melike
May 21, 2021, 5:36pm
1
Hi all, I have an input tensor of shape 12x3x3 which corresponds to 12 patches of size 3x3, can be explained as {Patch1, Patch2, …, Patch12}. I want to convert it to 1x12x9, keeping the patches as described below,
Patch1 Patch2 Patch3
I tried torch.unfold() with different parameters, but couldn’t find the right way. Any ideas?
tom
May 21, 2021, 7:31pm
2
How about inp.view(4, 3, 3, 3).permute(0, 2, 1, 3).reshape(1, 12, 9)?
(If anything, unfold would be used when getting into the other direction with overlapping patches.)
1 Like
melike
May 21, 2021, 8:50pm
3
Yes, that’s it, thank you. Much better than my desperate attempts. And for a different patch size ps, it should be,
inp.view(4, 3, ps, ps).permute(0, 2, 1, 3).reshape(1, 4 * ps, 3 * ps)