image preprocessing layers in keras_cv vs keras_hub vs keras ?

[matpalm]

guessing the plan is to move all preprocessing to run as backend agnositic ( using keras.ops )
under https://github.com/keras-team/keras/tree/master/keras/src/layers/preprocessing/image_preprocessing
and deprecate the image parts of https://github.com/keras-team/keras-cv/tree/master/keras_cv/src/layers/preprocessing ?

i see the keras_cv is being moved to keras_hub, so what does that mean for
https://github.com/keras-team/keras-hub/tree/master/keras_hub/src/layers/preprocessing ?

[fchollet]

Model-related functionality from KerasCV is being moved to KerasHub. However preprocessing and augmentation layers belong in core Keras (including bbox preprocessing functionality), so we're moving that here. For the time being you can keep using the KerasCV implementations though. Was there a particular layer you were missing?

[matpalm]

https://github.com/keras-team/keras-cv/blob/master/keras_cv/src/layers/preprocessing/mosaic.py is the main one; for a lot of datasets ( with yolo like models at least ) i see mosaic makes a BIG difference

since i'm interested in training relatively small models i often have free GPU capacity, so i'm motivated to include augmentations in the training graph directly where possible i.e. not as part of a loading tf.data pipeline. i also see with small models that augmentation on CPU can take a reasonable chunk of the wall clock time itself! so i'd definitely love to have them all based on keras.ops for training with the jax backend.

[fchollet]

That layer only appears to use TF ops that are readily available in Keras 3, so it should be simple to port it. Would you be willing to open a PR?

[matpalm]

Yeah sure, I'll give it go! ( Also suspect it might have a bug re: dropping bboxes occasionally so am keen to review the code deeper )

…

On Wed, 27 Nov 2024, 22:41 François Chollet, ***@***.***> wrote: That layer only appears to use TF ops that are readily available in Keras 3, so it should be simple to port it. Would you be willing to open a PR? — Reply to this email directly, view it on GitHub <#20551 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAAFILOYGS2GSM5LZ5R6XAL2CWVV3AVCNFSM6AAAAABSRPSXU6VHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTCMZZGM4DGMQ> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[james77777778]

FYI #20050

I previously developed some useful preprocessing/augmentation layers using keras.ops.
However, I don't have the bandwidth to port them to Keras.