This document contains a set of instructions on how to update your code towork with NumPy 2.0. It covers changes in NumPy’s Python and C APIs.
Note
Note that NumPy 2.0 also breaks binary compatibility - if you aredistributing binaries for a Python package that depends on NumPy’s C API,please see NumPy 2.0-specific advice.
Ruff plugin#
Many of the changes covered in the 2.0 release notes and in this migrationguide can be automatically adapted in downstream code with a dedicatedRuff rule, namely ruleNPY201.
You should install ruff>=0.4.8
and add the NPY201
rule to yourpyproject.toml
:
[tool.ruff.lint]select = ["NPY201"]
You can also apply the NumPy 2.0 rule directly from the command line:
$ ruff check path/to/code/ --select NPY201
Changes to NumPy data type promotion#
NumPy 2.0 changes promotion (the result of combining dissimilar data types)as per NEP 50. Please see the NEP for details on this change.It includes a table of example changes and a backwards compatibility section.
The largest backwards compatibility change is that the precision of scalarsis now preserved consistently. Two examples are:
np.float32(3) + 3.
now returns a float32 when it previously returneda float64.np.array([3], dtype=np.float32) + np.float64(3)
will now return a float64array. (The higher precision of the scalar is not ignored.)
For floating point values, this can lead to lower precision results whenworking with scalars. For integers, errors or overflows are possible.
To solve this, you may cast explicitly. Very often, it may also be a goodsolution to ensure you are working with Python scalars via int()
,float()
, or numpy_scalar.item()
.
To track down changes, you can enable emitting warnings for changed behavior(use warnings.simplefilter
to raise it as an error for a traceback):
np._set_promotion_state("weak_and_warn")
which is useful during testing. Unfortunately,running this may flag many changes that are irrelevant in practice.
Windows default integer#
The default integer used by NumPy is now 64bit on all 64bit systems (and32bit on 32bit system). For historic reasons related to Python 2 it waspreviously equivalent to the C long
type.The default integer is now equivalent to np.intp
.
Most end-users should not be affected by this change. Some operations willuse more memory, but some operations may actually become faster.If you experience issues due to calling a library written in a compiledlanguage it may help to explicitly cast to a long
, for example with:arr = arr.astype("long", copy=False)
.
Libraries interfacing with compiled code that are written in C, Cython, ora similar language may require updating to accommodate user input if theyare using the long
or equivalent type on the C-side.In this case, you may wish to use intp
and cast user input or supportboth long
and intp
(to better support NumPy 1.x as well).When creating a new integer array in C or Cython, the new NPY_DEFAULT_INT
macro will evaluate to either NPY_LONG
or NPY_INTP
depending on theNumPy version.
Note that the NumPy random API is not affected by this change.
C-API Changes#
Some definitions were removed or replaced due to being outdated orunmaintainable. Some new API definitions will evaluate differently atruntime between NumPy 2.0 and NumPy 1.x.Some are defined in numpy/_core/include/numpy/npy_2_compat.h
(for example NPY_DEFAULT_INT
) which can be vendored in full or partto have the definitions available when compiling against NumPy 1.x.
If necessary, PyArray_RUNTIME_VERSION >= NPY_2_0_API_VERSION
can beused to explicitly implement different behavior on NumPy 1.x and 2.0.(The compat header defines it in a way compatible with such use.)
Please let us know if you require additional workarounds here.
The PyArray_Descr
struct has been changed#
One of the most impactful C-API changes is that the PyArray_Descr
structis now more opaque to allow us to add additional flags and haveitemsizes not limited by the size of int
as well as allow improvingstructured dtypes in the future and not burden new dtypes with their fields.
Code which only uses the type number and other initial fields is unaffected.Most code will hopefully mainly access the ->elsize
field, when thedtype/descriptor itself is attached to an array (e.g. arr->descr->elsize
)this is best replaced with PyArray_ITEMSIZE(arr)
.
Where not possible, new accessor functions are required:
PyDataType_ELSIZE
andPyDataType_SET_ELSIZE
(note that the resultis nownpy_intp
and notint
).PyDataType_ALIGNMENT
PyDataType_FIELDS
,PyDataType_NAMES
,PyDataType_SUBARRAY
PyDataType_C_METADATA
Cython code should use Cython 3, in which case the change is transparent.(Struct access is available for elsize and alignment when compiling only forNumPy 2.)
For compiling with both 1.x and 2.x if you use these new accessors it isunfortunately necessary to either define them locally via a macro like:
#if NPY_ABI_VERSION < 0x02000000 #define PyDataType_ELSIZE(descr) ((descr)->elsize)#endif
or adding npy2_compat.h
into your code base and explicitly include itwhen compiling with NumPy 1.x (as they are new API).Including the file has no effect on NumPy 2.
Please do not hesitate to open a NumPy issue, if you require assistance orthe provided functions are not sufficient.
Custom User DTypes:Existing user dtypes must now use PyArray_DescrProto to definetheir dtype and slightly modify the code. See note in PyArray_RegisterDataType.
Functionality moved to headers requiring import_array()
#
If you previously included only ndarraytypes.h
you may find that somefunctionality is not available anymore and requires the inclusion ofndarrayobject.h
or similar.This include is also needed when vendoring npy_2_compat.h
into your owncodebase to allow use of the new definitions when compiling with NumPy 1.x.
Functionality which previously did not require import includes:
Functions to access dtype flags:
PyDataType_FLAGCHK
,PyDataType_REFCHK
, and the relatedNPY_BEGIN_THREADS_DESCR
.PyArray_GETITEM
andPyArray_SETITEM
.
Warning
It is important that the import_array()
mechanism is used to ensurethat the full NumPy API is accessible when using the npy_2_compat.h
header. In most cases your extension module probably already calls it.However, if not we have added PyArray_ImportNumPyAPI()
as a preferableway to ensure the NumPy API is imported. This function is light-weightwhen called multiple times so that you may insert it wherever it may beneeded (if you wish to avoid setting it up at module import).
Increased maximum number of dimensions#
The maximum number of dimensions (and arguments) was increased to 64. Thisaffects the NPY_MAXDIMS
and NPY_MAXARGS
macros.It may be good to review their use, and we generally encourage you tonot use these macros (especially NPY_MAXARGS
), so that a future version ofNumPy can remove this limitation on the number of dimensions.
NPY_MAXDIMS
was also used to signal axis=None
in the C-API, includingthe PyArray_AxisConverter
.The latter will return -2147483648
as an axis (the smallest integer value).Other functions may error withAxisError: axis 64 is out of bounds for array of dimension
in whichcase you need to pass NPY_RAVEL_AXIS
instead of NPY_MAXDIMS
.NPY_RAVEL_AXIS
is defined in the npy_2_compat.h
header and runtimedependent (mapping to 32 on NumPy 1.x and -2147483648
on NumPy 2.x).
Complex types - Underlying type changes#
The underlying C types for all of the complex types have been changed to usenative C99 types. While the memory layout of those types remains identicalto the types used in NumPy 1.x, the API is slightly different, since directfield access (like c.real
or c.imag
) is no longer possible.
It is recommended to use the functions npy_creal
and npy_cimag
(and the corresponding float and long double variants) to retrievethe real or imaginary part of a complex number, as these will work with bothNumPy 1.x and with NumPy 2.x. New functions npy_csetreal
andnpy_csetimag
, along with compatibility macros NPY_CSETREAL
andNPY_CSETIMAG
(and the corresponding float and long double variants),have been added for setting the real or imaginary part.
The underlying type remains a struct under C++ (all of the above still remainsvalid).
This has implications for Cython. It is recommened to always use the nativetypedefs cfloat_t
, cdouble_t
, clongdouble_t
rather than the NumPytypes npy_cfloat
, etc, unless you have to interface with C code writtenusing the NumPy types. You can still write cython code using the c.real
andc.imag
attributes (using the native typedefs), but you can no longer usein-place operators c.imag += 1
in Cython’s c++ mode.
Changes to namespaces#
In NumPy 2.0 certain functions, modules, and constants were moved or removedto make the NumPy namespace more user-friendly by removing unnecessary oroutdated functionality and clarifying which parts of NumPy are consideredprivate.Please see the tables below for guidance on migration. For most changes thismeans replacing it with a backwards compatible alternative.
Please refer to NEP 52 for more details.
Main namespace#
About 100 members of the main np
namespace have been deprecated, removed, ormoved to a new place. It was done to reduce clutter and establish only one way toaccess a given attribute. The table below shows members that have been removed:
removed member | migration guideline |
---|---|
add_docstring | It’s still available as |
add_newdoc | It’s still available as |
add_newdoc_ufunc | It’s an internal function and doesn’t have a replacement. |
alltrue | Use |
asfarray | Use |
byte_bounds | Now it’s available under |
cast | Use |
cfloat | Use |
clongfloat | Use |
compat | There’s no replacement, as Python 2 is no longer supported. |
complex_ | Use |
cumproduct | Use |
DataSource | It’s still available as |
deprecate | Emit |
deprecate_with_doc | Emit |
disp | Use your own printing function instead. |
fastCopyAndTranspose | Use |
find_common_type | Use |
get_array_wrap | |
float_ | Use |
geterrobj | Use the np.errstate context manager instead. |
Inf | Use |
Infinity | Use |
infty | Use |
issctype | Use |
issubclass_ | Use |
issubsctype | Use |
mat | Use |
maximum_sctype | Use a specific dtype instead. You should avoid relyingon any implicit mechanism and select the largest dtype ofa kind explicitly in the code. |
NaN | Use |
nbytes | Use |
NINF | Use |
NZERO | Use |
longcomplex | Use |
longfloat | Use |
lookfor | Search NumPy’s documentation directly. |
obj2sctype | Use |
PINF | Use |
product | Use |
PZERO | Use |
recfromcsv | Use |
recfromtxt | Use |
round_ | Use |
safe_eval | Use |
sctype2char | Use |
sctypes | Access dtypes explicitly instead. |
seterrobj | Use the np.errstate context manager instead. |
set_numeric_ops | For the general case, use |
set_string_function | Use |
singlecomplex | Use |
string_ | Use |
sometrue | Use |
source | Use |
tracemalloc_domain | It’s now available from |
unicode_ | Use |
who | Use an IDE variable explorer or |
If the table doesn’t contain an item that you were using but was removed in 2.0
,then it means it was a private member. You should either use the existing API or,in case it’s infeasible, reach out to us with a request to restore the removed entry.
The next table presents deprecated members, which will be removed in a release after 2.0
:
deprecated member | migration guideline |
---|---|
in1d | Use |
row_stack | Use |
trapz | Use |
Finally, a set of internal enums has been removed. As they weren’t used indownstream libraries we don’t provide any information on how to replace them:
[FLOATING_POINT_SUPPORT
, FPE_DIVIDEBYZERO
, FPE_INVALID
, FPE_OVERFLOW
,FPE_UNDERFLOW
, UFUNC_BUFSIZE_DEFAULT
, UFUNC_PYVALS_NAME
, CLIP
, WRAP
,RAISE
, BUFSIZE
, ALLOW_THREADS
, MAXDIMS
, MAY_SHARE_EXACT
,MAY_SHARE_BOUNDS
]
numpy.lib namespace#
Most of the functions available within np.lib
are also present in the mainnamespace, which is their primary location. To make it unambiguous how to access eachpublic function, np.lib
is now empty and contains only a handful of specialized submodules,classes and functions:
array_utils
,format
,introspect
,mixins
,npyio
andstride_tricks
submodules,Arrayterator
andNumpyVersion
classes,add_docstring
andadd_newdoc
functions,tracemalloc_domain
constant.
If you get an AttributeError
when accessing an attribute from np.lib
you shouldtry accessing it from the main np
namespace then. If an item is also missing fromthe main namespace, then you’re using a private member. You should either use the existingAPI or, in case it’s infeasible, reach out to us with a request to restore the removed entry.
numpy.core namespace#
The np.core
namespace is now officially private and has been renamed to np._core
.The user should never fetch members from the _core
directly - instead the mainnamespace should be used to access the attribute in question. The layout of the _core
module might change in the future without notice, contrary to public modules which adhereto the deprecation period policy. If an item is also missing from the main namespace,then you should either use the existing API or, in case it’s infeasible, reach out to uswith a request to restore the removed entry.
ndarray and scalar methods#
A few methods from np.ndarray
and np.generic
scalar classes have been removed.The table below provides replacements for the removed members:
expired member | migration guideline |
---|---|
newbyteorder | Use |
ptp | Use |
setitem | Use |
numpy.strings namespace#
A new numpy.strings namespace has been created, where most of the stringoperations are implemented as ufuncs. The old numpy.char namespace still isavailable, and, wherever possible, uses the new ufuncs for greater performance.We recommend using the strings functions going forward. Thechar namespace may be deprecated in the future.
Other changes#
Note about pickled files#
NumPy 2.0 is designed to load pickle files created with NumPy 1.26,and vice versa. For versions 1.25 and earlier loading NumPy 2.0pickle file will throw an exception.
Adapting to changes in the copy
keyword#
The copy keyword behavior changes inasarray, array and ndarray.__array__ may require these changes:
Code using
np.array(..., copy=False)
can in most cases be changed tonp.asarray(...)
. Older code tended to usenp.array
like this becauseit had less overhead than the defaultnp.asarray
copy-if-neededbehavior. This is no longer true, andnp.asarray
is the preferred function.For code that explicitly needs to pass
None
/False
meaning “copy ifneeded” in a way that’s compatible with NumPy 1.x and 2.x, seescipy#20172 for an exampleof how to do so.For any
__array__
method on a non-NumPy array-like object,dtype=None
andcopy=None
keywords must be added to the signature - this will work with olderNumPy versions as well (although older numpy versions will never pass incopy
keyword).If the keywords are added to the__array__
signature, then for:copy=True
and anydtype
value always return a new copy,copy=None
create a copy if required (for example bydtype
),copy=False
a copy must never be made. If a copy is needed to return a numpy arrayor satisfydtype
, then raise an exception (ValueError
).
Writing numpy-version-dependent code#
It should be fairly rare to have to write code that explicitly branches on thenumpy
version - in most cases, code can be rewritten to be compatible with1.x and 2.0 at the same time. However, if it is necessary, here is a suggestedcode pattern to use, using numpy.lib.NumpyVersion:
# example with AxisError, which is no longer available in# the main namespace in 2.0, and not available in the# `exceptions` namespace in <1.25.0 (example uses <2.0.0b1# for illustrative purposes):if np.lib.NumpyVersion(np.__version__) >= '2.0.0b1': from numpy.exceptions import AxisErrorelse: from numpy import AxisError
This pattern will work correctly including with NumPy release candidates, whichis important during the 2.0.0 release period.