Dataset handling + yet unagnostic details (Question/feature request)

[jchwenger]

Hi there,

First off, congratulations on the new edition of Deep Learning With Python, it's good to see the material refreshed!

A quick question came up (I'm updating course material based on the book): currently I see that quite a few of the examples still use tf.data.Dataset (for convenience I suppose, and inheriting from the previous edition), which is perhaps the single major break remaining from the agnostic direction of Keras 3. I found PyDataset somewhere in the doc: this is meant to be used to make dataset handling backend-agnostic as well? If so, wouldn't that be a good addition to the book, instead of tf?

I also notice that something like tf.reshape is used here and there, and even though there is a keras.layers.Reshape, I didn't find something like a keras.ops.reshape (that maybe would invoke tf.reshape, torch.Tensor.view and jax.numpy.reshape/jax.Array.reshape depending on the backend)? Silly me, obviously it's in keras.ops.reshape, can't believe I missed it, sorry about that.

Thanks for reading,
Best,
Jeremie