Metric learning (Triplet loss)

[Rusteam]

Would it make sense to add a metric learning task for cases when there're many classes and few images per class? They usually outperform simple classification methods.

If so, should it be integrated into Image Embedder (although I don't see vissl supporting this option) or created as a standalone task (possible using this package?