Changelog#

All notable changes to this project will be documented here.

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

v0.8.0 - 2025-08-02#

Additions

  • Add support for numpy == 2.2, numpy == 2.3, scipy == 1.15, and scipy == 1.16.

  • Add the hyperparameter residual results from retraining configurations for ResNet18 on ImageNet with different random seeds.

    • Add the data itself: data/resnet/resnet18_residual.results.jsonl.

    • Update its README to describe the data: data/resnet/README.md.

  • Add the hyperparameter tuning results from tuning AlexNet for ImageNet via random search.

    • Add the data itself: data/alexnet/alexnet_tuning.results.jsonl.

    • Add a license for the data: data/alexnet/LICENSE.

    • Add a README for the data: data/alexnet/README.md.

  • Add the hyperparameter tuning results from tuning ConvNeXt for ImageNet via random search.

    • Add the data itself: data/convnext/convnext-tiny_tuning.results.jsonl.

    • Add a license for the data: data/convnext/LICENSE.

    • Add a README for the data: data/convnext/README.md.

  • Add the hyperparameter ablation results from fixing successively more hyperparameters when tuning ResNet18 on ImageNet via random search.

    • Add the data itself: data/resnet/resnet18_ablation.results.jsonl.

    • Update its README to describe the data: data/resnet/README.md.

  • Add the hyperparameter tuning results from tuning small Llama models for SlimPajama via random search.

    • Add the data itself: data/llama/llama-33m_tuning.results.jsonl.

    • Add a license for the data: data/llama/LICENSE.

    • Add a README for the data: data/llama/README.md.

  • Add the hyperparameter residual results from retraining configurations for small Llama models on SlimPajama with different random seeds.

    • Add the data itself: data/llama/llama-33m_residual.results.jsonl.

    • Update its README to describe the data: data/llama/README.md.

  • Make opda.parametric.QuadraticDistribution.fit and opda.parametric.NoisyQuadraticDistribution.fit accept infinite observations in ys if those observations are censored by limits.

  • Add a generator parameter to experiments.analytic.get_approximation_parameters so that it’s possible to make it deterministic.

  • Add scikit-learn >= 1.5.2 as part of the experiments optional dependencies.

Changes

  • Rename hyperparameter tuning results files:

    • Rename data/deberta/deberta-base.results.jsonl to data/deberta/deberta-base_tuning.results.jsonl.

    • Rename data/deberta/deberta-v3-base.results.jsonl to data/deberta/deberta-v3-base_tuning.results.jsonl.

    • Rename data/resnet/resnet18_scaling.results.jsonl to data/resnet/resnet18_tuning.results.jsonl.

  • Upgrade packaging tools to setuptools >= 77.0, build == 1.2.2, and twine == 6.1.0.

  • Specify license using an SPDX license expression for project.license in pyproject.toml instead of the classifier: License :: OSI Approved :: Apache Software License.

  • Parallelize the testpackage CI job over all NumPy and SciPy versions in addition to Python versions.

  • Automatically adjust the number of histogram bins used in experiments.visualization.plot_pdf based on the sample size.

  • Standardize the figure sizes for plotting functions in experiments.visualization.

  • Improve the notation used in the notebooks and experiments.visualization.

Removals

  • Drop support for numpy == 1.23, numpy == 1.24, numpy == 1.25, and scipy == 1.10.

  • Drop support for Python 3.8.

  • Remove experiments.visualization.plot_random_search.

Fixes

  • Fix the support policy check in continuous integration:

    • Determine currently supported python versions more correctly.

    • Mention updating the setup documentation when versions change.

  • Suppress expected deprecation warnings that autograd fires when installed with certain versions of numpy.

  • Make opda.parametric.NoisyQuadraticDistribution.fit slower but also much more robust.

  • Fix flakiness in various tests.

  • Fix an error thrown by opda.parametric.QuadraticDistribution.fit and opda.parametric.NoisyQuadraticDistribution.fit when all ys are negative.

  • Fix a race condition when downloading and testing a new release in the “release” nox session.

Documentation

  • Improve the README for the DeBERTa hyperparameter tuning results (data/deberta/README.md).

  • Fix examples in docstrings for opda.parametric.QuadraticDistribution.fit and opda.parametric.NoisyQuadraticDistribution.fit that constrain the parameters using bounds on the validation score.

  • Add Jupyter notebooks developing, describing, and analyzing the parametric analysis:

    • Parametric Analysis

    • Validating the Parametric Analysis in Practice

    • Analyzing Variation Due to Random Seeds

    • Generalizing the Parametric Analysis Across Architectures

  • Begin linking to the changelog rather than copying it in release descriptions on GitHub.

v0.7.0 - 2024-11-11#

Additions

  • Add the hyperparameter tuning scaling results from tuning ResNet18 for ImageNet via random search.

    • Add the data itself: data/resnet/resnet18_scaling.results.jsonl.

    • Add a license for the data: data/resnet/LICENSE.

    • Add a README for the data: data/resnet/README.md.

  • Add support for scipy == 1.14 and numpy == 2.1.

  • Add the C_MIN and C_MAX class attributes to opda.parametric.QuadraticDistribution and opda.parametric.NoisyQuadraticDistribution.

  • Add opda.utils.normal_ppf, a fast function for computing the standard normal’s PPF (quantile function).

  • Check that c is finite when validating the arguments to opda.parametric.QuadraticDistribution and opda.parametric.NoisyQuadraticDistribution.

  • Add fit methods to the parametric distributions: opda.parametric.QuadraticDistribution.fit and opda.parametric.NoisyQuadraticDistribution.fit.

  • Add support for Python 3.13.

Changes

  • Simplify the implementation of opda.parametric.NoisyQuadraticDistribution by making scale the standard deviation of the corresponding normal random variable.

  • Upgrade packaging tools to build == 1.2.1 and twine == 5.1.1.

  • Skip some tests for opda.utils.beta_equal_tailed_interval, opda.utils.beta_highest_density_interval, and opda.utils.beta_equal_tailed_coverage when they fail due to an issue in scipy == 1.14.0 on Linux that causes spurious NaN values.

  • Use the point mass, noiseless, and normal approximations to speed up the ppf method of opda.parametric.NoisyQuadraticDistribution.

  • Disable lint rules:

    • B023: “Function definition does not bind loop variable {name}”.

    • PLW2901: “Outer {outer_kind} variable {name} overwritten by inner {inner_kind} target”.

    • SIM102: “Use a single if statement instead of nested if statements”.

  • Restrict the range of values for c supported by opda.parametric.QuadraticDistribution to 1 to 10.

Removals

  • Drop support for numpy == 1.21, numpy == 1.22, scipy == 1.8, and scipy == 1.9.

  • Remove the estimate_initial_parameters_and_bounds method from opda.parametric.QuadraticDistribution.

Fixes

  • Update tests for the __repr__ methods of opda.nonparametric.EmpiricalDistribution, opda.parametric.QuadraticDistribution, and opda.parametric.NoisyQuadraticDistribution so that they’re compatible with numpy == 2.0.

  • Fix flakiness in the test: NoisyQuadraticDistributionTestCase.test_ppf_is_inverse_of_cdf.

  • Fix the quantile_tuning_curve methods of opda.parametric.QuadraticDistribution and opda.parametric.NoisyQuadratricDistribution which gave incorrect output when q wasn’t equal to 0.5 and either minimize was True or minimize was None and convex was True.

Documentation

  • Correct the quantile function’s definition in the docstrings for the ppf methods of opda.parametric.QuadraticDistribution and opda.parametric.NoisyQuadraticDistribution.

  • Make docstring formatting more consistent:

    • Always use “Returns” and never “Return” for the docstring section header.

    • Always document optional types with “or None” coming at the end of the type description.

    • Always document each returned value with a separate subheading when a function or method returns multiple values.

  • Remove default values from parameters’ type descriptions.

v0.6.1 - 2024-04-03#

Additions

  • Add support for numpy == 2.0 and scipy == 1.13.

Changes

  • Disable lint rule PLR1714: “Consider merging multiple comparisons”.

Documentation

  • Split the Jupyter notebook Evaluating DeBERTaV3 with the Nonparametric Analysis into several smaller more focused notebooks:

    • Evaluating DeBERTaV3 with the Nonparametric Analysis

    • Choosing a Sample Size for the Nonparametric Analysis

    • Demonstrating the Exact Coverage of the Nonparametric Analysis

    • Studying Ablations of the Nonparametric Analysis

  • Add a section about how to analyze a hyperparameter (the number of epochs) to the Jupyter notebook Evaluating DeBERTaV3 with the Nonparametric Analysis.

  • In the Examples doc (docs/tutorial/examples.rst), improve the code, wording, and title of the model comparison example (previously titled Compare Models’ Tuning Curves, now titled Compare Models).

  • In the Examples doc (docs/tutorial/examples.rst), expand the Compare Models example with discussion on how to compare models with different training costs.

  • Fix incorrect markup in the Examples doc (docs/tutorial/examples.rst).

  • Add a new example showing how to analyze a hyperparameter in the Examples doc (docs/tutorial/examples.rst).

  • Add plots of DeBERTa and DeBERTaV3’s tuning curves with confidence bands using large sample sizes (1,024) to the Jupyter notebook Evaluating DeBERTaV3 with the Nonparametric Analysis.

v0.6.0 - 2024-03-04#

Additions

  • Add opda.approximation, a module for approximation-theoretic operations, with the following functions:

    • opda.approximation.lagrange_interpolate: Interpolate points with a polynomial.

    • opda.approximation.remez: Identify the reference corresponding to the minimax polynomial approximation of a function.

    • opda.approximation.minimax_polynomial_approximation: Evaluate the minimax polynomial approximation of a function.

    • opda.approximation.minimax_polynomial_coefficients: Compute the coefficients of the minimax polynomial approximation of a function.

    • opda.approximation.piecewise_polynomial_knots: Find the knots for the minimax piecewise polynomial approximation of a function.

  • Add opda.exceptions.NumericalError, an exception for numerical issues.

  • Add more tests for opda.parametric.QuadraticDistribution.

  • Add mean and variance attributes to opda.nonparametric.EmpiricalDistribution.

  • Add mean and variance attributes to opda.parametric.QuadraticDistribution.

  • Add opda.utils.normal_pdf, a fast function for computing the standard normal’s PDF.

  • Add opda.utils.normal_cdf, a fast function for computing the standard normal’s CDF.

  • Add opda.exceptions.IntegrationError, an exception for integration issues.

  • Add opda.parametric.NoisyQuadraticDistribution, a probability distribution representing a quadratic random variable plus normal noise.

  • Increase argument validation in opda.utils.dkw_epsilon.

  • Add more test cases for opda.utils.dkw_epsilon.

  • Validate that lightness is between 0 and 1 (inclusive) in experiments.visualization.color_with_lightness.

  • Validate that arguments are finite floats where appropriate.

Changes

  • Reparametrize parametric.QuadraticDistribution so c is the effective number of hyperparameters instead of half the number.

  • Completely disable the eradicate (ERA) lint rules.

  • Enable "py" as the primary domain in the documentation.

  • Always use numpy’s numeric types for scalar class attributes, instead of Python’s native numeric types.

  • Improve tests for parametric.QuadraticDistribution, making them more thorough, robust, and avoiding re-running redundant test cases.

  • Update the tests for parametric.QuadraticDistribution to cover the case when a == b.

  • Move the source repository from github.com/nalourie/opda to github.com/nicholaslourie/opda, and move the docs from nalourie.github.io/opda to nicholaslourie.github.io/opda. Update the project URLs in pyproject.toml and all the links throughout the repository to reflect these changes.

  • Require fraction is greater than 0 in opda.parametric.QuadraticDistribution.estimate_initial_parameters_and_bounds.

  • Throw an error if fraction is too small and thus causes opda.parametric.QuadraticDistribution.estimate_initial_parameters_and_bounds to try and form an estimate from an empty list.

Fixes

  • Fix parametric.QuadraticDistribution (the .pdf, .cdf, and .estimate_initial_parameters_and_bounds methods) for the case when a == b, in which case the distribution is an atom (point mass).

  • Fix opda.parametric.QuadraticDistribution.estimate_initial_parameters_and_bounds when convex is False and fraction is small enough so that the estimate should be based on an empty list. In this case, the method incorrectly uses all of ys. Instead, throw an error saying that fraction is too small (as it produces an empty list).

  • Avoid throwing an unnecessary warning in opda.utils.dkw_epsilon when confidence is 1.

Documentation

  • Improve the docstring for experiments.analytic.get_approximation_parameters.

  • Use inline math markup in docstrings.

  • Fix the equation in the docstring for opda.parametric.QuadraticDistribution.ppf. The infimum that defines the quantile function has as its domain the interval from a to b, not the entire real line.

  • Add “See Also” and “Notes” sections to the docstring for opda.parametric.QuadraticDistribution, matching the newly added docstring for opda.parametric.NoisyQuadraticDistribution.

  • Update all links to use github.com/nicholaslourie and nicholaslourie.github.io/opda in place of github.com/nalourie and nalourie.github.io/opda.

  • Document stricter dependent type constraints (e.g., non-negativity, finiteness) for function and method inputs and outputs.

  • Document range constraints for inputs and outputs more precisely and consistently (e.g., q is a float from 0 to 1 inclusive).

  • Improve the docstring for experiments.analytic.ellipse_volume.

  • Fix docstrings across the code base in order to consistently document when a value can take on either scalar (e.g., float) or array (e.g., array of floats) values.

v0.5.0 - 2024-01-15#

This version is the first uploaded to PyPI and available via pip!

Additions

  • Add a continuous integration job to ensure every pull request updates the changelog.

  • In the continuous integration job for building the packages, add a step to list the packages’ contents.

  • Add the “release” nox session for making new releases to PyPI.

Changes

  • Upgrade the development dependencies.

  • Upgrade the Development Status PyPI classifier for opda from 3 - Alpha to 4 - Beta.

Fixes

  • Fix flakiness in the test: EmpiricalDistributionTestCase.test_average_tuning_curve.

Documentation

  • Pin links to the source on GitHub to the commit that builds the documentation.

  • Move development documentation into the “Contributing” section of the sidebar and URL tree.

  • Omit from the documentation’s sidebar any project URLs that link to the documentation.

  • Add an announcement banner to the documentation when it’s built for an unreleased version.

  • Add a changelog (CHANGELOG.rst).

  • Document the project’s various conventions in the development docs.

  • Add the “Release” doc describing the release process.

  • Update the docs to suggest installing opda from PyPI rather than the source for regular usage.

v0.4.0 - 2024-01-10#

Additions

  • Add the package optional dependencies.

  • Add a build for “distribution” as opposed to “local” use. The distribution package contains only the opda library and not experiments.

    • Add a nox session for building the distribution package.

    • Add a continuous integration job to build the package and store it as an artifact on each pull request.

    • Add a continuous integration job to test the distribution package against all combinations of supported versions of major dependencies.

Changes

  • Increase retention for documentation build artifacts from 60 to 90 days in continuous integration.

  • Prune each set of optional dependencies.

  • Rename the tests optional dependencies to test.

  • Split the test session in nox into test, for testing the local project, and testpackage, for testing distribution packages.

  • In continuous integration, only test the local build against default versions of major dependencies, since we now build and test the distribution package against all combinations of supported versions.

Documentation

  • Document how to build and test the distribution package.

v0.3.0 - 2024-01-07#

Additions

  • Extend nonparametric.QuadraticDistribution.sample and nonparametric.EmpiricalDistribution.sample to return a scalar when size=None, and make it the default argument.

  • Add documentation builds via Sphinx:

    • Create a Sphinx setup for building the documentation.

    • Add tutorial-style documentation for users.

    • Add development documentation.

    • Automatically generate API reference documentation.

  • Add a GitHub Actions workflow for building and publishing the documentation to GitHub Pages.

  • Make tests backwards compatible with numpy >= 1.21.

  • Adjust package dependency requirements to allow numpy >= 1.21 and scipy >= 1.8.

  • Add ci optional dependencies for continuous integration.

  • Add nox for automating development tasks, testing against all supported major dependencies, and continuous integration.

  • Add a GitHub Actions workflow for continuous integration. Run it on each pull request as well as every calendar quarter. Use the workflow to:

    • Check opda’s major dependency versions are up-to-date.

    • Lint the project.

    • Build and test the documentation.

    • Test the project against all combinations of supported versions of major dependencies.

Changes

  • Always return scalars rather than 0 dimensional arrays from methods (nonparametric.EmpiricalDistribution.pmf and parametric.QuadraticDistribution.pdf).

  • Explicitly test that all methods of nonparametric.EmpiricalDistribution and parametric.QuadraticDistribution return scalars rather than 0D arrays.

  • Configure pytest to always use a non-interactive backend for matplotlib.

  • Update the project URLs in packaging.

  • Split out the experiments package’s dependencies as optional dependencies.

Fixes

  • Include src/experiments/default.mplstyle in the package data for the experiments package so the style can be used from non-editable installs.

  • Make tests more robust to changes in rounding errors across environments by replacing some equality checks with near equality.

Documentation

  • Remove broken references to the sections of numpy-style docstrings. Standard tooling doesn’t make these sections linkable.

  • Fix errors in the docstrings’ markup.

  • Use cross-references in the docs wherever possible and appropriate.

  • Use proper markup for citations.

  • Change the language from bash to console in code blocks.

  • Improve the modules’ docstrings.

  • Rewrite README.rst, adding a “Quickstart” section and moving much of the old content into new tutorial-style documentation built with Sphinx.

  • Document how to build and test the documentation.

  • Document how to setup and use nox for common development tasks.

v0.2.0 - 2023-12-16#

Additions

  • Add backwards compatibility for Python 3.8.

  • Add pyproject.toml for building the project, replacing the setup.py based build.

  • Add and increase argument validation in functions and methods.

  • Add the --all-levels pytest flag for running all tests.

  • Add new tests for nonparametric.EmpiricalDistribution and parametric.QuadraticDistribution.

  • Give all tuning curve methods a new parameter, minimize, for computing minimizing hyperparameter tuning curves.

    • nonparametric.EmpiricalDistribution methods: quantile_tuning_curve, average_tuning_curve, naive_tuning_curve, v_tuning_curve, and u_tuning_curve.

    • parametric.QuadraticDistribution methods: quantile_tuning_curve, and average_tuning_curve.

  • Add __repr__, __str__, and __eq__ methods to nonparamatric.EmpiricalDistribution and parametric.QuadraticDistribution.

  • Add a generator parameter to set the random seed in functions and methods using randomness (experiments.simulation.Simulation.run, experiments.visualization.plot_random_search, nonparametric.EmpiricalDistribution.confidence_bands, nonparametric.EmpiricalDistribution.sample, and parametric.QuadraticDistribution.sample).

  • Add the opda.random module to migrate off of numpy’s legacy API for random numbers while still enabling control of opda’s global random state via opda.random.set_seed.

  • Add the RandomTestCase class for making tests using randomness reproducible.

  • Configure ruff for linting the project.

Changes

  • Require pytest >= 6 for running tests.

  • Configure pytest to use the tests/ test path.

  • Use Private :: Do Not Upload classifier to prevent the package from being uploaded to PyPI.

  • Speed up coverage tests for nonparametric.EmpiricalDistribution.confidence_bands.

  • Rename optional dependencies from dev to tests.

  • Standardize the error messages for violating argument type constraints.

  • Expand existing tests to cover more cases for EmpiricalDistribution and QuadraticDistribution.

  • Rename exceptions.OptimizationException to exceptions.OptimizationError.

  • Use TypeError in place of ValueError for type errors.

  • Across all functions and methods, standardize which parameters are keyword-only. Reserve keyword-only status for rarely used arguments, such as implementation details like optimization tolerances.

  • Disallow None as an argument for the a and b parameters of nonparametric.EmpiricalDistribution.

Fixes

  • Fix flakiness in various tests.

  • Ensure utils.beta_highest_density_interval always returns an interval containing the mode, even for very small intervals.

  • Fix bug in nonparametric.EmpiricalDistribution.confidence_bands that caused coverage to be too high, especially given small samples.

  • Improve coverage tests for nonparametric.EmpiricalDistribution.confidence_bands so that they’re more sensitive and explicitly test small sample sizes.

  • Prevent warnings during expected use of various methods of QuadraticDistribution.

  • Suppress expected warnings in tests.

  • Fix parametric.QuadraticDistribution.quantile_tuning_curve which would throw an exception when the instance had convex=True.

  • Fix tests for parametric.QuadraticDistribution so that they actually check all intended cases.

Removals

  • Remove the setup.py based build and associated files (setup.py, setup.cfg, MANIFEST.in, and requirements.txt), replacing it with pyproject.toml.

Documentation

  • Add sections and improve markup in README.rst.

  • Add links to and citations for Show Your Work with Confidence.

  • Add sections, update content, and improve markup in existing docstrings.

  • Document development tools for the project.

  • Begin running doctests on all documentation.

    • Document how to run doctests in README.rst.

    • Set the random seed in documentation examples to make them testable.

    • Fix errors in examples discovered via doctests.

  • Document pip version requirements for editable installs in README.rst.

  • Document type constraints (e.g., non-negative integers as opposed to integers) in functions and methods’ docstrings.

  • Document the atol parameter of utils.beta_highest_density_interval and utils.highest_density_coverage.

v0.1.0 - 2023-11-14#

Additions

  • Initial release.