I’d say the easiest way to use and not make a mistake is to use PyTorch Lightning with
Trainer(use_amp=True).
This will train your model using 16-bit.
https://pytorch-lightning.readthedocs.io/en/0.6.0/trainer.html
1 Like