xref: /external/tensorflow/tensorflow/contrib/rnn/python/ops/rnn_cell.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.
# ==============================================================================

"""Module for constructing RNN Cells."""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import collections
import math

from tensorflow.contrib.compiler import jit
from tensorflow.contrib.layers.python.layers import layers
from tensorflow.contrib.rnn.python.ops import core_rnn_cell
from tensorflow.contrib.rnn.python.ops import core_rnn_cell_impl
from tensorflow.python.framework import dtypes
from tensorflow.python.framework import op_def_registry
from tensorflow.python.framework import ops
from tensorflow.python.ops import array_ops
from tensorflow.python.ops import clip_ops
from tensorflow.python.ops import init_ops
from tensorflow.python.ops import math_ops
from tensorflow.python.ops import nn_ops
from tensorflow.python.ops import variable_scope as vs
from tensorflow.python.platform import tf_logging as logging
from tensorflow.python.util import nest


def _get_concat_variable(name, shape, dtype, num_shards):
  """Get a sharded variable concatenated into one tensor."""
  sharded_variable = _get_sharded_variable(name, shape, dtype, num_shards)
  if len(sharded_variable) == 1:
    return sharded_variable[0]

  concat_name = name + "/concat"
  concat_full_name = vs.get_variable_scope().name + "/" + concat_name + ":0"
  for value in ops.get_collection(ops.GraphKeys.CONCATENATED_VARIABLES):
    if value.name == concat_full_name:
      return value

  concat_variable = array_ops.concat(sharded_variable, 0, name=concat_name)
  ops.add_to_collection(ops.GraphKeys.CONCATENATED_VARIABLES,
                        concat_variable)
  return concat_variable


def _get_sharded_variable(name, shape, dtype, num_shards):
  """Get a list of sharded variables with the given dtype."""
  if num_shards > shape[0]:
    raise ValueError("Too many shards: shape=%s, num_shards=%d" %
                     (shape, num_shards))
  unit_shard_size = int(math.floor(shape[0] / num_shards))
  remaining_rows = shape[0] - unit_shard_size * num_shards

  shards = []
  for i in range(num_shards):
    current_size = unit_shard_size
    if i < remaining_rows:
      current_size += 1
    shards.append(vs.get_variable(name + "_%d" % i, [current_size] + shape[1:],
                                  dtype=dtype))
  return shards


class CoupledInputForgetGateLSTMCell(core_rnn_cell.RNNCell):
  """Long short-term memory unit (LSTM) recurrent network cell.

  The default non-peephole implementation is based on:

    http://deeplearning.cs.cmu.edu/pdfs/Hochreiter97_lstm.pdf

  S. Hochreiter and J. Schmidhuber.
  "Long Short-Term Memory". Neural Computation, 9(8):1735-1780, 1997.

  The peephole implementation is based on:

    https://research.google.com/pubs/archive/43905.pdf

  Hasim Sak, Andrew Senior, and Francoise Beaufays.
  "Long short-term memory recurrent neural network architectures for
   large scale acoustic modeling." INTERSPEECH, 2014.

  The coupling of input and forget gate is based on:

    http://arxiv.org/pdf/1503.04069.pdf

  Greff et al. "LSTM: A Search Space Odyssey"

  The class uses optional peep-hole connections, and an optional projection
  layer.
  """

  def __init__(self, num_units, use_peepholes=False,
               initializer=None, num_proj=None, proj_clip=None,
               num_unit_shards=1, num_proj_shards=1,
               forget_bias=1.0, state_is_tuple=False,
               activation=math_ops.tanh):
    """Initialize the parameters for an LSTM cell.

    Args:
      num_units: int, The number of units in the LSTM cell
      use_peepholes: bool, set True to enable diagonal/peephole connections.
      initializer: (optional) The initializer to use for the weight and
        projection matrices.
      num_proj: (optional) int, The output dimensionality for the projection
        matrices.  If None, no projection is performed.
      proj_clip: (optional) A float value.  If `num_proj > 0` and `proj_clip` is
      provided, then the projected values are clipped elementwise to within
      `[-proj_clip, proj_clip]`.
      num_unit_shards: How to split the weight matrix.  If >1, the weight
        matrix is stored across num_unit_shards.
      num_proj_shards: How to split the projection matrix.  If >1, the
        projection matrix is stored across num_proj_shards.
      forget_bias: Biases of the forget gate are initialized by default to 1
        in order to reduce the scale of forgetting at the beginning of
        the training.
      state_is_tuple: If True, accepted and returned states are 2-tuples of
        the `c_state` and `m_state`.  By default (False), they are concatenated
        along the column axis.  This default behavior will soon be deprecated.
      activation: Activation function of the inner states.
    """
    if not state_is_tuple:
      logging.warn(
          "%s: Using a concatenated state is slower and will soon be "
          "deprecated.  Use state_is_tuple=True.", self)
    self._num_units = num_units
    self._use_peepholes = use_peepholes
    self._initializer = initializer
    self._num_proj = num_proj
    self._proj_clip = proj_clip
    self._num_unit_shards = num_unit_shards
    self._num_proj_shards = num_proj_shards
    self._forget_bias = forget_bias
    self._state_is_tuple = state_is_tuple
    self._activation = activation

    if num_proj:
      self._state_size = (
          core_rnn_cell.LSTMStateTuple(num_units, num_proj)
          if state_is_tuple else num_units + num_proj)
      self._output_size = num_proj
    else:
      self._state_size = (
          core_rnn_cell.LSTMStateTuple(num_units, num_units)
          if state_is_tuple else 2 * num_units)
      self._output_size = num_units

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

  @property
  def output_size(self):
    return self._output_size

  def __call__(self, inputs, state, scope=None):
    """Run one step of LSTM.

    Args:
      inputs: input Tensor, 2D, batch x num_units.
      state: if `state_is_tuple` is False, this must be a state Tensor,
        `2-D, batch x state_size`.  If `state_is_tuple` is True, this must be a
        tuple of state Tensors, both `2-D`, with column sizes `c_state` and
        `m_state`.
      scope: VariableScope for the created subgraph; defaults to "LSTMCell".

    Returns:
      A tuple containing:
      - A `2-D, [batch x output_dim]`, Tensor representing the output of the
        LSTM after reading `inputs` when previous state was `state`.
        Here output_dim is:
           num_proj if num_proj was set,
           num_units otherwise.
      - Tensor(s) representing the new state of LSTM after reading `inputs` when
        the previous state was `state`.  Same type and shape(s) as `state`.

    Raises:
      ValueError: If input size cannot be inferred from inputs via
        static shape inference.
    """
    sigmoid = math_ops.sigmoid

    num_proj = self._num_units if self._num_proj is None else self._num_proj

    if self._state_is_tuple:
      (c_prev, m_prev) = state
    else:
      c_prev = array_ops.slice(state, [0, 0], [-1, self._num_units])
      m_prev = array_ops.slice(state, [0, self._num_units], [-1, num_proj])

    dtype = inputs.dtype
    input_size = inputs.get_shape().with_rank(2)[1]
    if input_size.value is None:
      raise ValueError("Could not infer input size from inputs.get_shape()[-1]")
    with vs.variable_scope(scope or "coupled_input_forget_gate_lstm_cell",
                           initializer=self._initializer):
      concat_w = _get_concat_variable(
          "W", [input_size.value + num_proj, 3 * self._num_units],
          dtype, self._num_unit_shards)

      b = vs.get_variable(
          "B",
          shape=[3 * self._num_units],
          initializer=init_ops.zeros_initializer(),
          dtype=dtype)

      # j = new_input, f = forget_gate, o = output_gate
      cell_inputs = array_ops.concat([inputs, m_prev], 1)
      lstm_matrix = nn_ops.bias_add(math_ops.matmul(cell_inputs, concat_w), b)
      j, f, o = array_ops.split(value=lstm_matrix, num_or_size_splits=3, axis=1)

      # Diagonal connections
      if self._use_peepholes:
        w_f_diag = vs.get_variable(
            "W_F_diag", shape=[self._num_units], dtype=dtype)
        w_o_diag = vs.get_variable(
            "W_O_diag", shape=[self._num_units], dtype=dtype)

      if self._use_peepholes:
        f_act = sigmoid(f + self._forget_bias + w_f_diag * c_prev)
      else:
        f_act = sigmoid(f + self._forget_bias)
      c = (f_act * c_prev + (1 - f_act) * self._activation(j))

      if self._use_peepholes:
        m = sigmoid(o + w_o_diag * c) * self._activation(c)
      else:
        m = sigmoid(o) * self._activation(c)

      if self._num_proj is not None:
        concat_w_proj = _get_concat_variable(
            "W_P", [self._num_units, self._num_proj],
            dtype, self._num_proj_shards)

        m = math_ops.matmul(m, concat_w_proj)
        if self._proj_clip is not None:
          # pylint: disable=invalid-unary-operand-type
          m = clip_ops.clip_by_value(m, -self._proj_clip, self._proj_clip)
          # pylint: enable=invalid-unary-operand-type

    new_state = (core_rnn_cell.LSTMStateTuple(c, m) if self._state_is_tuple else
                 array_ops.concat([c, m], 1))
    return m, new_state


class TimeFreqLSTMCell(core_rnn_cell.RNNCell):
  """Time-Frequency Long short-term memory unit (LSTM) recurrent network cell.

  This implementation is based on:

    Tara N. Sainath and Bo Li
    "Modeling Time-Frequency Patterns with LSTM vs. Convolutional Architectures
    for LVCSR Tasks." submitted to INTERSPEECH, 2016.

  It uses peep-hole connections and optional cell clipping.
  """

  def __init__(self, num_units, use_peepholes=False,
               cell_clip=None, initializer=None,
               num_unit_shards=1, forget_bias=1.0,
               feature_size=None, frequency_skip=None):
    """Initialize the parameters for an LSTM cell.

    Args:
      num_units: int, The number of units in the LSTM cell
      use_peepholes: bool, set True to enable diagonal/peephole connections.
      cell_clip: (optional) A float value, if provided the cell state is clipped
        by this value prior to the cell output activation.
      initializer: (optional) The initializer to use for the weight and
        projection matrices.
      num_unit_shards: int, How to split the weight matrix.  If >1, the weight
        matrix is stored across num_unit_shards.
      forget_bias: float, Biases of the forget gate are initialized by default
        to 1 in order to reduce the scale of forgetting at the beginning
        of the training.
      feature_size: int, The size of the input feature the LSTM spans over.
      frequency_skip: int, The amount the LSTM filter is shifted by in
        frequency.
    """
    self._num_units = num_units
    self._use_peepholes = use_peepholes
    self._cell_clip = cell_clip
    self._initializer = initializer
    self._num_unit_shards = num_unit_shards
    self._forget_bias = forget_bias
    self._feature_size = feature_size
    self._frequency_skip = frequency_skip
    self._state_size = 2 * num_units
    self._output_size = num_units

  @property
  def output_size(self):
    return self._output_size

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

  def __call__(self, inputs, state, scope=None):
    """Run one step of LSTM.

    Args:
      inputs: input Tensor, 2D, batch x num_units.
      state: state Tensor, 2D, batch x state_size.
      scope: VariableScope for the created subgraph; defaults to
        "TimeFreqLSTMCell".

    Returns:
      A tuple containing:
      - A 2D, batch x output_dim, Tensor representing the output of the LSTM
        after reading "inputs" when previous state was "state".
        Here output_dim is num_units.
      - A 2D, batch x state_size, Tensor representing the new state of LSTM
        after reading "inputs" when previous state was "state".
    Raises:
      ValueError: if an input_size was specified and the provided inputs have
        a different dimension.
    """
    sigmoid = math_ops.sigmoid
    tanh = math_ops.tanh

    freq_inputs = self._make_tf_features(inputs)
    dtype = inputs.dtype
    actual_input_size = freq_inputs[0].get_shape().as_list()[1]
    with vs.variable_scope(scope or "time_freq_lstm_cell",
                           initializer=self._initializer):  # "TimeFreqLSTMCell"
      concat_w = _get_concat_variable(
          "W", [actual_input_size + 2*self._num_units, 4 * self._num_units],
          dtype, self._num_unit_shards)
      b = vs.get_variable(
          "B",
          shape=[4 * self._num_units],
          initializer=init_ops.zeros_initializer(),
          dtype=dtype)

      # Diagonal connections
      if self._use_peepholes:
        w_f_diag = vs.get_variable(
            "W_F_diag", shape=[self._num_units], dtype=dtype)
        w_i_diag = vs.get_variable(
            "W_I_diag", shape=[self._num_units], dtype=dtype)
        w_o_diag = vs.get_variable(
            "W_O_diag", shape=[self._num_units], dtype=dtype)

      # initialize the first freq state to be zero
      m_prev_freq = array_ops.zeros([int(inputs.get_shape()[0]),
                                     self._num_units], dtype)
      for fq in range(len(freq_inputs)):
        c_prev = array_ops.slice(state, [0, 2*fq*self._num_units],
                                 [-1, self._num_units])
        m_prev = array_ops.slice(state, [0, (2*fq+1)*self._num_units],
                                 [-1, self._num_units])
        # i = input_gate, j = new_input, f = forget_gate, o = output_gate
        cell_inputs = array_ops.concat([freq_inputs[fq], m_prev, m_prev_freq],
                                       1)
        lstm_matrix = nn_ops.bias_add(math_ops.matmul(cell_inputs, concat_w), b)
        i, j, f, o = array_ops.split(
            value=lstm_matrix, num_or_size_splits=4, axis=1)

        if self._use_peepholes:
          c = (sigmoid(f + self._forget_bias + w_f_diag * c_prev) * c_prev +
               sigmoid(i + w_i_diag * c_prev) * tanh(j))
        else:
          c = (sigmoid(f + self._forget_bias) * c_prev + sigmoid(i) * tanh(j))

        if self._cell_clip is not None:
          # pylint: disable=invalid-unary-operand-type
          c = clip_ops.clip_by_value(c, -self._cell_clip, self._cell_clip)
          # pylint: enable=invalid-unary-operand-type

        if self._use_peepholes:
          m = sigmoid(o + w_o_diag * c) * tanh(c)
        else:
          m = sigmoid(o) * tanh(c)
        m_prev_freq = m
        if fq == 0:
          state_out = array_ops.concat([c, m], 1)
          m_out = m
        else:
          state_out = array_ops.concat([state_out, c, m], 1)
          m_out = array_ops.concat([m_out, m], 1)
    return m_out, state_out

  def _make_tf_features(self, input_feat):
    """Make the frequency features.

    Args:
      input_feat: input Tensor, 2D, batch x num_units.

    Returns:
      A list of frequency features, with each element containing:
      - A 2D, batch x output_dim, Tensor representing the time-frequency feature
        for that frequency index. Here output_dim is feature_size.
    Raises:
      ValueError: if input_size cannot be inferred from static shape inference.
    """
    input_size = input_feat.get_shape().with_rank(2)[-1].value
    if input_size is None:
      raise ValueError("Cannot infer input_size from static shape inference.")
    num_feats = int((input_size - self._feature_size) / (
        self._frequency_skip)) + 1
    freq_inputs = []
    for f in range(num_feats):
      cur_input = array_ops.slice(input_feat, [0, f*self._frequency_skip],
                                  [-1, self._feature_size])
      freq_inputs.append(cur_input)
    return freq_inputs


class GridLSTMCell(core_rnn_cell.RNNCell):
  """Grid Long short-term memory unit (LSTM) recurrent network cell.

  The default is based on:
    Nal Kalchbrenner, Ivo Danihelka and Alex Graves
    "Grid Long Short-Term Memory," Proc. ICLR 2016.
    http://arxiv.org/abs/1507.01526

  When peephole connections are used, the implementation is based on:
    Tara N. Sainath and Bo Li
    "Modeling Time-Frequency Patterns with LSTM vs. Convolutional Architectures
    for LVCSR Tasks." submitted to INTERSPEECH, 2016.

  The code uses optional peephole connections, shared_weights and cell clipping.
  """

  def __init__(self, num_units, use_peepholes=False,
               share_time_frequency_weights=False,
               cell_clip=None, initializer=None,
               num_unit_shards=1, forget_bias=1.0,
               feature_size=None, frequency_skip=None,
               num_frequency_blocks=None,
               start_freqindex_list=None,
               end_freqindex_list=None,
               couple_input_forget_gates=False,
               state_is_tuple=False):
    """Initialize the parameters for an LSTM cell.

    Args:
      num_units: int, The number of units in the LSTM cell
      use_peepholes: (optional) bool, default False. Set True to enable
        diagonal/peephole connections.
      share_time_frequency_weights: (optional) bool, default False. Set True to
        enable shared cell weights between time and frequency LSTMs.
      cell_clip: (optional) A float value, default None, if provided the cell
        state is clipped by this value prior to the cell output activation.
      initializer: (optional) The initializer to use for the weight and
        projection matrices, default None.
      num_unit_shards: (optional) int, defualt 1, How to split the weight
        matrix. If > 1,the weight matrix is stored across num_unit_shards.
      forget_bias: (optional) float, default 1.0, The initial bias of the
        forget gates, used to reduce the scale of forgetting at the beginning
        of the training.
      feature_size: (optional) int, default None, The size of the input feature
        the LSTM spans over.
      frequency_skip: (optional) int, default None, The amount the LSTM filter
        is shifted by in frequency.
      num_frequency_blocks: [required] A list of frequency blocks needed to
        cover the whole input feature splitting defined by start_freqindex_list
        and end_freqindex_list.
      start_freqindex_list: [optional], list of ints, default None,  The
        starting frequency index for each frequency block.
      end_freqindex_list: [optional], list of ints, default None. The ending
        frequency index for each frequency block.
      couple_input_forget_gates: (optional) bool, default False, Whether to
        couple the input and forget gates, i.e. f_gate = 1.0 - i_gate, to reduce
        model parameters and computation cost.
      state_is_tuple: If True, accepted and returned states are 2-tuples of
        the `c_state` and `m_state`.  By default (False), they are concatenated
        along the column axis.  This default behavior will soon be deprecated.
    Raises:
      ValueError: if the num_frequency_blocks list is not specified
    """
    if not state_is_tuple:
      logging.warn("%s: Using a concatenated state is slower and will soon be "
                   "deprecated.  Use state_is_tuple=True.", self)
    self._num_units = num_units
    self._use_peepholes = use_peepholes
    self._share_time_frequency_weights = share_time_frequency_weights
    self._couple_input_forget_gates = couple_input_forget_gates
    self._state_is_tuple = state_is_tuple
    self._cell_clip = cell_clip
    self._initializer = initializer
    self._num_unit_shards = num_unit_shards
    self._forget_bias = forget_bias
    self._feature_size = feature_size
    self._frequency_skip = frequency_skip
    self._start_freqindex_list = start_freqindex_list
    self._end_freqindex_list = end_freqindex_list
    self._num_frequency_blocks = num_frequency_blocks
    self._total_blocks = 0
    if self._num_frequency_blocks is None:
      raise ValueError("Must specify num_frequency_blocks")

    for block_index in range(len(self._num_frequency_blocks)):
      self._total_blocks += int(self._num_frequency_blocks[block_index])
    if state_is_tuple:
      state_names = ""
      for block_index in range(len(self._num_frequency_blocks)):
        for freq_index in range(self._num_frequency_blocks[block_index]):
          name_prefix = "state_f%02d_b%02d" % (freq_index, block_index)
          state_names += ("%s_c, %s_m," % (name_prefix, name_prefix))
      self._state_tuple_type = collections.namedtuple(
          "GridLSTMStateTuple", state_names.strip(","))
      self._state_size = self._state_tuple_type(
              *([num_units, num_units] * self._total_blocks))
    else:
      self._state_tuple_type = None
      self._state_size = num_units * self._total_blocks * 2
    self._output_size = num_units * self._total_blocks * 2

  @property
  def output_size(self):
    return self._output_size

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

  @property
  def state_tuple_type(self):
    return self._state_tuple_type

  def __call__(self, inputs, state, scope=None):
    """Run one step of LSTM.

    Args:
      inputs: input Tensor, 2D, [batch, feature_size].
      state: Tensor or tuple of Tensors, 2D, [batch, state_size], depends on the
        flag self._state_is_tuple.
      scope: (optional) VariableScope for the created subgraph; if None, it
        defaults to "GridLSTMCell".

    Returns:
      A tuple containing:
      - A 2D, [batch, output_dim], Tensor representing the output of the LSTM
        after reading "inputs" when previous state was "state".
        Here output_dim is num_units.
      - A 2D, [batch, state_size], Tensor representing the new state of LSTM
        after reading "inputs" when previous state was "state".
    Raises:
      ValueError: if an input_size was specified and the provided inputs have
        a different dimension.
    """
    batch_size = int(inputs.get_shape()[0])
    freq_inputs = self._make_tf_features(inputs)
    with vs.variable_scope(scope or "grid_lstm_cell",
                           initializer=self._initializer):  # "GridLSTMCell"
      m_out_lst = []
      state_out_lst = []
      for block in range(len(freq_inputs)):
        m_out_lst_current, state_out_lst_current = self._compute(
            freq_inputs[block], block, state, batch_size,
            state_is_tuple=self._state_is_tuple)
        m_out_lst.extend(m_out_lst_current)
        state_out_lst.extend(state_out_lst_current)
      if self._state_is_tuple:
        state_out = self._state_tuple_type(*state_out_lst)
      else:
        state_out = array_ops.concat(state_out_lst, 1)
      m_out = array_ops.concat(m_out_lst, 1)
    return m_out, state_out

  def _compute(self, freq_inputs, block, state, batch_size,
               state_prefix="state",
               state_is_tuple=True):
    """Run the actual computation of one step LSTM.

    Args:
      freq_inputs: list of Tensors, 2D, [batch, feature_size].
      block: int, current frequency block index to process.
      state: Tensor or tuple of Tensors, 2D, [batch, state_size], it depends on
        the flag state_is_tuple.
      batch_size: int32, batch size.
      state_prefix: (optional) string, name prefix for states, defaults to
        "state".
      state_is_tuple: boolean, indicates whether the state is a tuple or Tensor.

    Returns:
      A tuple, containing:
      - A list of [batch, output_dim] Tensors, representing the output of the
        LSTM given the inputs and state.
      - A list of [batch, state_size] Tensors, representing the LSTM state
        values given the inputs and previous state.
    """
    sigmoid = math_ops.sigmoid
    tanh = math_ops.tanh
    num_gates = 3 if self._couple_input_forget_gates else 4
    dtype = freq_inputs[0].dtype
    actual_input_size = freq_inputs[0].get_shape().as_list()[1]

    concat_w_f = _get_concat_variable(
        "W_f_%d" % block, [actual_input_size + 2 * self._num_units,
                           num_gates * self._num_units],
        dtype, self._num_unit_shards)
    b_f = vs.get_variable(
        "B_f_%d" % block,
        shape=[num_gates * self._num_units],
        initializer=init_ops.zeros_initializer(),
        dtype=dtype)
    if not self._share_time_frequency_weights:
      concat_w_t = _get_concat_variable(
          "W_t_%d" % block, [actual_input_size + 2 * self._num_units,
                             num_gates * self._num_units],
          dtype, self._num_unit_shards)
      b_t = vs.get_variable(
          "B_t_%d" % block,
          shape=[num_gates * self._num_units],
          initializer=init_ops.zeros_initializer(),
          dtype=dtype)

    if self._use_peepholes:
      # Diagonal connections
      if not self._couple_input_forget_gates:
        w_f_diag_freqf = vs.get_variable(
            "W_F_diag_freqf_%d" % block, shape=[self._num_units], dtype=dtype)
        w_f_diag_freqt = vs.get_variable(
            "W_F_diag_freqt_%d"% block, shape=[self._num_units], dtype=dtype)
      w_i_diag_freqf = vs.get_variable(
          "W_I_diag_freqf_%d" % block, shape=[self._num_units], dtype=dtype)
      w_i_diag_freqt = vs.get_variable(
          "W_I_diag_freqt_%d" % block, shape=[self._num_units], dtype=dtype)
      w_o_diag_freqf = vs.get_variable(
          "W_O_diag_freqf_%d" % block, shape=[self._num_units], dtype=dtype)
      w_o_diag_freqt = vs.get_variable(
          "W_O_diag_freqt_%d" % block, shape=[self._num_units], dtype=dtype)
      if not self._share_time_frequency_weights:
        if not self._couple_input_forget_gates:
          w_f_diag_timef = vs.get_variable(
              "W_F_diag_timef_%d" % block, shape=[self._num_units], dtype=dtype)
          w_f_diag_timet = vs.get_variable(
              "W_F_diag_timet_%d" % block, shape=[self._num_units], dtype=dtype)
        w_i_diag_timef = vs.get_variable(
            "W_I_diag_timef_%d" % block, shape=[self._num_units], dtype=dtype)
        w_i_diag_timet = vs.get_variable(
            "W_I_diag_timet_%d" % block, shape=[self._num_units], dtype=dtype)
        w_o_diag_timef = vs.get_variable(
            "W_O_diag_timef_%d" % block, shape=[self._num_units], dtype=dtype)
        w_o_diag_timet = vs.get_variable(
            "W_O_diag_timet_%d" % block, shape=[self._num_units], dtype=dtype)

    # initialize the first freq state to be zero
    m_prev_freq = array_ops.zeros([batch_size, self._num_units], dtype)
    c_prev_freq = array_ops.zeros([batch_size, self._num_units], dtype)
    for freq_index in range(len(freq_inputs)):
      if state_is_tuple:
        name_prefix = "%s_f%02d_b%02d" % (state_prefix, freq_index, block)
        c_prev_time = getattr(state, name_prefix + "_c")
        m_prev_time = getattr(state, name_prefix + "_m")
      else:
        c_prev_time = array_ops.slice(
            state, [0, 2 * freq_index * self._num_units],
            [-1, self._num_units])
        m_prev_time = array_ops.slice(
            state, [0, (2 * freq_index + 1) * self._num_units],
            [-1, self._num_units])

      # i = input_gate, j = new_input, f = forget_gate, o = output_gate
      cell_inputs = array_ops.concat(
          [freq_inputs[freq_index], m_prev_time, m_prev_freq], 1)

      # F-LSTM
      lstm_matrix_freq = nn_ops.bias_add(math_ops.matmul(cell_inputs,
                                                         concat_w_f), b_f)
      if self._couple_input_forget_gates:
        i_freq, j_freq, o_freq = array_ops.split(
            value=lstm_matrix_freq, num_or_size_splits=num_gates, axis=1)
        f_freq = None
      else:
        i_freq, j_freq, f_freq, o_freq = array_ops.split(
            value=lstm_matrix_freq, num_or_size_splits=num_gates, axis=1)
      # T-LSTM
      if self._share_time_frequency_weights:
        i_time = i_freq
        j_time = j_freq
        f_time = f_freq
        o_time = o_freq
      else:
        lstm_matrix_time = nn_ops.bias_add(math_ops.matmul(cell_inputs,
                                                           concat_w_t), b_t)
        if self._couple_input_forget_gates:
          i_time, j_time, o_time = array_ops.split(
              value=lstm_matrix_time, num_or_size_splits=num_gates, axis=1)
          f_time = None
        else:
          i_time, j_time, f_time, o_time = array_ops.split(
              value=lstm_matrix_time, num_or_size_splits=num_gates, axis=1)

      # F-LSTM c_freq
      # input gate activations
      if self._use_peepholes:
        i_freq_g = sigmoid(i_freq +
                           w_i_diag_freqf * c_prev_freq +
                           w_i_diag_freqt * c_prev_time)
      else:
        i_freq_g = sigmoid(i_freq)
      # forget gate activations
      if self._couple_input_forget_gates:
        f_freq_g = 1.0 - i_freq_g
      else:
        if self._use_peepholes:
          f_freq_g = sigmoid(f_freq + self._forget_bias +
                             w_f_diag_freqf * c_prev_freq +
                             w_f_diag_freqt * c_prev_time)
        else:
          f_freq_g = sigmoid(f_freq + self._forget_bias)
      # cell state
      c_freq = f_freq_g * c_prev_freq + i_freq_g * tanh(j_freq)
      if self._cell_clip is not None:
        # pylint: disable=invalid-unary-operand-type
        c_freq = clip_ops.clip_by_value(c_freq, -self._cell_clip,
                                        self._cell_clip)
        # pylint: enable=invalid-unary-operand-type

      # T-LSTM c_freq
      # input gate activations
      if self._use_peepholes:
        if self._share_time_frequency_weights:
          i_time_g = sigmoid(i_time +
                             w_i_diag_freqf * c_prev_freq +
                             w_i_diag_freqt * c_prev_time)
        else:
          i_time_g = sigmoid(i_time +
                             w_i_diag_timef * c_prev_freq +
                             w_i_diag_timet * c_prev_time)
      else:
        i_time_g = sigmoid(i_time)
      # forget gate activations
      if self._couple_input_forget_gates:
        f_time_g = 1.0 - i_time_g
      else:
        if self._use_peepholes:
          if self._share_time_frequency_weights:
            f_time_g = sigmoid(f_time + self._forget_bias +
                               w_f_diag_freqf * c_prev_freq +
                               w_f_diag_freqt * c_prev_time)
          else:
            f_time_g = sigmoid(f_time + self._forget_bias +
                               w_f_diag_timef * c_prev_freq +
                               w_f_diag_timet * c_prev_time)
        else:
          f_time_g = sigmoid(f_time + self._forget_bias)
      # cell state
      c_time = f_time_g * c_prev_time + i_time_g * tanh(j_time)
      if self._cell_clip is not None:
        # pylint: disable=invalid-unary-operand-type
        c_time = clip_ops.clip_by_value(c_time, -self._cell_clip,
                                        self._cell_clip)
        # pylint: enable=invalid-unary-operand-type

      # F-LSTM m_freq
      if self._use_peepholes:
        m_freq = sigmoid(o_freq +
                         w_o_diag_freqf * c_freq +
                         w_o_diag_freqt * c_time) * tanh(c_freq)
      else:
        m_freq = sigmoid(o_freq) * tanh(c_freq)

      # T-LSTM m_time
      if self._use_peepholes:
        if self._share_time_frequency_weights:
          m_time = sigmoid(o_time +
                           w_o_diag_freqf * c_freq +
                           w_o_diag_freqt * c_time) * tanh(c_time)
        else:
          m_time = sigmoid(o_time +
                           w_o_diag_timef * c_freq +
                           w_o_diag_timet * c_time) * tanh(c_time)
      else:
        m_time = sigmoid(o_time) * tanh(c_time)

      m_prev_freq = m_freq
      c_prev_freq = c_freq
      # Concatenate the outputs for T-LSTM and F-LSTM for each shift
      if freq_index == 0:
        state_out_lst = [c_time, m_time]
        m_out_lst = [m_time, m_freq]
      else:
        state_out_lst.extend([c_time, m_time])
        m_out_lst.extend([m_time, m_freq])

    return m_out_lst, state_out_lst

  def _make_tf_features(self, input_feat, slice_offset=0):
    """Make the frequency features.

    Args:
      input_feat: input Tensor, 2D, [batch, num_units].
      slice_offset: (optional) Python int, default 0, the slicing offset is only
        used for the backward processing in the BidirectionalGridLSTMCell. It
        specifies a different starting point instead of always 0 to enable the
        forward and backward processing look at different frequency blocks.

    Returns:
      A list of frequency features, with each element containing:
      - A 2D, [batch, output_dim], Tensor representing the time-frequency
        feature for that frequency index. Here output_dim is feature_size.
    Raises:
      ValueError: if input_size cannot be inferred from static shape inference.
    """
    input_size = input_feat.get_shape().with_rank(2)[-1].value
    if input_size is None:
      raise ValueError("Cannot infer input_size from static shape inference.")
    if slice_offset > 0:
      # Padding to the end
      inputs = array_ops.pad(
          input_feat, array_ops.constant([0, 0, 0, slice_offset], shape=[2, 2],
                                         dtype=dtypes.int32),
          "CONSTANT")
    elif slice_offset < 0:
      # Padding to the front
      inputs = array_ops.pad(
          input_feat, array_ops.constant([0, 0, -slice_offset, 0], shape=[2, 2],
                                         dtype=dtypes.int32),
          "CONSTANT")
      slice_offset = 0
    else:
      inputs = input_feat
    freq_inputs = []
    if not self._start_freqindex_list:
      if len(self._num_frequency_blocks) != 1:
        raise ValueError("Length of num_frequency_blocks"
                         " is not 1, but instead is %d",
                         len(self._num_frequency_blocks))
      num_feats = int((input_size - self._feature_size) / (
          self._frequency_skip)) + 1
      if num_feats != self._num_frequency_blocks[0]:
        raise ValueError(
            "Invalid num_frequency_blocks, requires %d but gets %d, please"
            " check the input size and filter config are correct." % (
                self._num_frequency_blocks[0], num_feats))
      block_inputs = []
      for f in range(num_feats):
        cur_input = array_ops.slice(
            inputs, [0, slice_offset + f * self._frequency_skip],
            [-1, self._feature_size])
        block_inputs.append(cur_input)
      freq_inputs.append(block_inputs)
    else:
      if len(self._start_freqindex_list) != len(self._end_freqindex_list):
        raise ValueError("Length of start and end freqindex_list"
                         " does not match %d %d",
                         len(self._start_freqindex_list),
                         len(self._end_freqindex_list))
      if len(self._num_frequency_blocks) != len(self._start_freqindex_list):
        raise ValueError("Length of num_frequency_blocks"
                         " is not equal to start_freqindex_list %d %d",
                         len(self._num_frequency_blocks),
                         len(self._start_freqindex_list))
      for b in range(len(self._start_freqindex_list)):
        start_index = self._start_freqindex_list[b]
        end_index = self._end_freqindex_list[b]
        cur_size = end_index - start_index
        block_feats = int((cur_size - self._feature_size) / (
            self._frequency_skip)) + 1
        if block_feats != self._num_frequency_blocks[b]:
          raise ValueError(
              "Invalid num_frequency_blocks, requires %d but gets %d, please"
              " check the input size and filter config are correct." % (
                  self._num_frequency_blocks[b], block_feats))
        block_inputs = []
        for f in range(block_feats):
          cur_input = array_ops.slice(
              inputs, [0, start_index + slice_offset + f *
                       self._frequency_skip],
              [-1, self._feature_size])
          block_inputs.append(cur_input)
        freq_inputs.append(block_inputs)
    return freq_inputs


class BidirectionalGridLSTMCell(GridLSTMCell):
  """Bidirectional GridLstm cell.

  The bidirection connection is only used in the frequency direction, which
  hence doesn't affect the time direction's real-time processing that is
  required for online recognition systems.
  The current implementation uses different weights for the two directions.
  """

  def __init__(self, num_units, use_peepholes=False,
               share_time_frequency_weights=False,
               cell_clip=None, initializer=None,
               num_unit_shards=1, forget_bias=1.0,
               feature_size=None, frequency_skip=None,
               num_frequency_blocks=None,
               start_freqindex_list=None,
               end_freqindex_list=None,
               couple_input_forget_gates=False,
               backward_slice_offset=0):
    """Initialize the parameters for an LSTM cell.

    Args:
      num_units: int, The number of units in the LSTM cell
      use_peepholes: (optional) bool, default False. Set True to enable
        diagonal/peephole connections.
      share_time_frequency_weights: (optional) bool, default False. Set True to
        enable shared cell weights between time and frequency LSTMs.
      cell_clip: (optional) A float value, default None, if provided the cell
        state is clipped by this value prior to the cell output activation.
      initializer: (optional) The initializer to use for the weight and
        projection matrices, default None.
      num_unit_shards: (optional) int, defualt 1, How to split the weight
        matrix. If > 1,the weight matrix is stored across num_unit_shards.
      forget_bias: (optional) float, default 1.0, The initial bias of the
        forget gates, used to reduce the scale of forgetting at the beginning
        of the training.
      feature_size: (optional) int, default None, The size of the input feature
        the LSTM spans over.
      frequency_skip: (optional) int, default None, The amount the LSTM filter
        is shifted by in frequency.
      num_frequency_blocks: [required] A list of frequency blocks needed to
        cover the whole input feature splitting defined by start_freqindex_list
        and end_freqindex_list.
      start_freqindex_list: [optional], list of ints, default None,  The
        starting frequency index for each frequency block.
      end_freqindex_list: [optional], list of ints, default None. The ending
        frequency index for each frequency block.
      couple_input_forget_gates: (optional) bool, default False, Whether to
        couple the input and forget gates, i.e. f_gate = 1.0 - i_gate, to reduce
        model parameters and computation cost.
      backward_slice_offset: (optional) int32, default 0, the starting offset to
        slice the feature for backward processing.
    """
    super(BidirectionalGridLSTMCell, self).__init__(
        num_units, use_peepholes, share_time_frequency_weights, cell_clip,
        initializer, num_unit_shards, forget_bias, feature_size, frequency_skip,
        num_frequency_blocks, start_freqindex_list, end_freqindex_list,
        couple_input_forget_gates=False,
        state_is_tuple=True)
    self._backward_slice_offset = int(backward_slice_offset)
    state_names = ""
    for direction in ["fwd", "bwd"]:
      for block_index in range(len(self._num_frequency_blocks)):
        for freq_index in range(self._num_frequency_blocks[block_index]):
          name_prefix = "%s_state_f%02d_b%02d" % (direction, freq_index,
                                                  block_index)
          state_names += ("%s_c, %s_m," % (name_prefix, name_prefix))
    self._state_tuple_type = collections.namedtuple(
        "BidirectionalGridLSTMStateTuple", state_names.strip(","))
    self._state_size = self._state_tuple_type(
        *([num_units, num_units] * self._total_blocks * 2))
    self._output_size = 2 * num_units * self._total_blocks * 2

  def __call__(self, inputs, state, scope=None):
    """Run one step of LSTM.

    Args:
      inputs: input Tensor, 2D, [batch, num_units].
      state: tuple of Tensors, 2D, [batch, state_size].
      scope: (optional) VariableScope for the created subgraph; if None, it
        defaults to "BidirectionalGridLSTMCell".

    Returns:
      A tuple containing:
      - A 2D, [batch, output_dim], Tensor representing the output of the LSTM
        after reading "inputs" when previous state was "state".
        Here output_dim is num_units.
      - A 2D, [batch, state_size], Tensor representing the new state of LSTM
        after reading "inputs" when previous state was "state".
    Raises:
      ValueError: if an input_size was specified and the provided inputs have
        a different dimension.
    """
    batch_size = int(inputs.get_shape()[0])
    fwd_inputs = self._make_tf_features(inputs)
    if self._backward_slice_offset:
      bwd_inputs = self._make_tf_features(inputs, self._backward_slice_offset)
    else:
      bwd_inputs = fwd_inputs

    # Forward processing
    with vs.variable_scope(scope or "bidirectional_grid_lstm_cell",
                           initializer=self._initializer):
      with vs.variable_scope("fwd"):
        fwd_m_out_lst = []
        fwd_state_out_lst = []
        for block in range(len(fwd_inputs)):
          fwd_m_out_lst_current, fwd_state_out_lst_current = self._compute(
              fwd_inputs[block], block, state, batch_size,
              state_prefix="fwd_state", state_is_tuple=True)
          fwd_m_out_lst.extend(fwd_m_out_lst_current)
          fwd_state_out_lst.extend(fwd_state_out_lst_current)
      # Backward processing
      bwd_m_out_lst = []
      bwd_state_out_lst = []
      with vs.variable_scope("bwd"):
        for block in range(len(bwd_inputs)):
          # Reverse the blocks
          bwd_inputs_reverse = bwd_inputs[block][::-1]
          bwd_m_out_lst_current, bwd_state_out_lst_current = self._compute(
              bwd_inputs_reverse, block, state, batch_size,
              state_prefix="bwd_state", state_is_tuple=True)
          bwd_m_out_lst.extend(bwd_m_out_lst_current)
          bwd_state_out_lst.extend(bwd_state_out_lst_current)
    state_out = self._state_tuple_type(*(fwd_state_out_lst + bwd_state_out_lst))
    # Outputs are always concated as it is never used separately.
    m_out = array_ops.concat(fwd_m_out_lst + bwd_m_out_lst, 1)
    return m_out, state_out


# pylint: disable=protected-access
_linear = core_rnn_cell_impl._linear
# pylint: enable=protected-access


class AttentionCellWrapper(core_rnn_cell.RNNCell):
  """Basic attention cell wrapper.

  Implementation based on https://arxiv.org/abs/1409.0473.
  """

  def __init__(self, cell, attn_length, attn_size=None, attn_vec_size=None,
               input_size=None, state_is_tuple=False):
    """Create a cell with attention.

    Args:
      cell: an RNNCell, an attention is added to it.
      attn_length: integer, the size of an attention window.
      attn_size: integer, the size of an attention vector. Equal to
          cell.output_size by default.
      attn_vec_size: integer, the number of convolutional features calculated
          on attention state and a size of the hidden layer built from
          base cell state. Equal attn_size to by default.
      input_size: integer, the size of a hidden linear layer,
          built from inputs and attention. Derived from the input tensor
          by default.
      state_is_tuple: If True, accepted and returned states are n-tuples, where
        `n = len(cells)`.  By default (False), the states are all
        concatenated along the column axis.

    Raises:
      TypeError: if cell is not an RNNCell.
      ValueError: if cell returns a state tuple but the flag
          `state_is_tuple` is `False` or if attn_length is zero or less.
    """
    if not isinstance(cell, core_rnn_cell.RNNCell):
      raise TypeError("The parameter cell is not RNNCell.")
    if nest.is_sequence(cell.state_size) and not state_is_tuple:
      raise ValueError("Cell returns tuple of states, but the flag "
                       "state_is_tuple is not set. State size is: %s"
                       % str(cell.state_size))
    if attn_length <= 0:
      raise ValueError("attn_length should be greater than zero, got %s"
                       % str(attn_length))
    if not state_is_tuple:
      logging.warn(
          "%s: Using a concatenated state is slower and will soon be "
          "deprecated.  Use state_is_tuple=True.", self)
    if attn_size is None:
      attn_size = cell.output_size
    if attn_vec_size is None:
      attn_vec_size = attn_size
    self._state_is_tuple = state_is_tuple
    self._cell = cell
    self._attn_vec_size = attn_vec_size
    self._input_size = input_size
    self._attn_size = attn_size
    self._attn_length = attn_length

  @property
  def state_size(self):
    size = (self._cell.state_size, self._attn_size,
            self._attn_size * self._attn_length)
    if self._state_is_tuple:
      return size
    else:
      return sum(list(size))

  @property
  def output_size(self):
    return self._attn_size

  def __call__(self, inputs, state, scope=None):
    """Long short-term memory cell with attention (LSTMA)."""
    with vs.variable_scope(scope or "attention_cell_wrapper"):
      if self._state_is_tuple:
        state, attns, attn_states = state
      else:
        states = state
        state = array_ops.slice(states, [0, 0], [-1, self._cell.state_size])
        attns = array_ops.slice(
            states, [0, self._cell.state_size], [-1, self._attn_size])
        attn_states = array_ops.slice(
            states, [0, self._cell.state_size + self._attn_size],
            [-1, self._attn_size * self._attn_length])
      attn_states = array_ops.reshape(attn_states,
                                      [-1, self._attn_length, self._attn_size])
      input_size = self._input_size
      if input_size is None:
        input_size = inputs.get_shape().as_list()[1]
      inputs = _linear([inputs, attns], input_size, True)
      lstm_output, new_state = self._cell(inputs, state)
      if self._state_is_tuple:
        new_state_cat = array_ops.concat(nest.flatten(new_state), 1)
      else:
        new_state_cat = new_state
      new_attns, new_attn_states = self._attention(new_state_cat, attn_states)
      with vs.variable_scope("attn_output_projection"):
        output = _linear([lstm_output, new_attns], self._attn_size, True)
      new_attn_states = array_ops.concat(
          [new_attn_states, array_ops.expand_dims(output, 1)], 1)
      new_attn_states = array_ops.reshape(
          new_attn_states, [-1, self._attn_length * self._attn_size])
      new_state = (new_state, new_attns, new_attn_states)
      if not self._state_is_tuple:
        new_state = array_ops.concat(list(new_state), 1)
      return output, new_state

  def _attention(self, query, attn_states):
    conv2d = nn_ops.conv2d
    reduce_sum = math_ops.reduce_sum
    softmax = nn_ops.softmax
    tanh = math_ops.tanh

    with vs.variable_scope("attention"):
      k = vs.get_variable(
          "attn_w", [1, 1, self._attn_size, self._attn_vec_size])
      v = vs.get_variable("attn_v", [self._attn_vec_size])
      hidden = array_ops.reshape(attn_states,
                                 [-1, self._attn_length, 1, self._attn_size])
      hidden_features = conv2d(hidden, k, [1, 1, 1, 1], "SAME")
      y = _linear(query, self._attn_vec_size, True)
      y = array_ops.reshape(y, [-1, 1, 1, self._attn_vec_size])
      s = reduce_sum(v * tanh(hidden_features + y), [2, 3])
      a = softmax(s)
      d = reduce_sum(
          array_ops.reshape(a, [-1, self._attn_length, 1, 1]) * hidden, [1, 2])
      new_attns = array_ops.reshape(d, [-1, self._attn_size])
      new_attn_states = array_ops.slice(attn_states, [0, 1, 0], [-1, -1, -1])
      return new_attns, new_attn_states


class LayerNormBasicLSTMCell(core_rnn_cell.RNNCell):
  """LSTM unit with layer normalization and recurrent dropout.

  This class adds layer normalization and recurrent dropout to a
  basic LSTM unit. Layer normalization implementation is based on:

    https://arxiv.org/abs/1607.06450.

  "Layer Normalization"
  Jimmy Lei Ba, Jamie Ryan Kiros, Geoffrey E. Hinton

  and is applied before the internal nonlinearities.
  Recurrent dropout is base on:

    https://arxiv.org/abs/1603.05118

  "Recurrent Dropout without Memory Loss"
  Stanislau Semeniuta, Aliaksei Severyn, Erhardt Barth.
  """

  def __init__(self, num_units, forget_bias=1.0,
               input_size=None, activation=math_ops.tanh,
               layer_norm=True, norm_gain=1.0, norm_shift=0.0,
               dropout_keep_prob=1.0, dropout_prob_seed=None):
    """Initializes the basic LSTM cell.

    Args:
      num_units: int, The number of units in the LSTM cell.
      forget_bias: float, The bias added to forget gates (see above).
      input_size: Deprecated and unused.
      activation: Activation function of the inner states.
      layer_norm: If `True`, layer normalization will be applied.
      norm_gain: float, The layer normalization gain initial value. If
        `layer_norm` has been set to `False`, this argument will be ignored.
      norm_shift: float, The layer normalization shift initial value. If
        `layer_norm` has been set to `False`, this argument will be ignored.
      dropout_keep_prob: unit Tensor or float between 0 and 1 representing the
        recurrent dropout probability value. If float and 1.0, no dropout will
        be applied.
      dropout_prob_seed: (optional) integer, the randomness seed.
    """

    if input_size is not None:
      logging.warn("%s: The input_size parameter is deprecated.", self)

    self._num_units = num_units
    self._activation = activation
    self._forget_bias = forget_bias
    self._keep_prob = dropout_keep_prob
    self._seed = dropout_prob_seed
    self._layer_norm = layer_norm
    self._g = norm_gain
    self._b = norm_shift

  @property
  def state_size(self):
    return core_rnn_cell.LSTMStateTuple(self._num_units, self._num_units)

  @property
  def output_size(self):
    return self._num_units

  def _norm(self, inp, scope):
    shape = inp.get_shape()[-1:]
    gamma_init = init_ops.constant_initializer(self._g)
    beta_init = init_ops.constant_initializer(self._b)
    with vs.variable_scope(scope):
      # Initialize beta and gamma for use by layer_norm.
      vs.get_variable("gamma", shape=shape, initializer=gamma_init)
      vs.get_variable("beta", shape=shape, initializer=beta_init)
    normalized = layers.layer_norm(inp, reuse=True, scope=scope)
    return normalized

  def _linear(self, args):
    out_size = 4 * self._num_units
    proj_size = args.get_shape()[-1]
    weights = vs.get_variable("weights", [proj_size, out_size])
    out = math_ops.matmul(args, weights)
    if not self._layer_norm:
      bias = vs.get_variable("biases", [out_size])
      out = nn_ops.bias_add(out, bias)
    return out

  def __call__(self, inputs, state, scope=None):
    """LSTM cell with layer normalization and recurrent dropout."""

    with vs.variable_scope(scope or "layer_norm_basic_lstm_cell"):
      c, h = state
      args = array_ops.concat([inputs, h], 1)
      concat = self._linear(args)

      i, j, f, o = array_ops.split(value=concat, num_or_size_splits=4, axis=1)
      if self._layer_norm:
        i = self._norm(i, "input")
        j = self._norm(j, "transform")
        f = self._norm(f, "forget")
        o = self._norm(o, "output")

      g = self._activation(j)
      if (not isinstance(self._keep_prob, float)) or self._keep_prob < 1:
        g = nn_ops.dropout(g, self._keep_prob, seed=self._seed)

      new_c = (c * math_ops.sigmoid(f + self._forget_bias)
               + math_ops.sigmoid(i) * g)
      if self._layer_norm:
        new_c = self._norm(new_c, "state")
      new_h = self._activation(new_c) * math_ops.sigmoid(o)

      new_state = core_rnn_cell.LSTMStateTuple(new_c, new_h)
      return new_h, new_state


_REGISTERED_OPS = None


class CompiledWrapper(core_rnn_cell.RNNCell):
  """Wraps step execution in an XLA JIT scope."""

  def __init__(self, cell, compile_stateful=False):
    """Create CompiledWrapper cell.

    Args:
      cell: Instance of `RNNCell`.
      compile_stateful: Whether to compile stateful ops like initializers
        and random number generators (default: False).
    """
    self._cell = cell
    self._compile_stateful = compile_stateful

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

  @property
  def output_size(self):
    return self._cell.output_size

  def __call__(self, inputs, state, scope=None):
    if self._compile_stateful:
      compile_ops = True
    else:
      def compile_ops(node_def):
        global _REGISTERED_OPS
        if _REGISTERED_OPS is None:
          _REGISTERED_OPS = op_def_registry.get_registered_ops()
        return not _REGISTERED_OPS[node_def.op].is_stateful

    with jit.experimental_jit_scope(compile_ops=compile_ops):
      return self._cell(inputs, state, scope=scope)