NumPy 1.25.0 Release Notes

The NumPy 1.25.0 release continues the ongoing work to improve the handling and promotion of dtypes, increase the execution speed, and clarify the documentation. There has also been work to prepare for the future NumPy 2.0.0 release, resulting in a large number of new and expired deprecation. Highlights are:

• Support for MUSL, there are now MUSL wheels.

• Support the Fujitsu C/C++ compiler.

• Object arrays are now supported in einsum

• Support for inplace matrix multiplication (@=).

We will be releasing a NumPy 1.26 when Python 3.12 comes out. That is needed because distutils has been dropped by Python 3.12 and we will be switching to using meson for future builds. The next mainline release will be NumPy 2.0.0. We plan that the 2.0 series will still support downstream projects built against earlier versions of NumPy.

The Python versions supported in this release are 3.9-3.11.

Deprecations

• np.core.MachAr is deprecated. It is private API. In names defined in np.core should generally be considered private.

  (gh-22638)

• np.finfo(None) is deprecated.

  (gh-23011)

• np.round_ is deprecated. Use np.round instead.

  (gh-23302)

• np.product is deprecated. Use np.prod instead.

  (gh-23314)

• np.cumproduct is deprecated. Use np.cumprod instead.

  (gh-23314)

• np.sometrue is deprecated. Use np.any instead.

  (gh-23314)

• np.alltrue is deprecated. Use np.all instead.

  (gh-23314)

• Only ndim-0 arrays are treated as scalars. NumPy used to treat all arrays of size 1 (e.g., np.array([3.14])) as scalars. In the future, this will be limited to arrays of ndim 0 (e.g., np.array(3.14)). The following expressions will report a deprecation warning:

  a = np.array([3.14])
  float(a)  # better: a[0] to get the numpy.float or a.item()
  
  b = np.array([[3.14]])
  c = numpy.random.rand(10)
  c[0] = b  # better: c[0] = b[0, 0]
  

  (gh-10615)

• np.find_common_type is deprecated. numpy.find_common_type is now deprecated and its use should be replaced with either numpy.result_type or numpy.promote_types. Most users leave the second scalar_types argument to find_common_type as [] in which case np.result_type and np.promote_types are both faster and more robust. When not using scalar_types the main difference is that the replacement intentionally converts non-native byte-order to native byte order. Further, find_common_type returns object dtype rather than failing promotion. This leads to differences when the inputs are not all numeric. Importantly, this also happens for e.g. timedelta/datetime for which NumPy promotion rules are currently sometimes surprising.

  When the scalar_types argument is not [] things are more complicated. In most cases, using np.result_type and passing the Python values 0, 0.0, or 0j has the same result as using int, float, or complex in scalar_types.

  When scalar_types is constructed, np.result_type is the correct replacement and it may be passed scalar values like np.float32(0.0). Passing values other than 0, may lead to value-inspecting behavior (which np.find_common_type never used and NEP 50 may change in the future). The main possible change in behavior in this case, is when the array types are signed integers and scalar types are unsigned.

  If you are unsure about how to replace a use of scalar_types or when non-numeric dtypes are likely, please do not hesitate to open a NumPy issue to ask for help.

  (gh-22539)

Expired deprecations

• np.core.machar and np.finfo.machar have been removed.

  (gh-22638)

• +arr will now raise an error when the dtype is not numeric (and positive is undefined).

  (gh-22998)

• A sequence must now be passed into the stacking family of functions (stack, vstack, hstack, dstack and column_stack).

  (gh-23019)

• np.clip now defaults to same-kind casting. Falling back to unsafe casting was deprecated in NumPy 1.17.

  (gh-23403)

• np.clip will now propagate np.nan values passed as min or max. Previously, a scalar NaN was usually ignored. This was deprecated in NumPy 1.17.

  (gh-23403)

• The np.dual submodule has been removed.

  (gh-23480)

• NumPy now always ignores sequence behavior for an array-like (defining one of the array protocols). (Deprecation started NumPy 1.20)

  (gh-23660)

• The niche FutureWarning when casting to a subarray dtype in astype or the array creation functions such as asarray is now finalized. The behavior is now always the same as if the subarray dtype was wrapped into a single field (which was the workaround, previously). (FutureWarning since NumPy 1.20)

  (gh-23666)

• == and != warnings have been finalized. The == and != operators on arrays now always:

  • raise errors that occur during comparisons such as when the arrays have incompatible shapes (np.array([1, 2]) == np.array([1, 2, 3])).

  • return an array of all True or all False when values are fundamentally not comparable (e.g. have different dtypes). An example is np.array(["a"]) == np.array([1]).

    This mimics the Python behavior of returning False and True when comparing incompatible types like "a" == 1 and "a" != 1. For a long time these gave DeprecationWarning or FutureWarning.

  (gh-22707)

• Nose support has been removed. NumPy switched to using pytest in 2018 and nose has been unmaintained for many years. We have kept NumPy’s nose support to avoid breaking downstream projects who might have been using it and not yet switched to pytest or some other testing framework. With the arrival of Python 3.12, unpatched nose will raise an error. It is time to move on.

  Decorators removed:

  • raises

  • slow

  • setastest

  • skipif

  • knownfailif

  • deprecated

  • parametrize

  • _needs_refcount

  These are not to be confused with pytest versions with similar names, e.g., pytest.mark.slow, pytest.mark.skipif, pytest.mark.parametrize.

  Functions removed:

  • Tester

  • import_nose

  • run_module_suite

  (gh-23041)

• The numpy.testing.utils shim has been removed. Importing from the numpy.testing.utils shim has been deprecated since 2019, the shim has now been removed. All imports should be made directly from numpy.testing.

  (gh-23060)

• The environment variable to disable dispatching has been removed. Support for the NUMPY_EXPERIMENTAL_ARRAY_FUNCTION environment variable has been removed. This variable disabled dispatching with __array_function__.

  (gh-23376)

• Support for y= as an alias of out= has been removed. The fix, isposinf and isneginf functions allowed using y= as a (deprecated) alias for out=. This is no longer supported.

  (gh-23376)

Compatibility notes

• The busday_count method now correctly handles cases where the begindates is later in time than the enddates. Previously, the enddates was included, even though the documentation states it is always excluded.

  (gh-23229)

• When comparing datetimes and timedelta using np.equal or np.not_equal numpy previously allowed the comparison with casting="unsafe". This operation now fails. Forcing the output dtype using the dtype kwarg can make the operation succeed, but we do not recommend it.

  (gh-22707)

• When loading data from a file handle using np.load, if the handle is at the end of file, as can happen when reading multiple arrays by calling np.load repeatedly, numpy previously raised ValueError if allow_pickle=False, and OSError if allow_pickle=True. Now it raises EOFError instead, in both cases.

  (gh-23105)

np.pad with mode=wrap pads with strict multiples of original data

Code based on earlier version of pad that uses mode="wrap" will return different results when the padding size is larger than initial array.

np.pad with mode=wrap now always fills the space with strict multiples of original data even if the padding size is larger than the initial array.

(gh-22575)

Cython long_t and ulong_t removed

long_t and ulong_t were aliases for longlong_t and ulonglong_t and confusing (a remainder from of Python 2). This change may lead to the errors:

'long_t' is not a type identifier
'ulong_t' is not a type identifier

We recommend use of bit-sized types such as cnp.int64_t or the use of cnp.intp_t which is 32 bits on 32 bit systems and 64 bits on 64 bit systems (this is most compatible with indexing). If C long is desired, use plain long or npy_long. cnp.int_t is also long (NumPy’s default integer). However, long is 32 bit on 64 bit windows and we may wish to adjust this even in NumPy. (Please do not hesitate to contact NumPy developers if you are curious about this.)

(gh-22637)

Changed error message and type for bad axes argument to ufunc

The error message and type when a wrong axes value is passed to ufunc(..., axes=[...])` has changed. The message is now more indicative of the problem, and if the value is mismatched an AxisError will be raised. A TypeError will still be raised for invalid input types.

(gh-22675)

Array-likes that define __array_ufunc__ can now override ufuncs if used as where

If the where keyword argument of a numpy.ufunc is a subclass of numpy.ndarray or is a duck type that defines numpy.class.__array_ufunc__ it can override the behavior of the ufunc using the same mechanism as the input and output arguments. Note that for this to work properly, the where.__array_ufunc__ implementation will have to unwrap the where argument to pass it into the default implementation of the ufunc or, for numpy.ndarray subclasses before using super().__array_ufunc__.

(gh-23240)

Compiling against the NumPy C API is now backwards compatible by default

NumPy now defaults to exposing a backwards compatible subset of the C-API. This makes the use of oldest-supported-numpy unnecessary. Libraries can override the default minimal version to be compatible with using:

#define NPY_TARGET_VERSION NPY_1_22_API_VERSION

before including NumPy or by passing the equivalent -D option to the compiler. The NumPy 1.25 default is NPY_1_19_API_VERSION. Because the NumPy 1.19 C API was identical to the NumPy 1.16 one resulting programs will be compatible with NumPy 1.16 (from a C-API perspective). This default will be increased in future non-bugfix releases. You can still compile against an older NumPy version and run on a newer one.

For more details please see For downstream package authors.

(gh-23528)

New Features

np.einsum now accepts arrays with object dtype

The code path will call python operators on object dtype arrays, much like np.dot and np.matmul.

(gh-18053)

Add support for inplace matrix multiplication

It is now possible to perform inplace matrix multiplication via the @= operator.

>>> import numpy as np

>>> a = np.arange(6).reshape(3, 2)
>>> print(a)
[[0 1]
 [2 3]
 [4 5]]

>>> b = np.ones((2, 2), dtype=int)
>>> a @= b
>>> print(a)
[[1 1]
 [5 5]
 [9 9]]

(gh-21120)

Added NPY_ENABLE_CPU_FEATURES environment variable

Users may now choose to enable only a subset of the built CPU features at runtime by specifying the NPY_ENABLE_CPU_FEATURES environment variable. Note that these specified features must be outside the baseline, since those are always assumed. Errors will be raised if attempting to enable a feature that is either not supported by your CPU, or that NumPy was not built with.

(gh-22137)

NumPy now has an np.exceptions namespace

NumPy now has a dedicated namespace making most exceptions and warnings available. All of these remain available in the main namespace, although some may be moved slowly in the future. The main reason for this is to increase discoverability and add future exceptions.

(gh-22644)

np.linalg functions return NamedTuples

np.linalg functions that return tuples now return namedtuples. These functions are eig(), eigh(), qr(), slogdet(), and svd(). The return type is unchanged in instances where these functions return non-tuples with certain keyword arguments (like svd(compute_uv=False)).

(gh-22786)

String functions in np.char are compatible with NEP 42 custom dtypes

Custom dtypes that represent unicode strings or byte strings can now be passed to the string functions in np.char.

(gh-22863)

String dtype instances can be created from the string abstract dtype classes

It is now possible to create a string dtype instance with a size without using the string name of the dtype. For example, type(np.dtype('U'))(8) will create a dtype that is equivalent to np.dtype('U8'). This feature is most useful when writing generic code dealing with string dtype classes.

(gh-22963)

Fujitsu C/C++ compiler is now supported

Support for Fujitsu compiler has been added. To build with Fujitsu compiler, run:

python setup.py build -c fujitsu

SSL2 is now supported

Support for SSL2 has been added. SSL2 is a library that provides OpenBLAS compatible GEMM functions. To enable SSL2, it need to edit site.cfg and build with Fujitsu compiler. See site.cfg.example.

(gh-22982)

Improvements

NDArrayOperatorsMixin specifies that it has no __slots__

The NDArrayOperatorsMixin class now specifies that it contains no __slots__, ensuring that subclasses can now make use of this feature in Python.

(gh-23113)

Fix power of complex zero

np.power now returns a different result for 0^{non-zero} for complex numbers. Note that the value is only defined when the real part of the exponent is larger than zero. Previously, NaN was returned unless the imaginary part was strictly zero. The return value is either 0+0j or 0-0j.

(gh-18535)

New DTypePromotionError

NumPy now has a new DTypePromotionError which is used when two dtypes cannot be promoted to a common one, for example:

np.result_type("M8[s]", np.complex128)

raises this new exception.

(gh-22707)

np.show_config uses information from Meson

Build and system information now contains information from Meson. np.show_config now has a new optional parameter mode to help customize the output.

(gh-22769)

Fix np.ma.diff not preserving the mask when called with arguments prepend/append.

Calling np.ma.diff with arguments prepend and/or append now returns a MaskedArray with the input mask preserved.

Previously, a MaskedArray without the mask was returned.

(gh-22776)

Corrected error handling for NumPy C-API in Cython

Many NumPy C functions defined for use in Cython were lacking the correct error indicator like except -1 or except *. These have now been added.

(gh-22997)

Ability to directly spawn random number generators

numpy.random.Generator.spawn now allows to directly spawn new independent child generators via the numpy.random.SeedSequence.spawn mechanism. numpy.random.BitGenerator.spawn does the same for the underlying bit generator.

Additionally, numpy.random.BitGenerator.seed_seq now gives direct access to the seed sequence used for initializing the bit generator. This allows for example:

seed = 0x2e09b90939db40c400f8f22dae617151
rng = np.random.default_rng(seed)
child_rng1, child_rng2 = rng.spawn(2)

# safely use rng, child_rng1, and child_rng2

Previously, this was hard to do without passing the SeedSequence explicitly. Please see numpy.random.SeedSequence for more information.

(gh-23195)

numpy.logspace now supports a non-scalar base argument

The base argument of numpy.logspace can now be array-like if it is broadcastable against the start and stop arguments.

(gh-23275)

np.ma.dot() now supports for non-2d arrays

Previously np.ma.dot() only worked if a and b were both 2d. Now it works for non-2d arrays as well as np.dot().

(gh-23322)

Explicitly show keys of .npz file in repr

NpzFile shows keys of loaded .npz file when printed.

>>> npzfile = np.load('arr.npz')
>>> npzfile
NpzFile 'arr.npz' with keys arr_0, arr_1, arr_2, arr_3, arr_4...

(gh-23357)

NumPy now exposes DType classes in np.dtypes

The new numpy.dtypes module now exposes DType classes and will contain future dtype related functionality. Most users should have no need to use these classes directly.

(gh-23358)

Drop dtype metadata before saving in .npy or .npz files

Currently, a *.npy file containing a table with a dtype with metadata cannot be read back. Now, np.save and np.savez drop metadata before saving.

(gh-23371)

numpy.lib.recfunctions.structured_to_unstructured returns views in more cases

structured_to_unstructured now returns a view, if the stride between the fields is constant. Prior, padding between the fields or a reversed field would lead to a copy. This change only applies to ndarray, memmap and recarray. For all other array subclasses, the behavior remains unchanged.

(gh-23652)

Signed and unsigned integers always compare correctly

When uint64 and int64 are mixed in NumPy, NumPy typically promotes both to float64. This behavior may be argued about but is confusing for comparisons ==, <=, since the results returned can be incorrect but the conversion is hidden since the result is a boolean. NumPy will now return the correct results for these by avoiding the cast to float.

(gh-23713)

Performance improvements and changes

Faster np.argsort on AVX-512 enabled processors

32-bit and 64-bit quicksort algorithm for np.argsort gain up to 6x speed up on processors that support AVX-512 instruction set.

Thanks to Intel corporation for sponsoring this work.

(gh-23707)

Faster np.sort on AVX-512 enabled processors

Quicksort for 16-bit and 64-bit dtypes gain up to 15x and 9x speed up on processors that support AVX-512 instruction set.

Thanks to Intel corporation for sponsoring this work.

(gh-22315)

__array_function__ machinery is now much faster

The overhead of the majority of functions in NumPy is now smaller especially when keyword arguments are used. This change significantly speeds up many simple function calls.

(gh-23020)

ufunc.at can be much faster

Generic ufunc.at can be up to 9x faster. The conditions for this speedup:

• operands are aligned

• no casting

If ufuncs with appropriate indexed loops on 1d arguments with the above conditions, ufunc.at can be up to 60x faster (an additional 7x speedup). Appropriate indexed loops have been added to add, subtract, multiply, floor_divide, maximum, minimum, fmax, and fmin.

The internal logic is similar to the logic used for regular ufuncs, which also have fast paths.

Thanks to the D. E. Shaw group for sponsoring this work.

(gh-23136)

Faster membership test on NpzFile

Membership test on NpzFile will no longer decompress the archive if it is successful.

(gh-23661)

Changes

np.r_[] and np.c_[] with certain scalar values

In rare cases, using mainly np.r_ with scalars can lead to different results. The main potential changes are highlighted by the following:

>>> np.r_[np.arange(5, dtype=np.uint8), -1].dtype
int16  # rather than the default integer (int64 or int32)
>>> np.r_[np.arange(5, dtype=np.int8), 255]
array([  0,   1,   2,   3,   4, 255], dtype=int16)

Where the second example returned:

array([ 0,  1,  2,  3,  4, -1], dtype=int8)

The first one is due to a signed integer scalar with an unsigned integer array, while the second is due to 255 not fitting into int8 and NumPy currently inspecting values to make this work. (Note that the second example is expected to change in the future due to NEP 50; it will then raise an error.)

(gh-22539)

Most NumPy functions are wrapped into a C-callable

To speed up the __array_function__ dispatching, most NumPy functions are now wrapped into C-callables and are not proper Python functions or C methods. They still look and feel the same as before (like a Python function), and this should only improve performance and user experience (cleaner tracebacks). However, please inform the NumPy developers if this change confuses your program for some reason.

(gh-23020)

C++ standard library usage

NumPy builds now depend on the C++ standard library, because the numpy.core._multiarray_umath extension is linked with the C++ linker.

(gh-23601)