CycleCL

[guarin]

CycleCL is a new contrastive learning method for cyclic videos: https://arxiv.org/abs/2311.03402

It could be an interesting addition to the package. We would have to add the following:

•  Loading batches from a single video (ordered!)
• Finding negative and positive examples in a batch
• Loss -> It is basically Ntxentloss but generalized to accepting custom negative/positive pairs.
• Augmentations

[philippmwirth]

There already exists a triplet loss implementation in torch. The main contribution would be to implement a triplet sampler that works on multiple GPUs and then a thin wrapper to pass the triplets to the triplet loss.

[philippmwirth]

We'd also need to add a custom projection head (they use 3D convolution).

[guarin]

Note: We are not planning on adding this for now.