SciPy 1.5.0 Release Notes
Note Scipy 1.5.0
is not released yet!
SciPy 1.5.0
is the culmination of 6 months of hard work. It contains
many new features, numerous bug-fixes, improved test coverage and better
documentation. There have been a number of deprecations and 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
1.5.x branch, and on adding new features on the master branch.
This release requires Python 3.6+
and NumPy 1.14.5
or greater.
For running on PyPy, PyPy3 6.0+
and NumPy 1.15.0
are required.
Highlights of this release
- wrappers for more than a dozen new
LAPACK
routines are now available
inscipy.linalg.lapack
- Improved support for leveraging 64-bit integer size from linear algebra
backends - addition of the probability distribution for two-sided one-sample
Kolmogorov-Smirnov tests
New features
scipy.cluster
improvements
Initialization of scipy.cluster.vq.kmeans2
using minit="++"
had a
quadratic complexity in the number of samples. It has been improved, resulting
in a much faster initialization with quasi-linear complexity.
scipy.cluster.hierarchy.dendrogram
now respects the matplotlib
color
palette
scipy.fft
improvements
A new keyword-only argument plan
is added to all FFT functions in this
module. It is reserved for passing in a precomputed plan from libraries
providing a FFT backend (such as PyFFTW
and mkl-fft
), and it is
currently not used in SciPy.
scipy.integrate
improvements
scipy.interpolate
improvements
scipy.io
improvements
scipy.io.wavfile
error messages are more explicit about what's wrong, and
extraneous bytes at the ends of files are ignored instead of raising an error
when the data has successfully been read.
scipy.io.loadmat
gained a simplify_cells
parameter, which if set to
True
simplifies the structure of the return value if the .mat
file
contains cell arrays.
pathlib.Path
objects are now supported in scipy.io
Matrix Market I/O
functions
scipy.linalg
improvements
scipy.linalg.eigh
has been improved. Now various LAPACK
drivers can be
selected at will and also subsets of eigenvalues can be requested via
subset_by_value
keyword. Another keyword subset_by_index
is introduced.
Keywords turbo
and eigvals
are deprecated.
Similarly, standard and generalized Hermitian eigenvalue LAPACK
routines
?<sy/he>evx
are added and existing ones now have full _lwork
counterparts.
Wrappers for the following LAPACK
routines have been added to
scipy.linalg.lapack
:
?getc2
: computes the LU factorization of a general matrix with complete
pivoting?gesc2
: solves a linear system given an LU factorization from?getc2
?gejsv
: computes the singular value decomposition of a general matrix
with higher accuracy calculation of tiny singular values and their
corresponding singular vectors?geqrfp
: computes the QR factorization of a general matrix with
non-negative elements on the diagonal of R?gtsvx
: solves a linear system with general tridiagonal matrix?gttrf
: computes the LU factorization of a tridiagonal matrix?gttrs
: solves a linear system given an LU factorization from?gttrf
?ptsvx
: solves a linear system with symmetric positive definite
tridiagonal matrix?pttrf
: computes the LU factorization of a symmetric positive definite
tridiagonal matrix?pttrs
: solves a linear system given an LU factorization from?pttrf
?pteqr
: computes the eigenvectors and eigenvalues of a positive definite
tridiagonal matrix?tbtrs
: solves a linear system with a triangular banded matrix?csd
: computes the Cosine Sine decomposition of an orthogonal/unitary
matrix
Generalized QR factorization routines (?geqrf
) now have full _lwork
counterparts.
scipy.linalg.cossin
Cosine Sine decomposition of unitary matrices has been
added.
The function scipy.linalg.khatri_rao
, which computes the Khatri-Rao product,
was added.
The new function scipy.linalg.convolution_matrix
constructs the Toeplitz
matrix representing one-dimensional convolution.
scipy.ndimage
improvements
scipy.optimize
improvements
The finite difference numerical differentiation used in various minimize
methods that use gradients has several new features:
- 2-point, 3-point, or complex step finite differences can be used. Previously
only a 2-step finite difference was available. - There is now the possibility to use a relative step size, previously only an
absolute step size was available. - If the
minimize
method uses bounds the numerical differentiation strictly
obeys those limits. - The numerical differentiation machinery now makes use of a simple cache,
which in some cases can reduce the number of function evaluations. minimize
'smethod= 'powell'
now supports simple bound constraints
There have been several improvements to scipy.optimize.linprog
:
- The
linprog
benchmark suite has been expanded considerably. linprog
's dense pivot-based redundancy removal routine and sparse
presolve are faster- When
scikit-sparse
is available, solving sparse problems with
method='interior-point'
is faster
The caching of values when optimizing a function returning both value and
gradient together has been improved, avoiding repeated function evaluations
when using a HessianApproximation
such as BFGS
.
differential_evolution
can now use the modern np.random.Generator
as
well as the legacy np.random.RandomState
as a seed.
scipy.signal
improvements
A new optional argument include_nyquist
is added to freqz
functions in
this module. It is used for including the last frequency (Nyquist frequency).
scipy.signal.find_peaks_cwt
now accepts a window_size
parameter for the
size of the window used to calculate the noise floor.
scipy.sparse
improvements
Outer indexing is now faster when using a 2d column vector to select column
indices.
scipy.sparse.lil.tocsr
is faster
Fixed/improved comparisons between pydata sparse arrays and sparse matrices
BSR format sparse multiplication performance has been improved.
scipy.sparse.linalg.LinearOperator
has gained the new ndim
class
attribute
scipy.spatial
improvements
scipy.spatial.geometric_slerp
has been added to enable geometric
spherical linear interpolation on an n-sphere
scipy.spatial.SphericalVoronoi
now supports calculation of region areas in 2D
and 3D cases
The tree building algorithm used by cKDTree
has improved from quadratic
worst case time complexity to loglinear. Benchmarks are also now available for
building and querying of balanced/unbalanced kd-trees.
scipy.special
improvements
The following functions now have Cython interfaces in cython_special
:
scipy.special.erfinv
scipy.special.erfcinv
scipy.special.spherical_jn
scipy.special.spherical_yn
scipy.special.spherical_in
scipy.special.spherical_kn
scipy.special.log_softmax
has been added to calculate the logarithm of softmax
function. It provides better accuracy than log(scipy.special.softmax(x))
for
inputs that make softmax saturate.
scipy.stats
improvements
The function for generating random samples in scipy.stats.dlaplace
has been
improved. The new function is approximately twice as fast with a memory
footprint reduction between 25 % and 60 % (see gh-11069).
scipy.stats
functions that accept a seed for reproducible calculations using
random number generation (e.g. random variates from distributions) can now use
the modern np.random.Generator
as well as the legacy
np.random.RandomState
as a seed.
The axis
parameter was added to scipy.stats.rankdata
. This allows slices
of an array along the given axis to be ranked independently.
The axis
parameter was added to scipy.stats.f_oneway
, allowing it to
compute multiple one-way ANOVA tests for data stored in n-dimensional
arrays. The performance of f_oneway
was also improved for some cases.
The PDF and CDF methods for stats.geninvgauss
are now significantly faster
as the numerical integration to calculate the CDF uses a Cython based
LowLevelCallable
.
Moments of the normal distribution (scipy.stats.norm
) are now calculated using
analytical formulas instead of numerical integration for greater speed and
accuracy
Moments and entropy trapezoidal distribution (scipy.stats.trapz
) are now
calculated using analytical formulas instead of numerical integration for
greater speed and accuracy
Methods of the truncated normal distribution (scipy.stats.truncnorm
),
especially _rvs
, are significantly faster after a complete rewrite.
The fit
method of the Laplace distribution, scipy.stats.laplace
, now uses
the analytical formulas for the maximum likelihood estimates of the parameters.
Generation of random variates is now thread safe for all SciPy distributions.
3rd-party distributions may need to modify the signature of the _rvs()
method to conform to _rvs(self, ..., size=None, random_state=None)
. (A
one-time VisibleDeprecationWarning is emitted when using non-conformant
distributions.)
The Kolmogorov-Smirnov two-sided test statistic distribution
(scipy.stats.kstwo
) was added. Calculates the distribution of the K-S
two-sided statistic D_n
for a sample of size n, using a mixture of exact
and asymptotic algorithms.
The new function median_abs_deviation
replaces the deprecated
median_absolute_deviation
.
The wilcoxon
function now computes the p-value for Wilcoxon's signed rank
test using the exact distribution for inputs up to length 25. The function has
a new mode
parameter to specify how the p-value is to be computed. The
default is "auto"
, which uses the exact distribution for inputs up to length
25 and the normal approximation for larger inputs.
Added a new Cython-based implementation to evaluate guassian kernel estimates,
which should improve the performance of gaussian_kde
The winsorize
function now has a nan_policy
argument for refined
handling of nan
input values.
The binned_statistic_dd
function with statistic="std"
performance was
improved by ~4x.
scipy.stats.kstest(rvs, cdf,...)
now handles both one-sample and
two-sample testing. The one-sample variation uses scipy.stats.ksone
(or scipy.stats.kstwo
with back off to scipy.stats.kstwobign
) to calculate
the p-value. The two-sample variation, invoked if cdf
is array_like, uses
an algorithm described by Hodges to compute the probability directly, only
backing off to scipy.stats.kstwo
in case of overflow. The result in both
cases is more accurate p-values, especially for two-sample testing with
smaller (or quite different) sizes.
scipy.stats.maxwell
performance improvements include a 20 % speed up for
`fit()and 5 % for
pdf()``
scipy.stats.shapiro
and scipy.stats.jarque_bera
now return a named tuple
for greater consistency with other stats
functions
Deprecated features
scipy
deprecations
scipy.special
changes
The bdtr
, bdtrc
, and bdtri
functions are deprecating non-negative
non-integral n
arguments.
scipy.stats
changes
The function median_absolute_deviation
is deprecated. Use
median_abs_deviation
instead.
The use of the string "raw"
with the scale
parameter of iqr
is
deprecated. Use scale=1
instead.
Backwards incompatible changes
scipy.interpolate
changes
scipy.linalg
changes
The output signatures of ?syevr
, ?heevr
have been changed from
w, v, info
to w, v, m, isuppz, info
The order of output arguments w
, v
of <sy/he>{gv, gvd, gvx}
is
swapped.
scipy.signal
changes
The output length of scipy.signal.upfirdn
has been corrected, resulting
outputs may now be shorter for some combinations of up/down ratios and input
signal and filter lengths.
scipy.signal.resample
now supports a domain
keyword argument for
specification of time or frequency domain input.
scipy.stats
changes
Other changes
Improved support for leveraging 64-bit integer size from linear algebra backends
in several parts of the SciPy codebase.
Shims designed to ensure the compatibility of SciPy with Python 2.7 have now
been removed.
Many warnings due to unused imports and unused assignments have been addressed.
Many usage examples were added to function docstrings, and many input
validations and intuitive exception messages have been added throughout the
codebase.
Early stage adoption of type annotations in a few parts of the codebase
Authors
- @endolith
- Hameer Abbasi
- ADmitri +
- Wesley Alves +
- Berkay Antmen +
- Sylwester Arabas +
- Arne Küderle +
- Christoph Baumgarten
- Peter Bell
- Felix Berkenkamp
- Jordão Bragantini +
- Clemens Brunner +
- Evgeni Burovski
- Matthias Bussonnier +
- CJ Carey
- Derrick Chambers +
- Leander Claes +
- Christian Clauss
- Luigi F. Cruz +
- dankleeman
- Andras Deak
- Milad Sadeghi DM +
- jeremie du boisberranger +
- Stefan Endres
- Malte Esders +
- Leo Fang +
- felixhekhorn +
- Isuru Fernando
- Andrew Fowlie
- Lakshay Garg +
- Gaurav Gijare +
- Ralf Gommers
- Emmanuelle Gouillart +
- Kevin Green +
- Martin Grignard +
- Maja Gwozdz
- Sturla Molden
- gyu-don +
- Matt Haberland
- hakeemo +
- Charles Harris
- Alex Henrie
- Santi Hernandez +
- William Hickman +
- Till Hoffmann +
- Joseph T. Iosue +
- Anany Shrey Jain
- Jakob Jakobson
- Charles Jekel +
- Julien Jerphanion +
- Jiacheng-Liu +
- Christoph Kecht +
- Paul Kienzle +
- Reidar Kind +
- Dmitry E. Kislov +
- Konrad +
- Konrad0
- Takuya KOUMURA +
- Krzysztof Pióro
- Peter Mahler Larsen
- Eric Larson
- Antony Lee
- Gregory Lee +
- Gregory R. Lee
- Chelsea Liu
- Cong Ma +
- Kevin Mader +
- Maja Gwóźdź +
- Alex Marvin +
- Matthias Kümmerer
- Nikolay Mayorov
- Mazay0 +
- G. D. McBain
- Nicholas McKibben +
- Sabrina J. Mielke +
- Sebastian J. Mielke +
- Miloš Komarčević +
- Shubham Mishra +
- Santiago M. Mola +
- Grzegorz Mrukwa +
- Peyton Murray
- Andrew Nelson
- Nico Schlömer
- nwjenkins +
- odidev +
- Sambit Panda
- Vikas Pandey +
- Rick Paris +
- Harshal Prakash Patankar +
- Balint Pato +
- Matti Picus
- Ilhan Polat
- poom +
- Siddhesh Poyarekar
- Vladyslav Rachek +
- Bharat Raghunathan
- Manu Rajput +
- Tyler Reddy
- Andrew Reed +
- Lucas Roberts
- Ariel Rokem
- Heshy Roskes
- Matt Ruffalo
- Atsushi Sakai +
- Benjamin Santos +
- Christoph Schock +
- Lisa Schwetlick +
- Chris Simpson +
- Leo Singer
- Kai Striega
- Søren Fuglede Jørgensen
- Kale-ab Tessera +
- Seth Troisi +
- Robert Uhl +
- Paul van Mulbregt
- Vasiliy +
- Isaac Virshup +
- Pauli Virtanen
- Shakthi Visagan +
- Jan Vleeshouwers +
- Sam Wallan +
- Lijun Wang +
- Warren Weckesser
- Richard Weiss +
- wenhui-prudencemed +
- Eric Wieser
- Josh Wilson
- James Wright +
- Ruslan Yevdokymov +
- Ziyao Zhang +
A total of 129 people contributed to this release.
People with a "+" by their names contributed a patch for the first time.
This list of names is automatically generated, and may not be fully complete.