rnn_cell.py 文件源码

def __call__(self, inputs, state, scope=None):
    """Long short-term memory cell (LSTM)."""
    with vs.variable_scope(scope or type(self).__name__):
      # Parameters of gates are concatenated into one multiply for efficiency.
      if self._state_is_tuple:
        c, h = state
      else:
        c, h = array_ops.split(1, 2, state)
      concat = _linear([inputs, h], 4 * self._num_units, True)

      # i = input_gate, j = new_input, f = forget_gate, o = output_gate
      i, j, f, o = array_ops.split(1, 4, concat)

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


      eps = 1e-13
      temp = math_ops.div(math_ops.reduce_sum(math_ops.mul(c, new_c),1),math_ops.reduce_sum(math_ops.mul(c,c),1) + eps)

      dummy = array_ops.transpose(c)

      t1 = math_ops.mul(dummy, temp)
      t1 = array_ops.transpose(t1) 
      distract_c = new_c  -  t1

      new_h = self._activation(distract_c) * sigmoid(o)

      if self._state_is_tuple:
        new_state = LSTMStateTuple(new_c, new_h)
      else:
        new_state = array_ops.concat(1, [new_c, new_h])
      return new_h, new_state