CelebaA HQ 256 out of memory and NaN about nvae

[arash-vahdat]

Sorry for the slow reply. Are you sure your PyTorch version is the same as the recommended version? I did see a bit of different behavior in OOM with different PyTorch's version. But I did try these commands in the recommended version before releasing the code.

If the NaN issue persists, you can reduce the learning rate a bit to 8e-3 (--learning_rate=8e-3). That may help with instability.

[AlexZhurkevich]

Thank for the reply!
Yes I am running torch and torchvision of the same numeric versions (1.6, 0.7), although you are probably using plain 1.6 and 0.7 correspondingly which equate to ones that were compiled with CUDA 11. I however use 1.6.0+cu101 and 0.7.0+cu101 that were compiled with CUDA 10.1, this might be a source of the problem, I am upgrading CUDA on the following Monday, will report if it fixed the issue and will try your suggested learning rate, maybe it'll fix the problem. If nothing helps, Ill get back to ask more questions. Thank you for your time!