skip_rnn_cells.py 文件源码

def trainable_initial_state(self, batch_size):
        """
        Create a trainable initial state for the MultiSkipLSTMCell
        :param batch_size: number of samples per batch
        :return: list of SkipLSTMStateTuple
        """
        initial_states = []
        for idx in range(self._num_layers - 1):
            with tf.variable_scope('layer_%d' % (idx + 1)):
                with tf.variable_scope('initial_c'):
                    initial_c = rnn_ops.create_initial_state(batch_size, self._num_units[idx])
                with tf.variable_scope('initial_h'):
                    initial_h = rnn_ops.create_initial_state(batch_size, self._num_units[idx])
                initial_states.append(LSTMStateTuple(initial_c, initial_h))
        with tf.variable_scope('layer_%d' % self._num_layers):
            with tf.variable_scope('initial_c'):
                initial_c = rnn_ops.create_initial_state(batch_size, self._num_units[-1])
            with tf.variable_scope('initial_h'):
                initial_h = rnn_ops.create_initial_state(batch_size, self._num_units[-1])
            with tf.variable_scope('initial_update_prob'):
                initial_update_prob = rnn_ops.create_initial_state(batch_size, 1, trainable=False,
                                                                   initializer=tf.ones_initializer())
            with tf.variable_scope('initial_cum_update_prob'):
                initial_cum_update_prob = rnn_ops.create_initial_state(batch_size, 1, trainable=False,
                                                                       initializer=tf.zeros_initializer())
            initial_states.append(SkipLSTMStateTuple(initial_c, initial_h,
                                                            initial_update_prob, initial_cum_update_prob))
        return initial_states