Replace uses of `apex.amp` with PyTorch's amp in references

[fmassa]

Some reference scripts have added support for AMP via APEX a couple of years ago, see

vision/references/classification/train.py

Line 17 in 4bf6086

from apex import amp

for one example

PyTorch now natively supports AMP in the torch.cuda.amp namespace, so we should use PyTorch's AMP instead of apex for simplicity.

The replacement needs to happen on the following scripts:

• Classification
• Detection support amp training for detection models #4933
• Segmentation support amp training for segmention models #4994
• Video Classification

cc @datumbox

[datumbox]

Do you know if manually installing apex from the repo will still be needed? Cause if this step is removed it would be great.

[fmassa]

We won't need to install apex anymore, and everything will be handled by PyTorch now. We only used apex for amp in the reference scripts

[prabhat00155]

The name of the argument is apex. Now that we are going to use Pytorch's amp for mixed precision training, should we update this or keep it for BC?

[datumbox]

@prabhat00155 I don't think it's a problem to change the name. We don't have any models that actually use it for training and BC on the references is not our primary focus anyway. My view is that it's fine to update it.

[fmassa]

I second @datumbox comment, let's just change the argument name -- note that the reference scripts are not part of our API

[kbozas]

I am interested in the Video Classification bit.