tensorflow LSTM：张量变形，5维变4维，4维变5维

把5维张量变成4维张量进行卷积操作，再变回5维张量，类似于tf.keras.layers.TimeDistributed操作，手动操作如下：

import numpy as np
import tensorflow.compat.v1 as tf

# 定义维张量
time_step=6
input = tf.convert_to_tensor(np.random.random((4, time_step, 256, 256, 3)).astype(np.float64))

# 张量变形,5维变成4维
input = [tf.squeeze(t,axis=1) for t in tf.split(input,num_or_size_splits=time_step,axis=1)]
input = tf.concat(input,axis=0)

# 把4维张量再复原成5维张量
input_re = [tf.expand_dims(t,axis=1) for t in tf.split(input,num_or_size_splits=time_step,axis=0)]
input_re = tf.concat(input_re,axis=1)

# 验证是否相等
with tf.Session() as sess:
    input,input_re = sess.run([input,input_re])
    print(input_re==input)

输出：

[[[[[ True  True  True]
    [ True  True  True]
    [ True  True  True]
    ...
    [ True  True  True]
    [ True  True  True]
    [ True  True  True]]

   [[ True  True  True]
    [ True  True  True]
    [ True  True  True]
    ...
    [ True  True  True]
    [ True  True  True]
    [ True  True  True]]

...

结论：复原后与原来相同，说明操作成功

核心代码可整理如下：

def dim5_to_dim4(dim5_tensor,time_step):
    dim4_tensor = tf.concat([tf.squeeze(t, axis=1) for t in tf.split(dim5_tensor, num_or_size_splits=time_step, axis=1)],axis=0)
    return dim4_tensor

def dim4_to_dim5(dim4_tensor,time_step):
    dim5_tensor = tf.concat([tf.expand_dims(t, axis=1) for t in tf.split(dim4_tensor, num_or_size_splits=time_step, axis=0)],axis=1)
    return dim5_tensor