.add_weight(
        shape=(self.units,),
        name='input_gate_peephole_weights',
        initializer=self.kernel_initializer)
    self.forget_gate_peephole_weights = self.add_weight(
        shape=(self.units,),
        name='forget_gate_peephole_weights',
        initializer=self.kernel_initializer)
    self.output_gate_peephole_weights = self.add_weight(
        shape=(self.units,),
        name='output_gate_peephole_weights',
        initializer=self.kernel_initializer)

  def _compute_carry_and_output(self, x, h_tm1, c_tm1):
    x_i, x_f, x_c, x_o = x
    h_tm1_i, h_tm1_f, h_tm1_c, h_tm1_o = h_tm1
    i = self.recurrent_activation(
        x_i + K.dot(h_tm1_i, self.recurrent_kernel[:, :self.units]) +
        self.input_gate_peephole_weights * c_tm1)
    f = self.recurrent_activation(x_f + K.dot(
        h_tm1_f, self.recurrent_kernel[:, self.units:self.units * 2]) +
                                  self.forget_gate_peephole_weights * c_tm1)
    c = f * c_tm1 + i * self.activation(x_c + K.dot(
        h_tm1_c, self.recurrent_kernel[:, self.units * 2:self.units * 3]))
    o = self.recurrent_activation(
        x_o + K.dot(h_tm1_o, self.recurrent_kernel[:, self.units * 3:]) +
        self.output_gate_peephole_weights * c)
    return c, o

  def _compute_carry_and_output_fused(self, z, c_tm1):
    z0, z1, z2, z3 = z
    i = self.recurrent_activation(z0 +
                                  self.input_gate_peephole_weights * c_tm1)
    f = self.recurrent_activation(z1 +
                                  self.forget_gate_peephole_weights * c_tm1)
    c = f * c_tm1 + i * self.activation(z2)
    o = self.recurrent_activation(z3 + self.output_gate_peephole_weights * c)
    return c, o


@keras_export(v1=['keras.layers.LSTM'])
class LSTM(RNN):
  """Long Short-Term Memory layer - Hochreiter 1997.

   Note that this cell is not optimized for performance on GPU. Please use
  `tf.compat.v1.keras.layers.CuDNNLSTM` for better performance on GPU.

  Arguments:
    units: Positive integer, dimensionality of the output space.
    activation: Activation function to use.
      Default: hyperbolic tangent (`tanh`).
      If you pass `None`, no activation is applied
      (ie. "linear" activation: `a(x) = x`).
    recurrent_activation: Activation function to use
      for the recurrent step.
      Default: hard sigmoid (`hard_sigmoid`).
      If you pass `None`, no activation is applied
      (ie. "linear" activation: `a(x) = x`).
    use_bias: Boolean, whether the layer uses a bias vector.
    kernel_initializer: Initializer for the `kernel` weights matrix,
      used for the linear transformation of the inputs..
    recurrent_initializer: Initializer for the `recurrent_kernel`
      weights matrix,
      used for the linear transformation of the recurrent state.
    bias_initializer: Initializer for the bias vector.
    unit_forget_bias: Boolean.
      If True, add 1 to the bias of the forget gate at initialization.
      Setting it to true will also force `bias_initializer="zeros"`.
      This is recommended in [Jozefowicz et al., 2015](
        http://www.jmlr.org/proceedings/papers/v37/jozefowicz15.pdf).
    kernel_regularizer: Regularizer function applied to
      the `kernel` weights matrix.
    recurrent_regularizer: Regularizer function applied to
      the `recurrent_kernel` weights matrix.
    bias_regularizer: Regularizer function applied to the bias vector.
    activity_regularizer: Regularizer function applied to
      the output of the layer (its "activation")..
    kernel_constraint: Constraint function applied to
      the `kernel` weights matrix.
    recurrent_constraint: Constraint function applied to
      the `recurrent_kernel` weights matrix.
    bias_constraint: Constraint function applied to the bias vector.
    dropout: Float between 0 and 1.
      Fraction of the units to drop for
      the linear transformation of the inputs.
    recurrent_dropout: Float between 0 and 1.
      Fraction of the units to drop for
      the linear transformation of the recurrent state.
    implementation: Implementation mode, either 1 or 2.
      Mode 1 will structure its operations as a larger number of
      smaller dot products and additions, whereas mode 2 will
      batch them into fewer, larger operations. These modes will
      have different performance profiles on different hardware and
      for different applications.
    return_sequences: Boolean. Whether to return the last output.
      in the output sequence, or the full sequence.
    return_state: Boolean. Whether to return the last state
      in addition to the output.
    go_backwards: Boolean (default False).
      If True, process the input sequence backwards and return the
      reversed sequence.
    stateful: Boolean (default False). If True, the last state
      for each sample at index i in a batch will be used as initial
      state for the sample of index i in the following batch.
    unroll: Boolean (default False).
      If True, the network will be unrolled,
      else a symbolic loop will be used.
      Unrolling can speed-up a RNN,
      although it tends to be more memory-intensive.
      Unrolling is only suitable for short sequences.
    time_major: The shape format of the `inputs` and `outputs` tensors.
      If True, the inputs and outputs will be in shape
      `(timesteps, batch, ...)`, whereas in the False case, it will be
      `(batch, timesteps, ...)`. Using `time_major = True` is a bit more
      efficient because it avoids transposes at the beginning and end of the
      RNN calculation. However, most TensorFlow data is batch-major, so by
      default this function accepts input and emits output in batch-major
      form.

  Call arguments:
    inputs: A 3D tensor.
    mask: Binary tensor of shape `(samples, timesteps)` indicating whether
      a given timestep should be masked.
    training: Python boolean indicating whether the layer should behave in
      training mode or in inference mode. This argument is passed to the cell
      when calling it. This is only relevant if `dropout` or
      `recurrent_dropout` is used.
    initial_state: List of initial state tensors to be passed to the first
      call of the cell.
  """

  def __init__(self,
               units,
               activation='tanh',
               recurrent_activation='hard_sigmoid',
               use_bias=True,
               kernel_initializer='glorot_uniform',
               recurrent_initializer='orthogonal',
               bias_initializer='zeros',
               unit_forget_bias=True,
               kernel_regularizer=None,
               recurrent_regularizer=None,
               bias_regularizer=None,
               activity_regularizer=None,
               kernel_constraint=None,
               recurrent_constraint=None,
               bias_constraint=None,
               dropout=0.,
               recurrent_dropout=0.,
               implementation=1,
               return_sequences=False,
               return_state=False,
               go_backwards=False,
               stateful=False,
               unroll=False,
               **kwargs):
    if implementation == 0:
      logging.warning('`implementation=0` has been deprecated, '
                      'and now defaults to `implementation=1`.'
                      'Please update your layer call.')
    if 'enable_caching_device' in kwargs:
      cell_kwargs = {'enable_caching_device':
                     kwargs.pop('enable_caching_device')}
    else:
      cell_kwargs = {}
    cell = LSTMCell(
        units,
        activation=activation,
        recurrent_activation=recurrent_activation,
        use_bias=use_bias,
        kernel_initializer=kernel_initializer,
        recurrent_initializer=recurrent_initializer,
        unit_forget_bias=unit_forget_bias,
        bias_initializer=bias_initializer,
        kernel_regularizer=kernel_regularizer,
        recurrent_regularizer=recurrent_regularizer,
        bias_regularizer=bias_regularizer,
        kernel_constraint=kernel_constraint,
        recurrent_constraint=recurrent_constraint,
        bias_constraint=bias_constraint,
        dropout=dropout,
        recurrent_dropout=recurrent_dropout,
        implementation=implementation,
        dtype=kwargs.get('dtype'),
        trainable=kwargs.get('trainable', True),
        **cell_kwargs)
    super(LSTM, self).__init__(
        cell,
        return_sequences=return_sequences,
        return_state=return_state,
        go_backwards=go_backwards,
        stateful=stateful,
        unroll=unroll,
        **kwargs)
    self.activity_regularizer = regularizers.get(activity_regularizer)
    self.input_spec = [InputSpec(ndim=3)]

  def call(self, inputs, mask=None, training=None, initial_state=None):
    self._maybe_reset_cell_dropout_mask(self.cell)
    return super(LSTM, self).call(
        inputs, mask=mask, training=training, initial_state=initial_state)

  @property
  def units(self):
    return self.cell.units

  @property
  def activation(self):
    return self.cell.activation

  @property
  def recurrent_activation(self):
    return self.cell.recurrent_activation

  @property
  def use_bias(self):
    return self.cell.use_bias

  @property
  def kernel_initializer(self):
    return self.cell.kernel_initializer

  @property
  def recurrent_initializer(self):
    return self.cell.recurrent_initializer

  @property
  def bias_initializer(self):
    return self.cell.bias_initializer

  @property
  def unit_forget_bias(self):
    return self.cell.unit_forget_bias

  @property
  def kernel_regularizer(self):
    return self.cell.kernel_regularizer

  @property
  def recurrent_regularizer(self):
    return self.cell.recurrent_regularizer

  @property
  def bias_regularizer(self):
    return self.cell.bias_regularizer

  @property
  def kernel_constraint(self):
    return self.cell.kernel_constraint

  @property
  def recurrent_constraint(self):
    return self.cell.recurrent_constraint

  @property
  def bias_constraint(self):
    return self.cell.bias_constraint

  @property
  def dropout(self):
    return self.cell.dropout

  @property
  def recurrent_dropout(self):
    return self.cell.recurrent_dropout

  @property
  def implementation(self):
    return self.cell.implementation

  def get_config(self):
    config = {
        'units':
            self.units,
        'activation':
            activations.serialize(self.activation),
        'recurrent_activation':
            activations.serialize(self.recurrent_activation),
        'use_bias':
            self.use_bias,
        'kernel_initializer':
            initializers.serialize(self.kernel_initializer),
        'recurrent_initializer':
            initializers.serialize(self.recurrent_initializer),
        'bias_initializer':
            initializers.serialize(self.bias_initializer),
        'unit_forget_bias':
            self.unit_forget_bias,
        'kernel_regularizer':
            regularizers.serialize(self.kernel_regularizer),
        'recurrent_regularizer':
            regularizers.serialize(self.recurrent_regularizer),
        'bias_regularizer':
            regularizers.serialize(self.bias_regularizer),
        'activity_regularizer':
            regularizers.serialize(self.activity_regularizer),
        'kernel_constraint':
            constraints.serialize(self.kernel_constraint),
        'recurrent_constraint':
            constraints.serialize(self.recurrent_constraint),
        'bias_constraint':
            constraints.serialize(self.bias_constraint),
        'dropout':
            self.dropout,
        'recurrent_dropout':
            self.recurrent_dropout,
        'implementation':
            self.implementation
    }
    config.update(_config_for_enable_caching_device(self.cell))
    base_config = super(LSTM, self).get_config()
    del base_config['cell']
    return dict(list(base_config.items()) + list(config.items()))

  @classmethod
  def from_config(cls, config):
    if 'implementation' in config and config['implementation'] == 0:
      config['implementation'] = 1
    return cls(**config)


def _generate_dropout_mask(ones, rate, training=None, count=1):
  def dropped_inputs():
    return K.dropout(ones, rate)

  if count > 1:
    return [
        K.in_train_phase(dropped_inputs, ones, training=training)
        for _ in range(count)
    ]
  return K.in_train_phase(dropped_inputs, ones, training=training)


def _standardize_args(inputs, initial_state, constants, num_constants):
  """Standardizes `__call__` to a single list of tensor inputs.

  When running a model loaded from a file, the input tensors
  `initial_state` and `constants` can be passed to `RNN.__call__()` as part
  of `inputs` instead of by the dedicated keyword arguments. This method
  makes sure the arguments are separated and that `initial_state` and
  `constants` are lists of tensors (or None).

  Arguments:
    inputs: Tensor or list/tuple of tensors. which may include constants
      and initial states. In that case `num_constant` must be specified.
    initial_state: Tensor or list of tensors or None, initial states.
    constants: Tensor or list of tensors or None, constant tensors.
    num_constants: Expected number of constants (if constants are passed as
      part of the `inputs` list.

  Returns:
    inputs: Single tensor or tuple of tensors.
    initial_state: List of tensors or None.
    constants: List of tensors or None.
  """
  if isinstance(inputs, list):
    # There are several situations here:
    # In the graph mode, __call__ will be only called once. The initial_state
    # and constants could be in inputs (from file loading).
    # In the eager mode, __call__ will be called twice, once during
    # rnn_layer(inputs=input_t, constants=c_t, ...), and second time will be
    # model.fit/train_on_batch/predict with real np data. In the second case,
    # the inputs will contain initial_state and constants as eager tensor.
    #
    # For either case, the real input is the first item in the list, which
    # could be a nested structure itself. Then followed by initial_states, which
    # could be a list of items, or list of list if the initial_state is complex
    # structure, and finally followed by constants which is a flat list.
    assert initial_state is None and constants is None
    if num_constants:
      constants = inputs[-num_constants:]
      inputs = inputs[:-num_constants]
    if len(inputs) > 1:
      initial_state = inputs[1:]
      inputs = inputs[:1]

    if len(inputs) > 1:
      inputs = tuple(inputs)
    else:
      inputs = inputs[0]

  def to_list_or_none(x):
    if x is None or isinstance(x, list):
      return x
    if isinstance(x, tuple):
      return list(x)
    return [x]

  initial_state = to_list_or_none(initial_state)
  constants = to_list_or_none(constants)

  return inputs, initial_state, constants


def _is_multiple_state(state_size):
  """Check whether the state_size contains multiple states."""
  return (hasattr(state_size, '__len__') and
          not isinstance(state_size, tensor_shape.TensorShape))


def _generate_zero_filled_state_for_cell(cell, inputs, batch_size, dtype):
  if inputs is not None:
    batch_size = array_ops.shape(inputs)[0]
    dtype = inputs.dtype
  return _generate_zero_filled_state(batch_size, cell.state_size, dtype)


def _generate_zero_filled_state(batch_size_tensor, state_size, dtype):
  """Generate a zero filled tensor with shape [batch_size, state_size]."""
  if batch_size_tensor is None or dtype is None:
    raise ValueError(
        'batch_size and dtype cannot be None while constructing initial state: '
        'batch_size={}, dtype={}'.format(batch_size_tensor, dtype))

  def create_zeros(unnested_state_size):
    flat_dims = tensor_shape.as_shape(unnested_state_size).as_list()
    init_state_size = [batch_size_tensor] + flat_dims
    return array_ops.zeros(init_state_size, dtype=dtype)

  if nest.is_sequence(state_size):
    return nest.map_structure(create_zeros, state_size)
  else:
    return create_zeros(state_size)


def _caching_device(rnn_cell):
  """Returns the caching device for the RNN variable.

  This is useful for distributed training, when variable is not located as same
  device as the training worker. By enabling the device cache, this allows
  worker to read the variable once and cache locally, rather than read it every
  time step from remote when it is needed.

  Note that this is assuming the variable that cell needs for each time step is
  having the same value in the forward path, and only gets updated in the
  backprop. It is true for all the default cells (SimpleRNN, GRU, LSTM). If the
  cell body relies on any variable that gets updated every time step, then
  caching device will cause it to read the stall value.

  Args:
    rnn_cell: the rnn cell instance.
  """
  if context.executing_eagerly():
    # caching_device is not supported in eager mode.
    return None
  if not getattr(rnn_cell, '_enable_caching_device', False):
    return None
  # Don't set a caching device when running in a loop, since it is possible that
  # train steps could be wrapped in a tf.while_loop. In that scenario caching
  # prevents forward computations in loop iterations from re-reading the
  # updated weights.
  if control_flow_util.IsInWhileLoop(ops.get_default_graph()):
    logging.warn('Variable read device caching has been disabled because the '
                 'RNN is in tf.while_loop loop context, which will cause '
                 'reading stalled value in forward path. This could slow down '
                 'the training due to duplicated variable reads. Please '
                 'consider updating your code to remove tf.while_loop if '
                 'possible.')
    return None
  if rnn_cell._dtype_policy.should_cast_variables:
    logging.warn('Variable read device caching has been disabled since it '
                 'doesn\'t work with the mixed precision API. This is '
                 'likely to cause a slowdown for RNN training due to '
                 'duplicated read of variable for each timestep, which '
                 'will be significant in a multi remote worker setting. '
                 'Please consider disabling mixed precision API if '
                 'the performance has been affected.')
    return None
  # Cache the value on the device that access the variable.
  return lambda op: op.device


def _config_for_enable_caching_device(rnn_cell):
  """Return the dict config for RNN cell wrt to enable_caching_device field.

  Since enable_caching_device is a internal implementation detail for speed up
  the RNN variable read when running on the multi remote worker setting, we
  don't want this config to be serialized constantly in the JSON. We will only
  serialize this field when a none default value is used to create the cell.
  Args:
    rnn_cell: the RNN cell for serialize.

  Returns:
    A dict which contains the JSON config for enable_caching_device value or
    empty dict if the enable_caching_device value is same as the default value.
  """
  default_enable_caching_device = ops.executing_eagerly_outside_functions()
  if rnn_cell._enable_caching_device != default_enable_caching_device:
    return {'enable_caching_device': rnn_cell._enable_caching_device}
  return {}