New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
The fp32fft option #2
Comments
Hi, thanks for pointing out it. I just check the dtype here. It's true that the input and output are already float32. We have this option since we tried a It seems the |
Thanks for your timely answer! Do you find performance difference in your original linear -> fft -> global filter -> ifft structure, between using fp32fft and not using fp32fft? |
Oops, it seems the fft functions don't support complex fp16 tensors. My logs show fp32 is slightly better than fp16, but it may come from the differences between identical runs. The inputs should always be converted to fp32 when using fft functions. I have updated the code to avoid further confusion. |
Hello, I tested and it seems torch fft functions indeed don't support float16. But given that fft functions don't support fp16 inputs, how did you get a result with fp16, that is slightly worse than fp32? Thanks |
I suspect there is an inconsistency between my old logs and the actual implementation. Maybe I ran two identical experiments in this case and the differences may come from the randomness during training. Since the above-mentioned model and the fp32fft option are only used in our early experiments, I didn't re-run the experiments to check this result. I think the correct implementation is always converting the input to fp32/fp64 before using the fft functions, and I have removed the option from our code. So sorry for the confusion. |
Hello, thanks for your nice work!
I wonder what does the option fp32fft do. In my experiments the input and output to the fft function are already torch.float32, so I'm not sure why there is an option for converting to fp32. Thanks in advance
The text was updated successfully, but these errors were encountered: