Fabian-Robert Stöter
Fabian-Robert Stöter
@ayush055 - I was using one gpu per model (=source) - the dataloader is slow by default. Please use the wav version of the MUSDB18 dataset. and switch on https://github.com/sigsep/open-unmix-pytorch/blob/master/openunmix/data.py#L292...
Can you check your gpu utilization?
That's a known issue caused by the musdb dataloader. I might be able to fix this soon. In the meantime. Please use the the `trackfolder_fix` dataloader in the meantime. This...
Let me try this here and get back to you in the next week
@sevagh interesting, we have a [regression test](https://github.com/sigsep/open-unmix-pytorch/blob/master/tests/test_regression.py#L38-L73) that did compare the SDR against vs. 1.0, so that at least on that one track it doesn't seem to have changed. did...
@sevagh did you manage to figure out what the issue is?
I think it's because you evaluated on HQ as opposed to the normal musdb18.
@sevagh i think we can close this. I had to lower the regression test tolerance in #144 :-/
@sevagh yes, good point. Zenodo is painful when it comes to json since you can't download them directly or save them in the root dir... but i can update the...
it was split to allow torch.hub to be used with zenodo. For torch.hub a direct link to the pth file is required