In CUDA/Apex AMP, you set the optimization level: model, optimizer = amp.initialize(model, optimizer, opt_level="O1") In the examples I read on PyTorch’s website, I don’t see anything analogous to this. How is this accomplished?

Native AMP is similar to the recommended O1 level and doesn’t use any other opt_levels.

Torch.cuda.amp - How to set optimization level?

bfeeny August 14, 2020, 7:49pm 5

I just have one issue I have seen, which I posted here: