Exporting type aliases for DtypeLike, ShapeLike and ArrayLike

[shoyer]

These would indicate anything NumPy can coerce into the respective type:

• DtypeLike can be coerced into/with np.dtype
• ShapeLike can be coerced into Tuple[int, ...]
• ArrayLike can be coerced into np.ndarray with np.array

See https://github.com/numpy/numpy_stubs/pull/10 for (currently private) aliases _DtypeLike and _ShapeLike.

It would be nice to export these publicly for the benefit of downstream users.

The main complication is that this would likely also require adding these aliases to NumPy itself. Potentially, this could be unnecessary with from __future__ import annotations in Python 3.7 (PEP 564).

[JelleZijlstra]

Another option is for this repo to define a separate package called numpy_typing or similar that includes a .py defining these, so that users could do from numpy_typing import DtypeLike. With PEP 561 implemented, mypy should be able to pick up the definitions of the aliases from numpy_typing.

[shoyer]

@JelleZijlstra That's a great alternative! Yes, we should definitely try the separate numpy_typing package to avoid the need to make upstream changes in NumPy yet.

[flying-sheep]

numpy-mypy has those: https://github.com/machinalis/mypy-data/tree/master/numpy-mypy

[person142]

This can be done without adding a separate package and without from __future__ import annotations (though that is certainly the neatest solution); the annotation just needs to be 'np.ShapeLike' (e.g. a string).

[person142]

I think we resolved all the issues here in https://github.com/numpy/numpy-stubs/pull/66.