Dmytro Mishkin

@edgarriba what is happening with mypy test of `augmentation.container`? ``` =================================== FAILURES =================================== [271](https://github.com/kornia/kornia/runs/6656973962?check_suite_focus=true#step:4:272) _________________________________ test session _________________________________ [272](https://github.com/kornia/kornia/runs/6656973962?check_suite_focus=true#step:4:273) mypy exited with status 1. [273](https://github.com/kornia/kornia/runs/6656973962?check_suite_focus=true#step:4:274) ___________________ kornia/augmentation/container/augment.py ___________________ [274](https://github.com/kornia/kornia/runs/6656973962?check_suite_focus=true#step:4:275) 247:...

Could you please clean-up the thing first? 25 files diff is a bit weird.

@Oleksandra2020 I suggest to open new clean PR

@edgarriba how about this implementation? https://github.com/sunny2109/bilateral_filter_Pytorch/blob/main/bilateral_filter.py It seems one can just use unfold2d, no?

Wow, what a Christmas gift!

@rpautrat I have added some minor fixes, preserving commit history here: https://github.com/kornia/kornia/pull/1844 Thank you a lot for contributing!

Solved in #1665

@edgarriba I would prefer to not break compatability with LTS. Probably the best workaround would be to create one more GaussianBlurT with tensor sigma, or like that

I think, that if you run it with `torch.backends.cudnn.benchmark=True`, it would allow PyTorch to pickup the right alg. I am not sure if conv2d allows to select an algorithm, but...

Yes, but for others it would be even more complicated, we would have to have many ifs per backend per kernels size then I think that CPU conv is done...