search

NVlabs / stylegan3

Official PyTorch implementation of StyleGAN3
Other
6.38k stars 1.12k forks source link

how to know if there's a model collapse in training? #140

Open nicolai256 opened 2 years ago

nicolai256 commented 2 years ago

how do you recognize a model collapse? thanks :)

PDillis commented 2 years ago

It's fairly simple: absolute collapse is when all the outputs are exactly the same, irrespective of the input z. If this happens, then there's no other option but to restart the training from zero. A more usual form of collapse is when there are several images being generated by G, but they don't really differ by much. For example, this issue in StyleGAN2-ADA. Note that there is diversity in color, so the key there was to continue training to see if the model could get out of the collapse and that's exactly what happened (this is not always the case, but it's good to try).