Focal Loss

[sebastian-sz]

Focal loss was used to train RetinaNet.

From Papers with code:

A Focal Loss function addresses class imbalance during training in tasks like object detection.

Reference Keras Implementation:
https://keras.io/examples/vision/retinanet/#implementing-smooth-l1-loss-and-focal-loss-as-keras-custom-losses

What do you think about this component? I think that maybe it could also benefit other models.

[innat]

sigmoid focal loss : https://www.tensorflow.org/addons/api_docs/python/tfa/losses/SigmoidFocalCrossEntropy

[bhack]

https://discuss.tensorflow.org/t/focal-loss-function-for-multilabel-classification/8933

[sebastian-sz]

@bhack I understand that we would like to avoid duplication across repositories. I did not know there is an implementation already in Keras core (although it lacks the alpha parameter).

Should I close this issue?

[bhack]

@bhack I understand that we would like to avoid duplication across repositories. I did not know there is an implementation already in Keras core (although it lacks the alpha parameter).

Should I close this issue?

Personally I am ok to migrate things here.

[sebastian-sz]

@bhack Ok thanks.

[bhack]

@sebastian-sz Honestly It was the scope of #74 to do this in an orderly manner but it seems that in the end it is quite recurrent to never manage this kind of things in a clean way.

[LukeWood]

@bhack I understand that we would like to avoid duplication across repositories. I did not know there is an implementation already in Keras core (although it lacks the alpha parameter).

Should I close this issue?

No, this is a good issue to have open. TFA is not the same as the Keras API, even if some of the objects are in fact Keras Layer objects. TFA is a great place to look at reference implementations, but we should not expect users to install TFA to use KerasCV. Users should be able to coherently use KerasCV end to end using only KerasCV, Keras core, and TensorFlow core components (i.e. tf.data).

[quantumalaviya]

@LukeWood I can work on this!

[innat]

cc. @SamuelMarks
prev implementation on kcv
https://github.com/keras-team/keras-cv/blob/98eae129b6795bb4bfbc888e9e7e916f5d418da3/keras_cv/losses/focal_loss.py

Another resources.

• https://github.com/artemmavrin/focal-loss/tree/7a1810a968051b6acfedf2052123eb76ba3128c4/src/focal_loss

[SamuelMarks]

@innat I just changed the docstring… @tanzhenyu did the rest I think