Need pytorch nightly?

[agaldran]

Hi!
Thank you for making the Wasserstein loss extension available. Forgive me if this isn't an issue, I am not an expert user. I just wanted to comment that when I tried to run the extension in my computer (torch 1.1.0) I was getting this error in compilation time:

error: identifier "TORCH_CHECK" is undefined

After installing the latest pytorch-nightly everything seems to run smoothly, so I guess this may be a requirement?

[t-vi]

Thanks for raising this! I guess

#ifndef TORCH_CHECK
#define TORCH_CHECK AT_CHECK
#endif

would be a good idea, because that indeed changed after 1.1.

[Yozey]

Hello @t-vi ,

Thank you for sharing your awesome code for the Wassertein distance.

Could you please precise your CUDA version, pytorch version and CuDNN version please? I encountered 4 errors when I was compling the CUDA code and I suspected that it is due to the complier version.

1. error: static assertion failed with "ArgumentInfo is to be a POD struct"
2. error: no suitable constructor exists to convert from "c10::ScalarType" to "at::Type"
3. error: identifier "__shfl_xor" is undefined
4. error: identifier "TORCH_CHECK" is undefined (solved by your previous response)

I'm using CUDA8.0, CuDNN 7.1.1 and PyTorch 1.0.0.

[t-vi]

You need PyTorch 1.1 and the above hotfix or PyTorch nightlies. CuDNN should not matter at all, CUDA8 probably works.

[t-vi]

This is now all well with contemporary versions of PyTorch, so I'm closing it.