Upgrading numpy 1 to 2

Removes numpy.distutils, an expired deprecation.

Removes support for 2-dimensional vectors in numpy.cross.

Removes numpy._core.numerictypes.maximum_sctype.

Removes numpy.row_stack in favor of numpy.vstack.

Removes get_array_wrap.

Removes recfromtxt and recfromcsv from numpy.lib._npyio in favor of numpy.genfromtxt.

Removes the numpy.chararray re-export of numpy.char.chararray.

Removes the numpy.lib.math alias for the standard library math module.

Removes the data type alias 'a' in favor of 'S'.

Removes _add_newdoc_ufunc in favor of setting ufunc.__doc__ directly.

numpy.where no longer truncates out-of-range Python integers and raises OverflowError instead.

Removes support for scalar types in numpy.ctypeslib.as_ctypes.

Removes the deprecated MachAr runtime discovery mechanism.

Removes numpy.linalg.linalg and numpy.fft.helper public modules.

Removes the interpolation parameter from quantile and percentile functions.

Removes numpy.in1d in favor of numpy.isin.

Removes the numpy.ndindex.ndincr() method.

Removes the fix_imports parameter from numpy.save.

Removes the deprecated newshape parameter from numpy.reshape.

Removes deprecated numpy.trapz, disp, and the bias and ddof arguments of numpy.corrcoef.

Removes the deprecated delimitor parameter from numpy.ma.mrecords.fromtextfile().

Removes the style parameter from numpy.array2string and makes later arguments keyword-only; np.sum on a generator now raises TypeError.

Type-checkers no longer accept start as a keyword argument to numpy.arange, though it still works at runtime.

Removes the semi-public NPY_ALIGNMENT_REQUIRED macro.

Raises the default x86 cpu-baseline to x86-64-v2 microarchitecture, dropping SIMD support for pre-2009 processors.

numpy.typing.DTypeLike no longer accepts None; use DTypeLike | None instead.

Removes numpy/npy_1_7_deprecated_api.h and C macros like NPY_OWNDATA in favor of NPY_ARRAY_OWNDATA.

Removes aliases generate_divbyzero_error and generate_overflow_error in favor of npy_set_floatstatus_* functions.

Removes np.tostring, deprecated since 1.19.

datetime64 and timedelta64 tuple construction no longer accepts an event value; use a two-tuple or four-tuple.

Removes np.tostring, use tobytes instead.

concatenate() with axis=None now uses same-kind casting by default instead of unsafe.

np.bool scalars can no longer be interpreted as an index.

Parsing an integer via a float string is no longer supported.

Removes special matrix handling in np.outer; convert to an ndarray via matrix.A.

Removes the np.compat package source code.

Removes NEP 50 promotion state settings; _set_promotion_state and _get_promotion_state are removed and non-weak NPY_PROMOTION_STATE warns.

Hides API symbols NumPy adds by default, so the NumPy API can no longer be dynamically fetched from another library.

Removes many old shims and helper functions from npy_3kcompat.h.

Removes many non-recommended functions and aliases, reducing objects in the main namespace by about 10% and in numpy.lib by about 80%.

Removes many outdated C API functions and macros and hides private internals to ease future extensibility.

Changes the default integer type on Windows to int64 rather than int32, matching other platforms.

Removes np.geterrobj, np.seterrobj and the ufunc extobj= keyword argument; use the np.errstate() context manager instead.

Removes np.cast; use np.asarray(arg, dtype=dtype) instead.

Removes np.source; use inspect.getsource instead.

Removes np.lookfor.

Removes numpy.who; use a variable explorer in an IDE such as Spyder or Jupyter Notebook.

Warnings and exceptions in numpy.exceptions are no longer exposed in the main namespace.

Removes niche enums, expired members and functions such as ERR_*, SHIFT_*, np.fastCopyAndTranspose, np.kernel_version, np.numarray, np.oldnumeric and np.set_numeric_ops.

Replaces wildcard imports in numpy/__init__.py with explicit imports, removing members such as np.NINF, np.PINF, np.CLIP, np.WRAP, np.RAISE, np.BUFSIZE, np.MAXDIMS, np.add_docstring and…

Removes alias np.float_; use np.float64 instead.

Removes alias np.complex_; use np.complex128 instead.

Removes alias np.longfloat; use np.longdouble instead.

Removes alias np.singlecomplex; use np.complex64 instead.

Removes alias np.cfloat; use np.complex128 instead.

Removes alias np.longcomplex; use np.clongdouble instead.

Removes alias np.clongfloat; use np.clongdouble instead.

Removes alias np.string_; use np.bytes_ instead.

Removes alias np.unicode_; use np.str_ instead.

Removes alias np.Infinity; use np.inf instead.

Removes alias np.NaN; use np.nan instead.

Removes the alias np.mat; use np.asmatrix instead.

Removes np.issubclass_; use the issubclass builtin instead.

Removes np.asfarray; use np.asarray with a proper dtype instead.

Removes np.set_string_function; use np.set_printoptions with a formatter instead.

Removes np.issctype, np.maximum_sctype, np.obj2sctype, np.sctype2char, np.sctypes and np.issubsctype without replacement.

Removes deprecated np.deprecate and np.deprecate_with_doc; use DeprecationWarning instead.

Removes deprecated np.safe_eval; use ast.literal_eval instead.

Removes np.find_common_type; use numpy.promote_types or numpy.result_type instead.

Removes np.round_; use np.round instead.

Removes np.nbytes; use np.dtype(<dtype>).itemsize instead.

Removes np.compare_chararrays from the main namespace; use np.char.compare_chararrays instead.

Removes np.format_parser from the main namespace; use np.rec.format_parser instead.

Removes seven dtype string aliases from np.dtype: int0, uint0, void0, object0, str0, bytes0 and bool8.

Removes __array_prepare__; migrate to __array_ufunc__ or rely on __array_wrap__.

Removes the np.core.umath_tests submodule from the public API.

Removes the deprecated PyDataMem_SetEventHook; use tracemalloc and np.lib.tracemalloc_domain.

Removes set_numeric_ops and the C functions PyArray_SetNumericOps and PyArray_GetNumericOps.

Removes the deprecated fastCopyAndTranspose function and its C counterpart.

Removes np.msort; use np.sort(a, axis=0) instead.

Removes the PyArray_CGT, PyArray_CLT, PyArray_CGE, PyArray_CLE, PyArray_CEQ and PyArray_CNE macros.

Renames NPY_NTYPES to NPY_NTYPES_LEGACY since it excludes new built-in DTypes.

Removes the ->f slot from PyArray_Descr; use PyDataType_GetArrFuncs instead.

PyArray_GETITEM and PyArray_SETITEM now require importing the NumPy API table and are no longer in ndarraytypes.h.

Removes npy_interrupt.h and macros like NPY_SIGINT_ON.

Removes the noprefix.h header; missing symbols must be replaced with their prefixed counterparts.

Removes PyUFunc_GetPyVals, PyUFunc_handlefperr and PyUFunc_checkfperr.

Removes numpy/old_defines.h and its PyArray_CONSTANT style macros.

Removes the legacy_inner_loop_selector member of the ufunc struct.

Removes NPY_INTPLTR to avoid confusion.

Removes the advanced indexing MapIter and related C API, which was public but poorly tested and had one known user (Theano).

Removes PyArray_REFCNT and NPY_REFCOUNT; use Py_REFCNT instead.

Removes PyArray_MoveInto, PyArray_CastTo and PyArray_CastAnyTo; use PyArray_CopyInto or PyArray_CopyAnyInto.

Removes PyArray_FillObjectArray; create a new empty array or use PyArray_FillWithScalar().

Removes PyArray_CompareUCS4 and PyArray_CompareString; use standard C string comparison functions.

Removes PyArray_ISPYTHON, which is misleading, has no known use cases and is easy to replace.

Removes PyArray_FieldNames, which has unclear purpose and incorrect semantics in some use cases.

Removes PyArray_TypestrConvert; use explicit types or go via Python strings instead.

Removes PyDataType_GetDatetimeMetaData, which did nothing since at least NumPy 1.7.

Removes PyArray_GetCastFunc; use C casts directly for simple numeric types.

Complex types now use C99 complex types; direct field access like c.real or c.imag no longer works, use npy_math.h utilities instead.

The default NumPy integer dtype is now 64-bit on all 64-bit systems, including 64-bit Windows where it was historically 32-bit.

numpy.fft.helper is renamed to the private numpy.fft._helper; public functions should be accessed from numpy.fft.

numpy.linalg.linalg is renamed to the private numpy.linalg._linalg; public functions should be accessed from numpy.linalg.

NumPy now passes a copy keyword to the __array__ special method when it would be non-default; implementations should accept it.

np.can_cast no longer accepts Python int, float, or complex instances.

Casting a fixed-width byte string array (np.bytes_) to StringDType now raises TypeError when the bytes are not valid UTF-8.

Makes PyArray_StringDTypeObject an opaque struct under the abi3t stable ABI, so extensions cannot access its fields.

Passing None as dtype to np.finfo now raises a TypeError, an expired deprecation from 1.25.

bincount now raises TypeError for non-integer inputs.

linalg.eig and linalg.eigvals now always return complex arrays.

np.from_dlpack raises BufferError instead of RuntimeError for unsupported device, dtype, or too many dimensions.

datetime64 and timedelta64 addition, subtraction, and integer multiplication now raise OverflowError on int64 overflow or NaT results instead of silently wrapping.

numpy.ctypeslib.as_ctypes now only accepts numpy.ndarray and raises TypeError for scalar types.

Scalars now export __array_interface__ directly and are read-only instead of writeable.

np.meshgrid now always returns a tuple, including when sparse is true and copy is false.

For object arrays, .real and .imag now use elementwise getattr and return read-only results.

Raises TypeError when converting an array with ndim > 0 to a scalar.

Removes four undocumented ndarray.ctypes methods.

C extension modules use multi-phase initialisation and now raise ImportError if re-imported after deletion from sys.modules.

numpy.round now always returns a copy, including for integer inputs with decimals >= 0.

Boolean non-array indices to flatiter such as arr.flat[[True, True]] now raise an index error instead of being treated as integer indices.

Raises on np.conjugate of non-numeric types, deprecated since 1.13.

Raises when using np.bincount with minlength=None; use 0 instead.

Passing shape=None to functions with a non-optional shape argument now errors; use () instead.

Inexact matches for mode and searchside now raise, deprecated since 1.20.

Setting __array_finalize__ = None now errors, deprecated since 1.23.

np.fromfile and np.fromstring now error on bad data instead of guessing, deprecated since 1.18.

A dtype attribute on a class must now be a dtype instance rather than something parseable as a dtype.

Passing booleans as a partition index now errors, deprecated since 1.23.

Out-of-bounds indexes now raise an error even on empty arrays.

Disallows making a non-writeable array writeable when it has a base and does not own its data.

Unpickling a scalar with object dtype now raises an error.

The binary mode of fromstring now errors, use frombuffer instead.

Converting np.inexact or np.floating to a dtype now raises an error.

Converting np.complex, np.integer, np.signedinteger, np.unsignedinteger, np.generic to a dtype now raises an error.

The Python built-in round now errors for complex scalars; use np.round or scalar.round.

Using a length 1 tuple for the ufunc signature now errors; use dtype or fill the tuple with None.

Converting an empty array to bool now raises an error; use arr.size > 0 to check for elements.

numpy.cov now transposes single-row 2d design matrices when rowvar=False, returning an array instead of a scalar.

Disallows scalars and 0D arrays for numpy.nonzero and numpy.ndarray.nonzero.

Removes the internal set_string_function function and stubs out PyArray_SetStringFunction.

copyto now applies NEP 50 cast safety, changing behavior for Python int/float and NumPy scalar assignments, and aligns copyto, full, and full_like with NumPy 2 behavior.

NumPy 2.0.0 is the first major release since 2006, containing breaking changes to the Python and C APIs and an ABI break.

Includes an ABI break, type promotion rule changes, and API changes that may not have emitted deprecation warnings in 1.26.x.

Adopts NEP 50 type promotion, which can change output dtypes and lower precision for mixed-dtype operations.

Changes the maximum number of array dimensions from 32 to 64.

Binaries built against NumPy 1.xx using the C API will not work with NumPy 2.0 and will raise an ImportError about binary incompatibility.

Makes np.tracemalloc_domain available only from np.lib.

Makes np.recfromcsv and recfromtxt available only from np.lib.npyio.

Removes the experimental numpy.array_api submodule; use the main numpy namespace or the array-api-strict package.

Finalizes removal of the fasttake, fastclip and fastputmask ArrFuncs deprecation.

Finalizes removal of the deprecated PyArray_ScalarFromObject.

np.dtype(("f8", 1)) now returns a shape 1 subarray dtype rather than a non-subarray dtype.

Assigning to the .data attribute of an ndarray is disallowed and raises.

np.binary_repr(a, width) now raises if width is too small.

Using NPY_CHAR in PyArray_DescrFromType() now raises; use NPY_STRING, NPY_UNICODE or NPY_VSTRING.

Removes the f2py.compile() helper because it leaked memory and was experimental.

Resolves ambiguous broadcasting in np.solve when b has one fewer dimension than a, now compliant with the Array API.

Moves PyArray_MIN and PyArray_MAX from ndarraytypes.h to npy_math.h.

The C API now only exports static inline versions of the array accessors.

Moves dtype flag accessors to ndarrayobject.h, requiring import_array(); includes PyDataType_FLAGCHK, PyDataType_REFCHK and NPY_BEGIN_THREADS_DESCR.

Dtype flags on PyArray_Descr must be accessed via the PyDataType_FLAGS inline function for 1.x and 2.x compatibility.

Makes types and data arguments const-correct in generalized ufunc C API constructors, which may affect C++ code.

Makes the generalized ufunc C API const correct.

Increases NPY_MAXDIMS to 64 and NPY_MAXARGS, and introduces NPY_RAVEL_AXIS for axis=None usage.

Makes NPY_MAXARGS a runtime constant and changes the size of PyArrayMultiIterObject.

Requires changes for custom legacy user dtypes registered with PyArray_RegisterDataType due to an ABI break.

Requires structured dtype fields c_metadata, names, fields and subarray to be accessed through new functions such as PyDataType_NAMES.

Requires elsize and alignment descriptor fields to be accessed via PyDataType_ELSIZE, PyDataType_SET_ELSIZE and PyDataType_ALIGNMENT unless compiling only with NumPy 2 support.

Removes the NPY_MAX_ELSIZE macro, which only reflected builtin numeric types and served no internal purpose.

Makes PyArrayFlags_Type, PyArray_NewFlagsObject and PyArrayFlagsObject private; use the Python API if needed.

numpy/npy_common.h now includes complex.h, making complex a reserved keyword.

np.gradient() now returns a tuple rather than a list, making the return value immutable.

np.errstate can now only be entered once because it is fully context and thread-safe.

dtype.flags is now stored unsigned, so the aligned flag is positive (128) instead of negative (-128).

NumPy scalars now print with type information, e.g. np.float64(3.0) instead of 3.0, per NEP 51.

NumPy strings are now True when non-empty and False when empty, changing string-to-bool casts and np.nonzero behavior.

np.intp and np.uintp are redefined in C to match size_t/Py_ssize_t instead of intptr_t/uintptr_t, adding character codes nN.

Out-of-bound axis values now raise an error instead of behaving like axis=None, except for concatenate where it was deprecated.

A trailing comma in a dtype string now always creates a structured dtype, and parenthesized sub-array shapes like np.dtype("(2)i,") are deprecated.

Functions that returned a list of ndarrays now return a tuple of ndarrays instead.

np.unique with return_inverse=True now shapes unique_inverse so the input can be reconstructed directly.

any and all now return booleans for object arrays instead of a reduced object value.

Fixes a heap buffer overflow in timedelta to string casts.

Fixes a heap overflow in fixed-width string multiply.

Adds wheels for the newly released Python 3.15.0rc1 and supports Python versions 3.12-3.15.

Drops support for Python 3.11; this release supports Python versions 3.12 through 3.14.

NumPy now supports the stable ABI for free-threaded Python as described in PEP 803.

Supports Python versions 3.11 through 3.14.

Adds support for Python 3.14rc1.

Supports Python 3.10 through 3.13.

Supports Python versions 3.10 through 3.13.

Adds support for Python 3.13 and drops support for Python 3.9; Python 3.10-3.13 are supported.

Supports Python versions 3.9 through 3.12.

Deprecates numpy.char.chararray in favor of an ndarray with a string or bytes dtype.

numpy.take now checks casting of the result to out=out under the same-kind rule and warns when it fails; affects compress too.

Deprecates numpy.char.array and numpy.char.asarray in favor of numpy.array and numpy.asarray with a string or bytes dtype.

Deprecates setting the dtype attribute on arrays; use array.view(dtype=new_dtype) instead.

Deprecates setting the shape attribute on arrays; use np.reshape or np.ndarray.reshape instead.

Deprecates the generic unit in numpy.timedelta64 and operations implicitly relying on it; specify an explicit unit instead.

Deprecates resizing a NumPy array in place; use np.resize to create a resized array instead.

Deprecates numpy.fix; use numpy.trunc instead, which is faster and follows the Array API standard.

Deprecates numpy.ma.round_; use numpy.ma.round as a replacement.

Deprecates numpy.typename because its names were outdated and inconsistent; use numpy.dtype.name instead.

Deprecates non-integer inputs for numpy.triu_indices, numpy.tril_indices, numpy.tri, numpy.tril_indices_from and numpy.triu_indices_from.

Deprecates the dtype attribute setting path used by arr.view(dtype=...) for subclasses; subclasses should set _set_dtype.

Deprecates setting the strides attribute of arrays.

Deprecates passing out positionally to np.maximum and np.minimum.

Warns when align= is not a boolean in np.dtype().

Deprecates np.testing.assert_warns and np.testing.suppress_warnings.

Marks np.fix as pending deprecation in favor of numpy.trunc.

Marks in-place modification of ndarray.shape as pending deprecation.

Deprecates the numpy.lib.user_array.container class.

Float non-array indices to flatiter are now deprecated and will be removed in a future version.

0-dimensional boolean indices like arr.flat[True] are now deprecated and will be removed in a future version.

Deprecates numpy.typing.mypy_plugin in favor of platform-agnostic static type inference.

Deprecates the numpy.typing.NBitBase type, to be removed in a future version.

Deprecates _add_newdoc_ufunc; use ufunc.__doc__ = newdoc instead.

Deprecates the fix_imports keyword argument in numpy.save, which has been ignored since NumPy 1.17.

Deprecates passing non-integer inputs as the first argument of bincount, which are silently cast to integers.

Deprecates chararray in the main namespace; it can still be imported from np.char.chararray for now.

Deprecates np.compat because Python 2 is no longer supported.

Deprecates out-of-bounds Python integer conversion to integer arrays; use np.iinfo or np.array(value).astype(dtype).

Deprecates np.safe_eval; use ast.literal_eval instead.

Deprecates np.recfromcsv, np.recfromtxt, np.disp, np.get_array_wrap, np.maximum_sctype, np.deprecate and np.deprecate_with_doc.

Deprecates np.trapz; use np.trapezoid or a scipy.integrate function instead.

Deprecates np.in1d; use np.isin instead.

Deprecates the np.row_stack alias; use np.vstack directly.

__array_wrap__ now receives arr, context, return_scalar; implementations not accepting all three are deprecated.

Deprecates 2-dimensional vector arrays for np.cross; use 3-dimensional vectors instead.

Deprecates the np.dtype("a") alias for np.dtype(np.bytes_); use np.dtype("S") instead.

Deprecates keyword arguments x and y for assert_array_equal and assert_array_almost_equal; pass them positionally.

Deprecates n-D fft transforms with s not None and axes None, and s arrays containing None.

Plans to deprecate and remove the rcond keyword of numpy.linalg.pinv in favor of rtol; rcond default is used when neither is given.

Suggests moving from np.char to np.strings; np.char may be deprecated in the future.

Accessing numpy.core now emits a DeprecationWarning; the module is renamed to numpy._core.