Status of ragged tensors in tf nightly

[0x0L]

Hello,

It seems keras 3 (used in tf nightly) has dropped support for ragged tensors.

What are the plans for the future of ragged tensors in keras tf?

Thanks

[SuryanarayanaY]

Hi @0x0L ,

As Keras now become multi backend to support Pytorch, Jax as backend along with Tensorflow , there has been changes. Keras do have plan to support ragged tensors in future but not sure of exact time line.

Please have a look into Keras tickets #18467 and #18414 for more details.

Thanks!

[mihaimaruseac]

Meanwhile, you can still use tf.keras or directly the ragged tensors API from within TF though

[SuryanarayanaY]

@0x0L ,

If you are particularly looking to train a model with ragged input by using tf.keras, AFAIK, you can do it unitil TF2.14 version(Not sure about 2.15V yet). From Keras3(current tf-nightly keras package) onwards the Inputlayer(Input) don't accept ragged argument. I tested with a demo and attached gist here.

[0x0L]

@mihaimaruseac Actually no, it does not work in tf-nightly: tf.keras.Input does not support the ragged kwarg anymore as noted by @SuryanarayanaY (tf keras 2.15rc still has ragged support)

As mentioned in keras-team/keras#18467 they (keras) "may add it back later"... which does not sound too encouraging

[mihaimaruseac]

Try https://pypi.org/project/tf-keras-nightly/ instead of keras-nightly

[0x0L]

@mihaimaruseac Thank you so much !!

So I can expect ragged tensors to stay in tf.keras, is that right ?

[SuryanarayanaY]

Hi @0x0L ,

AFAIK, tf_keras will work for replacement of tf.keras . To use tf_keras with latest tf-nightly versions, we need to import tf_keras as keras and use keras instead of tf.keras since tf.keras still imports keras-nightly(i.e. Keras3) code instead of tf_keras code.

I tried the same demo above and it works importing tf_keras as keras and replace tf.keras with keras and attached gist here for reference.

Thanks to @mihaimaruseac for the inputs.

[google-ml-butler]

Are you satisfied with the resolution of your issue?
Yes
No