Better results when I having reimplement the "2D_FFTs" convolution with PyTorch

sascha-kirch / ML_Notebooks

Collection of machine learning related notebooks to share.

MIT License

17 stars 0 forks source link

Better results when I having reimplement the "2D_FFTs" convolution with PyTorch #1

Closed songyuc closed 2 years ago

songyuc commented 2 years ago

Hi, Sascha, Appreciate very much your great work on the blog of "Fourier Convolutions with Kernel Sizes of 1024x1024 and Larger"! I have got better results after reimplementing the "2D_FFTs" convolution with PyTorch, as,

Mean Absolute Error: 6.840202217972546e-07
Standard deviation: 6.106938030825404e-07

Thank you again for a lot of knowledge learned from your blog!

sascha-kirch commented 2 years ago

Hi songyuc, happy to hear that the article motivated you to implement it yourself!

Great results! The errors in both cases are very small and caused by the quntazition of values stored on the computer. Probably in pytorch you required less commands to implement the fourier convolution or used a datatype with higher accuracy.

All the best, Sascha