Recently, I tried the compGCN example in DGL, and I found something that may be improved.
- This line should be changed into
train_loss.append(tr_loss.item())for the efficiency of GPU.
- For users whose pytorch version >= 1.8, this line should be changed into
return th.fft.irfftn(th.conj(th.fft.rfftn(a, (-1))) * th.fft.rfftn(b, (-1)), (-1)), as the Fourier transformation APIs of pytorch have changes.