This NumPy release is marked by the removal of much technical debt: support for
Python 2 has been removed, many deprecations have been expired, and
documentation has been improved. The polishing of the random module continues
apace with bug fixes and better usability from Cython.
The Python versions supported for this release are 3.6-3.8. Downstream
developers should use Cython >= 0.29.16 for Python 3.8 support and
OpenBLAS >= 3.7 to avoid problems on the Skylake architecture.
Code compatibility with Python versions < 3.6 (including Python 2) was
dropped from both the python and C code. The shims in numpy.compat will
remain to support third-party packages, but they may be deprecated in a
future release. Note that 1.19.x will not compile with earlier versions of
Python due to the use of f-strings.
This concludes a deprecation from 1.9, where when an axis argument was
passed to a call to ~numpy.insert and ~numpy.delete on a 0d array, the
axis and obj argument and indices would be completely ignored.
In these cases, insert(arr, "nonsense", 42, axis=0) would actually overwrite the
entire array, while delete(arr, "nonsense", axis=0) would be arr.copy()
insert(arr, "nonsense", 42, axis=0)
delete(arr, "nonsense", axis=0)
Now passing axis on a 0d array raises ~numpy.AxisError.
This concludes deprecations from 1.8 and 1.9, where np.delete would ignore
both negative and out-of-bounds items in a sequence of indices. This was at
odds with its behavior when passed a single index.
Now out-of-bounds items throw IndexError, and negative items index from the
This concludes a deprecation from 1.9, where sequences of non-integers indices
were allowed and cast to integers. Now passing sequences of non-integral
indices raises IndexError, just like it does when passing a single
This concludes a deprecation from 1.8, where np.delete would cast boolean
arrays and scalars passed as an index argument into integer indices. The
behavior now is to treat boolean arrays as a mask, and to raise an error
on boolean scalars.
A bug in the generation of random variates for the Dirichlet distribution
with small ‘alpha’ values was fixed by using a different algorithm when
max(alpha) < 0.1. Because of the change, the stream of variates
generated by dirichlet in this case will be different from previous
max(alpha) < 0.1
The promotion of mixed scalars and arrays in PyArray_ConvertToCommonType
has been changed to adhere to those used by np.result_type.
This means that input such as (1000, np.array(, dtype=np.uint8)))
will now return uint16 dtypes. In most cases the behaviour is unchanged.
Note that the use of this C-API function is generally discouraged.
This also fixes np.choose to behave the same way as the rest of NumPy
in this respect.
(1000, np.array(, dtype=np.uint8)))
The fasttake and fastputmask slots are now never used and
must always be set to NULL. This will result in no change in behaviour.
However, if a user dtype should set one of these a DeprecationWarning
will be given.
np.ediff1d now uses the "same_kind" casting rule for
its additional to_end and to_begin arguments. This
ensures type safety except when the input array has a smaller
integer type than to_begin or to_end.
In rare cases, the behaviour will be more strict than it was
previously in 1.16 and 1.17. This is necessary to solve issues
with floating point NaN.
Objects with len(obj) == 0 which implement an “array-like” interface,
meaning an object implementing obj.__array__(),
obj.__array_interface__, obj.__array_struct__, or the python
buffer interface and which are also sequences (i.e. Pandas objects)
will now always retain there shape correctly when converted to an array.
If such an object has a shape of (0, 1) previously, it could
be converted into an array of shape (0,) (losing all dimensions
after the first 0).
len(obj) == 0
As part of the continued removal of Python 2 compatibility,
multiarray.int_asbuffer was removed. On Python 3, it threw a
NotImplementedError and was unused internally. It is expected that there
are no downstream use cases for this method with Python 3.
This module contained only the function get_exception(), which was used as:
e = get_exception()
Its purpose was to handle the change in syntax introduced in Python 2.6, from
except Exception, e: to except Exception as e:, meaning it was only
necessary for codebases supporting Python 2.5 and older.
except Exception, e:
except Exception as e:
numpy.issubdtype had a FutureWarning since NumPy 1.14 which
has expired now. This means that certain input where the second
argument was neither a datatype nor a NumPy scalar type
(such as a string or a python type like int or float)
will now be consistent with passing in np.dtype(arg2).type.
This makes the result consistent with expectations and leads to
a false result in some cases which previously returned true.
Output of the __round__ dunder method and consequently the Python
built-in round has been changed to be a Python int to be consistent
with calling it on Python float objects when called with no arguments.
Previously, it would return a scalar of the np.dtype that was passed in.
The former has changed to have the expected meaning of setting
numpy.ndarray.strides to (), while the latter continues to result in
strides being chosen automatically.
The C-level casts from strings were simplified. This changed
also fixes string to datetime and timedelta casts to behave
correctly (i.e. like Python casts using string_arr.astype("M8")
while previously the cast would behave like
This only affects code using low-level C-API to do manual casts
(not full array casts) of single scalar values or using e.g.
PyArray_GetCastFunc, and should thus not affect the vast majority
Small seeds (less than 2**96) were previously implicitly 0-padded out to
128 bits, the size of the internal entropy pool. When spawned, the spawn key
was concatenated before the 0-padding. Since the first spawn key is (0,),
small seeds before the spawn created the same states as the first spawned
SeedSequence. Now, the seed is explicitly 0-padded out to the internal
pool size before concatenating the spawn key. Spawned SeedSequences will
produce different results than in the previous release. Unspawned
SeedSequences will still produce the same results.
Calling np.array([[1, [1, 2, 3]]) will issue a DeprecationWarning as
per NEP 34. Users should explicitly use dtype=object to avoid the
np.array([[1, [1, 2, 3]])
0 is treated as a special case and is aliased to None in the functions:
In future, 0 will not be special cased, and will be treated as an array
length like any other integer.
The following C-API functions are probably unused and have been
In most cases PyArray_GetArrayParamsFromObject should be replaced
by converting to an array, while PyUFunc_GenericFunction can be
replaced with PyObject_Call (see documentation for details).
The super classes of scalar types, such as np.integer, np.generic,
or np.inexact will now give a deprecation warning when converted
to a dtype (or used in a dtype keyword argument).
The reason for this is that np.integer is converted to np.int_,
while it would be expected to represent any integer (e.g. also
int8, int16, etc.
For example, dtype=np.floating is currently identical to
dtype=np.float64, even though also np.float32 is a subclass of
Output of the __round__ dunder method and consequently the Python built-in
round has been deprecated on complex scalars. This does not affect
~numpy.ndarray.tobytes has existed since the 1.9 release, but until this
release ~numpy.ndarray.tostring emitted no warning. The change to emit a
warning brings NumPy in line with the builtin array.array methods of the
The following functions now accept a constant array of npy_intp:
Previously the caller would have to cast away the const-ness to call these
UFuncGenericFunction now expects pointers to const dimension and
strides as arguments. This means inner loops may no longer modify
either dimension or strides. This change leads to an
incompatible-pointer-types warning forcing users to either ignore
the compiler warnings or to const qualify their own loop signatures.
This allows the :attr:numpy.ufunc.identity attribute to be set on the
resulting ufunc, meaning it can be used for empty and multi-dimensional
calls to :meth:numpy.ufunc.reduce.
np.str_ arrays are always stored as UCS4, so the corresponding scalars
now expose this through the buffer interface, meaning
memoryview(np.str_('test')) now works.
A new kwarg, subok, was added to numpy.copy to allow users to toggle
the behavior of numpy.copy with respect to array subclasses. The default
value is False which is consistent with the behavior of numpy.copy for
previous numpy versions. To create a copy that preserves an array subclass with
numpy.copy, call np.copy(arr, subok=True). This addition better
documents that the default behavior of numpy.copy differs from the
numpy.ndarray.copy method which respects array subclasses by default.
out can be used to avoid creating unnecessary copies of the final product
computed by numpy.linalg.multidot.
The parameter keepdims was added to numpy.count_nonzero. The
parameter has the same meaning as it does in reduction functions such
as numpy.sum or numpy.mean.
The keyword argument equal_nan was added to numpy.array_equal.
equal_nan is a boolean value that toggles whether or not nan values are
considered equal in comparison (default is False). This matches API used in
related functions such as numpy.isclose and numpy.allclose.
Replace npy_cpu_supports which was a gcc specific mechanism to test support
of AVX with more general functions npy_cpu_init and npy_cpu_have, and
expose the results via a NPY_CPU_HAVE c-macro as well as a python-level
Use 64-bit integer size on 64-bit platforms in the fallback LAPACK library,
which is used when the system has no LAPACK installed, allowing it to deal with
linear algebra for large arrays.
Use AVX512 intrinsic to implement np.exp when input is np.float64,
which can improve the performance of np.exp with np.float64 input 5-7x
faster than before. The _multiarray_umath.so module has grown about 63 KB
On Linux NumPy has previously added support for madavise hugepages which can
improve performance for very large arrays. Unfortunately, on older Kernel
versions this led to peformance regressions, thus by default the support has
been disabled on kernels before version 4.6. To override the default, you can
use the environment variable:
or set it to 1 to force enabling support. Note that this only makes
a difference if the operating system is set up to use madvise
There is no longer a type error thrown when numpy.einsum is passed
a NumPy int64 array as its subscript list.
The ufunc ~numpy.logaddexp2 now has an identity of -inf, allowing it to
be called on empty sequences. This matches the identity of ~numpy.logaddexp.
A code path and test have been in the code since NumPy 0.4 for a two-argument
variant of __array__(dtype=None, context=None). It was activated when
calling ufunc(op) or ufunc.reduce(op) if op.__array__ existed.
However that variant is not documented, and it is not clear what the intention
was for its use. It has been removed.
In order to expose numpy.random.BitGenerator and
numpy.random.SeedSequence to Cython, the _bitgenerator module is now
public as numpy.random.bit_generator
c_distributions.pxd provides access to the c functions behind many of the
random distributions from Cython, making it convenient to use and extend them.
Previously, when passing method='eigh' or method='cholesky',
numpy.random.multivariate_normal produced samples from the wrong
distribution. This is now fixed.
This fix changes the stream produced from jumped MT19937 generators. It does
not affect the stream produced using RandomState or MT19937 that
are directly seeded.
The translation of the jumping code for the MT19937 contained a reversed loop
ordering. MT19937.jumped matches the Makoto Matsumoto’s original
implementation of the Horner and Sliding Window jump methods.