NumPy 0.9.6 is a bug-fix and optimization release with a few new features:
Breaking news: the original Win32 / Python 2.4 binary installer contained an unwanted runtime dependency on the win32all package, available from here. A re-build for this platform without this dependency is now available as numpy-0.9.6r1.win32-py2.4.exe.
New features (and changes)
- bigndarray removed and support for Python2.5 ssize_t added giving full support in Python2.5 to very-large arrays on 64-bit systems.
- Strides can be set more arbitrarily from Python (and checking is done to make sure memory won't be violated).
__array_finalize__ is now called for every array sub-class creation.
- kron and repmat functions added
- .round() method added for arrays
- rint, square, reciprocal, and ones_like ufuncs added.
- keyword arguments now possible for methods taking a single 'axis' argument
- Swig and Pyrex examples added in doc/swig and doc/pyrex
NumPy builds out of the box for cygwin
- Different unit testing naming schemes are now supported.
- memusage in numpy.distutils works for NT platforms
- numpy.lib.math functions now take vectors
- Most functions in oldnumeric now return intput class where possible
- x**n for integer n signficantly improved
array(<python scalar>) much faster
- .fill() method is much faster
- Output arrays to ufuncs works better.
- Several ma (Masked Array) fixes.
- umath code generation improved
- many fixes to optimized dot function
- fix for multiplying matrices with non-array objects
- scalartype fixes
- improvements to poly1d
- f2py fixed to handle character arrays in common blocks
- Scalar arithmetic improved to handle mixed-mode operation.
Make sure Python intYY types correspond exactly with C PyArray_INTYY