NumPy 1.9.0 Release Notes

This release supports Python 2.6 - 2.7 and 3.2 - 3.4.


  • Numerous performance improvements in various areas, most notably indexing and operations on small arrays are significantly faster. Indexing operations now also release the GIL.

  • Addition of nanmedian and nanpercentile rounds out the nanfunction set.

Dropped Support

  • The oldnumeric and numarray modules have been removed.

  • The doc/pyrex and doc/cython directories have been removed.

  • The doc/numpybook directory has been removed.

  • The numpy/testing/ file has been removed together with the importall function it contained.

Future Changes

  • The numpy/polynomial/ file will be removed in NumPy 1.10.0.

  • Default casting for inplace operations will change to ‘same_kind’ in Numpy 1.10.0. This will certainly break some code that is currently ignoring the warning.

  • Relaxed stride checking will be the default in 1.10.0

  • String version checks will break because, e.g., ‘1.9’ > ‘1.10’ is True. A NumpyVersion class has been added that can be used for such comparisons.

  • The diagonal and diag functions will return writeable views in 1.10.0

  • The S and/or a dtypes may be changed to represent Python strings instead of bytes, in Python 3 these two types are very different.

Compatibility notes

The diagonal and diag functions return readonly views.

In NumPy 1.8, the diagonal and diag functions returned readonly copies, in NumPy 1.9 they return readonly views, and in 1.10 they will return writeable views.

Special scalar float values don’t cause upcast to double anymore

In previous numpy versions operations involving floating point scalars containing special values NaN, Inf and -Inf caused the result type to be at least float64. As the special values can be represented in the smallest available floating point type, the upcast is not performed anymore.

For example the dtype of:

np.array([1.], dtype=np.float32) * float('nan')

now remains float32 instead of being cast to float64. Operations involving non-special values have not been changed.

Percentile output changes

If given more than one percentile to compute numpy.percentile returns an array instead of a list. A single percentile still returns a scalar. The array is equivalent to converting the list returned in older versions to an array via np.array.

If the overwrite_input option is used the input is only partially instead of fully sorted.

ndarray.tofile exception type

All tofile exceptions are now IOError, some were previously ValueError.

Invalid fill value exceptions

Two changes to

  • When the fill value is a string and the array type is not one of ‘OSUV’, TypeError is raised instead of the default fill value being used.

  • When the fill value overflows the array type, TypeError is raised instead of OverflowError.

Polynomial Classes no longer derived from PolyBase

This may cause problems with folks who depended on the polynomial classes being derived from PolyBase. They are now all derived from the abstract base class ABCPolyBase. Strictly speaking, there should be a deprecation involved, but no external code making use of the old baseclass could be found.

Using numpy.random.binomial may change the RNG state vs. numpy < 1.9

A bug in one of the algorithms to generate a binomial random variate has been fixed. This change will likely alter the number of random draws performed, and hence the sequence location will be different after a call to distribution.c::rk_binomial_btpe. Any tests which rely on the RNG being in a known state should be checked and/or updated as a result.

Random seed enforced to be a 32 bit unsigned integer

np.random.seed and np.random.RandomState now throw a ValueError if the seed cannot safely be converted to 32 bit unsigned integers. Applications that now fail can be fixed by masking the higher 32 bit values to zero: seed = seed & 0xFFFFFFFF. This is what is done silently in older versions so the random stream remains the same.

Argmin and argmax out argument

The out argument to np.argmin and np.argmax and their equivalent C-API functions is now checked to match the desired output shape exactly. If the check fails a ValueError instead of TypeError is raised.


Remove unnecessary broadcasting notation restrictions. np.einsum('ijk,j->ijk', A, B) can also be written as np.einsum('ij...,j->ij...', A, B) (ellipsis is no longer required on ‘j’)


The NumPy indexing has seen a complete rewrite in this version. This makes most advanced integer indexing operations much faster and should have no other implications. However some subtle changes and deprecations were introduced in advanced indexing operations:

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

  • Advanced indexing into one dimensional arrays used to have (undocumented) special handling regarding repeating the value array in assignments when the shape of the value array was too small or did not match. Code using this will raise an error. For compatibility you can use arr.flat[index] = values, which uses the old code branch. (for example a = np.ones(10); a[np.arange(10)] = [1, 2, 3])

  • The iteration order over advanced indexes used to be always C-order. In NumPy 1.9. the iteration order adapts to the inputs and is not guaranteed (with the exception of a single advanced index which is never reversed for compatibility reasons). This means that the result is undefined if multiple values are assigned to the same element. An example for this is arr[[0, 0], [1, 1]] = [1, 2], which may set arr[0, 1] to either 1 or 2.

  • Equivalent to the iteration order, the memory layout of the advanced indexing result is adapted for faster indexing and cannot be predicted.

  • All indexing operations return a view or a copy. No indexing operation will return the original array object. (For example arr[...])

  • In the future Boolean array-likes (such as lists of python bools) will always be treated as Boolean indexes and Boolean scalars (including python True) will be a legal boolean index. At this time, this is already the case for scalar arrays to allow the general positive = a[a > 0] to work when a is zero dimensional.

  • In NumPy 1.8 it was possible to use array(True) and array(False) equivalent to 1 and 0 if the result of the operation was a scalar. This will raise an error in NumPy 1.9 and, as noted above, treated as a boolean index in the future.

  • All non-integer array-likes are deprecated, object arrays of custom integer like objects may have to be cast explicitly.

  • The error reporting for advanced indexing is more informative, however the error type has changed in some cases. (Broadcasting errors of indexing arrays are reported as IndexError)

  • Indexing with more then one ellipsis (...) is deprecated.

Non-integer reduction axis indexes are deprecated

Non-integer axis indexes to reduction ufuncs like add.reduce or sum are deprecated.

promote_types and string dtype

promote_types function now returns a valid string length when given an integer or float dtype as one argument and a string dtype as another argument. Previously it always returned the input string dtype, even if it wasn’t long enough to store the max integer/float value converted to a string.

can_cast and string dtype

can_cast function now returns False in “safe” casting mode for integer/float dtype and string dtype if the string dtype length is not long enough to store the max integer/float value converted to a string. Previously can_cast in “safe” mode returned True for integer/float dtype and a string dtype of any length.

astype and string dtype

The astype method now returns an error if the string dtype to cast to is not long enough in “safe” casting mode to hold the max value of integer/float array that is being casted. Previously the casting was allowed even if the result was truncated.

npyio.recfromcsv keyword arguments change

npyio.recfromcsv no longer accepts the undocumented update keyword, which used to override the dtype keyword.

The doc/swig directory moved

The doc/swig directory has been moved to tools/swig.

The npy_3kcompat.h header changed

The unused simple_capsule_dtor function has been removed from npy_3kcompat.h. Note that this header is not meant to be used outside of numpy; other projects should be using their own copy of this file when needed.

Negative indices in C-Api sq_item and sq_ass_item sequence methods

When directly accessing the sq_item or sq_ass_item PyObject slots for item getting, negative indices will not be supported anymore. PySequence_GetItem and PySequence_SetItem however fix negative indices so that they can be used there.


When NpyIter_RemoveAxis is now called, the iterator range will be reset.

When a multi index is being tracked and an iterator is not buffered, it is possible to use NpyIter_RemoveAxis. In this case an iterator can shrink in size. Because the total size of an iterator is limited, the iterator may be too large before these calls. In this case its size will be set to -1 and an error issued not at construction time but when removing the multi index, setting the iterator range, or getting the next function.

This has no effect on currently working code, but highlights the necessity of checking for an error return if these conditions can occur. In most cases the arrays being iterated are as large as the iterator so that such a problem cannot occur.

This change was already applied to the 1.8.1 release.

zeros_like for string dtypes now returns empty strings

To match the zeros function zeros_like now returns an array initialized with empty strings instead of an array filled with ‘0’.

New Features

Percentile supports more interpolation options

np.percentile now has the interpolation keyword argument to specify in which way points should be interpolated if the percentiles fall between two values. See the documentation for the available options.

Generalized axis support for median and percentile

np.median and np.percentile now support generalized axis arguments like ufunc reductions do since 1.7. One can now say axis=(index, index) to pick a list of axes for the reduction. The keepdims keyword argument was also added to allow convenient broadcasting to arrays of the original shape.

Dtype parameter added to np.linspace and np.logspace

The returned data type from the linspace and logspace functions can now be specified using the dtype parameter.

More general np.triu and np.tril broadcasting

For arrays with ndim exceeding 2, these functions will now apply to the final two axes instead of raising an exception.

tobytes alias for tostring method

ndarray.tobytes and MaskedArray.tobytes have been added as aliases for tostring which exports arrays as bytes. This is more consistent in Python 3 where str and bytes are not the same.

Build system

Added experimental support for the ppc64le and OpenRISC architecture.

Compatibility to python numbers module

All numerical numpy types are now registered with the type hierarchy in the python numbers module.

increasing parameter added to np.vander

The ordering of the columns of the Vandermonde matrix can be specified with this new boolean argument.

unique_counts parameter added to np.unique

The number of times each unique item comes up in the input can now be obtained as an optional return value.

Support for median and percentile in nanfunctions

The np.nanmedian and np.nanpercentile functions behave like the median and percentile functions except that NaNs are ignored.

NumpyVersion class added

The class may be imported from numpy.lib and can be used for version comparison when the numpy version goes to 1.10.devel. For example:

>>> from numpy.lib import NumpyVersion
>>> if NumpyVersion(np.__version__) < '1.10.0'):
...     print('Wow, that is an old NumPy version!')

Allow saving arrays with large number of named columns

The numpy storage format 1.0 only allowed the array header to have a total size of 65535 bytes. This can be exceeded by structured arrays with a large number of columns. A new format 2.0 has been added which extends the header size to 4 GiB. will automatically save in 2.0 format if the data requires it, else it will always use the more compatible 1.0 format.

Full broadcasting support for np.cross

np.cross now properly broadcasts its two input arrays, even if they have different number of dimensions. In earlier versions this would result in either an error being raised, or wrong results computed.


Better numerical stability for sum in some cases

Pairwise summation is now used in the sum method, but only along the fast axis and for groups of the values <= 8192 in length. This should also improve the accuracy of var and std in some common cases.

Percentile implemented in terms of np.partition

np.percentile has been implemented in terms of np.partition which only partially sorts the data via a selection algorithm. This improves the time complexity from O(nlog(n)) to O(n).

Performance improvement for np.array

The performance of converting lists containing arrays to arrays using np.array has been improved. It is now equivalent in speed to np.vstack(list).

Performance improvement for np.searchsorted

For the built-in numeric types, np.searchsorted no longer relies on the data type’s compare function to perform the search, but is now implemented by type specific functions. Depending on the size of the inputs, this can result in performance improvements over 2x.

Optional reduced verbosity for np.distutils

Set numpy.distutils.system_info.system_info.verbosity = 0 and then calls to numpy.distutils.system_info.get_info('blas_opt') will not print anything on the output. This is mostly for other packages using numpy.distutils.

Covariance check in np.random.multivariate_normal

A RuntimeWarning warning is raised when the covariance matrix is not positive-semidefinite.

Polynomial Classes no longer template based

The polynomial classes have been refactored to use an abstract base class rather than a template in order to implement a common interface. This makes importing the polynomial package faster as the classes do not need to be compiled on import.

More GIL releases

Several more functions now release the Global Interpreter Lock allowing more efficient parallelization using the threading module. Most notably the GIL is now released for fancy indexing, np.where and the random module now uses a per-state lock instead of the GIL.

MaskedArray support for more complicated base classes

Built-in assumptions that the baseclass behaved like a plain array are being removed. In particalur, repr and str should now work more reliably.



Non-integer scalars for sequence repetition

Using non-integer numpy scalars to repeat python sequences is deprecated. For example np.float_(2) * [1] will be an error in the future.

select input deprecations

The integer and empty input to select is deprecated. In the future only boolean arrays will be valid conditions and an empty condlist will be considered an input error instead of returning the default.

rank function

The rank function has been deprecated to avoid confusion with numpy.linalg.matrix_rank.

Object array equality comparisons

In the future object array comparisons both == and np.equal will not make use of identity checks anymore. For example:

>>> a = np.array([np.array([1, 2, 3]), 1])
>>> b = np.array([np.array([1, 2, 3]), 1])
>>> a == b

will consistently return False (and in the future an error) even if the array in a and b was the same object.

The equality operator == will in the future raise errors like np.equal if broadcasting or element comparisons, etc. fails.

Comparison with arr == None will in the future do an elementwise comparison instead of just returning False. Code should be using arr is None.

All of these changes will give Deprecation- or FutureWarnings at this time.


The utility function npy_PyFile_Dup and npy_PyFile_DupClose are broken by the internal buffering python 3 applies to its file objects. To fix this two new functions npy_PyFile_Dup2 and npy_PyFile_DupClose2 are declared in npy_3kcompat.h and the old functions are deprecated. Due to the fragile nature of these functions it is recommended to instead use the python API when possible.

This change was already applied to the 1.8.1 release.