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 ?

[vidlb]

Hi @daspk04 , I've pushed releases candidate images of OTBTF including TF 2.18 and OTB 9.1.
We still need to update some python scripts to Keras 3, but if you already wanna try something with otb apps you may use one of the following images :

• registry.forgemia.inra.fr/orfeo-toolbox/otbtf:4.4.0-rc1-cpu
• registry.forgemia.inra.fr/orfeo-toolbox/otbtf:4.4.0-rc1-gpu

I've made a lot of changes in the build process, and hope I didn't break anything in the TF install, thus any feedback would be much appreciated !

Cheers