[OTX] Add SupCon recipes for classification task

[arrufat]

This PR contains recipes for training models using a SupCon-like style.

These recipes aim to increase the current classification baseline when the data is scarce.

Procedure:

• Train a model with two heads:
  • classification head with standard cross-entropy loss.
  • SupCon head with custom SelfSL loss (currently it uses the Barlow Twins loss).
• At inference, the SupCon head is discarded, hence incurring no overhead at runtime.

The recipes include:

• a data pipeline using TwoCrop transform (to generate two views of an image for the SelfSL head)
• a SupConClassifier (in charge of handling the two views of images)
• a SupConClsHead (the one that uses the dual-head approach)
• a SelfSL loss (the Barlow Twins loss)

So far, this method will be used in OTX when specifying the TrainType as SUPERVISEDCONTRASTIVE.

[jaegukhyun]

Genreally, LGTM. I left one comment

[jaegukhyun]

I left one comment, and I have a question. Where is the part for dividing learning rate by 2? I heard that is necessary because of changes in batch size

[arrufat]

I left one comment, and I have a question. Where is the part for dividing learning rate by 2? I heard that is necessary because of changes in batch size

As discussed with Wonju, for now, this logic is not implemented, we thought it would be better to do it manually and, once it because the default method for small datasets, add the logic in the code. But it's up for discussion.

[jaegukhyun]

LGTM, @harimkang, @supersoob, @JihwanEom anyone, could you check this PR once?

[harimkang]

If this works well, there seems to be no structural problem. LGTM.