Sources with different numbers of channels for sourcefolder dataset about open-unmix-pytorch

[faroit]

@sleglaive coming back to this, I am not sure how this can be fixed at the dataloader stage, when nb_channels=1 and a stereo file is loaded as we want the downmix to be done in the magnitude spectrogram domain. If we downmix in the waveform domain already the downmix in the spectrogram layer would never be called.

Instead, for heterogenous datasets, I would propose an upmix_to_stero=True switch for the dataloaders that would always upmix to stereo by copying the channels and later downmix in the spectrogram domain

@aliutkus additional thoughts?

[faroit]

Hi Simon, thanks for the issue. This is actually something that we could address as part of #33. I will leave this open until it's fixed.

Regarding a fix, do you agree that the downmix should be done in the magnitude spectrum domain?

[sleglaive]

Hi Fabian,
Yes it seems ok to me. Naturally, I would have done the downmix in the waveform domain, but actually I guess you chose the magnitude spectrum domain to avoid phase problems? It is then a good idea, and maybe better than selecting only one channel.

[faroit]

@sleglaive this is now addressed by #56 and merged into the dev branch. Maybe you can comment on this before we release the new version

[sleglaive]

@faroit sorry for the late reply, the solution you propose sounds great to me, thanks a lot!

[faroit]

Great. I will close this when the dev branch is merged