========================== NumPy 1.14.0 Release Notes

Numpy 1.14.0 is the result of seven months of work and contains a large number of bug fixes and new features, along with several changes with potential compatibility issues. The major change that users will notice are the stylistic changes in the way numpy arrays and scalars are printed, a change that will affect doctests. See below for details on how to preserve the old style printing when needed.

A major decision affecting future development concerns the schedule for dropping Python 2.7 support in the runup to 2020. The decision has been made to support 2.7 for all releases made in 2018, with the last release being designated a long term release with support for bug fixes extending through 2019. In 2019 support for 2.7 will be dropped in all new releases. More details can be found in NEP 12_.

This release supports Python 2.7 and 3.4 - 3.6.

.. _NEP 12: http://www.numpy.org/neps/nep-0014-dropping-python2.7-proposal.html

Highlights

• The np.einsum function uses BLAS when possible

• genfromtxt, loadtxt, fromregex and savetxt can now handle files with arbitrary Python supported encoding.

• Major improvements to printing of NumPy arrays and scalars.

New functions

• parametrize: decorator added to numpy.testing

• chebinterpolate: Interpolate function at Chebyshev points.

• format_float_positional and format_float_scientific : format floating-point scalars unambiguously with control of rounding and padding.

• PyArray_ResolveWritebackIfCopy and PyArray_SetWritebackIfCopyBase, new C-API functions useful in achieving PyPy compatibility.

Deprecations

• Using np.bool_ objects in place of integers is deprecated. Previously operator.index(np.bool_) was legal and allowed constructs such as [1, 2, 3][np.True_]. That was misleading, as it behaved differently from np.array([1, 2, 3])[np.True_].

• Truth testing of an empty array is deprecated. To check if an array is not empty, use array.size > 0.

• Calling np.bincount with minlength=None is deprecated. minlength=0 should be used instead.

• Calling np.fromstring with the default value of the sep argument is deprecated. When that argument is not provided, a broken version of np.frombuffer is used that silently accepts unicode strings and -- after encoding them as either utf-8 (python 3) or the default encoding (python 2) -- treats them as binary data. If reading binary data is desired, np.frombuffer should be used directly.

• The style option of array2string is deprecated in non-legacy printing mode.

• PyArray_SetUpdateIfCopyBase has been deprecated. For NumPy versions >= 1.14 use PyArray_SetWritebackIfCopyBase instead, see C API changes below for more details.

• The use of UPDATEIFCOPY arrays is deprecated, see C API changes below for details. We will not be dropping support for those arrays, but they are not compatible with PyPy.

Future Changes

• np.issubdtype will stop downcasting dtype-like arguments. It might be expected that issubdtype(np.float32, 'float64') and issubdtype(np.float32, np.float64) mean the same thing - however, there was an undocumented special case that translated the former into issubdtype(np.float32, np.floating), giving the surprising result of True.

  This translation now gives a warning that explains what translation is occurring. In the future, the translation will be disabled, and the first example will be made equivalent to the second.

• np.linalg.lstsq default for rcond will be changed. The rcond parameter to np.linalg.lstsq will change its default to machine precision times the largest of the input array dimensions. A FutureWarning is issued when rcond is not passed explicitly.

• a.flat.__array__() will return a writeable copy of a when a is non-contiguous. Previously it returned an UPDATEIFCOPY array when a was writeable. Currently it returns a non-writeable copy. See gh-7054 for a discussion of the issue.

• Unstructured void array's .item method will return a bytes object. In the future, calling .item() on arrays or scalars of np.void datatype will return a bytes object instead of a buffer or int array, the same as returned by bytes(void_scalar). This may affect code which assumed the return value was mutable, which will no longer be the case. A FutureWarning is now issued when this would occur.

Compatibility notes

The mask of a masked array view is also a view rather than a copy

There was a FutureWarning about this change in NumPy 1.11.x. In short, it is now the case that, when changing a view of a masked array, changes to the mask are propagated to the original. That was not previously the case. This change affects slices in particular. Note that this does not yet work properly if the mask of the original array is nomask and the mask of the view is changed. See gh-5580 for an extended discussion. The original behavior of having a copy of the mask can be obtained by calling the unshare_mask method of the view.

np.ma.masked is no longer writeable

Attempts to mutate the masked constant now error, as the underlying arrays are marked readonly. In the past, it was possible to get away with::

# emulating a function that sometimes returns np.ma.masked
val = random.choice([np.ma.masked, 10])
var_arr = np.asarray(val)
val_arr += 1  # now errors, previously changed np.ma.masked.data

np.ma functions producing fill_value s have changed

Previously, np.ma.default_fill_value would return a 0d array, but np.ma.minimum_fill_value and np.ma.maximum_fill_value would return a tuple of the fields. Instead, all three methods return a structured np.void object, which is what you would already find in the .fill_value attribute.

Additionally, the dtype guessing now matches that of np.array - so when passing a python scalar x, maximum_fill_value(x) is always the same as maximum_fill_value(np.array(x)). Previously x = long(1) on Python 2 violated this assumption.

a.flat.__array__() returns non-writeable arrays when a is non-contiguous

The intent is that the UPDATEIFCOPY array previously returned when a was non-contiguous will be replaced by a writeable copy in the future. This temporary measure is aimed to notify folks who expect the underlying array be modified in this situation that that will no longer be the case. The most likely places for this to be noticed is when expressions of the form np.asarray(a.flat) are used, or when a.flat is passed as the out parameter to a ufunc.

np.tensordot now returns zero array when contracting over 0-length dimension

Previously np.tensordot raised a ValueError when contracting over 0-length dimension. Now it returns a zero array, which is consistent with the behaviour of np.dot and np.einsum.

numpy.testing reorganized

This is not expected to cause problems, but possibly something has been left out. If you experience an unexpected import problem using numpy.testing let us know.

np.asfarray no longer accepts non-dtypes through the dtype argument

This previously would accept dtype=some_array, with the implied semantics of dtype=some_array.dtype. This was undocumented, unique across the numpy functions, and if used would likely correspond to a typo.

1D np.linalg.norm preserves float input types, even for arbitrary orders

Previously, this would promote to float64 when arbitrary orders were passed, despite not doing so under the simple cases::

>>> f32 = np.float32([[1, 2]])
>>> np.linalg.norm(f32, 2.0, axis=-1).dtype
dtype('float32')
>>> np.linalg.norm(f32, 2.0001, axis=-1).dtype
dtype('float64')  # numpy 1.13
dtype('float32')  # numpy 1.14

This change affects only float32 and float16 arrays.

count_nonzero(arr, axis=()) now counts over no axes, not all axes

Elsewhere, axis==() is always understood as "no axes", but count_nonzero had a special case to treat this as "all axes". This was inconsistent and surprising. The correct way to count over all axes has always been to pass axis == None.

__init__.py files added to test directories

This is for pytest compatibility in the case of duplicate test file names in the different directories. As a result, run_module_suite no longer works, i.e., python <path-to-test-file> results in an error.

.astype(bool) on unstructured void arrays now calls bool on each element

On Python 2, void_array.astype(bool) would always return an array of True, unless the dtype is V0. On Python 3, this operation would usually crash. Going forwards, astype matches the behavior of bool(np.void), considering a buffer of all zeros as false, and anything else as true. Checks for V0 can still be done with arr.dtype.itemsize == 0.

MaskedArray.squeeze never returns np.ma.masked

np.squeeze is documented as returning a view, but the masked variant would sometimes return masked, which is not a view. This has been fixed, so that the result is always a view on the original masked array. This breaks any code that used masked_arr.squeeze() is np.ma.masked, but fixes code that writes to the result of .squeeze().

Renamed first parameter of can_cast from from to from_

The previous parameter name from is a reserved keyword in Python, which made it difficult to pass the argument by name. This has been fixed by renaming the parameter to from_.

isnat raises TypeError when passed wrong type

The ufunc isnat used to raise a ValueError when it was not passed variables of type datetime or timedelta. This has been changed to raising a TypeError.

dtype.__getitem__ raises TypeError when passed wrong type

When indexed with a float, the dtype object used to raise ValueError.

User-defined types now need to implement __str__ and __repr__

Previously, user-defined types could fall back to a default implementation of __str__ and __repr__ implemented in numpy, but this has now been removed. Now user-defined types will fall back to the python default object.__str__ and object.__repr__.

Many changes to array printing, disableable with the new "legacy" printing mode

The str and repr of ndarrays and numpy scalars have been changed in a variety of ways. These changes are likely to break downstream user's doctests.

These new behaviors can be disabled to mostly reproduce numpy 1.13 behavior by enabling the new 1.13 "legacy" printing mode. This is enabled by calling np.set_printoptions(legacy="1.13"), or using the new legacy argument to np.array2string, as np.array2string(arr, legacy='1.13').

In summary, the major changes are:

• For floating-point types:

  • The repr of float arrays often omits a space previously printed in the sign position. See the new sign option to np.set_printoptions.
  • Floating-point arrays and scalars use a new algorithm for decimal representations, giving the shortest unique representation. This will usually shorten float16 fractional output, and sometimes float32 and float128 output. float64 should be unaffected. See the new floatmode option to np.set_printoptions.
  • Float arrays printed in scientific notation no longer use fixed-precision, and now instead show the shortest unique representation.
  • The str of floating-point scalars is no longer truncated in python2.

• For other data types:

  • Non-finite complex scalars print like nanj instead of nan*j.
  • NaT values in datetime arrays are now properly aligned.
  • Arrays and scalars of np.void datatype are now printed using hex notation.

• For line-wrapping:

  • The "dtype" part of ndarray reprs will now be printed on the next line if there isn't space on the last line of array output.
  • The linewidth format option is now always respected. The repr or str of an array will never exceed this, unless a single element is too wide.
  • The last line of an array string will never have more elements than earlier lines.
  • An extra space is no longer inserted on the first line if the elements are too wide.

• For summarization (the use of ... to shorten long arrays):

  • A trailing comma is no longer inserted for str. Previously, str(np.arange(1001)) gave '[ 0 1 2 ..., 998 999 1000]', which has an extra comma.
  • For arrays of 2-D and beyond, when ... is printed on its own line in order to summarize any but the last axis, newlines are now appended to that line to match its leading newlines and a trailing space character is removed.

• MaskedArray arrays now separate printed elements with commas, always print the dtype, and correctly wrap the elements of long arrays to multiple lines. If there is more than 1 dimension, the array attributes are now printed in a new "left-justified" printing style.

• recarray arrays no longer print a trailing space before their dtype, and wrap to the right number of columns.

• 0d arrays no longer have their own idiosyncratic implementations of str and repr. The style argument to np.array2string is deprecated.

• Arrays of bool datatype will omit the datatype in the repr.

• User-defined dtypes (subclasses of np.generic) now need to implement __str__ and __repr__.

Some of these changes are described in more detail below. If you need to retain the previous behavior for doctests or other reasons, you may want to do something like::

# FIXME: We need the str/repr formatting used in Numpy < 1.14.
try:
    np.set_printoptions(legacy='1.13')
except TypeError:
    pass

C API changes

PyPy compatible alternative to UPDATEIFCOPY arrays

UPDATEIFCOPY arrays are contiguous copies of existing arrays, possibly with different dimensions, whose contents are copied back to the original array when their refcount goes to zero and they are deallocated. Because PyPy does not use refcounts, they do not function correctly with PyPy. NumPy is in the process of eliminating their use internally and two new C-API functions,

• PyArray_SetWritebackIfCopyBase
• PyArray_ResolveWritebackIfCopy,

have been added together with a complementary flag, NPY_ARRAY_WRITEBACKIFCOPY. Using the new functionality also requires that some flags be changed when new arrays are created, to wit: NPY_ARRAY_INOUT_ARRAY should be replaced by NPY_ARRAY_INOUT_ARRAY2 and NPY_ARRAY_INOUT_FARRAY should be replaced by NPY_ARRAY_INOUT_FARRAY2. Arrays created with these new flags will then have the WRITEBACKIFCOPY semantics.

If PyPy compatibility is not a concern, these new functions can be ignored, although there will be a DeprecationWarning. If you do wish to pursue PyPy compatibility, more information on these functions and their use may be found in the c-api_ documentation and the example in how-to-extend_.

.. _c-api: https://github.com/numpy/numpy/blob/master/doc/source/reference/c-api.array.rst .. _how-to-extend: https://github.com/numpy/numpy/blob/master/doc/source/user/c-info.how-to-extend.rst

New Features

Encoding argument for text IO functions

genfromtxt, loadtxt, fromregex and savetxt can now handle files with arbitrary encoding supported by Python via the encoding argument. For backward compatibility the argument defaults to the special bytes value which continues to treat text as raw byte values and continues to pass latin1 encoded bytes to custom converters. Using any other value (including None for system default) will switch the functions to real text IO so one receives unicode strings instead of bytes in the resulting arrays.

External nose plugins are usable by numpy.testing.Tester

numpy.testing.Tester is now aware of nose plugins that are outside the nose built-in ones. This allows using, for example, nose-timer like so: np.test(extra_argv=['--with-timer', '--timer-top-n', '20']) to obtain the runtime of the 20 slowest tests. An extra keyword timer was also added to Tester.test, so np.test(timer=20) will also report the 20 slowest tests.

parametrize decorator added to numpy.testing

A basic parametrize decorator is now available in numpy.testing. It is intended to allow rewriting yield based tests that have been deprecated in pytest so as to facilitate the transition to pytest in the future. The nose testing framework has not been supported for several years and looks like abandonware.

The new parametrize decorator does not have the full functionality of the one in pytest. It doesn't work for classes, doesn't support nesting, and does not substitute variable names. Even so, it should be adequate to rewrite the NumPy tests.

chebinterpolate function added to numpy.polynomial.chebyshev

The new chebinterpolate function interpolates a given function at the Chebyshev points of the first kind. A new Chebyshev.interpolate class method adds support for interpolation over arbitrary intervals using the scaled and shifted Chebyshev points of the first kind.

Support for reading lzma compressed text files in Python 3

With Python versions containing the lzma module the text IO functions can now transparently read from files with xz or lzma extension.

sign option added to np.setprintoptions and np.array2string

This option controls printing of the sign of floating-point types, and may be one of the characters '-', '+' or ' '. With '+' numpy always prints the sign of positive values, with ' ' it always prints a space (whitespace character) in the sign position of positive values, and with '-' it will omit the sign character for positive values. The new default is '-'.

This new default changes the float output relative to numpy 1.13. The old behavior can be obtained in 1.13 "legacy" printing mode, see compatibility notes above.

hermitian option added to np.linalg.matrix_rank

The new hermitian option allows choosing between standard SVD based matrix rank calculation and the more efficient eigenvalue based method for symmetric/hermitian matrices.

threshold and edgeitems options added to np.array2string

These options could previously be controlled using np.set_printoptions, but now can be changed on a per-call basis as arguments to np.array2string.

concatenate and stack gained an out argument

A preallocated buffer of the desired dtype can now be used for the output of these functions.

Support for PGI flang compiler on Windows

The PGI flang compiler is a Fortran front end for LLVM released by NVIDIA under the Apache 2 license. It can be invoked by ::

python setup.py config --compiler=clang --fcompiler=flang install

There is little experience with this new compiler, so any feedback from people using it will be appreciated.

Improvements

Numerator degrees of freedom in random.noncentral_f need only be positive.

Prior to NumPy 1.14.0, the numerator degrees of freedom needed to be > 1, but the distribution is valid for values > 0, which is the new requirement.

The GIL is released for all np.einsum variations

Some specific loop structures which have an accelerated loop version did not release the GIL prior to NumPy 1.14.0. This oversight has been fixed.

The np.einsum function will use BLAS when possible and optimize by default

The np.einsum function will now call np.tensordot when appropriate. Because np.tensordot uses BLAS when possible, that will speed up execution. By default, np.einsum will also attempt optimization as the overhead is small relative to the potential improvement in speed.

f2py now handles arrays of dimension 0

f2py now allows for the allocation of arrays of dimension 0. This allows for more consistent handling of corner cases downstream.

numpy.distutils supports using MSVC and mingw64-gfortran together

Numpy distutils now supports using Mingw64 gfortran and MSVC compilers together. This enables the production of Python extension modules on Windows containing Fortran code while retaining compatibility with the binaries distributed by Python.org. Not all use cases are supported, but most common ways to wrap Fortran for Python are functional.

Compilation in this mode is usually enabled automatically, and can be selected via the --fcompiler and --compiler options to setup.py. Moreover, linking Fortran codes to static OpenBLAS is supported; by default a gfortran compatible static archive openblas.a is looked for.

np.linalg.pinv now works on stacked matrices

Previously it was limited to a single 2d array.

numpy.save aligns data to 64 bytes instead of 16

Saving NumPy arrays in the npy format with numpy.save inserts padding before the array data to align it at 64 bytes. Previously this was only 16 bytes (and sometimes less due to a bug in the code for version 2). Now the alignment is 64 bytes, which matches the widest SIMD instruction set commonly available, and is also the most common cache line size. This makes npy files easier to use in programs which open them with mmap, especially on Linux where an mmap offset must be a multiple of the page size.

NPZ files now can be written without using temporary files

In Python 3.6+ numpy.savez and numpy.savez_compressed now write directly to a ZIP file, without creating intermediate temporary files.

Better support for empty structured and string types

Structured types can contain zero fields, and string dtypes can contain zero characters. Zero-length strings still cannot be created directly, and must be constructed through structured dtypes::

str0 = np.empty(10, np.dtype([('v', str, N)]))['v']
void0 = np.empty(10, np.void)

It was always possible to work with these, but the following operations are now supported for these arrays:

• arr.sort()
• arr.view(bytes)
• arr.resize(...)
• pickle.dumps(arr)

Support for decimal.Decimal in np.lib.financial

Unless otherwise stated all functions within the financial package now support using the decimal.Decimal built-in type.

Float printing now uses "dragon4" algorithm for shortest decimal representation

The str and repr of floating-point values (16, 32, 64 and 128 bit) are now printed to give the shortest decimal representation which uniquely identifies the value from others of the same type. Previously this was only true for float64 values. The remaining float types will now often be shorter than in numpy 1.13. Arrays printed in scientific notation now also use the shortest scientific representation, instead of fixed precision as before.

Additionally, the str of float scalars scalars will no longer be truncated in python2, unlike python2 float\ s. np.double scalars now have a str and repr identical to that of a python3 float.

New functions np.format_float_scientific and np.format_float_positional are provided to generate these decimal representations.

A new option floatmode has been added to np.set_printoptions and np.array2string, which gives control over uniqueness and rounding of printed elements in an array. The new default is floatmode='maxprec' with precision=8, which will print at most 8 fractional digits, or fewer if an element can be uniquely represented with fewer. A useful new mode is floatmode="unique", which will output enough digits to specify the array elements uniquely.

Numpy complex-floating-scalars with values like inf*j or nan*j now print as infj and nanj, like the pure-python complex type.

The FloatFormat and LongFloatFormat classes are deprecated and should both be replaced by FloatingFormat. Similarly ComplexFormat and LongComplexFormat should be replaced by ComplexFloatingFormat.

void datatype elements are now printed in hex notation

A hex representation compatible with the python bytes type is now printed for unstructured np.void elements, e.g., V4 datatype. Previously, in python2 the raw void data of the element was printed to stdout, or in python3 the integer byte values were shown.

printing style for void datatypes is now independently customizable

The printing style of np.void arrays is now independently customizable using the formatter argument to np.set_printoptions, using the 'void' key, instead of the catch-all numpystr key as before.

Reduced memory usage of np.loadtxt

np.loadtxt now reads files in chunks instead of all at once which decreases its memory usage significantly for large files.

Changes

Multiple-field indexing/assignment of structured arrays

The indexing and assignment of structured arrays with multiple fields has changed in a number of ways, as warned about in previous releases.

First, indexing a structured array with multiple fields, e.g., arr[['f1', 'f3']], returns a view into the original array instead of a copy. The returned view will have extra padding bytes corresponding to intervening fields in the original array, unlike the copy in 1.13, which will affect code such as arr[['f1', 'f3']].view(newdtype).

Second, assignment between structured arrays will now occur "by position" instead of "by field name". The Nth field of the destination will be set to the Nth field of the source regardless of field name, unlike in numpy versions 1.6 to 1.13 in which fields in the destination array were set to the identically-named field in the source array or to 0 if the source did not have a field.

Correspondingly, the order of fields in a structured dtypes now matters when computing dtype equality. For example, with the dtypes ::

x = dtype({'names': ['A', 'B'], 'formats': ['i4', 'f4'], 'offsets': [0, 4]})
y = dtype({'names': ['B', 'A'], 'formats': ['f4', 'i4'], 'offsets': [4, 0]})

the expression x == y will now return False, unlike before. This makes dictionary based dtype specifications like dtype({'a': ('i4', 0), 'b': ('f4', 4)}) dangerous in python < 3.6 since dict key order is not preserved in those versions.

Assignment from a structured array to a boolean array now raises a ValueError, unlike in 1.13, where it always set the destination elements to True.

Assignment from structured array with more than one field to a non-structured array now raises a ValueError. In 1.13 this copied just the first field of the source to the destination.

Using field "titles" in multiple-field indexing is now disallowed, as is repeating a field name in a multiple-field index.

The documentation for structured arrays in the user guide has been significantly updated to reflect these changes.

Integer and Void scalars are now unaffected by np.set_string_function

Previously, unlike most other numpy scalars, the str and repr of integer and void scalars could be controlled by np.set_string_function. This is no longer possible.

0d array printing changed, style arg of array2string deprecated

Previously the str and repr of 0d arrays had idiosyncratic implementations which returned str(a.item()) and 'array(' + repr(a.item()) + ')' respectively for 0d array a, unlike both numpy scalars and higher dimension ndarrays.

Now, the str of a 0d array acts like a numpy scalar using str(a[()]) and the repr acts like higher dimension arrays using formatter(a[()]), where formatter can be specified using np.set_printoptions. The style argument of np.array2string is deprecated.

This new behavior is disabled in 1.13 legacy printing mode, see compatibility notes above.

Seeding RandomState using an array requires a 1-d array

RandomState previously would accept empty arrays or arrays with 2 or more dimensions, which resulted in either a failure to seed (empty arrays) or for some of the passed values to be ignored when setting the seed.

MaskedArray objects show a more useful repr

The repr of a MaskedArray is now closer to the python code that would produce it, with arrays now being shown with commas and dtypes. Like the other formatting changes, this can be disabled with the 1.13 legacy printing mode in order to help transition doctests.

The repr of np.polynomial classes is more explicit

It now shows the domain and window parameters as keyword arguments to make them more clear::

>>> np.polynomial.Polynomial(range(4))
Polynomial([0.,  1.,  2.,  3.], domain=[-1,  1], window=[-1,  1])