torchaudio.transforms.MelSpectrogram class in
torchaudio differentiable? As in, can I backpropagate through it? If so, can someone point to me towards some documentation on how this backpropagation works?
I think so, just do
requires_grad_ on your input. In the end it goes through
torchaudio.transforms.functional.spectrogram and uses the
torch.stft function. This calls
torch.fft (I think), which has a derivative defined.
There are several texts about how the inner parts of PyTorch work, I wrote something simple a long time ago and @ezyang has an awesome comprehensive tour of PyTorch internals.
Thanks a lot for the information! Especially the links about the inner workings of PyTorch!