Support a balanced image dataset generator for TripletLoss function

I have been advised by @fchollet to open the proposed solution in keras-cv instead of keras.

The problem when training a neural networ using TripletLoss is that as the number of classes increases, it is highely unlikely to produce postive pairs in every batch. This results in a division by zero and thus a NaN loss and no further training happens.

The proposed data generator makes sure that a user-defined number of classes and a user-defined number of samples per class is included in every batch. If follows most of the normal image_dataset_from_directory API

Please see this PR in Keras Core: https://github.com/keras-team/keras/pull/15990

[ma7555]

An example of the full TripletLoss workflow: Colab.

I guess I will close for now. If you feel this API is needed please reopen. Thanks!

[LukeWood]

There is a project called Tensorflow Similarity, which specifically provides an API for metric learning and self-supervised learning.

They have a module with different data samplers, designed to sample classes properly for these types of losses. Specifically they have the MultiShotMemorySampler which could be used to solve your problem. They have a Hello World notebook where they use this sampler.

These samplers subclass from tf.keras.utils.Sequence, so they can be used in any keras workflow.

@LukeWood How will KerasCV position itself in relation to TF Similarity regarding a SSL API?

Yeah, tf similarity is happy to user KerasCV apis as we create them. They’re going to adopt our CutMix and RandAugment once I finish it.