Roadmap for Static Type Hints in Low-level Algorithms

[chrisconlan]

Will scikit-learn accept pull requests for low-level code with static type hints? Does it ever plan to?

I don't know where the core team stands on static type hints in Python, but I think they make it 10x easier to grok code and convert users to contributors.

2018 seems like a great time to make this change, in light of Numpy dropping support for new features in Python 2.7. #11115

Supported version would need to change from Python (>= 2.7 or >= 3.4) to Python (>= 3.5) to support static type hints.

[rth]

That's an interesting question.

From what I understand, static typing won't be useful much (e.g. for mypy) as long as scipy and numpy don't have type annotations. Looking at numpy/numpy#7370 some annotations were added as a stand-alone package in https://github.com/numpy/numpy-stubs . If that works well, maybe a similar approach could be used for scikit-learn? Which mean that interested parties could start working on it (in a separate package) without waiting for #11115, I think...

[chrisconlan]

I think something like this could be done in Python 3.5+ without any extra dependencies:

import numpy as np
from typing import List, NewType

# Accepts either a one-dimensional array or a list of floats ...
Array1D = NewType('Array1D', np.ndarray, List[np.float64])

# Accepts either a two-dimensional array or a list of floats ...
Array2D = NewType('Array2D', np.ndarray, List[np.float64, np.float64])

# Define the function with appropriate type hints ...
def gradient_descent(some_input: Array2D, param_one: int, param_two:float) -> Array1D:
    # some logic ...

Just to clarify, this wouldn't help end users by catching compile-time typing errors. This would only be for readability.

I'd jump in on this for sure.

[chrisconlan]

P.S. I'm happy to see some big organizations supporting mypy. Type-annotated Python is definitely gonna be big: https://github.com/dropbox/mypy-PyCharm-plugin

[rth]

I guess, a start could be to some write stab files (.pyi) for the scikit-learn API (in a separate package) similarly to how it's done in https://github.com/numpy/numpy-stubs/blob/master/numpy-stubs/__init__.pyi for numpy. Some of it can be probably auto-generated.

There was some previous discussion about this in https://mail.python.org/pipermail/scikit-learn/2016-July/000306.html also see related discussion in #6599 (comment)

[rth]

I have been experimenting with mypy a bit, and in my experience results are mixed. It works quite well for pure python types and code relying on the stdlib where mypy stubs are built-in, though there are a few quirks and the general experience is still some way off from that of using a modern strongly typed language with type inference. For scientific computing, I think it is of limited use as numpy and scipy still don't have proper type annotations.

Still, I wonder if now that scikit-learn is Python 3.5+ only, starting to progressively add types at least on init parameters of estimators could be worthwhile. Given that keeping standalone stub files it in sync with the main repo would probably be somewhat difficult. cc @thomasjpfan

[jnothman]

I think type annotations for class parameters would be helpful.

[thomasjpfan]

Keeping stubs in sync would be hard to maintain. Placing them in __init__ using a new indentation style should work. We would need to come up with common union types to be use throughout the codebase, such as RandomStateType.

[adrinjalali]

Do we want to allow having hints for some input params and not the others?

[amueller]

I hadn't seen this issue before and I'm confused by the "low level algorithms" part. Our low level algorithms are all in Cython and already fully statically typed, right?

[adrinjalali]

Yeah I kinda assumed we don't really mean "low level" here. I guess the others did too.