Changelog

4.2.1 - 2026-02-04

Bug fixes:

Fixed methods _cross_categorical, _cross_sandwich, multiply, tocsr of CategoricalMatrix and function _sandwich_cat_cat_limited_rows_cols to operate on read-only buffers as well.

4.2.0 - 2026-01-13

New feature:

tabmat.from_formula() now also supports any dataframe supported by narwhals.

Other changes:

Require Python>=3.10.

4.1.5 - 2025-12-17

Other changes:

Restore building wheels for Intel-based macOS systems.

4.1.4 - 2025-12-16

Other changes:

tabmat.from_df() now avoids unnecessary copies of dense arrays, but still ensures that the results are contiguous (C or F order).
We now use narwhals’ v2 API for data frame handling.

4.1.3 - 2025-11-12

Bug fixes:

Fixed CategoricalMatrix.transpose_matvec() to operate on read-only buffers as well.
Fixed incorrect calculation of the shape of a CategoricalMatrix when initialized with zero categories and drop_first=True.

4.1.2 - 2025-07-17

Bug fixes:

Fixed a bug which caused issues when constructing tabmat matrices from existing ModelSpecs when they contained categorical columns with all levels dropped.
We can now treat dedicated pandas string series - which are the defaults for strings since pandas 2.3 - as categoricals.

4.1.1 - 2025-01-30

Bug fix:

A more robust DenseMatrix._get_col_stds() results in more accurate StandardizedMatrix.sandwich() results.

Other changes:

Build wheel for pypi on python 3.13.
Build and test with python 3.13 in CI.

4.1.0 - 2024-11-07

New feature:

Added a new function, tabmat.from_df(), to convert any dataframe supported by narwhals into a tabmat.SplitMatrix.

Other changes:

Allow CategoricalMatrix to be initialized directly with indices and categories.
Added checks for dimension and dtype mismatch in MatrixBasesandwich.sandwich().

Bug fix:

Fixed a bug in tabmat.CategoricalMatrix.standardize() that sometimes returned nan values for the standard deviation due to numerical instability if using np.float32 precision.

4.0.1 - 2024-06-25

Other changes:

Removed reference to the .A attribute and replaced it with .toarray().
Add support between formulaic and pandas 3.0.
Support pypi release for numpy 2.0

4.0.0 - 2024-04-23

Breaking changes:

To unify the API, DenseMatrix does not inherit from np.ndarray anymore. To convert a DenseMatrix to a np.ndarray, use DenseMatrix.unpack().
Similarly, SparseMatrix does not inherit from sps.csc_matrix anymore. To convert a SparseMatrix to a sps.csc_matrix, use SparseMatrix.unpack().

New features:

Added column name and term name metadata to MatrixBase objects. These are automatically populated when initializing a MatrixBase from a pandas.DataFrame. In addition, they can be accessed and modified via the MatrixBase.column_names and MatrixBase.term_names properties.
Added a formula interface for creating tabmat matrices from pandas data frames. See tabmat.from_formula() for details.
Added support for missing values in CategoricalMatrix by either creating a separate category for them or treating them as all-zero rows.
Added support for handling missing categorical values in pandas data frames.

Bug fix:

Added cython compiler directive legacy_implicit_noexcept = True to fix performance regression with cython 3.

Other changes:

Refactored the pre-commit hooks to use ruff.
Refactored CategoricalMatrix.transpose_matvec() to be deterministic when using OpenMP.
Adjusted transformation to sparse format in tabmat.from_pandas() to future changes in pandas.

3.1.13 - 2023-10-17

Other changes:

Pypi release is now done using trusted publisher.
Fix build and upload of x86_64 wheels on Linux.

3.1.12 - 2023-10-16

Other changes:

Fixed macos arm64 wheels with proper linkage.

3.1.11 - 2023-10-13

Other changes:

Improve the performance of from_pandas in the case of low-cardinality categorical variables.
Require Python>=3.9 in line with NEP 29
Build and test with Python 3.12 in CI.
Fixed macos arm64 wheels with proper linkage.

3.1.10 - 2023-06-23

Bug fixes:

We fixed a bug in the dense sandwich product, which would previously segfault for very large matrices.
Fixed the column order when initializing a SplitMatrix from a list containing other SplitMatrix objects.
Fixed getcol not respecting the drop_first attribute of a CategoricalMatrix.

3.1.9 - 2023-06-16

Other changes:

Support building on architectures that are unsupported by xsimd.

3.1.8 - 2023-06-13

Other changes:

The C++ types have been refactored. Loop indices are now using the Py_ssize_t type. Integers now have a templated type as well.
The documentation for matvec and matvec_transpose has been updated to reflect actual behavior.
Checks for dimension mismatch in matvec and matvec_transpose arguments have been added.
Remove upper pin on xsimd.

3.1.7 - 2022-03-28

Bug fix:

We fixed a bug in the cross sandwich product, which would previously segfault for very large matrices.

3.1.6 - 2022-03-27

Bug fix:

We fixed a bug in the dense sandwich product, which would previously segfault for very large F-contiguous matrices.

3.1.5 - 2022-03-20

Bug fix:

We fixed a bug in the dense matrix-vector and sandwich products, which would previously segfault for very large matrices.

3.1.4 - 2022-02-07

Bug fix:

Fixed the loading of jemalloc in Apple Silicon wheels.

3.1.3 - 2022-01-26

Other changes:

Build and upload wheels for Apple Silicon.

3.1.2 - 2022-07-01

Other changes:

Next attempt to build wheel for PyPI without march=native.

3.1.1 - 2022-07-01

Other changes:

Add Python 3.10 support to CI (remove Python 3.6).
Build wheel for PyPI without march=native.

3.1.0 - 2022-03-07

New feature

tabmat.CategoricalMatrix now accepts a drop_first argurment. This allows the user to drop the first column of a CategoricalMatrix to avoid multicollinearity problems in unregularized models.
tabmat.StandardizedMatrix and tabmat.MatrixBase now support the multiply method.

3.0.8 - 2022-01-03

Bug fix

Always use 64bit integers for indexing in tabmat.ext.sparse.sparse_sandwich() to avoid segmentation faults on very wide problems.

3.0.7 - 2021-11-23

Bug fix

Disable the use of static TLS in the Linux wheels to avoid issues with too small TLS on some distributions.

3.0.6 - 2021-11-11

Bug fix

We fixed a bug in tabmat.SplitMatrix.matvec(), where incorrect matrix vector products were computed when a SplitMatrix did not contain any dense components.

3.0.5 - 2021-11-05

Other changes

We are now specifying the run time dependencies in setup.py, so that missing dependencies are automatically installed from PyPI when installing tabmat via pip.

3.0.4 - 2021-11-03

Other changes

tabmat is now available on PyPI and will be automatically updated when a new release is published.

3.0.3 - 2021-10-15

Bug fix

We now support xsimd>=8 and support alternative jemalloc installations.

3.0.2 - 2021-10-14

Bug fix

Allow to link to alternatively suffixed jemalloc installation to work around #113 .

3.0.1 - 2021-10-07

Bug fix

The license was mistakenly left as proprietary. Corrected to BSD-3-Clause.

Other changes

ReadTheDocs integration.
CONTRIBUTING.md
Correct pyproject.toml to work with PEP-517

3.0.0 - 2021-10-07

Breaking changes:

The package has been renamed to tabmat. CELEBRATE!
The one_over_var_inf_to_val() function has been made private.
The csc_to_split() function has been re-named to tabmat.from_csc() to match the tabmat.from_pandas() function.
The tabmat.MatrixBase.get_col_means() and tabmat.MatrixBase.get_col_stds() methods have been made private.
The cross_sandwich() method has also been made private.

Bug fix

StandardizedMatrix.transpose_matvec() was giving the wrong answer when the out parameter was provided. This is now fixed.
SplitMatrix.__repr__() now calls the __repr__ method of component matrices instead of __str__.

Other changes

Optimized the tabmat.SparseMatrix.matvec() and tabmat.SparseMatrix.transpose_matvec() for when rows and cols are None.
Implemented CategoricalMatrix.__rmul__()
Reorganizing the documentation and updating the text to match the current API.
Enable indexing the rows of a CategoricalMatrix. Previously CategoricalMatrix.__getitem__() only supported column indexing.
Allow creating a SplitMatrix from a list of any MatrixBase objects including another SplitMatrix.
Reduced memory usage in tabmat.SplitMatrix.matvec().

2.0.3 - 2021-07-15

Bug fix

In SplitMatrix.sandwich(), when a col subset was specified, incorrect output was produced if the components of the indices array were not sorted. SplitMatrix.__init__() now checks for sorted indices and maintains sorted index lists when combining matrices.

Other changes

SplitMatrix.__init__() now filters out any empty matrices.
StandardizedMatrix.sandwich() passes rows=None and cols=None onwards to the underlying matrix instead of replacing them with full arrays of indices. This should improve performance slightly.
SplitMatrix.__repr__() now includes the type of the underlying matrix objects in the string output.

2.0.2 - 2021-06-24

Bug fix

Sparse matrices now accept 64-bit indices on Windows.

2.0.1 - 2021-06-20

Bug fix:

Split matrices now also work on Windows.

2.0.0 - 2021-06-17

Breaking changes:

We renamed several public functions to make them private. These include functions in tabmat.benchmark that are unlikely to be used outside of this package as well as

tabmat.dense_matrix._matvec_helper()

tabmat.sparse_matrix._matvec_helper().

tabmat.split_matrix._prepare_out_array().

Other changes:

We removed the dependency on sparse_dot_mkl. We now use scipy.sparse.csr_matvec() instead of sparse_dot_mkl.dot_product_mkl() on all platforms, because the former suffered from poor performance, especially on narrow problems. This also means that we removed the function tabmat.sparse_matrix._dot_product_maybe_mkl().
We updated the pre-commit hooks and made sure the code is line with the new hooks.

1.0.6 - 2020-04-26

Other changes:

We are now also making releases for Windows.

1.0.5 - 2020-04-26

Other changes:

Still trying.

1.0.4 - 2020-04-26

Other changes:

We are trying to make releases for Windows.

1.0.3 - 2020-04-21

Bug fixes:

Added a check that matrices are two-dimensional in the SplitMatrix.__init__
Replace np.int with np.int64 where appropriate due to NumPy deprecation of np.int.

1.0.2 - 2020-04-20

Other changes:

Added Python 3.9 support.
Use scipy.sparse dot product when MKL isn’t available.

1.0.1 - 2020-11-25

Bug fixes:

Handling for nulls when setting up a CategoricalMatrix
Fixes to make several functions work with both row and col restrictions and out

Other changes:

Added various tests and documentation improvements

1.0.0 - 2020-11-11

Breaking change:

Rename dot to matvec. Our dot function supports matrix-vector multiplication for every subclass, but only supports matrix-matrix multiplication for some. We therefore rename it to matvec in line with other libraries.

Bug fix:

Fix a bug in matvec for categorical components when the number of categories exceeds the number of rows.

0.0.6 - 2020-08-03

See git history.