NumPy 1.13.0 Release Notes

This release supports Python 2.7 and 3.4 - 3.6.

Highlights

  • Operations like a + b + c will reuse temporaries on some platforms, resulting in less memory use and faster execution.

  • Inplace operations check if inputs overlap outputs and create temporaries to avoid problems.

  • New __array_ufunc__ attribute provides improved ability for classes to override default ufunc behavior.

  • New np.block function for creating blocked arrays.

New functions

  • New np.positive ufunc.

  • New np.divmod ufunc provides more efficient divmod.

  • New np.isnat ufunc tests for NaT special values.

  • New np.heaviside ufunc computes the Heaviside function.

  • New np.isin function, improves on in1d.

  • New np.block function for creating blocked arrays.

  • New PyArray_MapIterArrayCopyIfOverlap added to NumPy C-API.

See below for details.

Deprecations

  • Calling np.fix, np.isposinf, and np.isneginf with f(x, y=out) is deprecated - the argument should be passed as f(x, out=out), which matches other ufunc-like interfaces.

  • Use of the C-API NPY_CHAR type number deprecated since version 1.7 will now raise deprecation warnings at runtime. Extensions built with older f2py versions need to be recompiled to remove the warning.

  • np.ma.argsort, np.ma.minimum.reduce, and np.ma.maximum.reduce should be called with an explicit axis argument when applied to arrays with more than 2 dimensions, as the default value of this argument (None) is inconsistent with the rest of numpy (-1, 0, and 0, respectively).

  • np.ma.MaskedArray.mini is deprecated, as it almost duplicates the functionality of np.MaskedArray.min. Exactly equivalent behaviour can be obtained with np.ma.minimum.reduce.

  • The single-argument form of np.ma.minimum and np.ma.maximum is deprecated. np.maximum. np.ma.minimum(x) should now be spelt np.ma.minimum.reduce(x), which is consistent with how this would be done with np.minimum.

  • Calling ndarray.conjugate on non-numeric dtypes is deprecated (it should match the behavior of np.conjugate, which throws an error).

  • Calling expand_dims when the axis keyword does not satisfy -a.ndim - 1 <= axis <= a.ndim, where a is the array being reshaped, is deprecated.

Future Changes

  • Assignment between structured arrays with different field names will change in NumPy 1.14. Previously, fields in the dst would be set to the value of the identically-named field in the src. In numpy 1.14 fields will instead be assigned ‘by position’: The n-th field of the dst will be set to the n-th field of the src array. Note that the FutureWarning raised in NumPy 1.12 incorrectly reported this change as scheduled for NumPy 1.13 rather than NumPy 1.14.

Build System Changes

  • numpy.distutils now automatically determines C-file dependencies with GCC compatible compilers.

Compatibility notes

Error type changes

  • numpy.hstack() now throws ValueError instead of IndexError when input is empty.

  • Functions taking an axis argument, when that argument is out of range, now throw np.AxisError instead of a mixture of IndexError and ValueError. For backwards compatibility, AxisError subclasses both of these.

Tuple object dtypes

Support has been removed for certain obscure dtypes that were unintentionally allowed, of the form (old_dtype, new_dtype), where either of the dtypes is or contains the object dtype. As an exception, dtypes of the form (object, [('name', object)]) are still supported due to evidence of existing use.

DeprecationWarning to error

See Changes section for more detail.

  • partition, TypeError when non-integer partition index is used.

  • NpyIter_AdvancedNew, ValueError when oa_ndim == 0 and op_axes is NULL

  • negative(bool_), TypeError when negative applied to booleans.

  • subtract(bool_, bool_), TypeError when subtracting boolean from boolean.

  • np.equal, np.not_equal, object identity doesn’t override failed comparison.

  • np.equal, np.not_equal, object identity doesn’t override non-boolean comparison.

  • Deprecated boolean indexing behavior dropped. See Changes below for details.

  • Deprecated np.alterdot() and np.restoredot() removed.

FutureWarning to changed behavior

See Changes section for more detail.

  • numpy.average preserves subclasses

  • array == None and array != None do element-wise comparison.

  • np.equal, np.not_equal, object identity doesn’t override comparison result.

dtypes are now always true

Previously bool(dtype) would fall back to the default python implementation, which checked if len(dtype) > 0. Since dtype objects implement __len__ as the number of record fields, bool of scalar dtypes would evaluate to False, which was unintuitive. Now bool(dtype) == True for all dtypes.

__getslice__ and __setslice__ are no longer needed in ndarray subclasses

When subclassing np.ndarray in Python 2.7, it is no longer _necessary_ to implement __*slice__ on the derived class, as __*item__ will intercept these calls correctly.

Any code that did implement these will work exactly as before. Code that invokes``ndarray.__getslice__`` (e.g. through super(...).__getslice__) will now issue a DeprecationWarning - .__getitem__(slice(start, end)) should be used instead.

Indexing MaskedArrays/Constants with ... (ellipsis) now returns MaskedArray

This behavior mirrors that of np.ndarray, and accounts for nested arrays in MaskedArrays of object dtype, and ellipsis combined with other forms of indexing.

C API changes

GUfuncs on empty arrays and NpyIter axis removal

It is now allowed to remove a zero-sized axis from NpyIter. Which may mean that code removing axes from NpyIter has to add an additional check when accessing the removed dimensions later on.

The largest followup change is that gufuncs are now allowed to have zero-sized inner dimensions. This means that a gufunc now has to anticipate an empty inner dimension, while this was never possible and an error raised instead.

For most gufuncs no change should be necessary. However, it is now possible for gufuncs with a signature such as (..., N, M) -> (..., M) to return a valid result if N=0 without further wrapping code.

PyArray_MapIterArrayCopyIfOverlap added to NumPy C-API

Similar to PyArray_MapIterArray but with an additional copy_if_overlap argument. If copy_if_overlap != 0, checks if input has memory overlap with any of the other arrays and make copies as appropriate to avoid problems if the input is modified during the iteration. See the documentation for more complete documentation.

New Features

__array_ufunc__ added

This is the renamed and redesigned __numpy_ufunc__. Any class, ndarray subclass or not, can define this method or set it to None in order to override the behavior of NumPy’s ufuncs. This works quite similarly to Python’s __mul__ and other binary operation routines. See the documentation for a more detailed description of the implementation and behavior of this new option. The API is provisional, we do not yet guarantee backward compatibility as modifications may be made pending feedback. See NEP 13 and documentation for more details.

New positive ufunc

This ufunc corresponds to unary +, but unlike + on an ndarray it will raise an error if array values do not support numeric operations.

New divmod ufunc

This ufunc corresponds to the Python builtin divmod, and is used to implement divmod when called on numpy arrays. np.divmod(x, y) calculates a result equivalent to (np.floor_divide(x, y), np.remainder(x, y)) but is approximately twice as fast as calling the functions separately.

np.isnat ufunc tests for NaT special datetime and timedelta values

The new ufunc np.isnat finds the positions of special NaT values within datetime and timedelta arrays. This is analogous to np.isnan.

np.heaviside ufunc computes the Heaviside function

The new function np.heaviside(x, h0) (a ufunc) computes the Heaviside function:

                   { 0   if x < 0,
heaviside(x, h0) = { h0  if x == 0,
                   { 1   if x > 0.

np.block function for creating blocked arrays

Add a new block function to the current stacking functions vstack, hstack, and stack. This allows concatenation across multiple axes simultaneously, with a similar syntax to array creation, but where elements can themselves be arrays. For instance:

>>> A = np.eye(2) * 2
>>> B = np.eye(3) * 3
>>> np.block([
...     [A,               np.zeros((2, 3))],
...     [np.ones((3, 2)), B               ]
... ])
array([[ 2.,  0.,  0.,  0.,  0.],
       [ 0.,  2.,  0.,  0.,  0.],
       [ 1.,  1.,  3.,  0.,  0.],
       [ 1.,  1.,  0.,  3.,  0.],
       [ 1.,  1.,  0.,  0.,  3.]])

While primarily useful for block matrices, this works for arbitrary dimensions of arrays.

It is similar to Matlab’s square bracket notation for creating block matrices.

isin function, improving on in1d

The new function isin tests whether each element of an N-dimensional array is present anywhere within a second array. It is an enhancement of in1d that preserves the shape of the first array.

Temporary elision

On platforms providing the backtrace function NumPy will try to avoid creating temporaries in expression involving basic numeric types. For example d = a + b + c is transformed to d = a + b; d += c which can improve performance for large arrays as less memory bandwidth is required to perform the operation.

axes argument for unique

In an N-dimensional array, the user can now choose the axis along which to look for duplicate N-1-dimensional elements using numpy.unique. The original behaviour is recovered if axis=None (default).

np.gradient now supports unevenly spaced data

Users can now specify a not-constant spacing for data. In particular np.gradient can now take:

  1. A single scalar to specify a sample distance for all dimensions.

  2. N scalars to specify a constant sample distance for each dimension. i.e. dx, dy, dz, …

  3. N arrays to specify the coordinates of the values along each dimension of F. The length of the array must match the size of the corresponding dimension

  4. Any combination of N scalars/arrays with the meaning of 2. and 3.

This means that, e.g., it is now possible to do the following:

>>> f = np.array([[1, 2, 6], [3, 4, 5]], dtype=np.float_)
>>> dx = 2.
>>> y = [1., 1.5, 3.5]
>>> np.gradient(f, dx, y)
[array([[ 1. ,  1. , -0.5], [ 1. ,  1. , -0.5]]),
 array([[ 2. ,  2. ,  2. ], [ 2. ,  1.7,  0.5]])]

Support for returning arrays of arbitrary dimensions in apply_along_axis

Previously, only scalars or 1D arrays could be returned by the function passed to apply_along_axis. Now, it can return an array of any dimensionality (including 0D), and the shape of this array replaces the axis of the array being iterated over.

.ndim property added to dtype to complement .shape

For consistency with ndarray and broadcast, d.ndim is a shorthand for len(d.shape).

Support for tracemalloc in Python 3.6

NumPy now supports memory tracing with tracemalloc module of Python 3.6 or newer. Memory allocations from NumPy are placed into the domain defined by numpy.lib.tracemalloc_domain. Note that NumPy allocation will not show up in tracemalloc of earlier Python versions.

NumPy may be built with relaxed stride checking debugging

Setting NPY_RELAXED_STRIDES_DEBUG=1 in the environment when relaxed stride checking is enabled will cause NumPy to be compiled with the affected strides set to the maximum value of npy_intp in order to help detect invalid usage of the strides in downstream projects. When enabled, invalid usage often results in an error being raised, but the exact type of error depends on the details of the code. TypeError and OverflowError have been observed in the wild.

It was previously the case that this option was disabled for releases and enabled in master and changing between the two required editing the code. It is now disabled by default but can be enabled for test builds.

Improvements

Ufunc behavior for overlapping inputs

Operations where ufunc input and output operands have memory overlap produced undefined results in previous NumPy versions, due to data dependency issues. In NumPy 1.13.0, results from such operations are now defined to be the same as for equivalent operations where there is no memory overlap.

Operations affected now make temporary copies, as needed to eliminate data dependency. As detecting these cases is computationally expensive, a heuristic is used, which may in rare cases result to needless temporary copies. For operations where the data dependency is simple enough for the heuristic to analyze, temporary copies will not be made even if the arrays overlap, if it can be deduced copies are not necessary. As an example,``np.add(a, b, out=a)`` will not involve copies.

To illustrate a previously undefined operation:

>>> x = np.arange(16).astype(float)
>>> np.add(x[1:], x[:-1], out=x[1:])

In NumPy 1.13.0 the last line is guaranteed to be equivalent to:

>>> np.add(x[1:].copy(), x[:-1].copy(), out=x[1:])

A similar operation with simple non-problematic data dependence is:

>>> x = np.arange(16).astype(float)
>>> np.add(x[1:], x[:-1], out=x[:-1])

It will continue to produce the same results as in previous NumPy versions, and will not involve unnecessary temporary copies.

The change applies also to in-place binary operations, for example:

>>> x = np.random.rand(500, 500)
>>> x += x.T

This statement is now guaranteed to be equivalent to x[...] = x + x.T, whereas in previous NumPy versions the results were undefined.

Partial support for 64-bit f2py extensions with MinGW

Extensions that incorporate Fortran libraries can now be built using the free MinGW toolset, also under Python 3.5. This works best for extensions that only do calculations and uses the runtime modestly (reading and writing from files, for instance). Note that this does not remove the need for Mingwpy; if you make extensive use of the runtime, you will most likely run into issues. Instead, it should be regarded as a band-aid until Mingwpy is fully functional.

Extensions can also be compiled using the MinGW toolset using the runtime library from the (moveable) WinPython 3.4 distribution, which can be useful for programs with a PySide1/Qt4 front-end.

Performance improvements for packbits and unpackbits

The functions numpy.packbits with boolean input and numpy.unpackbits have been optimized to be a significantly faster for contiguous data.

Fix for PPC long double floating point information

In previous versions of NumPy, the finfo function returned invalid information about the double double format of the longdouble float type on Power PC (PPC). The invalid values resulted from the failure of the NumPy algorithm to deal with the variable number of digits in the significand that are a feature of PPC long doubles. This release by-passes the failing algorithm by using heuristics to detect the presence of the PPC double double format. A side-effect of using these heuristics is that the finfo function is faster than previous releases.

Better default repr for ndarray subclasses

Subclasses of ndarray with no repr specialization now correctly indent their data and type lines.

More reliable comparisons of masked arrays

Comparisons of masked arrays were buggy for masked scalars and failed for structured arrays with dimension higher than one. Both problems are now solved. In the process, it was ensured that in getting the result for a structured array, masked fields are properly ignored, i.e., the result is equal if all fields that are non-masked in both are equal, thus making the behaviour identical to what one gets by comparing an unstructured masked array and then doing .all() over some axis.

np.matrix with booleans elements can now be created using the string syntax

np.matrix failed whenever one attempts to use it with booleans, e.g., np.matrix('True'). Now, this works as expected.

More linalg operations now accept empty vectors and matrices

All of the following functions in np.linalg now work when given input arrays with a 0 in the last two dimensions: det, slogdet, pinv, eigvals, eigvalsh, eig, eigh.

Bundled version of LAPACK is now 3.2.2

NumPy comes bundled with a minimal implementation of lapack for systems without a lapack library installed, under the name of lapack_lite. This has been upgraded from LAPACK 3.0.0 (June 30, 1999) to LAPACK 3.2.2 (June 30, 2010). See the LAPACK changelogs for details on the all the changes this entails.

While no new features are exposed through numpy, this fixes some bugs regarding “workspace” sizes, and in some places may use faster algorithms.

reduce of np.hypot.reduce and np.logical_xor allowed in more cases

This now works on empty arrays, returning 0, and can reduce over multiple axes. Previously, a ValueError was thrown in these cases.

Better repr of object arrays

Object arrays that contain themselves no longer cause a recursion error.

Object arrays that contain list objects are now printed in a way that makes clear the difference between a 2d object array, and a 1d object array of lists.

Changes

argsort on masked arrays takes the same default arguments as sort

By default, argsort now places the masked values at the end of the sorted array, in the same way that sort already did. Additionally, the end_with argument is added to argsort, for consistency with sort. Note that this argument is not added at the end, so breaks any code that passed fill_value as a positional argument.

average now preserves subclasses

For ndarray subclasses, numpy.average will now return an instance of the subclass, matching the behavior of most other NumPy functions such as mean. As a consequence, also calls that returned a scalar may now return a subclass array scalar.

array == None and array != None do element-wise comparison

Previously these operations returned scalars False and True respectively.

np.equal, np.not_equal for object arrays ignores object identity

Previously, these functions always treated identical objects as equal. This had the effect of overriding comparison failures, comparison of objects that did not return booleans, such as np.arrays, and comparison of objects where the results differed from object identity, such as NaNs.

Boolean indexing changes

  • Boolean array-likes (such as lists of python bools) are always treated as boolean indexes.

  • Boolean scalars (including python True) are legal boolean indexes and never treated as integers.

  • Boolean indexes must match the dimension of the axis that they index.

  • Boolean indexes used on the lhs of an assignment must match the dimensions of the rhs.

  • Boolean indexing into scalar arrays return a new 1-d array. This means that array(1)[array(True)] gives array([1]) and not the original array.

np.random.multivariate_normal behavior with bad covariance matrix

It is now possible to adjust the behavior the function will have when dealing with the covariance matrix by using two new keyword arguments:

  • tol can be used to specify a tolerance to use when checking that the covariance matrix is positive semidefinite.

  • check_valid can be used to configure what the function will do in the presence of a matrix that is not positive semidefinite. Valid options are ignore, warn and raise. The default value, warn keeps the the behavior used on previous releases.

assert_array_less compares np.inf and -np.inf now

Previously, np.testing.assert_array_less ignored all infinite values. This is not the expected behavior both according to documentation and intuitively. Now, -inf < x < inf is considered True for any real number x and all other cases fail.

assert_array_ and masked arrays assert_equal hide less warnings

Some warnings that were previously hidden by the assert_array_ functions are not hidden anymore. In most cases the warnings should be correct and, should they occur, will require changes to the tests using these functions. For the masked array assert_equal version, warnings may occur when comparing NaT. The function presently does not handle NaT or NaN specifically and it may be best to avoid it at this time should a warning show up due to this change.

offset attribute value in memmap objects

The offset attribute in a memmap object is now set to the offset into the file. This is a behaviour change only for offsets greater than mmap.ALLOCATIONGRANULARITY.

np.real and np.imag return scalars for scalar inputs

Previously, np.real and np.imag used to return array objects when provided a scalar input, which was inconsistent with other functions like np.angle and np.conj.

The polynomial convenience classes cannot be passed to ufuncs

The ABCPolyBase class, from which the convenience classes are derived, sets __array_ufun__ = None in order of opt out of ufuncs. If a polynomial convenience class instance is passed as an argument to a ufunc, a TypeError will now be raised.

Output arguments to ufuncs can be tuples also for ufunc methods

For calls to ufuncs, it was already possible, and recommended, to use an out argument with a tuple for ufuncs with multiple outputs. This has now been extended to output arguments in the reduce, accumulate, and reduceat methods. This is mostly for compatibility with __array_ufunc; there are no ufuncs yet that have more than one output.