search

diyiiyiii / StyTR-2

StyTr2 : Image Style Transfer with Transformers
352 stars 64 forks source link

About the patch partition. #15

Open Charlulote opened 2 years ago

Charlulote commented 2 years ago

It seems that the PatchEmbed module embeds images and divide it into pathes, which is implemented with dilated convolution and large kernel size. But how it realize the locality demonstrated in the Fig2 of the paper? That is, the patches in the framework figure are divided on the pixel grid continuously but I don't see any pixel partition operation in the code. Or is it just my misunderstanding of the paper?

kimbring2 commented 1 year ago

@Charlulote Same question here, I can not find the patch partition part in the code. However, that part is on paper.

I am not sure the Style Transfer can be done solely by the Self-Attention mechanism because of that part.