Kevin Musgrave

utils.inference: make default match_finder and knn_func use same distance metric

1

Also it might be possible to replace the functionality of ```match_finder``` with ```knn_func```

Add versioned documentation

2

See: https://github.com/jimporter/mike. This makes mkdocs support versioned docs.

Add wrapper for self supervised loss

2

A common use case is to have ```embeddings``` and ```ref_emb``` be augmented versions of each other. For most losses right now you have to create labels to indicate which ```embeddings```...

Add nearest-centroid accuracy as a default metric for AccuracyCalculator

1

- Compute centroid of each class in the reference set - Find nearest centroid for each query embedding. It gets a score of 1 if the nearest centroid is the...

Basically something like sklearn KNeighborsClassifier. We could also add nearest-centroid classification and others. See #342 and #343

The easiest way might be to make the trainers extend [Ignite](https://github.com/pytorch/ignite), [Catalyst](https://github.com/catalyst-team/catalyst), [Lightning](https://github.com/PyTorchLightning/pytorch-lightning) etc.

See #250

The docs should have the following type of explanation, plus a short note per loss function - If you pass pairs into a triplet loss, then triplets will be formed...

One option could involve overriding ```__getattribute()__```

If the number of mined pairs or triplets is small, then it doesn't make sense to compute all pairwise distances in the batch using compute_mat.