Release notes

• 1.22.0
  • Expired deprecations
    • Deprecated numeric style dtype strings have been removed
    • Expired deprecations for loads, ndfromtxt, and mafromtxt in npyio
  • Deprecations
    • Use delimiter rather than delimitor as kwarg in mrecords
    • Passing boolean kth values to (arg-)partition has been deprecated
    • The np.MachAr class has been deprecated
  • Compatibility notes
    • Distutils forces strict floating point model on clang
    • Removed floor division support for complex types
    • numpy.vectorize functions now produce the same output class as the base function
    • Python 3.7 is no longer supported
    • str/repr of complex dtypes now include space after punctuation
    • Corrected advance in PCG64DSXM and PCG64
    • Change in generation of random 32 bit floating point variates
  • C API changes
    • Masked inner-loops cannot be customized anymore
    • Experimental exposure of future DType and UFunc API
  • New Features
    • NEP 49 configurable allocators
    • Implementation of the NEP 47 (adopting the array API standard)
    • Generate C/C++ API reference documentation from comments blocks is now possible
    • Assign the platform-specific c_intp precision via a mypy plugin
    • Add NEP 47-compatible dlpack support
    • keepdims optional argument added to numpy.argmin, numpy.argmax
    • bit_count to compute the number of 1-bits in an integer
    • The ndim and axis attributes have been added to numpy.AxisError
    • Preliminary support for windows/arm64 target
    • Added support for LoongArch
    • A .clang-format file has been added
    • is_integer is now available to numpy.floating and numpy.integer
    • Symbolic parser for Fortran dimension specifications
    • ndarray, dtype and number are now runtime-subscriptable
  • Improvements
    • ctypeslib.load_library can now take any path-like object
    • Add smallest_normal and smallest_subnormal attributes to finfo
    • numpy.linalg.qr accepts stacked matrices as inputs
    • numpy.fromregex now accepts os.PathLike implementations
    • Add new methods for quantile and percentile
    • Missing parameters have been added to the nan<x> functions
    • Annotating the main Numpy namespace
    • Vectorize umath module using AVX-512
    • OpenBLAS v0.3.17
• 1.21.4
  • Contributors
  • Pull requests merged
• 1.21.3
  • Contributors
  • Pull requests merged
• 1.21.2
  • Contributors
  • Pull requests merged
• 1.21.1
  • Contributors
  • Pull requests merged
• 1.21.0
  • New functions
    • Add PCG64DXSM BitGenerator
  • Expired deprecations
  • Deprecations
    • The .dtype attribute must return a dtype
    • Inexact matches for numpy.convolve and numpy.correlate are deprecated
    • np.typeDict has been formally deprecated
    • Exceptions will be raised during array-like creation
    • Four ndarray.ctypes methods have been deprecated
  • Expired deprecations
    • Remove deprecated PolyBase and unused PolyError and PolyDomainError
  • Compatibility notes
    • Error type changes in universal functions
    • __array_ufunc__ argument validation
    • __array_ufunc__ and additional positional arguments
    • Validate input values in Generator.uniform
    • /usr/include removed from default include paths
    • Changes to comparisons with dtype=...
    • Changes to dtype and signature arguments in ufuncs
    • Ufunc signature=... and dtype= generalization and casting
    • Distutils forces strict floating point model on clang
  • C API changes
    • Use of ufunc->type_resolver and “type tuple”
  • New Features
    • Added a mypy plugin for handling platform-specific numpy.number precisions
    • Let the mypy plugin manage extended-precision numpy.number subclasses
    • New min_digits argument for printing float values
    • f2py now recognizes Fortran abstract interface blocks
    • BLAS and LAPACK configuration via environment variables
    • A runtime-subcriptable alias has been added for ndarray
  • Improvements
    • Arbitrary period option for numpy.unwrap
    • np.unique now returns single NaN
    • Generator.rayleigh and Generator.geometric performance improved
    • Placeholder annotations have been improved
  • Performance improvements
    • Improved performance in integer division of NumPy arrays
    • Improve performance of np.save and np.load for small arrays
  • Changes
    • numpy.piecewise output class now matches the input class
    • Enable Accelerate Framework
• 1.20.3
  • Contributors
  • Pull requests merged
• 1.20.2
  • Contributors
  • Pull requests merged
• 1.20.1
  • Highlights
  • Contributors
  • Pull requests merged
• 1.20.0
  • New functions
    • The random.Generator class has a new permuted function.
    • sliding_window_view provides a sliding window view for numpy arrays
    • numpy.broadcast_shapes is a new user-facing function
  • Deprecations
    • Using the aliases of builtin types like np.int is deprecated
    • Passing shape=None to functions with a non-optional shape argument is deprecated
    • Indexing errors will be reported even when index result is empty
    • Inexact matches for mode and searchside are deprecated
    • Deprecation of numpy.dual
    • outer and ufunc.outer deprecated for matrix
    • Further Numeric Style types Deprecated
    • The ndincr method of ndindex is deprecated
    • ArrayLike objects which do not define __len__ and __getitem__
  • Future Changes
    • Arrays cannot be using subarray dtypes
  • Expired deprecations
    • Financial functions removed
  • Compatibility notes
    • isinstance(dtype, np.dtype) and not type(dtype) is not np.dtype
    • Same kind casting in concatenate with axis=None
    • NumPy Scalars are cast when assigned to arrays
    • Array coercion changes when Strings and other types are mixed
    • Array coercion restructure
    • Writing to the result of numpy.broadcast_arrays will export readonly buffers
    • Numeric-style type names have been removed from type dictionaries
    • The operator.concat function now raises TypeError for array arguments
    • nickname attribute removed from ABCPolyBase
    • float->timedelta and uint64->timedelta promotion will raise a TypeError
    • numpy.genfromtxt now correctly unpacks structured arrays
    • mgrid, r_, etc. consistently return correct outputs for non-default precision input
    • Boolean array indices with mismatching shapes now properly give IndexError
    • Casting errors interrupt Iteration
    • f2py generated code may return unicode instead of byte strings
    • The first element of the __array_interface__["data"] tuple must be an integer
    • poly1d respects the dtype of all-zero argument
    • The numpy.i file for swig is Python 3 only.
    • Void dtype discovery in np.array
  • C API changes
    • The PyArray_DescrCheck macro is modified
    • Size of np.ndarray and np.void_ changed
  • New Features
    • where keyword argument for numpy.all and numpy.any functions
    • where keyword argument for numpy functions mean, std, var
    • norm=backward, forward keyword options for numpy.fft functions
    • NumPy is now typed
    • numpy.typing is accessible at runtime
    • New __f2py_numpy_version__ attribute for f2py generated modules.
    • mypy tests can be run via runtests.py
    • Negation of user defined BLAS/LAPACK detection order
    • Allow passing optimizations arguments to asv build
    • The NVIDIA HPC SDK nvfortran compiler is now supported
    • dtype option for cov and corrcoef
  • Improvements
    • Improved string representation for polynomials (__str__)
    • Remove the Accelerate library as a candidate LAPACK library
    • Object arrays containing multi-line objects have a more readable repr
    • Concatenate supports providing an output dtype
    • Thread safe f2py callback functions
    • numpy.core.records.fromfile now supports file-like objects
    • RPATH support on AIX added to distutils
    • Use f90 compiler specified by the command line args
    • Add NumPy declarations for Cython 3.0 and later
    • Make the window functions exactly symmetric
  • Performance improvements and changes
    • Enable multi-platform SIMD compiler optimizations
  • Changes
    • Changed behavior of divmod(1., 0.) and related functions
    • np.linspace on integers now uses floor
• 1.19.5
  • Contributors
  • Pull requests merged
• 1.19.4
  • Contributors
  • Pull requests merged
• 1.19.3
  • Contributors
  • Pull requests merged
• 1.19.2
  • Improvements
    • Add NumPy declarations for Cython 3.0 and later
  • Contributors
  • Pull requests merged
• 1.19.1
  • Contributors
  • Pull requests merged
• 1.19.0
  • Highlights
  • Expired deprecations
    • numpy.insert and numpy.delete can no longer be passed an axis on 0d arrays
    • numpy.delete no longer ignores out-of-bounds indices
    • numpy.insert and numpy.delete no longer accept non-integral indices
    • numpy.delete no longer casts boolean indices to integers
  • Compatibility notes
    • Changed random variate stream from numpy.random.Generator.dirichlet
    • Scalar promotion in PyArray_ConvertToCommonType
    • Fasttake and fastputmask slots are deprecated and NULL’ed
    • np.ediff1d casting behaviour with to_end and to_begin
    • Converting of empty array-like objects to NumPy arrays
    • Removed multiarray.int_asbuffer
    • numpy.distutils.compat has been removed
    • issubdtype no longer interprets float as np.floating
    • Change output of round on scalars to be consistent with Python
    • The numpy.ndarray constructor no longer interprets strides=() as strides=None
    • C-Level string to datetime casts changed
    • SeedSequence with small seeds no longer conflicts with spawning
  • Deprecations
    • Deprecate automatic dtype=object for ragged input
    • Passing shape=0 to factory functions in numpy.rec is deprecated
    • Deprecation of probably unused C-API functions
    • Converting certain types to dtypes is Deprecated
    • Deprecation of round for np.complexfloating scalars
    • numpy.ndarray.tostring() is deprecated in favor of tobytes()
  • C API changes
    • Better support for const dimensions in API functions
    • Const qualify UFunc inner loops
  • New Features
    • numpy.frompyfunc now accepts an identity argument
    • np.str_ scalars now support the buffer protocol
    • subok option for numpy.copy
    • numpy.linalg.multi_dot now accepts an out argument
    • keepdims parameter for numpy.count_nonzero
    • equal_nan parameter for numpy.array_equal
  • Improvements
  • Improve detection of CPU features
    • Use 64-bit integer size on 64-bit platforms in fallback lapack_lite
    • Use AVX512 intrinsic to implement np.exp when input is np.float64
    • Ability to disable madvise hugepages
    • numpy.einsum accepts NumPy int64 type in subscript list
    • np.logaddexp2.identity changed to -inf
  • Changes
    • Remove handling of extra argument to __array__
    • numpy.random._bit_generator moved to numpy.random.bit_generator
    • Cython access to the random distributions is provided via a pxd file
    • Fixed eigh and cholesky methods in numpy.random.multivariate_normal
    • Fixed the jumping implementation in MT19937.jumped
• 1.18.5
  • Contributors
  • Pull requests merged
• 1.18.4
  • Contributors
  • Pull requests merged
• 1.18.3
  • Highlights
  • Contributors
  • Pull requests merged
• 1.18.2
  • Contributors
  • Pull requests merged
• 1.18.1
  • Contributors
  • Pull requests merged
• 1.18.0
  • Highlights
  • New functions
    • Multivariate hypergeometric distribution added to numpy.random
  • Deprecations
    • np.fromfile and np.fromstring will error on bad data
    • Deprecate non-scalar arrays as fill values in ma.fill_value
    • Deprecate PyArray_As1D, PyArray_As2D
    • Deprecate np.alen
    • Deprecate the financial functions
    • The axis argument to numpy.ma.mask_cols and numpy.ma.mask_row is deprecated
  • Expired deprecations
  • Compatibility notes
    • numpy.lib.recfunctions.drop_fields can no longer return None
    • numpy.argmin/argmax/min/max returns NaT if it exists in array
    • np.can_cast(np.uint64, np.timedelta64, casting='safe') is now False
    • Changed random variate stream from numpy.random.Generator.integers
    • Add more ufunc loops for datetime64, timedelta64
    • Moved modules in numpy.random
  • C API changes
    • PyDataType_ISUNSIZED(descr) now returns False for structured datatypes
  • New Features
    • Add our own *.pxd cython import file
    • A tuple of axes can now be input to expand_dims
    • Support for 64-bit OpenBLAS
    • Add --f2cmap option to F2PY
  • Improvements
    • Different C numeric types of the same size have unique names
    • argwhere now produces a consistent result on 0d arrays
    • Add axis argument for random.permutation and random.shuffle
    • method keyword argument for np.random.multivariate_normal
    • Add complex number support for numpy.fromstring
    • numpy.unique has consistent axes order when axis is not None
    • numpy.matmul with boolean output now converts to boolean values
    • numpy.random.randint produced incorrect value when the range was 2**32
    • Add complex number support for numpy.fromfile
    • std=c99 added if compiler is named gcc
  • Changes
    • NaT now sorts to the end of arrays
    • Incorrect threshold in np.set_printoptions raises TypeError or ValueError
    • Warn when saving a dtype with metadata
    • numpy.distutils append behavior changed for LDFLAGS and similar
    • Remove numpy.random.entropy without a deprecation
    • Add options to quiet build configuration and build with -Werror
• 1.17.5
  • Contributors
  • Pull requests merged
• 1.17.4
  • Highlights
  • Contributors
  • Pull requests merged
• 1.17.3
  • Highlights
  • Compatibility notes
  • Contributors
  • Pull requests merged
• 1.17.2
  • Contributors
  • Pull requests merged
• 1.17.1
  • Contributors
  • Pull requests merged
• 1.17.0
  • Highlights
  • New functions
  • Deprecations
    • numpy.polynomial functions warn when passed float in place of int
    • Deprecate numpy.distutils.exec_command and temp_file_name
    • Writeable flag of C-API wrapped arrays
    • numpy.nonzero should no longer be called on 0d arrays
    • Writing to the result of numpy.broadcast_arrays will warn
  • Future Changes
    • Shape-1 fields in dtypes won’t be collapsed to scalars in a future version
  • Compatibility notes
    • float16 subnormal rounding
    • Signed zero when using divmod
    • MaskedArray.mask now returns a view of the mask, not the mask itself
    • Do not lookup __buffer__ attribute in numpy.frombuffer
    • out is buffered for memory overlaps in take, choose, put
    • Unpickling while loading requires explicit opt-in
    • Potential changes to the random stream in old random module
    • i0 now always returns a result with the same shape as the input
    • can_cast no longer assumes all unsafe casting is allowed
    • ndarray.flags.writeable can be switched to true slightly more often
  • C API changes
    • dimension or stride input arguments are now passed by npy_intp const*
  • New Features
    • New extensible numpy.random module with selectable random number generators
    • libFLAME
    • User-defined BLAS detection order
    • User-defined LAPACK detection order
    • ufunc.reduce and related functions now accept a where mask
    • Timsort and radix sort have replaced mergesort for stable sorting
    • packbits and unpackbits accept an order keyword
    • unpackbits now accepts a count parameter
    • linalg.svd and linalg.pinv can be faster on hermitian inputs
    • divmod operation is now supported for two timedelta64 operands
    • fromfile now takes an offset argument
    • New mode “empty” for pad
    • empty_like and related functions now accept a shape argument
    • Floating point scalars implement as_integer_ratio to match the builtin float
    • Structured dtype objects can be indexed with multiple fields names
    • .npy files support unicode field names
  • Improvements
    • Array comparison assertions include maximum differences
    • Replacement of the fftpack based fft module by the pocketfft library
    • Further improvements to ctypes support in numpy.ctypeslib
    • numpy.errstate is now also a function decorator
    • numpy.exp and numpy.log speed up for float32 implementation
    • Improve performance of numpy.pad
    • numpy.interp handles infinities more robustly
    • Pathlib support for fromfile, tofile and ndarray.dump
    • Specialized isnan, isinf, and isfinite ufuncs for bool and int types
    • isfinite supports datetime64 and timedelta64 types
    • New keywords added to nan_to_num
    • MemoryErrors caused by allocated overly large arrays are more descriptive
    • floor, ceil, and trunc now respect builtin magic methods
    • quantile now works on fraction.Fraction and decimal.Decimal objects
    • Support of object arrays in matmul
  • Changes
    • median and percentile family of functions no longer warn about nan
    • timedelta64 % 0 behavior adjusted to return NaT
    • NumPy functions now always support overrides with __array_function__
    • lib.recfunctions.structured_to_unstructured does not squeeze single-field views
    • clip now uses a ufunc under the hood
    • __array_interface__ offset now works as documented
    • Pickle protocol in savez set to 3 for force zip64 flag
    • Structured arrays indexed with non-existent fields raise KeyError not ValueError
• 1.16.6
  • Highlights
  • New functions
    • Allow matmul (@ operator) to work with object arrays.
  • Compatibility notes
    • Fix regression in matmul (@ operator) for boolean types
  • Improvements
    • Array comparison assertions include maximum differences
  • Contributors
  • Pull requests merged
• 1.16.5
  • Contributors
  • Pull requests merged
• 1.16.4
  • New deprecations
    • Writeable flag of C-API wrapped arrays
  • Compatibility notes
    • Potential changes to the random stream
  • Changes
    • numpy.lib.recfunctions.structured_to_unstructured does not squeeze single-field views
  • Contributors
  • Pull requests merged
• 1.16.3
  • Compatibility notes
    • Unpickling while loading requires explicit opt-in
  • Improvements
    • Covariance in random.mvnormal cast to double
  • Changes
    • __array_interface__ offset now works as documented
• 1.16.2
  • Compatibility notes
    • Signed zero when using divmod
  • Contributors
  • Pull requests merged
• 1.16.1
  • Contributors
  • Enhancements
  • Compatibility notes
  • New Features
    • divmod operation is now supported for two timedelta64 operands
  • Improvements
    • Further improvements to ctypes support in np.ctypeslib
    • Array comparison assertions include maximum differences
  • Changes
    • timedelta64 % 0 behavior adjusted to return NaT
• 1.16.0
  • Highlights
  • New functions
  • New deprecations
  • Expired deprecations
  • Future changes
  • Compatibility notes
    • f2py script on Windows
    • NaT comparisons
    • complex64/128 alignment has changed
    • nd_grid __len__ removal
    • np.unravel_index now accepts shape keyword argument
    • multi-field views return a view instead of a copy
  • C API changes
  • New Features
    • Integrated squared error (ISE) estimator added to histogram
    • max_rows keyword added for np.loadtxt
    • modulus operator support added for np.timedelta64 operands
  • Improvements
    • no-copy pickling of numpy arrays
    • build shell independence
    • np.polynomial.Polynomial classes render in LaTeX in Jupyter notebooks
    • randint and choice now work on empty distributions
    • linalg.lstsq, linalg.qr, and linalg.svd now work with empty arrays
    • Chain exceptions to give better error messages for invalid PEP3118 format strings
    • Einsum optimization path updates and efficiency improvements
    • numpy.angle and numpy.expand_dims now work on ndarray subclasses
    • NPY_NO_DEPRECATED_API compiler warning suppression
    • np.diff Added kwargs prepend and append
    • ARM support updated
    • Appending to build flags
    • Generalized ufunc signatures now allow fixed-size dimensions
    • Generalized ufunc signatures now allow flexible dimensions
    • np.clip and the clip method check for memory overlap
    • New value unscaled for option cov in np.polyfit
    • Detailed docstrings for scalar numeric types
    • __module__ attribute now points to public modules
    • Large allocations marked as suitable for transparent hugepages
    • Alpine Linux (and other musl c library distros) support
    • Speedup np.block for large arrays
    • Speedup np.take for read-only arrays
    • Support path-like objects for more functions
    • Better behaviour of ufunc identities during reductions
    • Improved conversion from ctypes objects
    • A new ndpointer.contents member
    • matmul is now a ufunc
    • Start and stop arrays for linspace, logspace and geomspace
    • CI extended with additional services
  • Changes
    • Comparison ufuncs will now error rather than return NotImplemented
    • Positive will now raise a deprecation warning for non-numerical arrays
    • NDArrayOperatorsMixin now implements matrix multiplication
    • The scaling of the covariance matrix in np.polyfit is different
    • maximum and minimum no longer emit warnings
    • Umath and multiarray c-extension modules merged into a single module
    • getfield validity checks extended
    • NumPy functions now support overrides with __array_function__
    • Arrays based off readonly buffers cannot be set writeable
• 1.15.4
  • Compatibility Note
  • Contributors
  • Pull requests merged
• 1.15.3
  • Compatibility Note
  • Contributors
  • Pull requests merged
• 1.15.2
  • Compatibility Note
  • Contributors
  • Pull requests merged
• 1.15.1
  • Compatibility Note
  • Contributors
  • Pull requests merged
• 1.15.0
  • Highlights
  • New functions
  • Deprecations
  • Future Changes
  • Compatibility notes
    • Compiled testing modules renamed and made private
    • The NpzFile returned by np.savez is now a collections.abc.Mapping
    • Under certain conditions, nditer must be used in a context manager
    • Numpy has switched to using pytest instead of nose for testing
    • Numpy no longer monkey-patches ctypes with __array_interface__
    • np.ma.notmasked_contiguous and np.ma.flatnotmasked_contiguous always return lists
    • np.squeeze restores old behavior of objects that cannot handle an axis argument
    • unstructured void array’s .item method now returns a bytes object
    • copy.copy and copy.deepcopy no longer turn masked into an array
    • Multifield Indexing of Structured Arrays will still return a copy
  • C API changes
    • New functions npy_get_floatstatus_barrier and npy_clear_floatstatus_barrier
    • Changes to PyArray_GetDTypeTransferFunction
  • New Features
    • np.gcd and np.lcm ufuncs added for integer and objects types
    • Support for cross-platform builds for iOS
    • return_indices keyword added for np.intersect1d
    • np.quantile and np.nanquantile
    • Build system
  • Improvements
    • np.einsum updates
    • np.ufunc.reduce and related functions now accept an initial value
    • np.flip can operate over multiple axes
    • histogram and histogramdd functions have moved to np.lib.histograms
    • histogram will accept NaN values when explicit bins are given
    • histogram works on datetime types, when explicit bin edges are given
    • histogram “auto” estimator handles limited variance better
    • The edges returned by histogram` and histogramdd now match the data float type
    • histogramdd allows explicit ranges to be given in a subset of axes
    • The normed arguments of histogramdd and histogram2d have been renamed
    • np.r_ works with 0d arrays, and np.ma.mr_ works with np.ma.masked
    • np.ptp accepts a keepdims argument, and extended axis tuples
    • MaskedArray.astype now is identical to ndarray.astype
    • Enable AVX2/AVX512 at compile time
    • nan_to_num always returns scalars when receiving scalar or 0d inputs
    • np.flatnonzero works on numpy-convertible types
    • np.interp returns numpy scalars rather than builtin scalars
    • Allow dtype field names to be unicode in Python 2
    • Comparison ufuncs accept dtype=object, overriding the default bool
    • sort functions accept kind='stable'
    • Do not make temporary copies for in-place accumulation
    • linalg.matrix_power can now handle stacks of matrices
    • Increased performance in random.permutation for multidimensional arrays
    • Generalized ufuncs now accept axes, axis and keepdims arguments
    • float128 values now print correctly on ppc systems
    • New np.take_along_axis and np.put_along_axis functions
• 1.14.6
  • Contributors
  • Pull requests merged
• 1.14.5
  • Contributors
  • Pull requests merged
• 1.14.4
  • Contributors
  • Pull requests merged
• 1.14.3
  • Contributors
  • Pull requests merged
• 1.14.2
  • Contributors
  • Pull requests merged
• 1.14.1
  • Contributors
  • Pull requests merged
• 1.14.0
  • Highlights
  • New functions
  • Deprecations
  • Future Changes
  • Compatibility notes
    • The mask of a masked array view is also a view rather than a copy
    • np.ma.masked is no longer writeable
    • np.ma functions producing fill_value s have changed
    • a.flat.__array__() returns non-writeable arrays when a is non-contiguous
    • np.tensordot now returns zero array when contracting over 0-length dimension
    • numpy.testing reorganized
    • np.asfarray no longer accepts non-dtypes through the dtype argument
    • 1D np.linalg.norm preserves float input types, even for arbitrary orders
    • count_nonzero(arr, axis=()) now counts over no axes, not all axes
    • __init__.py files added to test directories
    • .astype(bool) on unstructured void arrays now calls bool on each element
    • MaskedArray.squeeze never returns np.ma.masked
    • Renamed first parameter of can_cast from from to from_
    • isnat raises TypeError when passed wrong type
    • dtype.__getitem__ raises TypeError when passed wrong type
    • User-defined types now need to implement __str__ and __repr__
    • Many changes to array printing, disableable with the new “legacy” printing mode
  • C API changes
    • PyPy compatible alternative to UPDATEIFCOPY arrays
  • New Features
    • Encoding argument for text IO functions
    • External nose plugins are usable by numpy.testing.Tester
    • parametrize decorator added to numpy.testing
    • chebinterpolate function added to numpy.polynomial.chebyshev
    • Support for reading lzma compressed text files in Python 3
    • sign option added to np.setprintoptions and np.array2string
    • hermitian option added to``np.linalg.matrix_rank``
    • threshold and edgeitems options added to np.array2string
    • concatenate and stack gained an out argument
    • Support for PGI flang compiler on Windows
  • Improvements
    • Numerator degrees of freedom in random.noncentral_f need only be positive.
    • The GIL is released for all np.einsum variations
    • The np.einsum function will use BLAS when possible and optimize by default
    • f2py now handles arrays of dimension 0
    • numpy.distutils supports using MSVC and mingw64-gfortran together
    • np.linalg.pinv now works on stacked matrices
    • numpy.save aligns data to 64 bytes instead of 16
    • NPZ files now can be written without using temporary files
    • Better support for empty structured and string types
    • Support for decimal.Decimal in np.lib.financial
    • Float printing now uses “dragon4” algorithm for shortest decimal representation
    • void datatype elements are now printed in hex notation
    • printing style for void datatypes is now independently customizable
    • Reduced memory usage of np.loadtxt
  • Changes
    • Multiple-field indexing/assignment of structured arrays
    • Integer and Void scalars are now unaffected by np.set_string_function
    • 0d array printing changed, style arg of array2string deprecated
    • Seeding RandomState using an array requires a 1-d array
    • MaskedArray objects show a more useful repr
    • The repr of np.polynomial classes is more explicit
• 1.13.3
  • Contributors
  • Pull requests merged
• 1.13.2
  • Contributors
  • Pull requests merged
• 1.13.1
  • Pull requests merged
  • Contributors
• 1.13.0
  • Highlights
  • New functions
  • Deprecations
  • Future Changes
  • Build System Changes
  • Compatibility notes
    • Error type changes
    • Tuple object dtypes
    • DeprecationWarning to error
    • FutureWarning to changed behavior
    • dtypes are now always true
    • __getslice__ and __setslice__ are no longer needed in ndarray subclasses
    • Indexing MaskedArrays/Constants with ... (ellipsis) now returns MaskedArray
  • C API changes
    • GUfuncs on empty arrays and NpyIter axis removal
    • PyArray_MapIterArrayCopyIfOverlap added to NumPy C-API
  • New Features
    • __array_ufunc__ added
    • New positive ufunc
    • New divmod ufunc
    • np.isnat ufunc tests for NaT special datetime and timedelta values
    • np.heaviside ufunc computes the Heaviside function
    • np.block function for creating blocked arrays
    • isin function, improving on in1d
    • Temporary elision
    • axes argument for unique
    • np.gradient now supports unevenly spaced data
    • Support for returning arrays of arbitrary dimensions in apply_along_axis
    • .ndim property added to dtype to complement .shape
    • Support for tracemalloc in Python 3.6
    • NumPy may be built with relaxed stride checking debugging
  • Improvements
    • Ufunc behavior for overlapping inputs
    • Partial support for 64-bit f2py extensions with MinGW
    • Performance improvements for packbits and unpackbits
    • Fix for PPC long double floating point information
    • Better default repr for ndarray subclasses
    • More reliable comparisons of masked arrays
    • np.matrix with booleans elements can now be created using the string syntax
    • More linalg operations now accept empty vectors and matrices
    • Bundled version of LAPACK is now 3.2.2
    • reduce of np.hypot.reduce and np.logical_xor allowed in more cases
    • Better repr of object arrays
  • Changes
    • argsort on masked arrays takes the same default arguments as sort
    • average now preserves subclasses
    • array == None and array != None do element-wise comparison
    • np.equal, np.not_equal for object arrays ignores object identity
    • Boolean indexing changes
    • np.random.multivariate_normal behavior with bad covariance matrix
    • assert_array_less compares np.inf and -np.inf now
    • assert_array_ and masked arrays assert_equal hide less warnings
    • offset attribute value in memmap objects
    • np.real and np.imag return scalars for scalar inputs
    • The polynomial convenience classes cannot be passed to ufuncs
    • Output arguments to ufuncs can be tuples also for ufunc methods
• 1.12.1
  • Bugs Fixed
• 1.12.0
  • Highlights
  • Dropped Support
  • Added Support
  • Build System Changes
  • Deprecations
    • Assignment of ndarray object’s data attribute
    • Unsafe int casting of the num attribute in linspace
    • Insufficient bit width parameter to binary_repr
  • Future Changes
    • Multiple-field manipulation of structured arrays
  • Compatibility notes
    • DeprecationWarning to error
    • FutureWarning to changed behavior
    • power and ** raise errors for integer to negative integer powers
    • Relaxed stride checking is the default
    • The np.percentile ‘midpoint’ interpolation method fixed for exact indices
    • keepdims kwarg is passed through to user-class methods
    • bitwise_and identity changed
    • ma.median warns and returns nan when unmasked invalid values are encountered
    • Greater consistency in assert_almost_equal
    • NoseTester behaviour of warnings during testing
    • assert_warns and deprecated decorator more specific
    • C API
  • New Features
    • Writeable keyword argument for as_strided
    • axes keyword argument for rot90
    • Generalized flip
    • BLIS support in numpy.distutils
    • Hook in numpy/__init__.py to run distribution-specific checks
    • New nanfunctions nancumsum and nancumprod added
    • np.interp can now interpolate complex values
    • New polynomial evaluation function polyvalfromroots added
    • New array creation function geomspace added
    • New context manager for testing warnings
    • New masked array functions ma.convolve and ma.correlate added
    • New float_power ufunc
    • np.loadtxt now supports a single integer as usecol argument
    • Improved automated bin estimators for histogram
    • np.roll can now roll multiple axes at the same time
    • The __complex__ method has been implemented for the ndarrays
    • pathlib.Path objects now supported
    • New bits attribute for np.finfo
    • New signature argument to np.vectorize
    • Emit py3kwarnings for division of integer arrays
    • numpy.sctypes now includes bytes on Python3 too
  • Improvements
    • bitwise_and identity changed
    • Generalized Ufuncs will now unlock the GIL
    • Caches in np.fft are now bounded in total size and item count
    • Improved handling of zero-width string/unicode dtypes
    • Integer ufuncs vectorized with AVX2
    • Order of operations optimization in np.einsum
    • quicksort has been changed to an introsort
    • ediff1d improved performance and subclass handling
    • Improved precision of ndarray.mean for float16 arrays
  • Changes
    • All array-like methods are now called with keyword arguments in fromnumeric.py
    • Operations on np.memmap objects return numpy arrays in most cases
    • stacklevel of warnings increased
• 1.11.3
  • Contributors to maintenance/1.11.3
  • Pull Requests Merged
• 1.11.2
  • Pull Requests Merged
• 1.11.1
  • Fixes Merged
• 1.11.0
  • Highlights
  • Build System Changes
  • Future Changes
  • Compatibility notes
    • datetime64 changes
    • linalg.norm return type changes
    • polynomial fit changes
    • np.dot now raises TypeError instead of ValueError
    • FutureWarning to changed behavior
    • % and // operators
    • C API
    • object dtype detection for old-style classes
  • New Features
  • Improvements
    • np.gradient now supports an axis argument
    • np.lexsort now supports arrays with object data-type
    • np.ma.core.MaskedArray now supports an order argument
    • Memory and speed improvements for masked arrays
    • ndarray.tofile now uses fallocate on linux
    • Optimizations for operations of the form A.T @ A and A @ A.T
    • np.testing.assert_warns can now be used as a context manager
    • Speed improvement for np.random.shuffle
  • Changes
    • Pyrex support was removed from numpy.distutils
    • np.broadcast can now be called with a single argument
    • np.trace now respects array subclasses
    • np.dot now raises TypeError instead of ValueError
    • linalg.norm return type changes
  • Deprecations
    • Views of arrays in Fortran order
    • Invalid arguments for array ordering
    • Random number generator in the testing namespace
    • Random integer generation on a closed interval
  • FutureWarnings
    • Assigning to slices/views of MaskedArray
• 1.10.4
  • Compatibility notes
  • Issues Fixed
  • Merged PRs
• 1.10.3
• 1.10.2
  • Compatibility notes
    • Relaxed stride checking is no longer the default
    • Fix swig bug in numpy.i
    • Deprecate views changing dimensions in fortran order
  • Issues Fixed
  • Merged PRs
  • Notes
• 1.10.1
• 1.10.0
  • Highlights
  • Dropped Support
  • Future Changes
  • Compatibility notes
    • Default casting rule change
    • numpy version string
    • relaxed stride checking
    • Concatenation of 1d arrays along any but axis=0 raises IndexError
    • np.ravel, np.diagonal and np.diag now preserve subtypes
    • rollaxis and swapaxes always return a view
    • nonzero now returns base ndarrays
    • C API
    • recarray field return types
    • recarray views
    • ‘out’ keyword argument of ufuncs now accepts tuples of arrays
    • byte-array indices now raises an IndexError
    • Masked arrays containing objects with arrays
    • Median warns and returns nan when invalid values are encountered
    • Functions available from numpy.ma.testutils have changed
  • New Features
    • Reading extra flags from site.cfg
    • np.cbrt to compute cube root for real floats
    • numpy.distutils now allows parallel compilation
    • genfromtxt has a new max_rows argument
    • New function np.broadcast_to for invoking array broadcasting
    • New context manager clear_and_catch_warnings for testing warnings
    • cov has new fweights and aweights arguments
    • Support for the ‘@’ operator in Python 3.5+
    • New argument norm to fft functions
  • Improvements
    • np.digitize using binary search
    • np.poly now casts integer inputs to float
    • np.interp can now be used with periodic functions
    • np.pad supports more input types for pad_width and constant_values
    • np.argmax and np.argmin now support an out argument
    • More system C99 complex functions detected and used
    • np.loadtxt support for the strings produced by the float.hex method
    • np.isclose properly handles minimal values of integer dtypes
    • np.allclose uses np.isclose internally.
    • np.genfromtxt now handles large integers correctly
    • np.load, np.save have pickle backward compatibility flags
    • MaskedArray support for more complicated base classes
  • Changes
    • dotblas functionality moved to multiarray
    • stricter check of gufunc signature compliance
    • views returned from np.einsum are writeable
    • np.argmin skips NaT values
  • Deprecations
    • Array comparisons involving strings or structured dtypes
    • SafeEval
    • alterdot, restoredot
    • pkgload, PackageLoader
    • bias, ddof arguments to corrcoef
    • dtype string representation changes
• 1.9.2
  • Issues fixed
• 1.9.1
  • Issues fixed
• 1.9.0
  • Highlights
  • Dropped Support
  • Future Changes
  • Compatibility notes
    • The diagonal and diag functions return readonly views.
    • Special scalar float values don’t cause upcast to double anymore
    • Percentile output changes
    • ndarray.tofile exception type
    • Invalid fill value exceptions
    • Polynomial Classes no longer derived from PolyBase
    • Using numpy.random.binomial may change the RNG state vs. numpy < 1.9
    • Random seed enforced to be a 32 bit unsigned integer
    • Argmin and argmax out argument
    • Einsum
    • Indexing
    • Non-integer reduction axis indexes are deprecated
    • promote_types and string dtype
    • can_cast and string dtype
    • astype and string dtype
    • npyio.recfromcsv keyword arguments change
    • The doc/swig directory moved
    • The npy_3kcompat.h header changed
    • Negative indices in C-Api sq_item and sq_ass_item sequence methods
    • NDIter
    • zeros_like for string dtypes now returns empty strings
  • New Features
    • Percentile supports more interpolation options
    • Generalized axis support for median and percentile
    • Dtype parameter added to np.linspace and np.logspace
    • More general np.triu and np.tril broadcasting
    • tobytes alias for tostring method
    • Build system
    • Compatibility to python numbers module
    • increasing parameter added to np.vander
    • unique_counts parameter added to np.unique
    • Support for median and percentile in nanfunctions
    • NumpyVersion class added
    • Allow saving arrays with large number of named columns
    • Full broadcasting support for np.cross
  • Improvements
    • Better numerical stability for sum in some cases
    • Percentile implemented in terms of np.partition
    • Performance improvement for np.array
    • Performance improvement for np.searchsorted
    • Optional reduced verbosity for np.distutils
    • Covariance check in np.random.multivariate_normal
    • Polynomial Classes no longer template based
    • More GIL releases
    • MaskedArray support for more complicated base classes
    • C-API
  • Deprecations
    • Non-integer scalars for sequence repetition
    • select input deprecations
    • rank function
    • Object array equality comparisons
    • C-API
• 1.8.2
  • Issues fixed
• 1.8.1
  • Issues fixed
  • Changes
    • NDIter
    • Optional reduced verbosity for np.distutils
  • Deprecations
    • C-API
• 1.8.0
  • Highlights
  • Dropped Support
  • Future Changes
  • Compatibility notes
    • NPY_RELAXED_STRIDES_CHECKING
    • Binary operations with non-arrays as second argument
    • Function median used with overwrite_input only partially sorts array
    • Fix to financial.npv
    • Runtime warnings when comparing NaN numbers
  • New Features
    • Support for linear algebra on stacked arrays
    • In place fancy indexing for ufuncs
    • New functions partition and argpartition
    • New functions nanmean, nanvar and nanstd
    • New functions full and full_like
    • IO compatibility with large files
    • Building against OpenBLAS
    • New constant
    • New modes for qr
    • New invert argument to in1d
    • Advanced indexing using np.newaxis
    • C-API
    • runtests.py
  • Improvements
    • IO performance improvements
    • Performance improvements to pad
    • Performance improvements to isnan, isinf, isfinite and byteswap
    • Performance improvements via SSE2 vectorization
    • Performance improvements to median
    • Overridable operand flags in ufunc C-API
  • Changes
    • General
    • C-API Array Additions
    • C-API Ufunc Additions
    • C-API Developer Improvements
  • Deprecations
    • General
  • Authors
• 1.7.2
  • Issues fixed
• 1.7.1
  • Issues fixed
• 1.7.0
  • Highlights
  • Compatibility notes
  • New features
    • Reduction UFuncs Generalize axis= Parameter
    • Reduction UFuncs New keepdims= Parameter
    • Datetime support
    • Custom formatter for printing arrays
    • New function numpy.random.choice
    • New function isclose
    • Preliminary multi-dimensional support in the polynomial package
    • Ability to pad rank-n arrays
    • New argument to searchsorted
    • Build system
    • C API
  • Changes
    • General
    • Casting Rules
  • Deprecations
    • General
    • C-API
• 1.6.2
  • Issues fixed
    • numpy.core
    • numpy.lib
    • numpy.distutils
    • numpy.random
  • Changes
    • numpy.f2py
    • numpy.poly
• 1.6.1
  • Issues Fixed
• 1.6.0
  • Highlights
  • New features
    • New 16-bit floating point type
    • New iterator
    • Legendre, Laguerre, Hermite, HermiteE polynomials in numpy.polynomial
    • Fortran assumed shape array and size function support in numpy.f2py
    • Other new functions
  • Changes
    • default error handling
    • numpy.distutils
    • numpy.testing
    • C API
  • Deprecated features
  • Removed features
    • numpy.fft
    • numpy.memmap
    • numpy.lib
    • numpy.ma
    • numpy.distutils
• 1.5.0
  • Highlights
    • Python 3 compatibility
    • PEP 3118 compatibility
  • New features
    • Warning on casting complex to real
    • Dot method for ndarrays
    • linalg.slogdet function
    • new header
  • Changes
    • polynomial.polynomial
    • polynomial.chebyshev
    • histogram
    • correlate
• 1.4.0
  • Highlights
  • New features
    • Extended array wrapping mechanism for ufuncs
    • Automatic detection of forward incompatibilities
    • New iterators
    • New polynomial support
    • New C API
    • New ufuncs
    • New defines
    • Testing
    • Reusing npymath
    • Improved set operations
  • Improvements
  • Deprecations
  • Internal changes
    • Use C99 complex functions when available
    • split multiarray and umath source code
    • Separate compilation
    • Separate core math library
• 1.3.0
  • Highlights
    • Python 2.6 support
    • Generalized ufuncs
    • Experimental Windows 64 bits support
  • New features
    • Formatting issues
    • Nan handling in max/min
    • Nan handling in sign
    • New ufuncs
    • Masked arrays
    • gfortran support on windows
    • Arch option for windows binary
  • Deprecated features
    • Histogram
  • Documentation changes
  • New C API
    • Multiarray API
    • Ufunc API
    • New defines
    • Portable NAN, INFINITY, etc…
  • Internal changes
    • numpy.core math configuration revamp
    • umath refactor
    • Improvements to build warnings
    • Separate core math library
    • CPU arch detection