NumPy 1.15.0 Release Notes

NumPy 1.15.0 is a release with an unusual number of cleanups, many deprecations of old functions, and improvements to many existing functions. Please read the detailed descriptions below to see if you are affected.

For testing, we have switched to pytest as a replacement for the no longer maintained nose framework. The old nose based interface remains for downstream projects who may still be using it.

The Python versions supported by this release are 2.7, 3.4-3.7. The wheels are linked with OpenBLAS v0.3.0, which should fix some of the linalg problems reported for NumPy 1.14.

Highlights

  • NumPy has switched to pytest for testing.

  • A new numpy.printoptions context manager.

  • Many improvements to the histogram functions.

  • Support for unicode field names in python 2.7.

  • Improved support for PyPy.

  • Fixes and improvements to numpy.einsum.

New functions

  • numpy.gcd and numpy.lcm, to compute the greatest common divisor and least common multiple.

  • numpy.ma.stack, the numpy.stack array-joining function generalized to masked arrays.

  • numpy.quantile function, an interface to percentile without factors of 100

  • numpy.nanquantile function, an interface to nanpercentile without factors of 100

  • numpy.printoptions, a context manager that sets print options temporarily for the scope of the with block:

    >>> with np.printoptions(precision=2):
    ...     print(np.array([2.0]) / 3)
    [0.67]
    
  • numpy.histogram_bin_edges, a function to get the edges of the bins used by a histogram without needing to calculate the histogram.

  • C functions npy_get_floatstatus_barrier and npy_clear_floatstatus_barrier have been added to deal with compiler optimization changing the order of operations. See below for details.

Deprecations

  • Aliases of builtin pickle functions are deprecated, in favor of their unaliased pickle.<func> names:

    • numpy.loads

    • numpy.core.numeric.load

    • numpy.core.numeric.loads

    • numpy.ma.loads, numpy.ma.dumps

    • numpy.ma.load, numpy.ma.dump - these functions already failed on python 3 when called with a string.

  • Multidimensional indexing with anything but a tuple is deprecated. This means that the index list in ind = [slice(None), 0]; arr[ind] should be changed to a tuple, e.g., ind = [slice(None), 0]; arr[tuple(ind)] or arr[(slice(None), 0)]. That change is necessary to avoid ambiguity in expressions such as arr[[[0, 1], [0, 1]]], currently interpreted as arr[array([0, 1]), array([0, 1])], that will be interpreted as arr[array([[0, 1], [0, 1]])] in the future.

  • Imports from the following sub-modules are deprecated, they will be removed at some future date.

    • numpy.testing.utils

    • numpy.testing.decorators

    • numpy.testing.nosetester

    • numpy.testing.noseclasses

    • numpy.core.umath_tests

  • Giving a generator to numpy.sum is now deprecated. This was undocumented behavior, but worked. Previously, it would calculate the sum of the generator expression. In the future, it might return a different result. Use np.sum(np.from_iter(generator)) or the built-in Python sum instead.

  • Users of the C-API should call PyArrayResolveWriteBackIfCopy or PyArray_DiscardWritbackIfCopy on any array with the WRITEBACKIFCOPY flag set, before deallocating the array. A deprecation warning will be emitted if those calls are not used when needed.

  • Users of nditer should use the nditer object as a context manager anytime one of the iterator operands is writeable, so that numpy can manage writeback semantics, or should call it.close(). A RuntimeWarning may be emitted otherwise in these cases.

  • The normed argument of np.histogram, deprecated long ago in 1.6.0, now emits a DeprecationWarning.

Future Changes

  • NumPy 1.16 will drop support for Python 3.4.

  • NumPy 1.17 will drop support for Python 2.7.

Compatibility notes

Compiled testing modules renamed and made private

The following compiled modules have been renamed and made private:

  • umath_tests -> _umath_tests

  • test_rational -> _rational_tests

  • multiarray_tests -> _multiarray_tests

  • struct_ufunc_test -> _struct_ufunc_tests

  • operand_flag_tests -> _operand_flag_tests

The umath_tests module is still available for backwards compatibility, but will be removed in the future.

The NpzFile returned by np.savez is now a collections.abc.Mapping

This means it behaves like a readonly dictionary, and has a new .values() method and len() implementation.

For python 3, this means that .iteritems(), .iterkeys() have been deprecated, and .keys() and .items() now return views and not lists. This is consistent with how the builtin dict type changed between python 2 and python 3.

Under certain conditions, nditer must be used in a context manager

When using an numpy.nditer with the "writeonly" or "readwrite" flags, there are some circumstances where nditer doesn’t actually give you a view of the writable array. Instead, it gives you a copy, and if you make changes to the copy, nditer later writes those changes back into your actual array. Currently, this writeback occurs when the array objects are garbage collected, which makes this API error-prone on CPython and entirely broken on PyPy. Therefore, nditer should now be used as a context manager whenever it is used with writeable arrays, e.g., with np.nditer(...) as it: .... You may also explicitly call it.close() for cases where a context manager is unusable, for instance in generator expressions.

Numpy has switched to using pytest instead of nose for testing

The last nose release was 1.3.7 in June, 2015, and development of that tool has ended, consequently NumPy has now switched to using pytest. The old decorators and nose tools that were previously used by some downstream projects remain available, but will not be maintained. The standard testing utilities, assert_almost_equal and such, are not be affected by this change except for the nose specific functions import_nose and raises. Those functions are not used in numpy, but are kept for downstream compatibility.

Numpy no longer monkey-patches ctypes with __array_interface__

Previously numpy added __array_interface__ attributes to all the integer types from ctypes.

np.ma.notmasked_contiguous and np.ma.flatnotmasked_contiguous always return lists

This is the documented behavior, but previously the result could be any of slice, None, or list.

All downstream users seem to check for the None result from flatnotmasked_contiguous and replace it with []. Those callers will continue to work as before.

np.squeeze restores old behavior of objects that cannot handle an axis argument

Prior to version 1.7.0, numpy.squeeze did not have an axis argument and all empty axes were removed by default. The incorporation of an axis argument made it possible to selectively squeeze single or multiple empty axes, but the old API expectation was not respected because axes could still be selectively removed (silent success) from an object expecting all empty axes to be removed. That silent, selective removal of empty axes for objects expecting the old behavior has been fixed and the old behavior restored.

unstructured void array’s .item method now returns a bytes object

.item now returns a bytes object instead of a buffer or byte array. This may affect code which assumed the return value was mutable, which is no longer the case.

copy.copy and copy.deepcopy no longer turn masked into an array

Since np.ma.masked is a readonly scalar, copying should be a no-op. These functions now behave consistently with np.copy().

Multifield Indexing of Structured Arrays will still return a copy

The change that multi-field indexing of structured arrays returns a view instead of a copy is pushed back to 1.16. A new method numpy.lib.recfunctions.repack_fields has been introduced to help mitigate the effects of this change, which can be used to write code compatible with both numpy 1.15 and 1.16. For more information on how to update code to account for this future change see the “accessing multiple fields” section of the user guide.

C API changes

New functions npy_get_floatstatus_barrier and npy_clear_floatstatus_barrier

Functions npy_get_floatstatus_barrier and npy_clear_floatstatus_barrier have been added and should be used in place of the npy_get_floatstatus``and ``npy_clear_status functions. Optimizing compilers like GCC 8.1 and Clang were rearranging the order of operations when the previous functions were used in the ufunc SIMD functions, resulting in the floatstatus flags being checked before the operation whose status we wanted to check was run. See #10339.

Changes to PyArray_GetDTypeTransferFunction

PyArray_GetDTypeTransferFunction now defaults to using user-defined copyswapn / copyswap for user-defined dtypes. If this causes a significant performance hit, consider implementing copyswapn to reflect the implementation of PyArray_GetStridedCopyFn. See #10898.

New Features

np.gcd and np.lcm ufuncs added for integer and objects types

These compute the greatest common divisor, and lowest common multiple, respectively. These work on all the numpy integer types, as well as the builtin arbitrary-precision Decimal and long types.

Support for cross-platform builds for iOS

The build system has been modified to add support for the _PYTHON_HOST_PLATFORM environment variable, used by distutils when compiling on one platform for another platform. This makes it possible to compile NumPy for iOS targets.

This only enables you to compile NumPy for one specific platform at a time. Creating a full iOS-compatible NumPy package requires building for the 5 architectures supported by iOS (i386, x86_64, armv7, armv7s and arm64), and combining these 5 compiled builds products into a single “fat” binary.

return_indices keyword added for np.intersect1d

New keyword return_indices returns the indices of the two input arrays that correspond to the common elements.

np.quantile and np.nanquantile

Like np.percentile and np.nanpercentile, but takes quantiles in [0, 1] rather than percentiles in [0, 100]. np.percentile is now a thin wrapper around np.quantile with the extra step of dividing by 100.

Build system

Added experimental support for the 64-bit RISC-V architecture.

Improvements

np.einsum updates

Syncs einsum path optimization tech between numpy and opt_einsum. In particular, the greedy path has received many enhancements by @jcmgray. A full list of issues fixed are:

  • Arbitrary memory can be passed into the greedy path. Fixes gh-11210.

  • The greedy path has been updated to contain more dynamic programming ideas preventing a large number of duplicate (and expensive) calls that figure out the actual pair contraction that takes place. Now takes a few seconds on several hundred input tensors. Useful for matrix product state theories.

  • Reworks the broadcasting dot error catching found in gh-11218 gh-10352 to be a bit earlier in the process.

  • Enhances the can_dot functionality that previous missed an edge case (part of gh-11308).

np.flip can operate over multiple axes

np.flip now accepts None, or tuples of int, in its axis argument. If axis is None, it will flip over all the axes.

histogram and histogramdd functions have moved to np.lib.histograms

These were originally found in np.lib.function_base. They are still available under their un-scoped np.histogram(dd) names, and to maintain compatibility, aliased at np.lib.function_base.histogram(dd).

Code that does from np.lib.function_base import * will need to be updated with the new location, and should consider not using import * in future.

histogram will accept NaN values when explicit bins are given

Previously it would fail when trying to compute a finite range for the data. Since the range is ignored anyway when the bins are given explicitly, this error was needless.

Note that calling histogram on NaN values continues to raise the RuntimeWarning s typical of working with nan values, which can be silenced as usual with errstate.

histogram works on datetime types, when explicit bin edges are given

Dates, times, and timedeltas can now be histogrammed. The bin edges must be passed explicitly, and are not yet computed automatically.

histogram “auto” estimator handles limited variance better

No longer does an IQR of 0 result in n_bins=1, rather the number of bins chosen is related to the data size in this situation.

The edges returned by histogram` and histogramdd now match the data float type

When passed np.float16, np.float32, or np.longdouble data, the returned edges are now of the same dtype. Previously, histogram would only return the same type if explicit bins were given, and histogram would produce float64 bins no matter what the inputs.

histogramdd allows explicit ranges to be given in a subset of axes

The range argument of numpy.histogramdd can now contain None values to indicate that the range for the corresponding axis should be computed from the data. Previously, this could not be specified on a per-axis basis.

The normed arguments of histogramdd and histogram2d have been renamed

These arguments are now called density, which is consistent with histogram. The old argument continues to work, but the new name should be preferred.

np.r_ works with 0d arrays, and np.ma.mr_ works with np.ma.masked

0d arrays passed to the r_ and mr_ concatenation helpers are now treated as though they are arrays of length 1. Previously, passing these was an error. As a result, numpy.ma.mr_ now works correctly on the masked constant.

np.ptp accepts a keepdims argument, and extended axis tuples

np.ptp (peak-to-peak) can now work over multiple axes, just like np.max and np.min.

MaskedArray.astype now is identical to ndarray.astype

This means it takes all the same arguments, making more code written for ndarray work for masked array too.

Enable AVX2/AVX512 at compile time

Change to simd.inc.src to allow use of AVX2 or AVX512 at compile time. Previously compilation for avx2 (or 512) with -march=native would still use the SSE code for the simd functions even when the rest of the code got AVX2.

nan_to_num always returns scalars when receiving scalar or 0d inputs

Previously an array was returned for integer scalar inputs, which is inconsistent with the behavior for float inputs, and that of ufuncs in general. For all types of scalar or 0d input, the result is now a scalar.

np.flatnonzero works on numpy-convertible types

np.flatnonzero now uses np.ravel(a) instead of a.ravel(), so it works for lists, tuples, etc.

np.interp returns numpy scalars rather than builtin scalars

Previously np.interp(0.5, [0, 1], [10, 20]) would return a float, but now it returns a np.float64 object, which more closely matches the behavior of other functions.

Additionally, the special case of np.interp(object_array_0d, ...) is no longer supported, as np.interp(object_array_nd) was never supported anyway.

As a result of this change, the period argument can now be used on 0d arrays.

Allow dtype field names to be unicode in Python 2

Previously np.dtype([(u'name', float)]) would raise a TypeError in Python 2, as only bytestrings were allowed in field names. Now any unicode string field names will be encoded with the ascii codec, raising a UnicodeEncodeError upon failure.

This change makes it easier to write Python 2/3 compatible code using from __future__ import unicode_literals, which previously would cause string literal field names to raise a TypeError in Python 2.

Comparison ufuncs accept dtype=object, overriding the default bool

This allows object arrays of symbolic types, which override == and other operators to return expressions, to be compared elementwise with np.equal(a, b, dtype=object).

sort functions accept kind='stable'

Up until now, to perform a stable sort on the data, the user must do:

>>> np.sort([5, 2, 6, 2, 1], kind='mergesort')
[1, 2, 2, 5, 6]

because merge sort is the only stable sorting algorithm available in NumPy. However, having kind=’mergesort’ does not make it explicit that the user wants to perform a stable sort thus harming the readability.

This change allows the user to specify kind=’stable’ thus clarifying the intent.

Do not make temporary copies for in-place accumulation

When ufuncs perform accumulation they no longer make temporary copies because of the overlap between input an output, that is, the next element accumulated is added before the accumulated result is stored in its place, hence the overlap is safe. Avoiding the copy results in faster execution.

linalg.matrix_power can now handle stacks of matrices

Like other functions in linalg, matrix_power can now deal with arrays of dimension larger than 2, which are treated as stacks of matrices. As part of the change, to further improve consistency, the name of the first argument has been changed to a (from M), and the exceptions for non-square matrices have been changed to LinAlgError (from ValueError).

Increased performance in random.permutation for multidimensional arrays

permutation uses the fast path in random.shuffle for all input array dimensions. Previously the fast path was only used for 1-d arrays.

Generalized ufuncs now accept axes, axis and keepdims arguments

One can control over which axes a generalized ufunc operates by passing in an axes argument, a list of tuples with indices of particular axes. For instance, for a signature of (i,j),(j,k)->(i,k) appropriate for matrix multiplication, the base elements are two-dimensional matrices and these are taken to be stored in the two last axes of each argument. The corresponding axes keyword would be [(-2, -1), (-2, -1), (-2, -1)]. If one wanted to use leading dimensions instead, one would pass in [(0, 1), (0, 1), (0, 1)].

For simplicity, for generalized ufuncs that operate on 1-dimensional arrays (vectors), a single integer is accepted instead of a single-element tuple, and for generalized ufuncs for which all outputs are scalars, the (empty) output tuples can be omitted. Hence, for a signature of (i),(i)->() appropriate for an inner product, one could pass in axes=[0, 0] to indicate that the vectors are stored in the first dimensions of the two inputs arguments.

As a short-cut for generalized ufuncs that are similar to reductions, i.e., that act on a single, shared core dimension such as the inner product example above, one can pass an axis argument. This is equivalent to passing in axes with identical entries for all arguments with that core dimension (e.g., for the example above, axes=[(axis,), (axis,)]).

Furthermore, like for reductions, for generalized ufuncs that have inputs that all have the same number of core dimensions and outputs with no core dimension, one can pass in keepdims to leave a dimension with size 1 in the outputs, thus allowing proper broadcasting against the original inputs. The location of the extra dimension can be controlled with axes. For instance, for the inner-product example, keepdims=True, axes=[-2, -2, -2] would act on the inner-product example, keepdims=True, axis=-2 would act on the one-but-last dimension of the input arguments, and leave a size 1 dimension in that place in the output.

float128 values now print correctly on ppc systems

Previously printing float128 values was buggy on ppc, since the special double-double floating-point-format on these systems was not accounted for. float128s now print with correct rounding and uniqueness.

Warning to ppc users: You should upgrade glibc if it is version <=2.23, especially if using float128. On ppc, glibc’s malloc in these version often misaligns allocated memory which can crash numpy when using float128 values.

New np.take_along_axis and np.put_along_axis functions

When used on multidimensional arrays, argsort, argmin, argmax, and argpartition return arrays that are difficult to use as indices. take_along_axis provides an easy way to use these indices to lookup values within an array, so that:

np.take_along_axis(a, np.argsort(a, axis=axis), axis=axis)

is the same as:

np.sort(a, axis=axis)

np.put_along_axis acts as the dual operation for writing to these indices within an array.