F.S.Fire

add BN for generated Q, K, V

> Thanks for your reply but I don't think the difference is clear. > > > 1. The outlooker in VOLO is a new attention mechanism that targets at encoding...

You may need to check two things: 1. compile ddf 2. move the "*.so" files from the build/lib* folder to the ddf folder.

If you use python implementation, it will be very slow, because python implementation will introduce too many "for" loops.

It seems that you have not set up the ddf and moved the *.so files to the ddf folder.

Please check several things: 1. Have you run 'python setup.py install' to build the '.so' files? 2. Have you move out the generated '.so' files from the 'build' folder? 3....

Could you please tell me your CUDA version, nvcc version, and GPU type?

Oh, I am not sure whether my code supports CUDA 11 or not. So, I recommend using CUDA 10.1 in README

I have not implemented the Cuda code for the non-squared kernel. You may need to modify the Cuda implementation of ddf.

Please check your downloaded files. You can use the following python code to print them: ``` import torch model = torch.load('./ddf_mul_resnet50/ddf_mul_resnet50.pth.tar') print(model) ``` If everything is fine, you will see...