Unable to reproduce the original paper effect

[SEMLLYCAT]

Thank you for your work.
Hello, I tried to use your source code to train HDemucs, adopted 8 3090s, the parameters did not change, and trained 360 epochs. The test results are as follows:
Test/nsdr: 5.5024
Test/nsrd_bass:5.4150
Test/nsrd_drums: 6.2600
Test/nsdr_other: 3.5410
Test/nsrd_vocals: 6.7935
paper result:
Test/nsdr: 7.68
Test/nsrd_bass:6.06
Test/nsrd_drums: 8.24
Test/nsdr_other: 5.59
Test/nsrd_vocals: 8.13
I would like to ask you whether you have reproduced the original close result, whether there are other points I did not notice, whether 360epoch is not enough. Maybe I should change monitor loss to nsdr

[KinWaiCheuk]

Sorry for the late reply. We are also unable to reproduce the same result as reported in the paper.

[turian]

Not saying this will fix everything but
a) I think you should be monitoring NSDR
b) HTDemucs (and maybe HDemucs) use EMA, if I'm not mistaken?