Pretrained weights for new SCNet Tran model. It's model where LSTM blocks replaced with transformer blocks. Weights were obtained on MUSDB18 dataset only.
Multisong dataset metrics
Num overlap: 2
Instr drums sdr: 10.8707 (Std: 4.9156)
Instr drums bleedless: 21.1680 (Std: 6.7914)
Instr drums fullness: 14.7370 (Std: 4.3909)
Instr drums l1_freq: 34.3470 (Std: 11.9350)
Instr bass sdr: 10.9921 (Std: 4.7437)
Instr bass bleedless: 19.7168 (Std: 9.4653)
Instr bass fullness: 18.9573 (Std: 13.3901)
Instr bass l1_freq: 50.0550 (Std: 15.6142)
Instr other sdr: 5.6386 (Std: 3.1793)
Instr other bleedless: 15.4912 (Std: 5.5533)
Instr other fullness: 16.9250 (Std: 4.4698)
Instr other l1_freq: 30.2603 (Std: 9.8843)
Instr vocals sdr: 8.4272 (Std: 3.2268)
Instr vocals bleedless: 18.7054 (Std: 5.8792)
Instr vocals fullness: 15.2973 (Std: 4.1997)
Instr vocals l1_freq: 31.0975 (Std: 10.2741)
Metric avg sdr : 8.9821
Metric avg bleedless : 18.7704
Metric avg fullness : 16.4791
Metric avg l1_freq : 36.4400
MUSDB18 dataset metrics
Num overlap: 2
Instr drums sdr: 10.8139 (Std: 2.9125)
Instr drums l1_freq: 39.1727 (Std: 10.2864)
Instr drums fullness: 17.6692 (Std: 6.2694)
Instr drums bleedless: 25.0680 (Std: 10.1685)
Instr bass sdr: 8.0731 (Std: 4.6414)
Instr bass l1_freq: 55.2888 (Std: 13.7685)
Instr bass fullness: 14.9235 (Std: 9.6630)
Instr bass bleedless: 20.8816 (Std: 10.5097)
Instr other sdr: 6.8419 (Std: 2.8909)
Instr other l1_freq: 35.2425 (Std: 11.2693)
Instr other fullness: 19.2561 (Std: 6.6855)
Instr other bleedless: 20.1550 (Std: 7.3377)
Instr vocals sdr: 9.9799 (Std: 3.5204)
Instr vocals l1_freq: 45.7778 (Std: 11.0381)
Instr vocals fullness: 20.0345 (Std: 7.0781)
Instr vocals bleedless: 17.8214 (Std: 7.0943)
Metric avg sdr : 8.9272
Metric avg l1_freq : 43.8704
Metric avg fullness : 17.9708
Metric avg bleedless : 20.9815