Uninstall the PyPI keras so CI only imports the checkout

[MarcosAsh]

tensorflow and keras-mlx both pull keras in from PyPI, and that one has no mlx backend. This uninstalls it after installing keras-mlx, so if anything ends up importing keras from outside the checkout it fails instead of quietly running the wrong one.

numerical_test.py still needs PYTHONPATH=. since running a script puts its own folder on sys.path, not the cwd.

[hertschuh]

We should find a good pattern and do it across the repos.

A couple of questions:

• What problem is this addressing?
• Why did you need PYTHONPATH=. for the numerical test and not the other one?

My idea was to not install keras so that you would know that it's using the local source, which is why keras is not in requirements.txt, which is a little weird in some way.

What I didn't realize is that pip install -e . would install keras. But that's solvable.

• We can do pip uninstall keras after
• We can do pip install --no-deps -e . in the keras_mlx folder
• We can do pip install -e . in the keras folder first so that keras is already installed, but that might have some caveats

I'm not a big fan of this

import os, keras; assert os.path.realpath(keras.__file__).startswith(os.path.realpath('../keras'))

I'd rather make sure keras is not installed at all. But again, I'm not sure why you needed PYTHONPATH=.

[MarcosAsh]

Thanks for the review! The PyPI keras comes from tensorflow. pytest from ./keras gets the checkout because the cwd is on its path, but running numerical_test.py directly doesn't, so it picked up PyPI keras instead. I'll switch to --no-deps plus pip uninstall -y keras and drop the assert, makes sense to keep keras uninstalled.

[hertschuh]

I'll switch to --no-deps plus pip uninstall -y keras and drop the assert, makes sense to keep keras uninstalled.

I don't think you need both. The bulleted list was list of options to pick one from.

[MarcosAsh]

Ah got it, I read it as steps! Kept just the uninstall, tensorflow already pulls keras in from the requirements before keras-mlx gets installed, so --no-deps on its own wouldn't change anything. Just pushed.