Enable torch.complile to work with classification

[harimkang]

Summary

		torch.compile	Small		Medium			Large
Task	Model	Enable	Test/Accuracy	Train/e2e_time	Test/Accuracy	Train/e2e_time	Train/iter_time	Test/Accuracy	Train/e2e_time	Train/iter_time
MULTI_CLASS_CLS	EfficientNet-B0	X	1.0	10	0.7819	87	0.0724	0.8055	210	0.0839
		O	1.0	85	0.7842	273	0.0720	0.8425	513	0.0858
	DeiT-Tiny	X	1.0	10	0.7998	89	0.0612	0.8633	336	0.0820
		O	1.0	36	0.8108	136	0.0635	0.8620	418	0.0814

How to test

otx train ~~~ --model.torch_compile True

Checklist

• I have added unit tests to cover my changes.​
• I have added integration tests to cover my changes.​
• I have ran e2e tests and there is no issues.
• I have added the description of my changes into CHANGELOG in my target branch (e.g., CHANGELOG in develop).​
• I have updated the documentation in my target branch accordingly (e.g., documentation in develop).
• I have linked related issues.

License

• I submit my code changes under the same Apache License that covers the project.
  Feel free to contact the maintainers if that's a concern.
• I have updated the license header for each file (see an example below).

# Copyright (C) 2024 Intel Corporation
# SPDX-License-Identifier: Apache-2.0

[eunwoosh]

Thanks for your work. LGTM, but how about adding an integration test for this?
And it seems that there is no gain for speed referring experiment result. Do you know why?

[harimkang]

Thanks for your work. LGTM, but how about adding an integration test for this? And it seems that there is no gain for speed referring experiment result. Do you know why?

This will require more experimentation. Currently, it uses a lot of time at 0 epoch to do the model compile. In my opinion, this time is relatively high for the current use-case of OTX Classification. Also, custom models may not be optimized in most cases. I'll have to look into this further and experiment with improving it later. For now, this PR is about getting compile just to work.

[sungchul2]

LGTM.
Related to following results, as you said, it seems to need more experiments because there is no gain for iter_time.
I know compiling time is required and it can be bottleneck for e2e time, but official results show improved training performance and I think very short iter_time is required to reduce e2e time as the official results said.
And I searched some custom experimental results and usually torch.compile showed enhanced results about large batch size or large input resolution.
As far as I know, we use (maybe?) large batch size (=64) but small input resolution (=224) for classification.
I'd like to suggest extensive experiments to find which case affects critical damage to training performance.