About Contrastive Loss in NeMo

[EomSooHwan]

I am seeking some details regarding the implementation of ContrastiveLoss within the NeMo framework. Specifically, I would like to understand if ContrastiveLoss specifically refers to Contrastive Predictive Coding (CPC) and its derivatives. I am asking this because contrastive loss can sometimes denote contrastive loss between the representation from different views (or augmentation), and I find that the name could be misleading for some. It would be nice to have a reference paper on the losses that Contrastive Loss is intended to implement.

Also, if ContrastiveLoss in NeMo does not support such multi-view contrastive learning, is there any recipe for such?