NumPy 1.8.0 Release Notes#
This release supports Python 2.6 -2.7 and 3.2 - 3.3.
Highlights#
New, no 2to3, Python 2 and Python 3 are supported by a common code base.
New, gufuncs for linear algebra, enabling operations on stacked arrays.
New, inplace fancy indexing for ufuncs with the
.atmethod.New,
partitionfunction, partial sorting via selection for fast median.New,
nanmean,nanvar, andnanstdfunctions skipping NaNs.New,
fullandfull_likefunctions to create value initialized arrays.New,
PyUFunc_RegisterLoopForDescr, better ufunc support for user dtypes.Numerous performance improvements in many areas.
Dropped Support#
Support for Python versions 2.4 and 2.5 has been dropped,
Support for SCons has been removed.
Future Changes#
The Datetime64 type remains experimental in this release. In 1.9 there will probably be some changes to make it more usable.
The diagonal method currently returns a new array and raises a FutureWarning. In 1.9 it will return a readonly view.
Multiple field selection from an array of structured type currently returns a new array and raises a FutureWarning. In 1.9 it will return a readonly view.
The numpy/oldnumeric and numpy/numarray compatibility modules will be removed in 1.9.
Compatibility notes#
The doc/sphinxext content has been moved into its own github repository, and is included in numpy as a submodule. See the instructions in doc/HOWTO_BUILD_DOCS.rst for how to access the content.
The hash function of numpy.void scalars has been changed. Previously the pointer to the data was hashed as an integer. Now, the hash function uses the tuple-hash algorithm to combine the hash functions of the elements of the scalar, but only if the scalar is read-only.
Numpy has switched its build system to using ‘separate compilation’ by default. In previous releases this was supported, but not default. This should produce the same results as the old system, but if you’re trying to do something complicated like link numpy statically or using an unusual compiler, then it’s possible you will encounter problems. If so, please file a bug and as a temporary workaround you can re-enable the old build system by exporting the shell variable NPY_SEPARATE_COMPILATION=0.
For the AdvancedNew iterator the oa_ndim flag should now be -1 to indicate
that no op_axes and itershape are passed in. The oa_ndim == 0
case, now indicates a 0-D iteration and op_axes being NULL and the old
usage is deprecated. This does not effect the NpyIter_New or
NpyIter_MultiNew functions.
The functions nanargmin and nanargmax now return np.iinfo[‘intp’].min for the index in all-NaN slices. Previously the functions would raise a ValueError for array returns and NaN for scalar returns.
NPY_RELAXED_STRIDES_CHECKING#
There is a new compile time environment variable
NPY_RELAXED_STRIDES_CHECKING. If this variable is set to 1, then
numpy will consider more arrays to be C- or F-contiguous – for
example, it becomes possible to have a column vector which is
considered both C- and F-contiguous simultaneously. The new definition
is more accurate, allows for faster code that makes fewer unnecessary
copies, and simplifies numpy’s code internally. However, it may also
break third-party libraries that make too-strong assumptions about the
stride values of C- and F-contiguous arrays. (It is also currently
known that this breaks Cython code using memoryviews, which will be
fixed in Cython.) THIS WILL BECOME THE DEFAULT IN A FUTURE RELEASE, SO
PLEASE TEST YOUR CODE NOW AGAINST NUMPY BUILT WITH:
NPY_RELAXED_STRIDES_CHECKING=1 python setup.py install
You can check whether NPY_RELAXED_STRIDES_CHECKING is in effect by running:
np.ones((10, 1), order="C").flags.f_contiguous
This will be True if relaxed strides checking is enabled, and
False otherwise. The typical problem we’ve seen so far is C code
that works with C-contiguous arrays, and assumes that the itemsize can
be accessed by looking at the last element in the PyArray_STRIDES(arr)
array. When relaxed strides are in effect, this is not true (and in
fact, it never was true in some corner cases). Instead, use
PyArray_ITEMSIZE(arr).
For more information check the “Internal memory layout of an ndarray” section in the documentation.
Binary operations with non-arrays as second argument#
Binary operations of the form <array-or-subclass> * <non-array-subclass>
where <non-array-subclass> declares an __array_priority__ higher than
that of <array-or-subclass> will now unconditionally return
NotImplemented, giving <non-array-subclass> a chance to handle the
operation. Previously, NotImplemented would only be returned if
<non-array-subclass> actually implemented the reversed operation, and after
a (potentially expensive) array conversion of <non-array-subclass> had been
attempted. (bug, pull request)
Function median used with overwrite_input only partially sorts array#
If median is used with overwrite_input option the input array will now only be partially sorted instead of fully sorted.
Fix to financial.npv#
The npv function had a bug. Contrary to what the documentation stated, it
summed from indexes 1 to M instead of from 0 to M - 1. The
fix changes the returned value. The mirr function called the npv function,
but worked around the problem, so that was also fixed and the return value
of the mirr function remains unchanged.
Runtime warnings when comparing NaN numbers#
Comparing NaN floating point numbers now raises the invalid runtime
warning. If a NaN is expected the warning can be ignored using np.errstate.
E.g.:
with np.errstate(invalid='ignore'):
operation()
New Features#
Support for linear algebra on stacked arrays#
The gufunc machinery is now used for np.linalg, allowing operations on stacked arrays and vectors. For example:
>>> a
array([[[ 1., 1.],
[ 0., 1.]],
[[ 1., 1.],
[ 0., 1.]]])
>>> np.linalg.inv(a)
array([[[ 1., -1.],
[ 0., 1.]],
[[ 1., -1.],
[ 0., 1.]]])
In place fancy indexing for ufuncs#
The function at has been added to ufunc objects to allow in place
ufuncs with no buffering when fancy indexing is used. For example, the
following will increment the first and second items in the array, and will
increment the third item twice: numpy.add.at(arr, [0, 1, 2, 2], 1)
This is what many have mistakenly thought arr[[0, 1, 2, 2]] += 1 would do,
but that does not work as the incremented value of arr[2] is simply copied
into the third slot in arr twice, not incremented twice.
New functions partition and argpartition#
New functions to partially sort arrays via a selection algorithm.
A partition by index k moves the k smallest element to the front of
an array. All elements before k are then smaller or equal than the value
in position k and all elements following k are then greater or equal
than the value in position k. The ordering of the values within these
bounds is undefined.
A sequence of indices can be provided to sort all of them into their sorted
position at once iterative partitioning.
This can be used to efficiently obtain order statistics like median or
percentiles of samples.
partition has a linear time complexity of O(n) while a full sort has
O(n log(n)).
New functions nanmean, nanvar and nanstd#
New nan aware statistical functions are added. In these functions the results are what would be obtained if nan values were omitted from all computations.
New functions full and full_like#
New convenience functions to create arrays filled with a specific value; complementary to the existing zeros and zeros_like functions.
IO compatibility with large files#
Large NPZ files >2GB can be loaded on 64-bit systems.
Building against OpenBLAS#
It is now possible to build numpy against OpenBLAS by editing site.cfg.
New constant#
Euler’s constant is now exposed in numpy as euler_gamma.
New modes for qr#
New modes ‘complete’, ‘reduced’, and ‘raw’ have been added to the qr factorization and the old ‘full’ and ‘economic’ modes are deprecated. The ‘reduced’ mode replaces the old ‘full’ mode and is the default as was the ‘full’ mode, so backward compatibility can be maintained by not specifying the mode.
The ‘complete’ mode returns a full dimensional factorization, which can be useful for obtaining a basis for the orthogonal complement of the range space. The ‘raw’ mode returns arrays that contain the Householder reflectors and scaling factors that can be used in the future to apply q without needing to convert to a matrix. The ‘economic’ mode is simply deprecated, there isn’t much use for it and it isn’t any more efficient than the ‘raw’ mode.
New invert argument to in1d#
The function in1d now accepts a invert argument which, when True, causes the returned array to be inverted.
Advanced indexing using np.newaxis#
It is now possible to use np.newaxis/None together with index
arrays instead of only in simple indices. This means that
array[np.newaxis, [0, 1]] will now work as expected and select the first
two rows while prepending a new axis to the array.
C-API#
New ufuncs can now be registered with builtin input types and a custom output type. Before this change, NumPy wouldn’t be able to find the right ufunc loop function when the ufunc was called from Python, because the ufunc loop signature matching logic wasn’t looking at the output operand type. Now the correct ufunc loop is found, as long as the user provides an output argument with the correct output type.
runtests.py#
A simple test runner script runtests.py was added. It also builds Numpy via
setup.py build and can be used to run tests easily during development.
Improvements#
IO performance improvements#
Performance in reading large files was improved by chunking (see also IO compatibility).
Performance improvements to pad#
The pad function has a new implementation, greatly improving performance for all inputs except mode= (retained for backwards compatibility). Scaling with dimensionality is dramatically improved for rank >= 4.
Performance improvements to isnan, isinf, isfinite and byteswap#
isnan, isinf, isfinite and byteswap have been improved to take advantage of compiler builtins to avoid expensive calls to libc. This improves performance of these operations by about a factor of two on gnu libc systems.
Performance improvements via SSE2 vectorization#
Several functions have been optimized to make use of SSE2 CPU SIMD instructions.
- Float32 and float64:
base math (add, subtract, divide, multiply)
sqrt
minimum/maximum
absolute
- Bool:
logical_or
logical_and
logical_not
This improves performance of these operations up to 4x/2x for float32/float64 and up to 10x for bool depending on the location of the data in the CPU caches. The performance gain is greatest for in-place operations.
In order to use the improved functions the SSE2 instruction set must be enabled at compile time. It is enabled by default on x86_64 systems. On x86_32 with a capable CPU it must be enabled by passing the appropriate flag to the CFLAGS build variable (-msse2 with gcc).
Performance improvements to median#
median is now implemented in terms of partition instead of sort which reduces its time complexity from O(n log(n)) to O(n). If used with the overwrite_input option the array will now only be partially sorted instead of fully sorted.
Overridable operand flags in ufunc C-API#
When creating a ufunc, the default ufunc operand flags can be overridden via the new op_flags attribute of the ufunc object. For example, to set the operand flag for the first input to read/write:
PyObject *ufunc = PyUFunc_FromFuncAndData(…); ufunc->op_flags[0] = NPY_ITER_READWRITE;
This allows a ufunc to perform an operation in place. Also, global nditer flags can be overridden via the new iter_flags attribute of the ufunc object. For example, to set the reduce flag for a ufunc:
ufunc->iter_flags = NPY_ITER_REDUCE_OK;
Changes#
General#
The function np.take now allows 0-d arrays as indices.
The separate compilation mode is now enabled by default.
Several changes to np.insert and np.delete:
Previously, negative indices and indices that pointed past the end of the array were simply ignored. Now, this will raise a Future or Deprecation Warning. In the future they will be treated like normal indexing treats them – negative indices will wrap around, and out-of-bound indices will generate an error.
Previously, boolean indices were treated as if they were integers (always referring to either the 0th or 1st item in the array). In the future, they will be treated as masks. In this release, they raise a FutureWarning warning of this coming change.
In Numpy 1.7. np.insert already allowed the syntax np.insert(arr, 3, [1,2,3]) to insert multiple items at a single position. In Numpy 1.8. this is also possible for np.insert(arr, [3], [1, 2, 3]).
Padded regions from np.pad are now correctly rounded, not truncated.
C-API Array Additions#
Four new functions have been added to the array C-API.
PyArray_Partition
PyArray_ArgPartition
PyArray_SelectkindConverter
PyDataMem_NEW_ZEROED
C-API Ufunc Additions#
One new function has been added to the ufunc C-API that allows to register an inner loop for user types using the descr.
PyUFunc_RegisterLoopForDescr
C-API Developer Improvements#
The PyArray_Type instance creation function tp_new now
uses tp_basicsize to determine how much memory to allocate.
In previous releases only sizeof(PyArrayObject) bytes of
memory were allocated, often requiring C-API subtypes to
reimplement tp_new.
Deprecations#
The ‘full’ and ‘economic’ modes of qr factorization are deprecated.
General#
The use of non-integer for indices and most integer arguments has been deprecated. Previously float indices and function arguments such as axes or shapes were truncated to integers without warning. For example arr.reshape(3., -1) or arr[0.] will trigger a deprecation warning in NumPy 1.8., and in some future version of NumPy they will raise an error.