[Dependencies] Update tensorflow and numpy

[timlrx]

Is there an existing issue for this?

• I have searched the existing issues

Description

Currently v2.x still uses an older version of TensorFlow and Numpy for backward compatibility issues. Suggest considering bumping the version in the next release update.

The latest version of Tensorflow only supports Numpy 1, so we would have to wait for the next version (apparently releasing around the end of the month) for Numpy 2 support - https://blog.tensorflow.org/2024/07/whats-new-in-tensorflow-217.html

For Tensorflow, moving from 2.14.0 to 2.17.0 is a major change.

Tensorflow 2.16.0+ uses keras 3 instead of Keras 2 (see compatibility matrix at the end over here - https://keras.io/getting_started/). Keras 3 model format (.keras) is not compatible with Keras 2 (.keras)...

It also supports TensorFlow format (SavedModels and .h5) but through a lower level API - keras.layers.TFSMLayer(saved_model, call_endpoint='serving_default') that does not expose predict and predict_prob methods out of the box.

To bump the version, we will have to also update the TensorFlow serializer to support the other formats. Suggest to take this up in a subsequent release (can look at broader support for more commonly used packages e.g. PyTorch as well) and we can bump Numpy and Tensorflow then.