Unparameterized `np.ndarray` typings produce "Type of ... is partially unknown" Pyright type errors.

[DylanLukes]

Problem

The type np.ndarray is stubbed in this library as:

class ndarray(_ArrayOrScalarCommon, Generic[_ShapeType, _DType_co]):
 ...

Throughout these stubs, the type np.ndarray is used without provided type parameters, seemingly with the expectation that this is treated as np.ndarray[Any, Any] (or more properly np.ndarray[object. object]). However, Pyright in strict mode alternately interprets this as np.ndarray[Unknown, Unknown].

As a result, every method that involves np.ndarray or a type alias which includes it produces a partially-unknown-type error:

Example Reproduction

For example, if we take the following simple file example.py...

import numpy as np

from typings.sklearn.linear_model import LinearRegression


def example():
 x = np.array([1, 2, 3, 4, 5])
 y = np.array([2, 4, 6, 8, 10])

 linreg = LinearRegression()
 linreg.fit(x, y)

❯ pyright src/path/to/example.py

src/path/to/example.py
 src/path/to/example/example.py:11:5 - error: Type of "fit" is partially unknown
 Type of "fit" is "(X: ndarray[Unknown, Unknown] | DataFrame | spmatrix | Buffer | _SupportsArray[dtype[Any]] | _NestedSequence[_SupportsArray[dtype[Any]]] | bool | int | float | complex | str | bytes | _NestedSequence[bool | int | float | complex | str | bytes], y: ndarray[Unknown, Unknown] | DataFrame | spmatrix | Buffer | _SupportsArray[dtype[Any]] | _NestedSequence[_SupportsArray[dtype[Any]]] | bool | int | float | complex | str | bytes | _NestedSequence[bool | int | float | complex | str | bytes], sample_weight: Buffer | _SupportsArray[dtype[Any]] | _NestedSequence[_SupportsArray[dtype[Any]]] | bool | int | float | complex | str | bytes | _NestedSequence[bool | int | float | complex | str | bytes] | None = None) -> LinearRegression" 

In this case, the error occurs because the type of fit is:

 def fit(
 self: LinearRegression_Self,
 X: MatrixLike | ArrayLike,
 y: MatrixLike | ArrayLike,
 sample_weight: None | ArrayLike = None,
 ) -> LinearRegression_Self:
 ...

And in turn MatrixLike is a (private) typealias that resolves to:

MatrixLike = np.ndarray | pd.DataFrame | spmatrix

Resolution

At least for this example, changing that type alias as follows resolves the type error.

MatrixLike = np.ndarray | pd.DataFrame | spmatrix

---

System Details:

OS: MacOS Sonoma 14.1.2
Python: CPython 3.12.1
Pyright: 1.1.358

Pyright configuration:

[tool.pyright]
include = ["./src", "./tests"]
stubPath = "./typings"

typeCheckingMode = "strict"
reportMissingImports = true
reportMissingTypeStubs = true

pythonVersion = "3.12"

[DylanLukes]

Two other things:

1. It may be the case that np.ndarray[???, PythonScalar] where ??? is something more appropriate for _ShapeType is a better choice. It appears that Numpy defined NDArray = ndarray[Any, dtype[_ScalarType_co]].
2. There are many other places in the codebase where bare np.ndarray are used, this is just one that covers a lot of cases on account of being a widely used type alias.

I've also tried:

PythonScalar = str | int | float | bool

ArrayLike = numpy.typing.ArrayLike
MatrixLike = numpy.typing.NDArray[PythonScalar] | pd.DataFrame | spmatrix

But this also produces unknown types in ndarray[Any, dtype[Unknown]] on account of:

_ScalarType_co = TypeVar("_ScalarType_co", bound=generic, covariant=True) # <-- this bit
_DType = TypeVar("_DType", bound=dtype[Any])
_DType_co = TypeVar("_DType_co", covariant=True, bound=dtype[Any])

NDArray = ndarray[Any, dtype[_ScalarType_co]]

So, np.ndarray[object, object] it is...

[debonte]

When I change MatrixLike to:

MatrixLike = np.ndarray[object, object] | pd.DataFrame | spmatrix

Pylance gives me a diagnostic on the second type argument:

Type "object" cannot be assigned to type variable "_DType_co@ndarray"
 Type "object" is incompatible with bound type "dtype[Any]" for type variable "_DType_co@ndarray"
 "object" is incompatible with "dtype[Any]" Pylance[reportInvalidTypeArguments]

How about using Any which would be equivalent to what we have today but would eliminate the reportUnknownMemberType diagnostic?

MatrixLike = np.ndarray[Any, Any] | pd.DataFrame | spmatrix

Feel free to submit a PR.

[heejaechang]

we might need better documentation for these errors? (https://microsoft.github.io/pyright/#/configuration?id=type-check-diagnostics-settings)

reportUnknownParameterType [boolean or string, optional]: Generate or suppress diagnostics for input or return parameters for functions or methods that have an unknown type. The default value for this setting is "none".

reportUnknownArgumentType [boolean or string, optional]: Generate or suppress diagnostics for call arguments for functions or methods that have an unknown type. The default value for this setting is "none".

reportUnknownLambdaType [boolean or string, optional]: Generate or suppress diagnostics for input or return parameters for lambdas that have an unknown type. The default value for this setting is "none".

reportUnknownVariableType [boolean or string, optional]: Generate or suppress diagnostics for variables that have an unknown type. The default value for this setting is "none".

reportUnknownMemberType [boolean or string, optional]: Generate or suppress diagnostics for class or instance variables that have an unknown type. The default value for this setting is "none".

this doc might help to understand about these errors

https://microsoft.github.io/pyright/#/typed-libraries?id=examples-of-known-ambiguous-and-unknown-types

[Avasam]

FWIW, the report is correct: it means that the type annotations you're using aren't complete enough to ensure type-safety.
Replacing the missing generics with Any just hides that unsafe-ness when it's not true that "any type matching the generic" really fits.
Moreso those "type is partially unknown" errors only show up in pyright/pylance in strict mode (or if you explicitly enable them).

Unfortunately numpy arrays and shape types are really hard to get right in libraries without being overly-restrictive ^^"

---

All that to say: Not a bug, but rather simply incomplete stubs. And your checker is correct in telling you that code is not type-safe in strict mode.