F2 Score metric for object detection.

[innat]

The F2-score was used to evaluate the performance in TensorFlow - Help Protect the Great Barrier Reef. It might be a good addition in keras-cv.

[bhack]

https://github.com/tensorflow/addons/blob/master/tensorflow_addons/metrics/f_scores.py

[quantumalaviya]

Wouldn't it be better to just write a general FScore metric and change the beta values?

[LukeWood]

great issue, we can add this. I agree with @quantumalaviya that we can have a general FScore and tune beta; though it may be a valuable issue to also offer a F2Score class that just subclasses FScore and provides a beta of 2.

[bhack]

F2Score class that just subclasses FScore and provides a beta of 2.

...I think this is like my previous comment as:

class F1Score(FBetaScore):

[markub3327]

FBetaScore should be accept the sparsely defined labels.

For example:

# check for sparse labels
if tf.rank(y_true) < tf.rank(y_pred):
    y_true = tf.one_hot(y_true, self.num_classes)

[bhack]

@markub3327 Do you want to port the component and your PR on this repo?

[markub3327]

@bhack Yes, I can.

[bhack]

@LukeWood Can you assign this to @markub3327 ?

[markub3327]

@bhack

tensorflow_addons/metrics/f_scores.py to keras_cv/metrics/f_scores.py and my PR too?

[bhack]

Yes and if you want also F2 with the related tests.
Note, that you could need or not to adapt/refactor these a bit for this repo standards, we have few metrics integrated at https://github.com/keras-team/keras-cv/tree/master/keras_cv/metrics/coco

[markub3327]

@bhack Okey, let's do it.

[innat]

FBetaScore should be accept the sparsely defined labels.

For example:

# check for sparse labels
if tf.rank(y_true) < tf.rank(y_pred):
    y_true = tf.one_hot(y_true, self.num_classes)

Pleaae note that, in keras, it may require to write two separate classes for sparse and one hot encoded target vector.

[markub3327]

@innat Or, it can be argument:

sparse=True or sparse=False

[bhack]

@innat Or, it can be argument:

sparse=True or sparse=False

@markub3327 Check the thread at #371 (comment)

[markub3327]

@bhack Ok

[pure-rgb]

Any update on this?
Already asked. keras-team/keras#15811

[tanzhenyu]

I think we should add F1 score to Keras, and keep F2 score to KerasCV?

[tanmaypandey7]

@tanzhenyu why don't we keep both f1 and f2 scores to keras itself? Won't it be a bit confusing to the user, if he needs to import the f1 metric from keras and f2 from keras-cv?

Also, i would like to work/help on this PR if there is nobody working on it currently.
cc: @markub3327

[innat]

keras-team/keras#17516

[bhack]

#423

[LukeWood]

closing as this most likely fits in keras core

[innat]

@LukeWood
Could u please move this ticket to keras?

[LukeWood]

@LukeWood Could u please move this ticket to keras?

I do not have Keras ownership O_o. Can we re-open there with a link here?