Metric learning

[AlekseySh]

🚀 Feature Request

It would be great to have metric learning pipeline in Catalyst. The goal of metric learning is to get the model that can map some objects into vector space where similar objects will have close representation and different objects will have far representations. The typical example is the person reidentefication task - reid.

Now we only have several triplet loss implementations in contrib. But there are a lot of missing puzzle details like:

• Sampler that can sample batches with both positive and negatives examples in sufficient quantity.
• TripletSelector hat can provide politics for triplets selection: all triplets, hard triplets, k-hardest triplets and so on.
• RetrievalMetric (like cmc-score) that can evaluate the model: usually it is calculated after splitting validation dataset into query and gallery parts. and so on

It would be great to see all of this as a consistent framework. As datasets for debugging and example we can use: omniglot (objects are symbols) or mars (objects are people) or any other suitable dataset.

Alternative for reid: deep reid library.

[github-actions[bot]]

Hi! Thank you for your contribution! Great first issue!

[Scitator]

Hi @AlekseySh , Honestly speaking, we already have a few Metric Learning experiments with @PUSSYMIPT :)

Nevertheless, let's discuss your proposal a bit more. What samplers and metrics would you like to implement?

[AlekseySh]

Discussed on the call.

Next step: prepare proposal with @PUSSYMIPT

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

[Scitator]

@stale nope :)

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

[Scitator]

okay, now looks like it's done :)