Pytorch nn.functional.unfold()的简单理解与用法

unfold的作用就是手动实现(卷积中)的滑动窗口操作，也就是只有卷，没有积

ret = F.unfold(inp, size) 

• inp：输入Tensor，必须是四维的(B, C, H, W)
• size：元组，表示滑动窗口大小
• ret：输出三维Tensor(a, b, c)，a表示输入的batch数，b表示滑动窗口的大小，c表示滑动窗口滑动的次数

例如，对于一张1×1×5×5的如下特征图：

[[[[  1,  2,  3,  4,  5],
   [  6,  7,  8,  9, 10],
   [ 11, 12, 13, 14, 15],
   [ 16, 17, 18, 19, 20],
   [ 21, 22, 23, 24, 25]]]]

对其进行2×2的滑动窗口操作，过程是这样的：

1 2  ->  2 3  ->  3 4
6 7      7 8      8 9  ...

编写代码如下：

import torch
from torch.nn import functional as F 

x = torch.Tensor([[[[  1,  2,  3,  4,  5],
   					[  6,  7,  8,  9, 10],
   					[ 11, 12, 13, 14, 15],
   					[ 16, 17, 18, 19, 20],
   					[ 21, 22, 23, 24, 25]]]])
x = F.unfold(x, (2, 2)) 
print(x)
print(x.size())

输出如下：

tensor([[[ 1,  2,  3,  4,  6,  7,  8,  9, 11, 12, 13, 14, 16, 17, 18, 19],
         [ 2,  3,  4,  5,  7,  8,  9, 10, 12, 13, 14, 15, 17, 18, 19, 20],
         [ 6,  7,  8,  9, 11, 12, 13, 14, 16, 17, 18, 19, 21, 22, 23, 24],
         [ 7,  8,  9, 10, 12, 13, 14, 15, 17, 18, 19, 20, 22, 23, 24, 25]]])
torch.Size([1, 4, 16])    

行数为4，即对应着2×2的滑动窗口大小；而每一列的元素为滑动窗口依次所覆盖的内容，一共滑动了16次，因此有16列。