Fast Fourier transform support in PyTorch

[AashishV (Aashish Venkatesh)]

I am trying to implement Multimodal Compact Bilinear Pooling for Visual Question Answering and Visual Grounding. For bilinear pooling, they use FFT and IFFT. But I couldn’t find documentation for FFT in PyTorch, will FFT operation be added to the library later or is it already implemented and not available in the documentation.

[QuantScientist (Solomon K)]

What about https://github.com/locuslab/pytorch_fft?

[AashishV (Aashish Venkatesh)]

@QuantScientist, @smth It has an autograd issue, https://github.com/locuslab/pytorch_fft/issues/4.

[r9y9 (Ryuichi Yamamoto)]

The autograd issue seems to be fixed now.

[kediamanav (Manav Kedia)]

@AashishV were you able to implement it?

[AashishV (Aashish Venkatesh)]

No, I just saw the issue was solved yesterday. Also, I am working on different projects now. So, I will not be able to do it for some time now.

[varunagrawal]

I have implemented a model here. Currently training it to see if results match.

[SimonW (Simon Wang)]

fft and ifft will be available in next release

[guyue_hu]

hi, how could I useFFT in the unstable version now, because it could not be install by pip or conda?thank you!

[SimonW (Simon Wang)]

You can install from source or wait for a couple days for the next release.