Skip to content

CHANGELOG.md

Latest commit

 

History

History
85 lines (61 loc) · 6.56 KB

CHANGELOG.md

File metadata and controls

85 lines (61 loc) · 6.56 KB
Raw

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[3.2.2] - 2026-07-06

Changed

  • Lowered the minimum supported Python version from 3.11 to 3.10 (requires-python = ">=3.10, <3.15"). The code base uses no Python-3.11-only features; this broadens compatibility for downstream packages that still support Python 3.10. Python 3.10 is now included in the CI test matrices. No functional changes.

Fixed

  • The fold-index parameters (validation_indices / val_indices) are now annotated with a new IndexArray alias that _enable_jax_typing() broadens to admit jax.Array, matching the Array/Scalar aliases introduced in 3.2.1. Previously they were annotated as npt.NDArray[np.int_] only, so running the test suite with typeguard instrumentation enabled failed on the jax.jit/jax.vmap trace tests (the indices are abstract tracers under tracing). Runtime behavior is unchanged; this only affects runtime type checking.
  • CI now actually enables typeguard instrumentation during tests: the test command passed --typeguard-packages=cvmatrix/ (a path, with a trailing slash) instead of the package name cvmatrix, which silently instrumented nothing and masked the annotation issue above.

[3.2.1] - 2026-07-01

Changed

  • import cvmatrix no longer imports JAX, even when JAX is installed. JAX was previously imported at module load time to broaden the array/scalar type hints so that backend="jax" values satisfy runtime type checking; it is now imported lazily, only when the JAX backend is actually resolved (backend="jax"). The array/scalar type aliases start NumPy-only and are broadened in-place to also admit jax.Array the first time the JAX backend is used (annotations are deferred via from __future__ import annotations, so typeguard resolves them against the broadened aliases at call time). This keeps NumPy-only import paths — and downstream packages that use only the NumPy backend — free of the JAX import cost. The numpy and jax backends are behavior- and result-identical to 3.2.0.

[3.2.0] - 2026-06-29

Added

  • Optional JAX backend for CVMatrix via backend="jax" (a Literal["numpy", "jax"]; default remains "numpy"). All array operations are routed through a resolved array namespace (numpy or jax.numpy), so the per-fold training_XTX/training_XTY/training_XTX_XTY/training_statistics computations can be traced by jax.jit and batched with jax.vmap on CPU/GPU/TPU. Install with cvmatrix[jax].
  • Backend-neutral, trace-safe standard-deviation clamping (maximum/where instead of boolean-mask assignment). Degenerate-fold validation (ValueErrors) still fires for eager (numpy or concrete jax) execution; under jax.jit/jax.vmap tracing the check is deferred to a host-side pre-flight by the caller.

Changed

  • Switched build and development tooling from Poetry to uv: the project now uses PEP 621 [project] metadata, the hatchling build backend, and a PEP 735 [dependency-groups] dev group (poetry.lock is replaced by uv.lock). The published package, its runtime dependencies, and the jax extra are unchanged.

Notes

  • The numpy backend is byte-identical to previous releases (verified against the existing test suite) and performance-neutral.

[3.1.6] - 2025

Changed

  • CVMatrix and Partitioner can now be imported with from cvmatrix import CVMatrix, Partitioner or import cvmatrix followed by cvmatrix.CVMatrix and cvmatrix.Partitioner.

[3.0.0] - 2025

Added

  • Partitioner class for managing cross-validation folds

Changed

  • CVMatrix no longer internally stores a dictionairy of validation indices. This responsibility is offloaded to Partitioner so CVMatrix can be pickled more efficiently when used in a multiprocessing context such as by the ikpls package [1].

[2.0.0] - 2024

Added

  • Weighted matrix products $\mathbf{X}^{\mathbf{T}}\mathbf{W}\mathbf{Y}$ support without increasing time or space complexity
  • Generalization of the fast cross-validation algorithms by Engstrøm and Jensen to handle weighted cases
  • Support for weighted centering and weighted scaling based on training set statistical moments
  • All 16 (12 unique) combinations of weighted column-wise centering and scaling for X and Y matrices

Changed

  • Extended algorithms to correctly handle weighted cases while maintaining efficiency
  • Enhanced CVMatrix class to accept sample weights

Notes

  • The weighted extension maintains the same computational complexity as the unweighted algorithms
  • Formal description of the weighted generalization to be announced in upcoming publication

[1.0.0] - Initial Release

Added

  • Implementation of fast cross-validation algorithms by Engstrøm and Jensen [2] for computation of training set $\mathbf{X}^{\mathbf{T}}\mathbf{X}$ and $\mathbf{X}^{\mathbf{T}}\mathbf{Y}$
  • CVMatrix class for efficient kernel matrix computation in cross-validation settings
  • Support for arbitrary row-wise preprocessing of data
  • Support for column-wise centering and scaling of X and Y based on training set statistics
  • Methods for computing:
    • training_XTX_XTY(): Both training set kernel matrices and statistics
    • training_XTX(): Training set $\mathbf{X}^{\mathbf{T}}\mathbf{X}$ and X statistics
    • training_XTY(): Training set $\mathbf{X}^{\mathbf{T}}\mathbf{Y}$ and statistics
  • Comprehensive test suite and benchmarks
  • Documentation at cvmatrix.readthedocs.io

Features

  • Correct handling of column-wise centering and scaling without data leakage
  • Efficient computation of training set moments (means and standard deviations)
  • Support for all combinations of centering and scaling options for X and Y
  • Significantly faster than naive implementations for cross-validation scenarios

References

  1. Engstrøm, O.-C. G. and Jensen, M. H. (2025). Fast Partition-Based Cross-Validation With Centering and Scaling for $\mathbf{X}^\mathbf{T}\mathbf{X}$ and $\mathbf{X}^\mathbf{T}\mathbf{Y}$
  2. IKPLS. Fast CPU and GPU Python implementations of Improved Kernel Partial Least Squares (PLS) by Dayal and MacGregor (1997) and Fast Partition-Based Cross-Validation With Centering and Scaling for XTX and XTY by Engstrøm and Jensen (2025). This package also includes options to use sample weights for PLS modeling.