LSTM GRU 得到所有的state 而不是最后一个state

def custom_net(cell, inputs, init_state, timesteps, time_major=False, scope='custom_net_0'):
    # convert to time major format
    if not time_major:
        inputs_tm = tf.transpose(inputs, [1, 0, -1],name="input_time_major")
    # collection of states and outputs
    states, outputs = [init_state], []

    with tf.variable_scope(scope):

        for i in range(timesteps):
            if i > 0:
                tf.get_variable_scope().reuse_variables()
            output, state = cell(inputs_tm[i], states[-1])
            outputs.append(output)
            states.append(state)

return tf.stack(outputs), tf.stack(states[1:])

代码摘自​​https://github.com/ai-guild/r-net/blob/master/lib/recurrence.py​​