Keras `__init__.py` structure isn't readable by a static type checker about keras HOT 5 OPEN

Sorry I should have said, the code runs fine, it just can't be interpreted by a static type checker. Our users are confused by this because they can execute their code just fine, but Pylance doesn't think there's a module called keras.models. This makes it harder for them to write code using keras.models because there's no intellisense for it anymore.

from keras.

@mattdangerw @sampathweb can we resolve this?

from keras.

The code added in this commit dynamically adds the path to api to the sys path. This isn't something a static type checker can handle.

Basically, publicly-exported APIs are in keras/api/, rather than in the top-level keras/ folder. This is not unlike, e.g. TF APIs, which live in tf._api.v2. Can PyLance only handle packages where the directory structure matches the public namespace?

from keras.

Can PyLance only handle packages where the directory structure matches the public namespace?

We should be able to handle anything that's through static imports. The old code was importing all of the api modules, so Pylance could tell that those were imported in the __init__.py.

The code in the commit I referenced changed to add the api path to the sys.path at runtime. We don't execute any python code, so we can't actually figure out that api should be on the sys.path.

It's kind of similar to how we don't support PEP 660 because we can't execute the dynamic code in the .pth files.

Here's a similar example:
microsoft/pyright#6272

from keras.

We also have a problem with tensorflow too when it imports keras like this:

from tensorflow.keras import activations, initializers, layers, losses, metrics, models, optimizers, regularizers

This is because keras is added to the tensorflow package dynamically.

from keras.