Typed automodel and type checking during tests.

[gabrieldemarmiesse]

Linked to #949

[gabrieldemarmiesse]

Do not merge, I need to add a test to ensure the types are correct.

[gabrieldemarmiesse]

We need to merge #960 before for this PR to work.

[gabrieldemarmiesse]

About how this work: Typeguard has a pytest plugin. This plugin will hook into all autokeras imports and run the typechecked decorator.

Meaning that during the tests, every time a function with type hints is called, the arguments are checked (as opposed to the default behavior in python, where type hints do nothing).

If the type hints are wrong, here is the result: https://github.com/keras-team/autokeras/runs/440879952

[gabrieldemarmiesse]

Ready for review

[haifeng-jin]

Thanks for the PR!

[haifeng-jin]

Please change Node to Input (ak.nodes.Input)

[haifeng-jin]

Change Block to Head (ak.engine.head.Head)

[haifeng-jin]

why are we pinning the version of this one?
Is it stable?

[gabrieldemarmiesse]

It's stable, I'm just a bit too careful sometimes haha

[haifeng-jin]

It should be Union[Head, Node, list]. I think you accidentally removed Node.
Another concern is how will it show on the docs by autogen.py?
Will it be head_module.Head directly? That would be confusing for the reader.

[gabrieldemarmiesse]

Thanks, I'll fix it.

Good question, I'll dig into it. I'm planning to change the code generation too to display the types in the description, not the signature as it's too cumbersome and hurts readability.

[gabrieldemarmiesse]

Here is what it looks like in the docs:

autokeras.AutoModel(
    inputs: Union[autokeras.nodes.Input, List[autokeras.nodes.Input]],
    outputs: Union[autokeras.engine.head.Head, list],
    name: str = "auto_model",
    max_trials: int = 100,
    directory: Union[str, pathlib.Path, None] = None,
    objective: str = "val_loss",
    tuner: Union[str, Type[autokeras.engine.tuner.AutoTuner]] = "greedy",
    overwrite: bool = False,
    seed: Optional[int] = None,
)

I can see two things to fix:

• Fist, use the aliases in the type hints instead of the complete class path.
• Second, display those info in the arguments descriptions only.

I'll open two issue in keras-autodoc. I'll fix the aliases issue first, and I believe we can merge this PR once it's done to avoid blocking the typing of autokeras.

The second issue is more or less making the docs beautiful, which is important, but not critical.

[gabrieldemarmiesse]

It looks like that now:

autokeras.AutoModel(
    inputs: Union[autokeras.Input, List[autokeras.Input]],
    outputs: Union[autokeras.Head, autokeras.Node, list],
    name: str = "auto_model",
    max_trials: int = 100,
    directory: Union[str, pathlib.Path, None] = None,
    objective: str = "val_loss",
    tuner: Union[str, Type[autokeras.engine.tuner.AutoTuner]] = "greedy",
    overwrite: bool = False,
    seed: Optional[int] = None,
)

AutoTuner is not listed in the docs, so there is no aliases, this is why keras-autodoc can't use an alias for it.

Do we plan to make AutoTuner public in the future?

[haifeng-jin]

Yes, we should make it public for anyone would like to override to implement their own tuner. I will merge this PR now.

Would you also help us make typedapi part of the setup.py tests dependencies? That would make the community easier to setup the environment. Thank you!

[gabrieldemarmiesse]

Sure thing, I'll upload it on pypi when I have the time :)