recurrent.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__):  # "BasicLSTMCell"
            # Parameters of gates are concatenated into one multiply for efficiency.
            if self._state_is_tuple:
                c, h = state
            else:
                c, h = tf.split(1, 2, state)
            concat = _linear([inputs, h], 4 * self._num_units, True, device=self._device)

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

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

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