using rnn_cell inside of tf.while getting ValueError: The two structures don't have the same number of elements

Question

给定 data = tf.placeholder(tf.float32, [2, None, 3]) (batch_size * time_step * feature_size)，理想情况下我想 tf.unstack(data, axis = 1) 得到许多张量，每个张量都有[2,3] 的形状，所以稍后将它们送入带有 for 循环的 rnn，例如

for rnn_input in rnn_inputs:
    state = rnn_cell(rnn_input, state)

使用像 tf.nn.dynamic_rnn 这样的高级 API 是不合适的 table 所以我创建了一个像

这样的变通方法

import tensorflow as tf


data = tf.placeholder(tf.float32, [2, None, 3])

step_number = tf.placeholder(tf.int32, None)

loop_counter_inital = tf.constant(0)

initi_state = tf.zeros([2,3], tf.float32)

def while_condition(loop_counter, rnn_states):
    return loop_counter < step_number

def while_body(loop_counter, rnn_states):
    loop_counter_current = loop_counter

    current_states = tf.gather_nd(data, tf.stack([tf.range(0, 2), tf.zeros([2], tf.int32)+loop_counter_current], axis=1))     

    cell = tf.nn.rnn_cell.BasicRNNCell(3)

    rnn_states = cell(current_states, rnn_states)

    return [loop_counter_current, rnn_states]


_, _states = tf.while_loop(while_condition, while_body, 
                   loop_vars=[loop_counter_inital, initi_state], 
                   shape_invariants=[loop_counter_inital.shape, tf.TensorShape([2, 3])])

with tf.Session() as sess:    

    sess.run(tf.global_variables_initializer())

    print (sess.run(_states, feed_dict={data:[[[3,1,6],[4,1,2]],[[5,8,1],[0,5,2]]], step_number:2 }))

想法是遍历 data 的每个二维张量中的每一行，以获得每个时间步长的特征。我收到一个错误

First structure (2 elements): [<tf.Tensor 'while/Identity:0' shape=() dtype=int32>, <tf.Tensor 'while/Identity_1:0' shape=(2, 3) dtype=float32>]

Second structure (3 elements): [<tf.Tensor 'while/Identity:0' shape=() dtype=int32>, (<tf.Tensor 'while/basic_rnn_cell/Tanh:0' shape=(2, 3) dtype=float32>, <tf.Tensor 'while/basic_rnn_cell/Tanh:0' shape=(2, 3) dtype=float32>)]

似乎有一些相关的帖子。 None 确实有效。有人可以帮忙吗？

Answer 1

您需要知道每个 BasicRNNCell 都将使用签名 (output, next_state) = call(input, state) 实现 call()。这意味着您的结果是形状列表 ((?,unit),(?,unit))。所以你需要做如下操作。

rnn_states = cell(current_states, rnn_states)[1]

你这里也犯了一个错误。您忘记在 loop_counter_current.

中加 1

return [loop_counter_current+1, rnn_states]

添加

第一个结构体表示你传入的参数loop_vars的初始值，其中包含loop_counter_inital和initi_state的初始值。所以它的结构对应如下

[
<tf.Tensor 'while/Identity:0' shape=() dtype=int32>  #---> loop_counter_inital
, <tf.Tensor 'while/Identity_1:0' shape=(2, 3) dtype=float32>  #---> initi_state
]

第二个结构体表示循环后的参数loop_vars。根据之前的错误，其结果对应如下。

[
<tf.Tensor 'while/Identity:0' shape=() dtype=int32>  #---> loop_counter_inital
, (<tf.Tensor 'while/basic_rnn_cell/Tanh:0' shape=(2, 3) dtype=float32>  #---> output
, <tf.Tensor 'while/basic_rnn_cell/Tanh:0' shape=(2, 3) dtype=float32>)  #---> initi_state
]

using rnn_cell inside of tf.while getting ValueError: The two structures don't have the same number of elements

using rnn_cell inside of tf.while getting ValueError: The two structures don't have the same number of elements

python

tensorflow

recurrent-neural-network