Has anyone tried mixed precision training?

[ggsggs]

I was looking at the training script of YOLOv3 from Ultralytics repo and they use mixed-precision training using https://github.com/NVIDIA/apex.

Has anyone tried implementing in the training scripts?

Quoting the example from Nvidia:

Three lines enable Amp:

# Added after model and optimizer construction
 model, optimizer = amp.initialize(model, optimizer, flags...)
 ...
 # loss.backward() changed to:
 with amp.scale_loss(loss, optimizer) as scaled_loss:
     scaled_loss.backward()

[xingyizhou]

I haven't.