numpy.typing 2.2 breaks compatibility with type annotations for pydantic_numpy #59
Description
Not sure where to report really, but I'll start the discussion here.
So, prior to 2.2. we've had the following definition of numpy.typing.NDArray:
NDArray = ndarray[Any, dtype[_ScalarType_co]]
which is quite handy to use in our codebase to type return values of the functions,
e.g. npt.NDArray[np.uint8], but as you could see it doesn't carry any knowledge on number of dimensions used by this array.
in 2.2 they've changed the definition of NDArray to:
NDArray: TypeAlias = ndarray[_Shape, dtype[_ScalarType_co]]
where _Shape
_Shape: TypeAlias = tuple[int, ...]
and I can't see any concrete 1D, 2D, 3D array counterparts in numpy.typing
so then in 2.2 you get the following static analysis (pyright) error when you pass NDArray[np.uint8] to
numpy field in pydantic model:
[redacted] - error: Argument of type "NDArray[uint8]" cannot be assigned to parameter "img" of type "NpStrict3DArrayUint8" in function "__init__"
"ndarray[_Shape, dtype[uint8]]" is not assignable to "ndarray[tuple[int, int, int], dtype[uint8]]"
Type parameter "_ShapeT_co@ndarray" is covariant, but "_Shape" is not a subtype of "tuple[int, int, int]"
"tuple[int, ...]" is not assignable to "tuple[int, int, int]"
Tuple size mismatch; expected 3 but received indeterminate (reportArgumentType)
which is kinda make sense, as we're narrowing type to concrete dimesions.
and to me type definition prior to 2.2 should also trigger the same array (narrowing Any to tuple[int, int, int]), but it's not :P
So I'm basically looking for any words of wisdom here.
Is that something to report to pyright even?
Or we should have our own definitions of NDArray2D, NDArray3D?
Or we should relax static types in the lib itself?