Corrections for gradient centralization example.

[albertoesmp]

The gradient centralization example used the model trained without gradient centralization (NGC) when applying gradient centralization (GC). Thus, the results were better because the GC model wasn't trained from scratch but started from the trained NGC model. This fix uses independent models for GC and NGC and thus yields a fair comparison between both strategies.

[google-cla]

Thanks for your pull request! It looks like this may be your first contribution to a Google open source project. Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA).

View this failed invocation of the CLA check for more information.

For the most up to date status, view the checks section at the bottom of the pull request.

[JyotinderSingh]

@albertoesmp Thank you for this fix, could you please resolve the CLA issue mentioned in the comment above?

[gemini-code-assist]

Code Review

This pull request correctly addresses a significant issue in the example where the same model instance was reused for training with and without Gradient Centralization (GC), leading to an unfair comparison. The introduction of a make_model() factory function is an excellent fix for this.

However, my review has uncovered a more critical issue that persists after these changes. The GCRMSprop optimizer is implemented by overriding the get_gradients method. This method is part of a legacy Keras API and is no longer called during the training loop in Keras 3. As a result, the Gradient Centralization logic is never actually applied. The example is currently comparing a model with a standard RMSprop optimizer against another model with the exact same standard RMSprop optimizer.

To fix this, the optimizer customization needs to be updated to the Keras 3 API, likely by overriding the compute_gradients method. I've left detailed comments with a suggested implementation in the relevant files. Addressing this is crucial for the example to correctly demonstrate the effects of Gradient Centralization.