SciPy 1.3.0 Release Notes#
SciPy 1.3.0 is the culmination of 5 months of hard work. It contains
many new features, numerous bug-fixes, improved test coverage and better
documentation. There have been some API changes
in this release, which are documented below. All users are encouraged to
upgrade to this release, as there are a large number of bug-fixes and
optimizations. Before upgrading, we recommend that users check that
their own code does not use deprecated SciPy functionality (to do so,
run your code with python -Wd
and check for DeprecationWarning
s).
Our development attention will now shift to bug-fix releases on the
1.3.x branch, and on adding new features on the master branch.
This release requires Python 3.5+ and NumPy 1.13.3 or greater.
For running on PyPy, PyPy3 6.0+ and NumPy 1.15.0 are required.
Highlights of this release#
Three new
stats
functions, a rewrite ofpearsonr
, and an exact computation of the Kolmogorov-Smirnov two-sample test.A new Cython API for bounded scalar-function root-finders in
scipy.optimize
.Substantial
CSR
andCSC
sparse matrix indexing performance improvements.Added support for interpolation of rotations with continuous angular rate and acceleration in
RotationSpline
.
New features#
scipy.interpolate
improvements#
A new class CubicHermiteSpline
is introduced. It is a piecewise-cubic
interpolator which matches observed values and first derivatives. Existing
cubic interpolators CubicSpline
, PchipInterpolator
and
Akima1DInterpolator
were made subclasses of CubicHermiteSpline
.
scipy.io
improvements#
For the Attribute-Relation File Format (ARFF) scipy.io.arff.loadarff
now supports relational attributes.
scipy.io.mmread
can now parse Matrix Market format files with empty lines.
scipy.linalg
improvements#
Added wrappers for ?syconv
routines, which convert a symmetric matrix
given by a triangular matrix factorization into two matrices and vice versa.
scipy.linalg.clarkson_woodruff_transform
now uses an algorithm that leverages
sparsity. This may provide a 60-90 percent speedup for dense input matrices.
Truly sparse input matrices should also benefit from the improved sketch
algorithm, which now correctly runs in O(nnz(A))
time.
Added new functions to calculate symmetric Fiedler matrices and
Fiedler companion matrices, named scipy.linalg.fiedler
and
scipy.linalg.fiedler_companion
, respectively. These may be used
for root finding.
scipy.ndimage
improvements#
Gaussian filter performances may improve by an order of magnitude in
some cases, thanks to the removal of a dependence on np.polynomial
. This
may impact scipy.ndimage.gaussian_filter
for example.
scipy.optimize
improvements#
The scipy.optimize.brute
minimizer obtained a new keyword workers
, which
can be used to parallelize computation.
A Cython API for bounded scalar-function root-finders in scipy.optimize
is available in a new module scipy.optimize.cython_optimize
via cimport
.
This API may be used with nogil
and prange
to loop
over an array of function arguments to solve for an array of roots more
quickly than with pure Python.
'interior-point'
is now the default method for linprog
, and
'interior-point'
now uses SuiteSparse for sparse problems when the
required scikits (scikit-umfpack and scikit-sparse) are available.
On benchmark problems (gh-10026), execution time reductions by factors of 2-3
were typical. Also, a new method='revised simplex'
has been added.
It is not as fast or robust as method='interior-point'
, but it is a faster,
more robust, and equally accurate substitute for the legacy
method='simplex'
.
differential_evolution
can now use a Bounds
class to specify the
bounds for the optimizing argument of a function.
scipy.optimize.dual_annealing
performance improvements related to
vectorization of some internal code.
scipy.signal
improvements#
Two additional methods of discretization are now supported by
scipy.signal.cont2discrete
: impulse
and foh
.
scipy.signal.firls
now uses faster solvers.
scipy.signal.detrend
now has a lower physical memory footprint in some
cases, which may be leveraged using the new overwrite_data
keyword argument.
scipy.signal.firwin
pass_zero
argument now accepts new string arguments
that allow specification of the desired filter type: 'bandpass'
,
'lowpass'
, 'highpass'
, and 'bandstop'
.
scipy.signal.sosfilt
may have improved performance due to lower retention
of the global interpreter lock (GIL) in the algorithm.
scipy.sparse
improvements#
A new keyword was added to csgraph.dijsktra
that
allows users to query the shortest path to ANY of the passed-in indices,
as opposed to the shortest path to EVERY passed index.
scipy.sparse.linalg.lsmr
performance has been improved by roughly 10 percent
on large problems.
Improved performance and reduced physical memory footprint of the algorithm
used by scipy.sparse.linalg.lobpcg
.
CSR
and CSC
sparse matrix fancy indexing performance has been
improved substantially.
scipy.spatial
improvements#
scipy.spatial.ConvexHull
now has a good
attribute that can be used
alongsize the QGn
Qhull options to determine which external facets of a
convex hull are visible from an external query point.
scipy.spatial.cKDTree.query_ball_point
has been modernized to use some newer
Cython features, including GIL handling and exception translation. An issue
with return_sorted=True
and scalar queries was fixed, and a new mode named
return_length
was added. return_length
only computes the length of the
returned indices list instead of allocating the array every time.
scipy.spatial.transform.RotationSpline
has been added to enable interpolation
of rotations with continuous angular rates and acceleration.
scipy.stats
improvements#
Added a new function to compute the Epps-Singleton test statistic,
scipy.stats.epps_singleton_2samp
, which can be applied to continuous and
discrete distributions.
New functions scipy.stats.median_absolute_deviation and scipy.stats.gstd
(geometric standard deviation) were added. The scipy.stats.combine_pvalues
method now supports pearson
, tippett
and mudholkar_george
pvalue
combination methods.
The scipy.stats.ortho_group
and scipy.stats.special_ortho_group
rvs(dim)
functions’ algorithms were updated from a O(dim^4)
implementation to a O(dim^3)
which gives large speed improvements
for dim>100
.
A rewrite of scipy.stats.pearsonr
to use a more robust algorithm,
provide meaningful exceptions and warnings on potentially pathological input,
and fix at least five separate reported issues in the original implementation.
Improved the precision of hypergeom.logcdf
and hypergeom.logsf
.
Added exact computation for Kolmogorov-Smirnov (KS) two-sample test, replacing
the previously approximate computation for the two-sided test stats.ks_2samp.
Also added a one-sided, two-sample KS test, and a keyword alternative
to
stats.ks_2samp.
Backwards-incompatible changes#
scipy.interpolate
changes#
Functions from scipy.interpolate
(spleval
, spline
, splmake
,
and spltopp
) and functions from scipy.misc
(bytescale
,
fromimage
, imfilter
, imread
, imresize
, imrotate
,
imsave
, imshow
, toimage
) have been removed. The former set has
been deprecated since v0.19.0 and the latter has been deprecated since v1.0.0.
Similarly, aliases from scipy.misc
(comb
, factorial
,
factorial2
, factorialk
, logsumexp
, pade
, info
, source
,
who
) which have been deprecated since v1.0.0 are removed.
SciPy documentation for
v1.1.0
can be used to track the new import locations for the relocated functions.
scipy.linalg
changes#
For pinv
, pinv2
, and pinvh
, the default cutoff values are changed
for consistency (see the docs for the actual values).
scipy.optimize
changes#
The default method for linprog
is now 'interior-point'
. The method’s
robustness and speed come at a cost: solutions may not be accurate to
machine precision or correspond with a vertex of the polytope defined
by the constraints. To revert to the original simplex method,
include the argument method='simplex'
.
scipy.stats
changes#
Previously, ks_2samp(data1, data2)
would run a two-sided test and return
the approximated p-value. The new signature, ks_2samp(data1, data2,
alternative="two-sided", method="auto")
, still runs the two-sided test by
default but returns the exact p-value for small samples and the approximated
value for large samples. method="asymp"
would be equivalent to the
old version but auto
is the better choice.
Other changes#
Our tutorial has been expanded with a new section on global optimizers.
There has been a rework of the stats.distributions
tutorials.
scipy.optimize
now correctly sets the convergence flag of the result to
CONVERR
, a convergence error, for bounded scalar-function root-finders
if the maximum iterations has been exceeded, disp
is false, and
full_output
is true.
scipy.optimize.curve_fit
no longer fails if xdata
and ydata
dtypes
differ; they are both now automatically cast to float64
.
scipy.ndimage
functions including binary_erosion
, binary_closing
, and
binary_dilation
now require an integer value for the number of iterations,
which alleviates a number of reported issues.
Fixed normal approximation in case zero_method == "pratt"
in
scipy.stats.wilcoxon
.
Fixes for incorrect probabilities, broadcasting issues and thread-safety
related to stats distributions setting member variables inside _argcheck()
.
scipy.optimize.newton
now correctly raises a RuntimeError
in the following cases: when default
arguments are used and if a derivative of value zero is obtained (which is a special case of failing to converge).
A draft toolchain roadmap is now available, laying out a compatibility plan including Python versions, C standards, and NumPy versions.