TripletLoss

[LukeWood]

No description provided.

[bhack]

https://github.com/tensorflow/addons/blob/master/tensorflow_addons/losses/triplet.py

/cc @seanpmorgan @yarri-oss

[chjort]

Triplet loss is also implemented in TF Similarity.
https://github.com/tensorflow/similarity/blob/master/tensorflow_similarity/losses/triplet_loss.py

We could move these contrastive losses to Keras-CV if it is cool with TF Similarity.

However, I do think we need to think about the implementation design of contrastive losses compared to regular losses.
Contrastive losses are often used in self-supervised learning, where loss is computed for similarities between two different sets of images, and each set with its own labels. Therefore the standard loss(y_true, y_pred) signature is not well suited for contrastive losses. It would need something like loss(y_true_q, y_true_k, y_pred_q, y_pred_k).

See for example this figure from MOCO paper

Example of self-supervised losses like this is SimCLR, BYOL, Barlow Twins, MOCO.

Otherwise, we need to design our contrastive losses to expect y_pred to be a similarity matrix and y_true to be a pair-wise label matrix.

[LukeWood]

Great points… triplet in particular this is really tricky!!! This will need some thought.

SimCLR at least the loss still just has two inputs so it’s not as bad

[LukeWood]

The KerasCV team is attempting to narrow the package scope in order to more efficiently use our limited resources. As such, we are only accepting contributions directly related to items on the 6 month roadmap for the package.

If you feel this feature could be particularly useful, or would like to contribute it - please feel free to re-open this issue and ping LukeWood, ianstenbit or tanzhenyu

Feel free to read more on the roadmap:
https://github.com/keras-team/keras-cv/blob/master/.github/ROADMAP.md

This is already in tf similarity too.