chenmnz / cf-vit Goto Github PK

你好，请问这个实验的设置，您是修改的stride size还是image的H、W从而进行不同tokens的比较，patch size的大小是固定的吗？这部分实验的代码可以分享一下嘛

Where is the file for feature reuse?

Hi @ChenMnZ , thanks for your excellent work! Wolud you please guide me how to set the confidence threshould η, and the start epoch number (which is set to 200 in your paper)? Because I cannot find any config file to do it, I want to learn more about these settings, thanks.

Hello! @ChenMnZ , thanks for your excellent work!
But I have some problems reproducing the results. I use cf-deit-s-7x7-80.8.pth to test the code. When I set $\eta=1$, I get the result like: Acc@1 78.231 Acc@5 93.945, and $\eta=0.5$ the result is Acc@1 78.010 Acc@5 93.657.
I want to know if there are any potential reasons? Thanks!

Thanks for sharing you work.
In you paper, there are two 'vision transformer' blue rectangles in the framework picture. Do they use same network and share same parameters?

May I ask if the XX [0] and XX [1] in your input represent two inputs of the original image

你好，请问这里设置self.informative_selection = False，导致训练的时候并没有进行Informative Region Identification.这里我是不是应该修改成True？

Hello, thank you very much for your excellent work.
I can not find "from deit.datasets import get_post_process".Where is the code of "get_post_process"?

Hello, thank you very much for your excellent work. When I tried to reproduce the code following the steps you provided and set the model as 'cf_deit_small', I encountered the following error: "RuntimeError: Unknown model (cf_deit_small)". After debugging, I found that the function "register_model" in "registry.py" does not include 'cf_deit_small'. Could you please let me know if there is any important part that I might have overlooked?

In fine stage, you further split a patch into 2*2 patches. So, the token's dim decreases to 1/4 and it doesn't match the dim in original ViT. For example, a patch is 16,16,3, after splitting a patch is 8,8,3. How do you solve this problem?