[WIP] Add Numpy Backend

[ariG23498]

This PR will add NumPy backend to KerasCore

[fchollet]

Thanks for the PR!

[fchollet]

Thanks for the PR -- it looks good overall, what test failures do you see?

[ariG23498]

Every test realted to the Trainer fails (as NumPy should not have a trainer module). It would be great if someone from the team could help me understand whether the failing tests should be skipped or I should rather work on them to make them pass.

[AakashKumarNain]

I can take a look, and will try to help

[fchollet]

Every test realted to the Trainer fails (as NumPy should not have a trainer module). It would be great if someone from the team could help me understand whether the failing tests should be skipped or I should rather work on them to make them pass.

Normally every test that is "supposed" to fail should fail with a NotImplementedError exception. That's the way to tell.

[ariG23498]

Todos:

• Testing on metrics -- Confusion Metric to be specific
• Testing on layers
• Testing on ops -- All the image related utilities

[ariG23498]

@AakashKumarNain if you want to take up the image.py file that would be of immense help! I am having a hard time to aligning the resize methods with PIL.

[ariG23498]

@fchollet would it be okay if I introduced the numpy test in the GitHub Actions as well?

And to do that I think I would just need to add numpy here

keras-core/.github/workflows/actions.yml

Line 14 in ce9b511

backend: [tensorflow, jax, torch]

Am I correct here?

[ariG23498]

@fchollet some pointers:

• In the current state only tests for rnn, lstm, and gru fail. How would you like me to proceed with this?
• For the image.py I have use JAX (also suggested by @AakashKumarNain) against PIL. To begin with PIL cannot handle float32 with RGB images, conversions are lossy and inherently tests always fail. Using JAX also makes sense here as the conv operation in numpy already uses JAX (making it a dependency). [WIP] Add Numpy Backend #87 (comment)
• [WIP] Add Numpy Backend #87 (comment)

[AakashKumarNain]

It is a known issue that resizing with PIL though most correct, is impossible to reproduce in opencv or any other library per se. Every library has very minor difference for different interpolations.

[fchollet]

Let's carry on in the other PR!