I tried this the other day. There's several issues regardless of using the pinned versions in requirements.txt and trying to use debian bullseye and python 3.10.10 as implied by the provided dockerfile. At least these I noted:
- fix deprecated np.bool usage
- module 'scipy.signal' has no attribute 'hann' --> change to scipy.signal.windows.hann
- librosa.feature.melspectrogram(y, change to y=y
- needs CUDA 11
I got it to work without gpu acceleration cause of the last point. If I come back to this I'll probably just look at using demucs, librosa etc. directly (polymath.py is only ~600 lines gluing things together) and hopefully get it working with latest nvidia drivers.
Also if I have days to let this compile maybe I could get CUDA 11 from this https://aur.archlinux.org/packages/cuda-11.7