在张量流中动态地将2d张量分成多个张量

Question

给定2d张量（矩阵），我想将它分成几个大小相等的小张量。您可以将其视为最大池的预处理。例如，

1 2 3 4 5 6 7 8 
2 3 4 5 6 7 8 9 
3 4 5 6 7 8 9 10 
4 5 6 7 8 9 10 11 

考虑到动态的desired_size 2 * 4，输出应该是：

1 2 3 4 
2 3 4 5 

5 6 7 8 
6 7 8 9 

3 4 5 6 
4 5 6 7 

7 8 9 10 
8 9 10 11 

我已经研究了一段时间slice和gather。但我仍然不知道如何去做。你能告诉我如何得到那个吗？提前致谢！

Answer 1

你可以使用tf.extract_image_patches，即使原来有些冗长：

import numpy as np 
import tensorflow as tf 

x = tf.constant(np.arange(8) + np.arange(1,5)[:,np.newaxis]) 
e = tf.extract_image_patches(x[tf.newaxis,:,:,tf.newaxis], 
    [1, 2, 4, 1], [1, 2, 4, 1], [1, 1, 1, 1], padding='VALID') 
e = tf.reshape(e, [-1, 2, 4]) 
sess = tf.InteractiveSession() 
e.eval() 

# returns 
# array([[[ 1, 2, 3, 4], 
#   [ 2, 3, 4, 5]], 
#  [[ 5, 6, 7, 8], 
#   [ 6, 7, 8, 9]], 
#  [[ 3, 4, 5, 6], 
#   [ 4, 5, 6, 7]], 
#  [[ 7, 8, 9, 10], 
#   [ 8, 9, 10, 11]]]) 

Answer 2

我系有tf.split()：

num_splits = 2 
desired_size = (2, 4) 
A = tf.constant(a) 

C = tf.concat(tf.split(A, desired_size[0], 0),1) 
D = tf.reshape(tf.concat(tf.split(C, num_splits*desired_size[0], 1), 0), (-1, desired_size[0], desired_size[1])) 

#The result 
[[[ 1 2 3 4] 
[ 2 3 4 5]] 

[[ 5 6 7 8] 
[ 6 7 8 9]] 

[[ 3 4 5 6] 
[ 4 5 6 7]] 

[[ 7 8 9 10] 
[ 8 9 10 11]]] 

# For num_splits = 4, desired_size = (2, 2) you get 
[[[ 1 2] 
[ 2 3]] 

[[ 3 4] 
[ 4 5]] 

[[ 5 6] 
[ 6 7]] 

[[ 7 8] 
[ 8 9]] 

[[ 3 4] 
[ 4 5]] 

[[ 5 6] 
[ 6 7]] 

[[ 7 8] 
[ 8 9]] 

[[ 9 10] 
[10 11]]]