MSFR loss

[misakshoyan]

Hi @INVOKERer.
In MIMO-UNET, where pytorch version is 1.4, the FFT for MSFR loss is calculated as
torch.rfft(image, signal_ndim=2, normalized=False, onesided=False)

Could you, please, tell me how do you port it to torch.fft.rfft in pytorch 1.8?

[INVOKERer]

You can use torch.fft.rfft2/fft2/fftn/rfftn to calculate MSFR. We use torch.fft.fft2 to calculate MSFR.

[misakshoyan]

torch.fft.fft2 or torch.fft.rfft2?
And, could you, please, share your one line code with me if it is possible?

[INVOKERer]

torch.mean(abs(torch.fft.fft2(x) - torch.fft.fft2(y)))

[misakshoyan]

Thank You!

[misakshoyan]

I get the following warning message after loss.backward():
"Warning: Casting complex values to real discards the imaginary part (function operator())"

Is this expected?

When I use rfft2 instead of fft2,
"torch.mean(abs(torch.fft.rfft2(x) - torch.fft.rfft2(y)))"
warning disappears.

[misakshoyan]

I think this is a bug of pytorch, the warning message should not be given:
pytorch/pytorch#59524

But code should work correct.