search

sp-uhh / sgmse

Score-based Generative Models (Diffusion Models) for Speech Enhancement and Dereverberation
MIT License
454 stars 69 forks source link

Training loss #39

Closed yangdongchao closed 2 months ago

yangdongchao commented 7 months ago

Great work! Can you share the training loss logging? I am trying to reproduce it, but I find the training loss is very large: loss=1.64e+03. It is normal? I training 20 epoch, but find it does not work now.

julius-richter commented 2 months ago

This is what typical loss curves look like:

loss_curves