a small question about the loss function

lucidrains / audiolm-pytorch

Implementation of AudioLM, a SOTA Language Modeling Approach to Audio Generation out of Google Research, in Pytorch

MIT License

2.33k stars 249 forks source link

Closed PB20000090 closed 1 year ago

PB20000090 commented 1 year ago

In soundstream.py line 58: def hinge_gen_loss(fake) return -fake.mean(） May I ask if this should return (F.relu(1 - fake)).mean()? Thanks

lucidrains commented 1 year ago

so i'm taking the formulation used in stylegan2, which in turn uses this paper (paperswithcode explanation)