Detailed SciPy Roadmap#
Most of this roadmap is intended to provide a high-level view on what is most needed per SciPy submodule in terms of new functionality, bug fixes, etc. Besides important “business as usual” changes, it contains ideas for major new features - those are marked as such, and are expected to take significant dedicated effort. Things not mentioned in this roadmap are not necessarily unimportant or out of scope, however we (the SciPy developers) want to provide to our users and contributors a clear picture of where SciPy is going and where help is needed most.
Note
This is the detailed roadmap. A very high-level overview with only the most important ideas is SciPy Roadmap.
General#
This roadmap will be evolving together with SciPy. Updates can be submitted as pull requests. For large or disruptive changes you may want to discuss those first on the scipy-dev forum.
API changes#
In general, we want to evolve the API to remove known warts as much as possible, however as much as possible without breaking backwards compatibility.
Test coverage#
Test coverage of code added in the last few years is quite good, and we aim for a high coverage for all new code that is added. However, there is still a significant amount of old code for which coverage is poor. Bringing that up to the current standard is probably not realistic, but we should plug the biggest holes.
Besides coverage there is also the issue of correctness - older code may have a
few tests that provide decent statement coverage, but that doesn’t necessarily
say much about whether the code does what it says on the box. Therefore code
review of some parts of the code (stats
, signal
and ndimage
in
particular) is necessary.
Documentation#
The documentation is in good shape. Expanding of current docstrings - adding examples, references, and better explanations - should continue. Most modules also have a tutorial in the reference guide that is a good introduction, however there are a few missing or incomplete tutorials - this should be fixed.
Benchmarks#
The asv
-based benchmark system is in reasonable shape. It is quite easy to
add new benchmarks, however running the benchmarks is not very intuitive.
Making this easier is a priority.
Use of Cython#
Cython’s old syntax for using NumPy arrays should be removed and replaced with Cython memoryviews.
Binary sizes of extensions built from Cython code are large, and compile times
are long. We should aim to combine extension modules where possible (e.g.,
stats._boost
contains many extension modules now), and limit the use of
Cython to places where it’s the best choice. Note that conversion of Cython
to C++ is ongoing in scipy.special
.
Use of Pythran#
Pythran is still an optional build dependency, and can be disabled with
-Duse-pythran=false
. The aim is to make it a hard dependency - for that to
happen it must be clear that the maintenance burden is low enough.
Use of venerable Fortran libraries#
SciPy owes a lot of its success to relying on wrapping well established Fortran libraries (QUADPACK, FITPACK, ODRPACK, ODEPACK etc). Some of these libraries are aging well, others less so. We should audit our use of these libraries with respect to the maintenance effort, the functionality, and the existence of (possibly partial) alternatives, including those inside SciPy.
Continuous integration#
Continuous integration currently covers 32/64-bit Windows, macOS on x86-64/arm, 32/64-bit Linux on x86, and Linux on aarch64 - as well as a range of versions of our dependencies and building release quality wheels. Reliability of CI has not been good recently (H1 2023), due to the large amount of configurations to support and some CI jobs needing an overhaul. We aim to reduce build times by removing the remaining distutils-based jobs when we drop that build system and make the set of configurations in CI jobs more orthogonal.
Size of binaries#
SciPy binaries are quite large (e.g. an unzipped manylinux wheel for 1.7.3 is
39 MB on PyPI and 122 MB after installation), and this can be problematic - for
example for use in AWS Lambda, which has a 250 MB size limit. We aim to keep
binary size as low as possible; when adding new compiled extensions, this needs
checking. Stripping of debug symbols in multibuild
can perhaps be improved
(see this issue).
An effort should be made to slim down where possible, and not add new large
files. In the future, things that are being considered (very tentatively) and
may help are separating out the bundled` libopenblas
and removing support
for long double
.
Modules#
cluster#
dendrogram
needs a rewrite, it has a number of hard to fix open issues and
feature requests.
constants#
This module is basically done, low-maintenance and without open issues.
fft#
This module is in good shape.
integrate#
Needed for ODE solvers:
Documentation is pretty bad, needs fixing
A new ODE solver interface (
solve_ivp
) was added in SciPy 1.0.0. In the future we can consider (soft-)deprecating the older API.
The numerical integration functions are in good shape. Support for integrating complex-valued functions and integrating multiple intervals (see gh-3325) could be added.
interpolate#
Spline fitting: we need spline fitting routines with better user control. This includes
user-selectable alternatives for the smoothing criteria (manual, cross-validation etc); gh-16653 makes a start in this direction;
several strategies for knot placement, both manual and automatic (using algorithms by Dierckx, de Boor, possibly other).
Once we have a reasonably feature complete set, we can start taking a long look at the future of the venerable FITPACK Fortran library, which currently is the only way of constructing smoothing splines in SciPy.
Scalability and performance: For the FITPACK-based functionality, the data size is limited by 32-bit Fortran integer size (for non-ILP64 builds). For N-D scattered interpolators (which are QHull based) and N-D regular grid interpolators we need to check performance on large data sets and improve where lacking (gh-16483 makes progress in this direction).
Ideas for new features: NURBS support could be added.
io#
wavfile:
PCM float will be supported, for anything else use
audiolab
or other specialized libraries.Raise errors instead of warnings if data not understood.
Other sub-modules (matlab, netcdf, idl, harwell-boeing, arff, matrix market) are in good shape.
linalg#
scipy.linalg
is in good shape.
Needed:
Reduce duplication of functions with
numpy.linalg
, make APIs consistent.get_lapack_funcs
should always useflapack
Wrap more LAPACK functions
One too many funcs for LU decomposition, remove one
Ideas for new features:
Add type-generic wrappers in the Cython BLAS and LAPACK
Make many of the linear algebra routines into gufuncs
BLAS and LAPACK
The Python and Cython interfaces to BLAS and LAPACK in scipy.linalg
are one
of the most important things that SciPy provides. In general scipy.linalg
is in good shape, however we can make a number of improvements:
Library support. Our released wheels now ship with OpenBLAS, which is currently the only feasible performant option (ATLAS is too slow, MKL cannot be the default due to licensing issues, Accelerate support is dropped because Apple doesn’t update Accelerate anymore). OpenBLAS isn’t very stable though, sometimes its releases break things and it has issues with threading (currently the only issue for using SciPy with PyPy3). We need at the very least better support for debugging OpenBLAS issues, and better documentation on how to build SciPy with it. An option is to use BLIS for a BLAS interface (see numpy gh-7372).
Support for newer LAPACK features. In SciPy 1.2.0 we increased the minimum supported version of LAPACK to 3.4.0. Now that we dropped Python 2.7, we can increase that version further (MKL + Python 2.7 was the blocker for >3.4.0 previously) and start adding support for new features in LAPACK.
misc#
scipy.misc
will be removed as a public module. Most functions in it have
been moved to another submodule or deprecated. The few that are left:
derivative
,central_diff_weight
: remove, possibly replacing them with more extensive functionality for numerical differentiation.ascent
,face
,electrocardiogram
: remove or move to the appropriate subpackages (e.g.scipy.ndimage
,scipy.signal
).
ndimage#
Underlying ndimage
is a powerful interpolation engine. Users come
with an expectation of one of two models: a pixel model with (1,
1)
elements having centers (0.5, 0.5)
, or a data point model,
where values are defined at points on a grid. Over time, we’ve become
convinced that the data point model is better defined and easier to
implement, but this should be clearly communicated in the documentation.
More importantly, still, SciPy implements one variant of this data
point model, where datapoints at any two extremes of an axis share a
spatial location under periodic wrapping mode. E.g., in a 1D array,
you would have x[0]
and x[-1]
co-located. A very common
use-case, however, is for signals to be periodic, with equal spacing
between the first and last element along an axis (instead of zero
spacing). Wrapping modes for this use-case were added in
gh-8537, next the
interpolation routines should be updated to use those modes.
This should address several issues, including gh-1323, gh-1903, gh-2045
and gh-2640.
The morphology interface needs to be standardized:
binary dilation/erosion/opening/closing take a “structure” argument, whereas their grey equivalent take size (has to be a tuple, not a scalar), footprint, or structure.
a scalar should be acceptable for size, equivalent to providing that same value for each axis.
for binary dilation/erosion/opening/closing, the structuring element is optional, whereas it’s mandatory for grey. Grey morphology operations should get the same default.
other filters should also take that default value where possible.
odr#
This module is in reasonable shape, although it could use a bit more maintenance. No major plans or wishes here.
optimize#
Overall this module is in good shape. Two good global optimizers were added in 1.2.0; large-scale optimizers is still a gap that could be filled. Other things that are needed:
Many ideas for additional functionality (e.g. integer constraints) in
linprog
, see gh-9269.Add functionality to the benchmark suite to compare results more easily (e.g. with summary plots).
deprecate the
fmin_*
functions in the documentation,minimize
is preferred.scipy.optimize
has an extensive set of benchmarks for accuracy and speed of the global optimizers. That has allowed adding new optimizers (shgo
anddual_annealing
) with significantly better performance than the existing ones. Theoptimize
benchmark system itself is slow and hard to use however; we need to make it faster and make it easier to compare performance of optimizers via plotting performance profiles.
signal#
Convolution and correlation: (Relevant functions are convolve, correlate,
fftconvolve, convolve2d, correlate2d, and sepfir2d.) Eliminate the overlap with
ndimage (and elsewhere). From numpy
, scipy.signal
and scipy.ndimage
(and anywhere else we find them), pick the “best of class” for 1-D, 2-D and n-d
convolution and correlation, put the implementation somewhere, and use that
consistently throughout SciPy.
B-splines: (Relevant functions are gauss_spline, cspline1d, qspline1d, cspline2d, qspline2d, cspline1d_eval, and spline_filter.) Move the good stuff to interpolate (with appropriate API changes to match how things are done in interpolate), and eliminate any duplication.
Filter design: merge firwin and firwin2 so firwin2 can be removed.
Continuous-Time Linear Systems: Further improve the performance of ltisys
(fewer internal transformations between different representations). Fill gaps in lti
system conversion functions.
Second Order Sections: Make SOS filtering equally capable as existing methods. This includes ltisys objects, an lfiltic equivalent, and numerically stable conversions to and from other filter representations. SOS filters could be considered as the default filtering method for ltisys objects, for their numerical stability.
Wavelets: what’s there now doesn’t make much sense. Continuous wavelets only at the moment - decide whether to completely rewrite or remove them. Discrete wavelet transforms are out of scope (PyWavelets does a good job for those).
sparse#
The sparse matrix formats are mostly feature-complete, however the main issue
is that they act like numpy.matrix
(which will be deprecated in NumPy at
some point).
What we want is sparse arrays that act like numpy.ndarray
. Initial
support for a new set of classes (csr_array
et al.) was added in SciPy
1.8.0
and stabilized in 1.12.0
when construction functions for
arrays were added. Support for 1-D array is expected in 1.13.0
.
Next steps toward sparse array support:
- Extend sparse array API to 1-D arrays.
Support for COO, CSR and DOK formats.
CSR 1D support for min-max, indexing, arithmetic.
Help other libraries convert to sparse arrays from sparse matrices. Create transition guide and helpful scripts to flag code that needs further examination. NetworkX, scikit-learn and scikit-image are in progress or have completed conversion to sparse arrays.
After sparse array code is mature (~1 release cycle?) add deprecation warnings for sparse matrix.
Work with NumPy on deprecation/removal of
numpy.matrix
.Deprecate and then remove sparse matrix in favor of sparse array.
- Start API shift of construction function names (
diags
,block
, etc.) Note: as a whole, the construction functions undergo two name shifts. Once to move from matrix creation to new functions for array creation (i.e.
eye
->eye_array
). Then a second move to change names to match thearray_api
name (i.e.eye_array
toeye
) after sparse matrices are removed. We will keep the*_array
versions for a long time as (maybe hidden) aliases.
- Start API shift of construction function names (
Add construction function names matching
array_api
names.Deprecate the transition construction function names.
An alternative (more ambitious, and unclear if it will materialize)
plan is being worked on in pydata/sparse.
To support that effort we aim to support PyData Sparse in all functions that
accept sparse arrays. Support for pydata/sparse
in scipy.sparse.linalg
and scipy.sparse.csgraph
is mostly complete.
Regarding the different sparse matrix formats: there are a lot of them. These should be kept, but improvements/optimizations should go into CSR/CSC, which are the preferred formats. LIL may be the exception, it’s inherently inefficient. It could be dropped if DOK is extended to support all the operations LIL currently provides.
sparse.csgraph#
This module is in good shape.
sparse.linalg#
There are a significant number of open issues for _arpack
and lobpcg
.
_propack
is new in 1.8.0, TBD how robust it will turn out to be.
_isolve
:
callback keyword is inconsistent
tol keyword is broken, should be relative tol
Fortran code not re-entrant (but we don’t solve, maybe reuse from PyKrilov)
_dsolve
:
add license-compatible sparse Cholesky or incomplete Cholesky
add license-compatible sparse QR
improve interface to SuiteSparse UMFPACK
add interfaces to SuiteSparse CHOLMOD and SPQR
spatial#
QHull wrappers are in good shape, as is KDTree
.
A rewrite of spatial.distance
metrics in C++ is in progress - this should
improve performance, make behaviour (e.g., for various non-float64 input
dtypes) more consistent, and fix a few remaining issues with definitions of the
math implement by a few of the metrics.
special#
Though there are still a lot of functions that need improvements in precision, probably the only show-stoppers are hypergeometric functions, parabolic cylinder functions, and spheroidal wave functions. Three possible ways to handle this:
Get good double-precision implementations. This is doable for parabolic cylinder functions (in progress). I think it’s possible for hypergeometric functions, though maybe not in time. For spheroidal wavefunctions this is not possible with current theory.
Port Boost’s arbitrary precision library and use it under the hood to get double precision accuracy. This might be necessary as a stopgap measure for hypergeometric functions; the idea of using arbitrary precision has been suggested before by @nmayorov and in gh-5349. Likely necessary for spheroidal wave functions, this could be reused: radelman/scattering.
Add clear warnings to the documentation about the limits of the existing implementations.
stats#
The scipy.stats
subpackage aims to provide fundamental statistical
methods as might be covered in standard statistics texts such as Johnson’s
“Miller & Freund’s Probability and Statistics for Engineers”, Sokal & Rohlf’s
“Biometry”, or Zar’s “Biostatistical Analysis”. It does not seek to duplicate
the advanced functionality of downstream packages (e.g. StatsModels,
LinearModels, PyMC3); instead, it can provide a solid foundation on which
they can build. (Note that these are rough guidelines, not strict rules.
“Advanced” is an ill-defined and subjective term, and “advanced” methods
may also be included in SciPy, especially if no other widely used and
well-supported package covers the topic. Also note that some duplication
with downstream projects is inevitable and not necessarily a bad thing.)
In addition to the items described in the SciPy Roadmap, the following improvements will help SciPy better serve this role.
Add fundamental and widely used hypothesis tests, such as:
post hoc tests (e.g. Dunnett’s test)
the various types of analysis of variance (ANOVA):
two-way ANOVA (single replicate, uniform number of replicates, variable number of replicates)
multiway ANOVA (i.e. generalize two-way ANOVA)
nested ANOVA
analysis of covariance (ANCOVA)
Also, provide an infrastructure for implementing hypothesis tests.
Add additional tools for meta-analysis
Add tools for survival analysis
Speed up random variate sampling (method
rvs
) of distributions, leveragingscipy.stats.sampling
where appropriateExpand QMC capabilities and performance
Enhance the fit method of the continuous probability distributions:
Expand the options for fitting to include:
maximal product spacings
method of L-moments / probability weighted moments
Include measures of goodness-of-fit in the results
Handle censored data (e.g. merge gh-13699)
Implement additional widely used continuous and discrete probability distributions, e.g. mixture distributions.
Improve the core calculations provided by SciPy’s probability distributions so they can robustly handle wide ranges of parameter values. Specifically, replace many of the PDF and CDF methods from the Fortran library CDFLIB used in
scipy.special
with Boost implementations as in gh-13328.
In addition, we should:
Continue work on making the function signatures of
stats
andstats.mstats
more consistent, and add tests to ensure that that remains the case.Improve statistical tests: return confidence intervals for the test statistic, and implement exact p-value calculations - considering the possibility of ties - where computationally feasible.