numpy.typing

Typing (numpy.typing)

Large parts of the NumPy API have Pep 484-style type annotations. In addition a number of type aliases are available to users, most prominently the two below:

• ArrayLike: objects that can be converted to arrays

• DTypeLike: objects that can be converted to dtypes

Mypy plugin

Differences from the runtime NumPy API

NumPy is very flexible. Trying to describe the full range of possibilities statically would result in types that are not very helpful. For that reason, the typed NumPy API is often stricter than the runtime NumPy API. This section describes some notable differences.

ArrayLike

The ArrayLike type tries to avoid creating object arrays. For example,

>>> np.array(x**2 for x in range(10))
array(<generator object <genexpr> at ...>, dtype=object)

is valid NumPy code which will create a 0-dimensional object array. Type checkers will complain about the above example when using the NumPy types however. If you really intended to do the above, then you can either use a # type: ignore comment:

>>> np.array(x**2 for x in range(10))  # type: ignore

or explicitly type the array like object as Any:

>>> from typing import Any
>>> array_like: Any = (x**2 for x in range(10))
>>> np.array(array_like)
array(<generator object <genexpr> at ...>, dtype=object)

ndarray

It's possible to mutate the dtype of an array at runtime. For example, the following code is valid:

>>> x = np.array([1, 2])
>>> x.dtype = np.bool

This sort of mutation is not allowed by the types. Users who want to write statically typed code should instead use the numpy.ndarray.view method to create a view of the array with a different dtype.

DTypeLike

The DTypeLike type tries to avoid creation of dtype objects using dictionary of fields like below:

>>> x = np.dtype({"field1": (float, 1), "field2": (int, 3)})

Although this is valid NumPy code, the type checker will complain about it, since its usage is discouraged. Please see : Data type objects <arrays.dtypes>

Number precision

The precision of numpy.number subclasses is treated as a invariant generic parameter (see NBitBase), simplifying the annotating of processes involving precision-based casting.

>>> from typing import TypeVar
>>> import numpy as np
>>> import numpy.typing as npt

>>> T = TypeVar("T", bound=npt.NBitBase)
>>> def func(a: np.floating[T], b: np.floating[T]) -> np.floating[T]:
...     ...

Consequently, the likes of float16, float32 and float64 are still sub-types of floating, but, contrary to runtime, they're not necessarily considered as sub-classes.

from typing import overload
   import numpy as np

   @overload
   def phase(x: np.complex64) -> np.float32: ...
   @overload
   def phase(x: np.complex128) -> np.float64: ...
   @overload
   def phase(x: np.clongdouble) -> np.longdouble: ...
   def phase(x: np.complexfloating) -> np.floating:
       ...

Timedelta64

The timedelta64 class is not considered a subclass of signedinteger, the former only inheriting from generic while static type checking.

0D arrays

During runtime numpy aggressively casts any passed 0D arrays into their corresponding generic instance. Until the introduction of shape typing (see Pep 646) it is unfortunately not possible to make the necessary distinction between 0D and >0D arrays. While thus not strictly correct, all operations that can potentially perform a 0D-array -> scalar cast are currently annotated as exclusively returning an ndarray.

If it is known in advance that an operation will perform a 0D-array -> scalar cast, then one can consider manually remedying the situation with either typing.cast or a # type: ignore comment.

Record array dtypes

The dtype of numpy.recarray, and the routines.array-creation.rec functions in general, can be specified in one of two ways:

• Directly via the dtype argument.

• With up to five helper arguments that operate via numpy.rec.format_parser: formats, names, titles, aligned and byteorder.

These two approaches are currently typed as being mutually exclusive, i.e. if dtype is specified than one may not specify formats. While this mutual exclusivity is not (strictly) enforced during runtime, combining both dtype specifiers can lead to unexpected or even downright buggy behavior.

API