What are c_state and m_state in Tensorflow LSTM?

Question

Tensorflow r0.12's documentation for tf.nn.rnn_cell.LSTMCell describes this as the init:

tf.nn.rnn_cell.LSTMCell.__call__(inputs, state, scope=None)

where state is as follows:

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.

What aare c_state and m_state and how do they fit into LSTMs? I cannot find reference to them anywhere in the documentation.

Here is a link to that page in the documentation.

Answer 1

I agree that the documentation is unclear. Looking at tf.nn.rnn_cell.LSTMCell.__call__ clarifies (I took the code from TensorFlow 1.0.0):

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 "lstm_cell".

    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.
    """
    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 "lstm_cell",
                           initializer=self._initializer) as unit_scope:
      if self._num_unit_shards is not None:
        unit_scope.set_partitioner(
            partitioned_variables.fixed_size_partitioner(
                self._num_unit_shards))
      # i = input_gate, j = new_input, f = forget_gate, o = output_gate
      lstm_matrix = _linear([inputs, m_prev], 4 * self._num_units, bias=True,
                            scope=scope)
      i, j, f, o = array_ops.split(
          value=lstm_matrix, num_or_size_splits=4, axis=1)

      # Diagonal connections
      if self._use_peepholes:
        with vs.variable_scope(unit_scope) as projection_scope:
          if self._num_unit_shards is not None:
            projection_scope.set_partitioner(None)
          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)

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

      if self._num_proj is not None:
        with vs.variable_scope("projection") as proj_scope:
          if self._num_proj_shards is not None:
            proj_scope.set_partitioner(
                partitioned_variables.fixed_size_partitioner(
                    self._num_proj_shards))
          m = _linear(m, self._num_proj, bias=False, scope=scope)

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

The key lines are:

c = (sigmoid(f + self._forget_bias) * c_prev + sigmoid(i) *
         self._activation(j))

and

m = sigmoid(o) * self._activation(c)

and

new_state = (LSTMStateTuple(c, m) 

If you compare the code to compute c and m with the LSTM equations (see below), you can see it corresponds to the cell state (typically denoted with c) and hidden state (typically denoted with h), respectively:

new_state = (LSTMStateTuple(c, m) indicates that the first element of the returned state tuple is c (cell state a.k.a. c_state), and the second element of the returned state tuple is m (hidden state a.k.a. m_state).