Jaccard loss

[innat]

Another useful loss function for semantic segmentation tasks. It was available in keras-contrib, losses/jaccard.py

[qlzh727]

If I understand the jaccard correctly, it is like IoU metric in keras-team/keras. In that sense, we might just add it to keras repo, since usually metric and loss just share the same backend function.

[bhack]

@qlzh727 It was old time https://github.com/keras-team/keras-contrib/blob/master/keras_contrib/losses/jaccard.py

[bhack]

I meant is it ready now for Keras? As if It was in contrib it was not accepted in Keras at that time.

[qlzh727]

The IoU metric is in keras-team/keras at https://github.com/keras-team/keras/blob/9e8bd934b088b4724221520828146a358609caa6/keras/metrics/metrics.py#L2568. If the jaccard loss is same/similar as IoU metric, then I think it might be fit for keras-team/keras. @fchollet

[bhack]

Ok in that case I suppose that we can move the issue between repos under the same org, or not?

[qlzh727]

I think the issue can be moved between repos in the same org

[innat]

@qlzh727 It would be great to have this loss function from core keras. But as it's mostly cv-related loss function, isn't it fitter to have it from keras_cv; if cv-related stuff moves to keras_cv from keras in the future? Also, if we have Dice loss from keras_cv and Jaccard loss from keras, it might be a confusing UX, IM0.

[qlzh727]

I will check with API owner for this question, and see where we want to place this loss. Will update later this week.

[bhack]

Just for the record: keras-team/keras-contrib#533

[qlzh727]

Moved back to keras-cv. I think we would like this to be more cv specific, eg for bounding box etc. We will keep this in keras-cv for now, until we feel there are generic use case support in core keras. (based on the discussion we have within keras-team.)

[qlzh727]

Feel free to send PR to keras-cv/losses.

[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