About patch split in the fine stage about cf-vit HOT 2 CLOSED

In our approach, we maintain a fixed patch dimension while varying the number of embedded tokens by resizing input images. This process is equivalent to directly changing the patch dimension but has the advantage of facilitating parameter sharing in the patch embedding layer.

For example, suppose the original input image has a resolution of 224x224 and a patch dimension of 16x16. During the coarse stage, we downsize the input image to 112x112. During the fine stage, we use the original 224x224 images. We then select informative tokens from the embed result of the 224x224 resolution and concatenate them with uninformative tokens from the embed result of the 112x112 resolution.

The code corresponding to our approach can be found in lines 211-240 of the models_deit.py file.

from cf-vit.

Thank u. I get it

from cf-vit.