James Fury

I means the CNNs you mentioned in the paper's section 3.2 is for transform T, but actually is done by Fc layer, right? The CNNs is just for reducing dimensions.

Does someone have?

same question,why the step 1000 needs so big consume?

even though the Occupation rate before is only 50%, after the step 1000 ,it still out of memory. I change the code in the utils, Device('cuda') -> Device('cpu') it seems...

> > 首先命令行运行pip install torchfile安装torchfile, > > from torch.utils.serialization import load_lua --》import torchfile, > > 您好，我按照您说的，pip install torchfile，然后将from torch.utils.serialization import load_lua替换为import torchfile，然后将代码中load_lua改为torchfile.load，最后还是遇到了错误，TypeError：‘NoneType’ object is not callable，请问您是没有遇到么，怎么解决的啊 请问你解决了吗？

I want to know the code how to call the get_cls_model function in the cls_cvt.py

By the way , the` self.proj = nn.Linear(dim_out, dim_out) `Means FFN only projection with same dimension?

Thanks, there are so many linear projections that aren't be mentioned by paper.

> @askerlee, I think it is part of depthwise separable convolutions. Depthwise convolutions followed by pointwise projections. No, I think the proj_q\k\v are exactly the things the paper does not...

By the way the `self.proj = nn.Linear(dim_out, dim_out)` Means FFN only projection with same dimension?