How to train with fp16 in torch.cuda.amp?

How can I train a model with only fp16?
the same operation with apex opt_level=“03” not mixed precision

The deprecated apex.amp opt_level="O3" was using “pure” FP16, so you can just call .half() on your model and input data in your training script.

if so, should I remove the and gradscale?
also is there any way to converge model? the model only predicts zero or one in binary classification…

Yes, you won’t need to use autocast, since you are explicitly skipping the mixed-precision training util. Note that pure FP16 training is generally not stable for training.