ArrayLike limitations

[OriolAbril]

I am not completely sure how things have changed (if at all) between the different releases, but the current generation of type hints for ArrayLike can result is a broken type hint.

While NDArray or np.array support specifying the type with [dtype], ArrayLike does not. Attempting it results in an error:

np.typing.ArrayLike[float]
# TypeError: typing.Union[collections.abc.Buffer, numpy._typing._array_like._SupportsArray[numpy.dtype[typing.Any]], numpy._typing._nested_sequence._NestedSequence[numpy._typing._array_like._SupportsArray[numpy.dtype[typing.Any]]], complex, bytes, str, numpy._typing._nested_sequence._NestedSequence[complex | bytes | str]]
# is not a generic class

The examples folder clearly shows this behaviour and IIUC test_natlang_array is testing this happens.

Note: if necessary, I will try to generate a minimal example of behaviour (or different error) through pyi files. The error above happens in the command line or when attempting to import a function with that inline hint.

[lagru]

Ah, so ArrayLike can't be indexed / contain a dtype. I'll have to think a bit on how best to address this.

To be clear, the TypeError that you mention is happening at runtime? How pressing is this bug for you? It seems to me one can fix this by simply restricting usage to array-like (without the of part).

Thanks, for reporting!

[OriolAbril]

It is not pressing at all. As far as I can tell from my limited type hint knowledge, if the ArrayLike[float] are in pyi files it is generally parsed correctly by language servers so I see ArrayLike[float] in the help windows. mypy seems to ignore it silently in our pre-commit check but we have many things silenced so I am not sure this information is very useful.

I realized about the issue because I tried to refine a type hint using an inline one slightly different than the docstring info in which case I get the error above at import time.

---

Regarding ideal behaviour, I would very much prefer being able to use array-like of float in the docstrings and have the type hint ignore the of float part when generating the type hint. I think in the majority of cases things that are array-like are also quite flexible or agnostic when it comes to dtype, but there are several cases where it is relevant information and the array-like of dtype syntax is used very consistently. The main example that comes to mind are statistical distributions; all* distributions for numpy.random.Generator use array_like of floats or array_like of ints for the parameters of the distribution.

*: I haven't checked all of them exhaustively but I use many distributions and very frequently and out of the ones I use, the only exception I remember is multivatiate_normal.