Closed anvuew closed 3 weeks ago
Yes, but in this case memory requirements grows a lot. I think it must be done as option.
Also for me this code doesn't work correctly - metrics stays relatively at the same place (I did several experiments). I need to understand what is wrong.
For low memory version, I guess using only 1or 2 resolutions instead of 5 in the multiresolution STFT loss could free up some memory.
should low_mem roformers be added here? https://github.com/ZFTurbo/Music-Source-Separation-Training/blob/7e2cc6ecd134f0d108c8c80b0d03ed48f8c549d7/train.py#L274