在张量流中将二维张量动态划分为多个张量

Question

给定一个二维张量(矩阵(，我想把它分成几个大小相等的小张量。您可以将其视为最大池化的预处理。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

给定 2 * 4 的动态desired_size，输出应为：

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

我学习slice和gather已经有一段时间了。但我仍然不知道该怎么做。你能告诉我如何得到它吗？提前感谢！

Answer 1

您可以使用tf.extract_image_patches，即使结果有些冗长：

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

我和tf.split()并列：

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]]]