About training my own MDX net model

[cro888888]

I would like to work out a MDX-net model for singling out only Tomori's voice in MyGO based off currently available code here. I have some data but UVR.py and separate.py is just too huge and too much params for me to find where is the actually training/eval part, can someone who knows the code highlight where I should look? thx

Edit:I found separate.py line 487 but still can't find like all the actually essential params like spectrogram dimensions etc, from netron I'm able to determine input size but scaling problem still exists