This is the first minor release since the 1.0 release since October 2006. There are a few major changes, which introduce some minor API breakage. In addition this release includes tremendous improvements in terms of bug-fixing, testing, and documentation.

Major Changes

New implementation of masked arrays

MaskedArray is now a subclass of ndarray and the _data section can now be any subclass of ndarray. Apart from a couple of issues, the behavior of the new MaskedArray class reproduces the old one. API changes in the new MaskedArray implementation can be found here: http://scipy.org/scipy/numpy/wiki/MaskedArrayApiChanges

The old ma.py interface has been moved to the oldnumeric compatibility layer (i.e., numpy/oldnumeric/ma.py).

Complete overhaul of the IO system started

The IO code in both NumPy and SciPy is undergoing a major reworking. NumPy will provide basic code for reading and writing NumPy arrays, while SciPy will house file readers and writers for various data formats (data, audio, video, images, matlab, excel, etc.). As a first step in this direction:

• NumPy has a new standard binary file format (.npy/.npz) for arrays/groups_of_arrays. This is the new default method of storing arrays; pickling arrays is discouraged.

Support for numscons

Numscons extends SCons (http://www.scons.org/) to provide an modern and flexible build system for NumPy. While still under active development, Numscons all ready replaces most of the functionalities of numpy.distutils and provides much additional functionality. SCons is a software construction tool designed to replace traditional tools like make. One of the most attractive features of SCons is that its configuration files are actually Python scripts, and has a sane API for configuration test. Some of the advantages of numscons are: fine-grained control of compilation flags (up to per compiled object file), possibility of building ctypes-based extensions, automatic dependency handling, autoconf-like tests.

Note that the current build system still uses numpy.distutils by default and NumPy does not include numscons. This release does, however, include hook in numpy.distutils to call SCons, the SCons scripts, as well as some refactoring of existing code to make it easier to use numscons. You can find out more information about numscons (and where to get it) on NumScons page.

Semantic change for histogram

The histogram function is undergoing some changes to improve its consistency with the other histogram functions (histogram2d and histogramdd) and resolve long standing confusion regarding the handling of outliers. Specifically, the past behavior for histogram was to store upper outliers in the rightmost bin and ignore lower outliers. This behavior was a side effect of bins being specified by their left edges only, implicitly placing the rightmost bin edge at infinity. This made the interpretation of the density potentially confusing and the overall behavior irritating for new users.

In version 1.3, histogram will accept and return all the bin edges, including the rightmost edge, so that

>>> hist, edges = histogram(a, n)

will return arrays of length n (hist) and n+1 (edges) respectively. Also, values lying outside of the user specified range (lower and upper outliers) will not be tallied.

In order to minimize code breakage, in the current release the histogram function retains its old behavior, however

• A warning is raised at the first invocation of histogram, notifying users of the upcoming changes.
• Use of the range keyword will raise a warning about the coming changes to outliers handling.
• Use of explicit bin edges will raise a warning notifying users of the coming change regarding the edges definition.
• Use of explicit bin edges and normed=True will raise an error.

The new behavior for histogram, along with added support for weights, is available using the new=True argument. In version 1.1, new is set to False by default. In version 1.2, new will default to True, and in 1.3, it should disappear altogether. Setting new=True will allow users to call histogram without being annoyed by warnings.

New Features

New binary installers for Microsoft Windows

We provide new win32 installer aimed at solving the recurring problem of non-working atlas on different sets of CPU. This installer simply checks which cpu you have, and installs the appropriate NumPy accordingly (without atlas if your cpu is not supported).

Technical details:

• ATLAS is 3.8.0, and built with cygwin. Built on Core 2 duo for SSE3, pentium 4 xeon for SSE2.
• BLAS/LAPACK are built with mingw.
• only SSE3 and SSE2 are supported. If you don't have at least sse2, it will not use ATLAS, but netlib blas/lapack instead.
• the numpy installers are prepared with build_wininst for each different supported architecture
• the numpy installers are then bundled in another installer, based on nsis.
• Nsis is an open source installer, based on a small scripting language, and is extensible by plug-ins. I wrote a plug-in to detect the cpu type, and the installer will simply execute a different numpy installer depending on the cpu.

Official binary installers for MacOS X

For the first time, we provide an official Universal Mac binary.

Technical details:

• Built on OSX 10.5.2, Intel Core 2 Duo
• Using XCode 3.0 with gcc 4.0.1 and the Accelerate.framework 1.4.2, gfortran 4.2.1 and Python 2.5.2 from python.org (MacPython?)
• Used bdist_mpkg 0.4.3 to make the distribution

Since it uses the MacPython?, it will install NumPy under here:

/Library/Frameworks/Python.framework/Versions/2.5/lib/python2.5/site-packages

Median

This release introduces 'axis' support for numpy.median(). This is not just useful on its own, but it is also a first step in making median() consistent with mean(), min(), max(), var(), std(), sum(), prod(), etc.

While our goal is to harmonize the median() call signature and functionality with mean() and the rest, we also need to minimize API breakage for this release. So for 1.1, the new call signature is:

def median(a, axis=0, out=None, overwrite_input=False)

This returns the median of the array elements where the median is taken over the first axis of the array by default, otherwise over the specified axis. Then in the 1.2 release, we will allow API breakage and will, thus, change it to:

def median(a, axis=None, out=None, overwrite_input=False)

Var and Std

ddof keyword

Currently var and std are computed using 1/N. Traditionalists often like to use 1/(N-1) in these calculations. Some analysis methods suggest using 1/(N-alpha) for a wide-variety of alpha's depending on what you are trying to accomplish. In 1.0.5, var, and std methods and functions now take a ddof keyword. The divisor used in these calculations will be (N-ddof). The default is still 0, leaving the calculation unchanged.

complex data-types

For complex-valued data-types, the var and std calculations have been changed so that var = real(sum(x*conj(x)))/N and std = sqrt(var).

Bug-fixes

#69

Why numpy headers are installed using add_data_dir and not add_headers?

#419

Make web page links to NumPy bug report (ticket) system

#420

Add numpy.floatc for C float type

#430

numpy fails with PyWin build 210 (PyWin32-210.win32-py2.5.exe)

#460

better integration of matrices using properties

#502

Unbiased estimates of variance/ std deviation

#551

numpy.ndarray messed up after unpickling

#555

setting random seed does not work with a numpy.int64

#558

'axis' support for numpy.median()

#562

Reference leaks

#570

Vectorize leaks

#571

Support universal gfortran compiler on OS X

#573

numpy does not build when LDFLAGS are set

#581

random.set_state does not reset state of random.standard_normal

#583

fromiter: memory error on invalid dtype

#587

f2py fails on INTEGER(8) return statement in f90 code

#589

problem installing with latest version of Intel MKL

#593

Avoiding polluting public namespace in numpy headers

#594

Major slowdown when mixing array with Python's list

#596

setuptools' develop mode broken

#597

assigning negative # to a uint64 array element gives wrong answer on 32-bit machine

#601

function for computing powers of a matrix

#603

numpy 1.0.3.1, python crash when indexing an array with a int32 ndarray (see example)

#607

segfault from particular indexing of empty array

#608

<ufunc 'absolute'> should clear the sign bit.

#609

numpy/linalg/lapack_litemodule.c:lapack_lite_zgeqrf int vs. long

#610

Let numpy.interp accept scalar values as first argument

#611

[Vista - NPY_OWNDATA] data created in C code deleted on return to python

#612

NumPy 1.0.4 fails to compile on Cygwin

#613

resize on an empty array fails in some cases, very weird

#614

SegFault/double free with simple array mask operation

#615

numpy.array([])[arange(10000)>=0] = 123 segfaults

#616

arange: incorrect output byteorder

#617

numpy.random.randint error with negative numbers

#618

numpy freebsd fix

#620

Error in the typemaps in numpy.i

#621

wrapping f95 fails

#622

function for computing the condition number

#623

loadtxt fails with record arrays

#624

Statistics-related array method are not self-consistent.

#625

ImportError: No module named test_ufunc

#626

percentile() and clamp()

#627

linalg.svd takes 100% CPU and never ends

#628

polyfit uses 100% CPU and does not stop (Windows XP, Python 2.5.1 NumPy 1.0.4)

#630

If float('123.45') works, so should numpy.float32('123.45')

#631

incorrect description of numpy.linalg.pinv in linalg/info.py

#632

numpy.histogram fails with bin=<list>

#633

segfault in assignment to pyobject array

#634

masked array with dtype bool, sum to bool instead of int.

#635

numpy-1.0.4 crashes when doing linalg.inv on P3 Windows machine. Invalid op code.

#636

Faster array version of ndindex

#637

typo in scalarmathmodule.c.src for ctype_negative

#638

var should take absolute value for complex numbers.

#639

test_format.py:283: ImportError: No module named nose.tools

#641

numpy.sqrt(numpy.array([-1.0], dtype=numpy.complex256)) segfaults

#643

Tracebacks in Turkish locales

#644

distutils/cpuinfo fails to detect some nocona based hardware

#645

maskedarray extrema operations break on non-array types

#646

incorrect calculation on int64 array

#647

Bus Error with object arrays on big endian Solaris system

#648

Bus Error with object arrays on big endian system

#649

numpy.linalg.lstsq segfaults on i686 32-bit and ia64 linux

#650

fromfile and fromstring differ in parsing ascii integer strings

#651

typos in _compiled_base.c function arr_bincount

#652

linalg.eig takes 100% cpu and never ends

#653

v1.0.4 incorrectly identifies Xeon as 64-bit

#654

Numpy tests failure on solaris

#656

numpy.histogram fails if bins is a sequence

#657

x.flat[:] interprets non-native byteorder arrays incorrectly

#658

Segfault on reshape

#660

Crash in numpy.unique1D (release only)

#661

maximum handles nan improperly

#662

eigh hangup

#663

Support for multi formatting elements in savetxt

#664

more accurate representation of polynomials

#665

100% cpu error consolidation, tickets 627, 628, 652

#667

_dotblas.so only built if ATLAS is installed

#668

Explain why numpy headers are installed using add_data_dir

#670

cmp_arg_types bug

#671

linalg ImportError on Solaris 10 systems

#672

Fatal Python error: PyEval_RestoreThread: NULL tstate

#674

Impossible to rename object fields

#676

fortran-order .flatten() fails for multidimensional arrays

#677

format string / substitutable parameter mismatch in numpy/core/src/arrayobject.c

#678

fromfile() : reading past end of file causes unhandled exception

#679

ERROR: check_MyPyLong_AsUnsignedLongLong (test_regression.TestRegression)

#680

Bug in numpy/core/defmatrix.py

#682

fromfile/tofile broken under Windows

#683

MyPyLong_AsUnsignedLongLong fails on 64bit *nix platforms

#684

Intp takes 2 arguments on all platforms except Windows x86 64 MSVC

#685

Intp takes 2 arguments on all platforms except Windows x86 64 MSVC

#686

Intp takes 2 arguments on all platforms except Windows x86 64 MSVC

#687

site.cfg: ConfigParser in Python 2.6 does not allow section 'DEFAULT'

#688

Numpy and scipy should give informative error messages when run from source dir

#689

SyntaxError while installing v1.0.5 on MacTel running OSX v10.4.11

#690

No test for r4806

#691

No test for r4819

#692

No test for r4822

#693

No test for r4826

#694

No test for r4827 (possibly)

#696

No test for r4853

#697

ValueError: shape mismatch: objects cannot be broadcast to a single shape

#699

corrcoef incorrect behaviour

#700

Test for mean fails, incorrect processing of weights

#702

Memory leak in generic 1d array

#703

ma.masked_where can make a defective array

#704

Building SciPy crashes if no FORTRAN compiler available

#705

[ numpy-Bugs-601052 ] SVD did not converge

#706

[ numpy-Bugs-601052 ] SVD did not converge

#708

Setuptools call to numpy trunk causes numpy.scons error

#710

Reference count error, in array_reshape?

#711

object array refcount bug

#712

scalar math test failure on solaris

#713

numpy.any segfaults for object arrays with >=10000 elements

#715

The interactive mode of setup.py incorrect lower-cases prefix settings

#716

Inconsistent behavior, i686 vs x86_64

#717

numpy.loadtxt fails when missing values are present

#719

No test for 'fromregex' in r4960

#720

Changeset 4945 requires docstrings and tests

#723

gfortran installed in Program Files is not found

#724

numpy.ma.std and var should match numpy.std and var behaviour

#726

r4980 doesn't test assert_ functions with nan arrays

#728

numpy.r_ incorrectly casts with mixed types

#731

numpy.linalg.eigvals can't take nested lists while numpy.linalg.eig can

#732

error string on IndexError is wrong

#733

check_object_casting failing on Python 2.6

#734

[PATCH] interactive docstring search (lookfor)

#735

FAIL: test_record (numpy.lib.tests.test_io.Testloadtxt) -- On Solaris 10

#736

Inconsistent integer conversion from strings

#737

String to integer conversion inconsistent.

#738

Test trying to write file in wrong location

#739

Please don't use yellow in the output of setup.py

#742

Don't require bz2 or gzip

#743

Array content zeroed by |= even though exception raised

#744

Loss of dimension bug in numpy.ma.__getitem__ for arrays of shape (n, 1) or (1, n) or (n, 1, 1, ...)

#746

Segfault calling logical_not on own object

#747

Reference counting bug in call to ufunc object loops.

#748

Ifft pads incorrectly

#749

incorrect behaviour when using mask to index masked array

#751

distutils does not seem to handle the ALL section in site.cfg

#753

numpy.ma.morestats requires scipy

#759

spurious underflow warning in numpy.lib.machar

#760

scalar indexing of matrices -> deprecation warning

#761

operator += fails silently when adding arrays with diffenent base types

#767

numpy.i typo in error messages

#770

numpy.core.tests.test_multiarray.TestView failures on big-endian machines

#776

array resize shifts values in array

#777

array resize shifts values in array

#778

Markup errors in trunk/numpy/ma/API_CHANGES.txt

#779

Markup errors in trunk/numpy/ma/API_CHANGES.txt

#780

Markup errors in trunk/numpy/ma/API_CHANGES.txt

#781

Markup errors in trunk/numpy/ma/API_CHANGES.txt

#782

Markup errors in trunk/numpy/ma/API_CHANGES.txt

#783

Markup errors in trunk/numpy/ma/API_CHANGES.txt

#788

astype sometimes fails to return a copy

#789

Segfault if ndarray.compress can't safely cast into the "out" array

#790

import error with latest numpy

#791

Possible bug: std/var not properly calling array_finalize

#792

set -D_FORTIFY_SOURCE=1 instead of 2

#793

recarray.list() raises RunTimeError