Scientific-Python Developer Summit#
(May 22-26, 2023, Seattle WA) – The first Scientific Python Developer Summit provided an opportunity for core developers from the scientific Python ecosystem to come together to:
- improve joint infrastructure
- better coordinate core projects
- work on a shared strategic plan
One of the focuses of the summit was Sparse Arrays, and specifically their implementation in SciPy. This post attempts to recap what happened with “sparse” at the summit and a glimpse of plans for our continuing work. The Sparse Array working group holds open follow-up meetings, currently scheduled every two weeks, to continue the momentum and move this project forward.
At the Summit, we focused on improving the newly added Sparse Array API in SciPy, that lets users manipulate sparse data with NumPy semantics (before, SciPy used NumPy’s 2D-only Matrix API, but that is slated for deprecation). Our goal at the summit was to give focused energy to the effort, bring new people on board, and connect downstream users with the development effort. We also worked to create a working group for this project that would last beyond the summit itself.
The specific PRs and Issues involved in
scipy.sparse are detailed in the
Summit 2023 scipy.sparse Report,
with more detailed description appearing in the
Some big picture take-aways are:
- Reorganized how to check for matrix/array/format info. This involved
formatattribute describing which format of sparse storage is used, changing functions
isspmatrixas well as shifting the class hierarchy to allow easy
isinstancechecking. The interface going forward includes:
issparse(A): True when a sparse array or matrix.
isinstance(A, sparray): True when a sparse array.
isspmatrix(A): True when a sparse matrix. To check the format of a sparse array or matrix use
A.format == "csr"or similar.
- Made decisions about how to approach the “creation functions” for sparse arrays.
The big-picture approach is to introduce new functions with an
_arraysuffix which construct sparse arrays. The old names will continue to create sparse matrix until post-deprecation removal. Some specific changes made include:
- Add the creation function
diags_array(A)(and planned for
- Create a
sparse.linalg.matrix_powerfunction for positive integer matrix power of a sparse array
- Add the creation function
- Made progress toward 1D sparse arrays. The data structures for 1d may be quite different from 2d.
coo_arrayallowed exploration of possible n-d arrays, though that is not a short-term goal.
- Explored feasibility and usefulness of defining
__array_*__protocols for sparse arrays
- Made clearer distinction between private and public methods for sparse arrays
- Improved documentation for sparse arrays
Our goal is to have a working set of sparse array construction functions
and a 1d sparse array class (focusing on
coo_array first) in plenty of
time for intensive testing before SciPy v1.12. This will then allow us to
focus on creating migration documents and tools as well as helping downstream
libraries make the shift to sparse arrays. We hope to enable the removal of
deprecated sparse matrix interfaces in favor of the array interface. For this
to happen we will need most downstream users to shift to the sparse array API.
We intend to help them do that.
Our work continues with a community call every two weeks on Fridays. Near term work is to:
- Continue improving sparse creation functions: diags, eye, random and others.
- Deprecate some matrix-specific functionality
- General performance improvements
- Adapting scikit-learn to support sparse arrays (to be discussed with scikit-learn’s maintainers)