Skip to content

Latest commit

 

History

History
292 lines (224 loc) · 15.8 KB

CHANGELOG.md

File metadata and controls

292 lines (224 loc) · 15.8 KB

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.

Unreleased

Added

  • Contributor-facing features:
    • Automated testing of the built package before release (gchq#854)

Fixed

Changed

Removed

Deprecated

0.3.1

Added

  • Added an analytical test for RPCholesky, and associated documentation in tests.unit.test_solvers. (gchq#822)
  • Added a unit test for RPCholesky to check whether the coreset has duplicates. (gchq#836)

Fixed

  • RPCholesky.reduce in coreax.solvers.coresubset now computes the iteration step correctly. (gchq#825)
  • RPCholesky.reduce in coreax.solvers.coresubset now does not produce duplicate points in the coreset.(gchq#836)
  • Fixed the example examples.david_map_reduce_weighted to prevent errors when downsampling is enabled, and to make it run faster. (gchq#821)
  • Build includes sub-packages. (gchq#845)

Changed

  • Test dependency from opencv-python to opencv-python-headless. (gchq#848)
  • Updated installation instructions in README. (gchq#848)

0.3.0 - [YANKED]

Yanked due to build failure.

Added

  • Added Kernel Stein Discrepancy divergence in coreax.metrics.KSD.(gchq#659)
  • Added the coreax.solvers.recombination module, which provides the following new solvers:
    • RecombinationSolver: an abstract base class for recombination solvers.
    • CaratheodoryRecombination: a simple deterministic approach to solving recombination problems.
    • TreeRecombination: an advanced deterministic approach that utilises CaratheodoryRecombination, but is faster for solving all but the smallest recombination problems.(gchq#504)
  • Added supervised coreset construction algorithm in coreax.solvers.GreedyKernelPoints.(gchq#686)
  • Added coreax.kernels.PowerKernel to replace repeated calls of coreax.kernels.ProductKernel within the ** magic method of coreax.kernel.ScalarValuedKernel.(gchq#708)
  • Added scalar-valued kernel functions coreax.kernels.PoissonKernel and coreax.kernels.MaternKernel.(#742)
  • Added progress_bar attribute to coreax.score_matching.SlicedScoreMatching to enable or disable tqdm progress bar terminal output. Defaults to disabled (False).(gchq#761)
  • Added analytical tests for kernel herding, and associated documentation in tests.unit.test_solvers.(gchq#794)
  • Added CI workflow for performance testing.
  • Added array dimensions to type annotations using jaxtyping.(gchq#746)
  • Added integration test for coreax.solver.recombination.TreeRecombination.(gchq#798)

Fixed

  • Fixed MapReduce in coreax.solvers.composite.py to keep track of the indices.(gchq#779)
  • Fixed negative weights on coreax.weights.qp.(gchq#698)

Changed

  • Refactored coreax.inverses.py functionality into coreax.least_squares.py:
    • coreax.inverses.RegularisedInverseApproximator replaced by coreax.least_squares.RegularisedLeastSquaresSolver.
    • coreax.inverses.LeastSquaresApproximator replaced by coreax.least_squares.MinimalEuclideanNormSolver.
    • coreax.inverses.RandomisedEigendecompositionApproximator replaced by coreax.least_squares.RandomisedEigendecompositionSolver.(gchq#700)
  • Refactoring of coreax.kernel.py into coreax.kernels sub-package:
    • kernels.util.py holds utility functions relating to kernels e.g. median_heuristic.
    • kernels.base.py holds the base kernel class ScalarValuedKernel (renamed from Kernel), as well as the base composite classes UniCompositeKernel (renamed from CompositeKernel), DuoCompositeKernel (renamed from PairedKernel) and the derived duo-composite kernels AdditiveKernel and ProductKernel
    • coreax.kernels.scalar_valued.py holds all currently implemented scalar valued kernels e.g. SquaredExponentialKernel. (gchq#708)
  • Refactored coreax.weights.py to make weight solvers generic on data type.(gchq#709)

Removed

  • coreax.weights.MMD - deprecated alias for coreax.weights.MMDWeightsOptimiser; deprecated since version 0.2.0.(gchq#784)
  • coreax.weights.SBQ - deprecated alias for coreax.weights.SBQWeightsOptimiser; deprecated since version 0.2.0.(gchq#784)
  • coreax.util.squared_distance_pairwise - deprecated alias for coreax.util.pairwise(squared_distance); deprecated since version 0.2.0.(gchq#784)
  • coreax.util.pairwise_difference - deprecated alias for coreax.util.pairwise(difference); deprecated since version 0.2.0.(gchq#784)

Deprecated

  • All uses of coreax.kernel.Kernel should be replaced with coreax.kernels.base.ScalarValuedKernel.(gchq#708)
  • All uses of coreax.kernel.UniCompositeKernel should be replaced with coreax.kernels.base.CompositeKernel.(gchq#708)
  • All uses of coreax.kernel.PairedKernel should be replaced with coreax.kernels.base.DuoCompositeKernel.(gchq#708)
  • All uses of coreax.kernel.AdditiveKernel should be replaced with coreax.kernels.base.AdditiveKernel.(gchq#708)
  • All uses of coreax.kernel.ProductKernel should be replaced with coreax.kernels.base.ProductKernel.(gchq#708)
  • All uses of coreax.kernel.LinearKernel should be replaced with coreax.kernels.scalar_valued.LinearKernel.(gchq#708)
  • All uses of coreax.kernel.PolynomialKernel should be replaced with coreax.kernels.scalar_valued.PolynomialKernel.(gchq#708)
  • All uses of coreax.kernel.SquaredExponentialKernel should be replaced with coreax.kernels.scalar_valued.SquaredExponentialKernel.(gchq#708)
  • All uses of coreax.kernel.ExponentialKernel should be replaced with coreax.kernels.scalar_valued.ExponentialKernel.(gchq#708)
  • All uses of coreax.kernel.RationalQuadraticKernel should be replaced with coreax.kernels.scalar_valued.RationalQuadraticKernel.(gchq#708)
  • All uses of coreax.kernel.PeriodicKernel should be replaced with coreax.kernels.scalar_valued.PeriodicKernel.(gchq#708)
  • All uses of coreax.kernel.LocallyPeriodicKernel should be replaced with coreax.kernels.scalar_valued.LocallyPeriodicKernel.(gchq#708)
  • All uses of coreax.kernel.LaplacianKernel should be replaced with coreax.kernels.scalar_valued.LaplacianKernel.(gchq#708)
  • All uses of coreax.kernel.SteinKernel should be replaced with coreax.kernels.scalar_valued.SteinKernel.(gchq#708)
  • All uses of coreax.kernel.PCIMQKernel should be replaced with coreax.kernels.scalar_valued.PCIMQKernel.(gchq#708)
  • All uses of coreax.util.median_heuristic should be replaced with coreax.kernels.util.median_heuristic.(gchq#708)

0.2.1

Added

  • Pyright to development tools (code does not pass yet)

Fixed

  • Nitpicks in documentation build
  • Incorrect package version number

Changed

  • Augmented unroll parameter to be consistent with block size in MMD metric

0.2.0

Added

  • Badge to README to show code coverage percentage.
  • Support for Python 3.12.
  • Added a deterministic, iterative, and greedy coreset algorithm which targets the Kernelised Stein Discrepancy via coreax.solvers.coresubset.SteinThinning.
  • Added a stochastic, iterative, and greedy coreset algorithm which approximates the Gramian of a given kernel function via coreax.solvers.coresubset.RPCholesky.
  • Added coreax.util.sample_batch_indices that allows one to sample an array of indices for batching.
  • Added kernel classes coreax.kernel.AdditiveKernel and coreax.kernel.ProductKernel that allow for arbitrary composition of positive semi-definite kernels to produce new positive semi-definite kernels.
  • Added additional kernel functions: coreax.kernel.Linear, coreax.kernel.Polynomial, coreax.kernel.RationalQuadratic, coreax.kernel.Periodic, coreax.kernel.LocallyPeriodic.
  • Added capability to approximate the inverses of arrays via least-squares (coreax.inverses.LeastSquaresApproximator) or randomised eigendecomposition (coreax.inverses.RandomisedEigendecompositionApproximator) all inheriting from coreax.inverses.RegularisedInverseApproximator,
  • Refactor of package to a functional style to allow for JIT-compilation of the codebase in the largest possible scope:
    • Added data classes coreax.data.Data and coreax.data.SupervisedData that draw distinction between supervised and unsupervised datasets, and handle weighted data. Replaces coreax.data.DataReader and coreax.data.ArrayData.
    • Added coreax.solvers.base.Solver to replace functionality in coreax.refine.py, coreax.coresubset.py and coreax.reduction.py. In particular, coreax.solvers.base.CoresubsetSolver parents coresubset algorithms, coreax.solvers.base.RefinementSolver parents coresubset algorithms which support refinement post-reduction, coreax.solvers.base.ExplicitSizeSolver parents all coreset algorithms which return a coreset of a specific size.
    • coreax.reduction.MapReduce functionality moved to coreax.solvers.composite.MapReduce, now JIT-compilable via promise described in coreax.solvers.base.PaddingInvariantSolver.
    • Moved all coresubset algorithms in coreax.coresubset.py to coreax.solvers.coresubset.py.
    • All coreset algorithms now return a coreax.coreset.Coreset rather than modifying a coreax.reduction.Coreset in-place.
  • Use Equinox instead of manually constructing pytrees.

Fixed

  • Wording improvements in README.
  • Documentation now builds without warnings.
  • GitHub workflow runs automatically after Pre-commit autoupdate.

Changed

  • Documentation has been rearranged.
  • Renamed coreax.weights.MMD to coreax.weights.MMDWeightsOptimiser and added deprecation warning.
  • Renamed coreax.weights.SBQ to coreax.weights.SBQWeightsOptimiser and added deprecation warning.
  • requirements-*.txt will no longer be updated frequently, thereby providing stable versions.
  • Single requirements files covering all supported Python versions.
  • All references to kernel_matrix_row_{sum,mean} have been replaced with Gramian row-mean.
  • coreax.networks.ScoreNetwork now allows the user to specify number of hidden layers.
  • Classes in weights.py and score_matching.py now inherit from equinox.Module.
  • Performance tests replaced by jit_variants tests, which checks whether a function has been compiled for reuse.
  • Replace some pygrep-hooks with ruff equivalents.
  • Use Pytest fixtures instead of unittest style.

Removed

  • Bash script to run integration tests has been removed. pytest tests/integration should now work as expected.
  • Tests for coreax.kernels.Kernel.{calculate, update}_kernel_matrix_row_sum.
  • coreax.util.KernelComputeType; use Callable[[ArrayLike, ArrayLike], Array] instead.
  • coreax.kernels.Kernel.calculate_kernel_matrix_row_{sum,mean}; use coreax.kernels.Kernel.gramian_row_mean.
  • coreax.kernels.Kernel.updated_kernel_matrix_row_sum; use coreax.kernels.Kernel.gramian_row_mean if possible.
  • coreax.data.DataReader and coreax.data.ArrayData; use coreax.data.Data and coreax.data.SupervisedData.
  • coreax.refine.py and coreax.coresubset.py removed; use coreax.solvers.base.RefinementSolver or coreax.solvers.base.CoresubsetSolver to define coreset algorithms in coreax.solvers.coresubset.
  • coreax.reduction removed, use coreax.solvers.base.ExplicitSizeSolver in place of coreax.reduction.SizeReduce and coreax.solvers.composite.MapReduce in place of coreax.reduction.MapReduce. Use coreax.coreset.Coreset and coreax.coreset.Coresubset in place of coreax.reduction.Coreset.

Deprecated

  • All uses of coreax.weights.MMD should be replaced with coreax.weights.MMDWeightsOptimiser.
  • All uses of coreax.weights.SBQ should be replaced with coreax.weights.SBQWeightsOptimiser.
  • All uses of coreax.util.squared_distance_pairwise should be replaced with coreax.util.pairwise(squared_distance).
  • All uses of coreax.util.pairwise_difference should be replaced with coreax.util.pairwise(difference).

0.1.0

Added

  • Base Coreax package using Object-Oriented Programming incorporating:
    • coreset methods: kernel herding, random sample
    • reduction strategies: size reduce, map reduce
    • kernels: squared exponential, Laplacian, PCIMQ, Stein
    • refinement: regular, reverse, random
    • metrics: MMD
    • approximations of kernel matrix row sum mean: random, ANNchor, Nystrom
    • weights optimisers: SBQ, MMD
    • score matching: sliced score matching, kernel density estimation
    • I/O: array data not requiring any preprocessing
  • Near-complete unit test coverage.
  • Example scripts for coreset generation, which may be called as integration tests.
  • Bash script to run integration tests in sequence to avoid Jax errors.
  • Detailed documentation for the Coreax package published to Read the Docs.
  • README.md including an overview of what coresets are, setup instructions, a how-to guide, example applications and an overview of features coming soon.
  • Support for Python 3.9-3.11.
  • Project configuration and dependencies through pyproject.toml.
  • Requirements files providing a pinned set of dependencies that are known to work for each supported Python version.
  • Mark Coreax as typed.
  • This changelog to make it easier for users and contributors to see precisely what notable changes have been made between each release of the project.
  • FAQ.md to address any commonly asked questions.
  • Contributor guidelines, code of conduct, license and security policy.
  • Git configuration.
  • GitHub Actions to run unit tests on Windows, macOS and Ubuntu for supported Python versions.
  • Pre-commit checks to run the following, also checked by GitHub Actions:
    • black
    • isort
    • pylint
    • cspell spell check with custom dictionaries for library names, people and miscellaneous
    • pyroma
    • pydocstyle
    • assorted file format and encoding checks

Deprecated

  • Look-before-you-leap validation of all input to public functions