Support for TF 2.16.0

[daspk04]

Hi @remicres !

The last week tensorflow released the 2.16.0-rc version. One interesting point is that Keras 3 will be the default version. Keras 3 seems quite interesting, it supports multi framework (tesorflow, pytorch, jax).

Use data pipelines from any source. The Keras 3 fit()/evaluate()/predict() routines are compatible with tf.data.Dataset objects, with PyTorch DataLoader objects, with NumPy arrays, Pandas dataframes — regardless of the backend you're using. You can train a Keras 3 + TensorFlow model on a PyTorch DataLoader or train a Keras 3 + PyTorch model on a tf.data.Dataset.

So then I'm assuming that we would be able to run directly any model written in pytorch with OTBTF as well?

[remicres]

Hi @Pratyush1991 ,

It looks like, yes... thanks for the information, I have to dig into that. If somebody would be kind enough to provide us a minimal working example, it would be so great.

Like

• train a pytorch model from otbtf.dataset
• perform an inference from a SavedModel created from a trained pytorch model
• ...

[daspk04]

Thanks @remicres!

I might be able do it. At-least based on current OTBTF version:

• I can create a simple pytorch model which would just be trained based the patches! Actually tfrecords should work as well as TorchData supports it. c.f
• Possibly then that model can be tested with TF 2.16.0 ?