v0.10.0

v0.10.0 - 💥📚🐞🆕 New valuation interface, improved docs, new methods, breaking changes and tons of improvements

After lots of work, bug-fixing, bug-introducing, fixing again, and a good measure of bike shedding, we bring a major update putting us closer to the final APIs. The main goals of this release were to improve usability, documentation, and extensibility.

• We have added a new module pydvl.valuation. The pydvl.value module is deprecated and will be removed in the next release. The new interface allows for a more consistent and flexible way to define and use valuation methods. It also simplifies experimentation, manipulation of results and data, as well as parallelization.
• We have many improvements to the influence module including several new methods and approximations.
• The whole documentation has been improved and consolidated, with detailed method descriptions and examples. See pydvl.org.

Added

• Simple result serialization to resume computation of values PR #666
• Simple memory monitor / reporting PR #663
• New stopping criterion MaxSamples PR #661
• Introduced UtilityModel and two implementations IndicatorUtilityModel and DeepSetsUtilityModel for data utility learning PR #650
• Introduced the concept of ResultUpdater in order to allow samplers to declare the proper strategy to use by valuations PR #641
• Added Banzhaf precomputed values to some games. PR #641
• Introduced new IndexIterations, for consistent usage across all PowersetSamplers PR #641
• Added run_removal_experiment for easy removal experiments PR #636
• Refactor Classwise Shapley valuation with the interfaces and sampler architecture PR #616
• Refactor KNN Shapley values with the new interface PR #610 PR #645
• Refactor MSR Banzhaf semivalues with the new sampler architecture. PR #605 PR #641
• Refactor group-testing shapley values with new sampler architecture PR #602
• Refactor least-core data valuation methods with more supported sampling methods and consistent interface. PR #580
• Refactor Owen-Shapley valuation with new sampler architecture. Enable use of OwenSamplers with all semi-values PR #597 PR #641
• New method InverseHarmonicMeanInfluence, implementation for the paper DataInf: Efficiently Estimating Data Influence in LoRA-tuned LLMs and Diffusion Models PR #582
• Add new backend implementations for influence computation to account for block-diagonal approximations PR #582
• Extend DirectInfluence with block-diagonal and Gauss-Newton approximation PR #591
• Extend LissaInfluence with block-diagonal and Gauss-Newton approximation PR #593
• Extend NystroemSketchInfluence with block-diagonal and Gauss-Newton approximation PR #596
• Extend ArnoldiInfluence with block-diagonal and Gauss-Newton approximation PR #598
• Extend CgInfluence with block-diagonal and Gauss-Newton approximation PR #601

Fixed

• Fixed show_warnings=False not being respected in subprocesses. Introduced suppress_warninigs decorator for more flexibility PR #647 PR #662
• Fixed several bugs in diverse stopping criteria, including: iteration counts, computing completion, resetting, nested composition PR #641 PR #650
• Fixed all weights of all samplers to ensure that mix-and-matching samplers and semi-value methods always works, for all possible combinations PR #641
• Fixed a bug whereby progress bars would not report the last step and remain incomplete PR #641
• Fixed the analysis of the adult dataset in the Data-OOB notebook PR #636
• Replace np.float_ with np.float64 and np.alltrue with np.all, as the old aliases are removed in NumPy 2.0 PR #604
• Fix a bug in pydvl.utils.numeric.random_subset where 1 - q was used instead of q as the probability of an element being sampled PR #597
• Fix a bug in the calculation of variance estimates for MSR Banzhaf PR #605
• Fix a bug in KNN Shapley values. See Issue 613 for details.
• Backport the KNN Shapley fix to the value module PR #633

Changed

• Slicing, comparing and setting of ValuationResult behave in a more natural and consistent way PR #660 PR #666
• Switched all semi-value coefficients and sampler weights to log-space in order to avoid overflows PR #643
• Updated and rewrote some of the MSR banzhaf notebook PR #641
• Updated Least-Core notebook PR #641
• Updated Shapley spotify notebook PR #628
• Updated Data Utility notebook PR #650
• Restructured and generalized StratifiedSampler to allow using heuristics, thus subsuming Variance-Reduced stratified sampling into a unified framework. Implemented the heuristics proposed in that paper PR #641
• Uniformly distribute test points across processes for KNNShapley. Fail for GroupedDataset PR #632
• Introduced the concept of logical vs data indices for Dataset, and GroupedDataset, fixing inconsistencies in how the latter operates on indices. Also, both now return objects of the same type when slicing. PR #631 PR #648
• Use tighter bounds for the calculation of the minimal sample size that guarantees an epsilon-delta approximation in group testing (Jia et al. 2023) PR #602
• Dropped black, isort and pylint from the CI pipeline, in favour of ruff PR #633
• Breaking Changes
  • Changed DataOOBValuation to only accept bagged models PR #636
  • Dropped support for python 3.8 after EOL PR #633 - Rename parameter hessian_regularization of DirectInfluence to regularization and change the type annotation to allow for block-wise regularization parameters PR #591
  • Rename parameter hessian_regularization of LissaInfluence to regularization and change the type annotation to allow for block-wise regularization parameters PR #593
  • Remove parameter h0 from init of LissaInfluence PR #593
  • Rename parameter hessian_regularization of NystroemSketchInfluence to regularization and change the type annotation to allow for block-wise regularization parameters PR #596
  • Renaming of parameters of ArnoldiInfluence, hessian_regularization -> regularization (modify type annotation), rank_estimate -> rank PR #598
  • Remove functions remove obsolete functions lanczos_low_rank_hessian_approximation, model_hessian_low_rank
    from influence.torch.functional PR #598
  • Renaming of parameters of CgInfluence, hessian_regularization -> regularization (modify type annotation), pre_conditioner -> preconditioner, use_block_cg -> solve_simultaneously PR #601
  • Remove parameter x0 from CgInfluence PR #601
  • Rename module influence.torch.pre_conditioner -> influence.torch.preconditioner PR #601
  • Refactor preconditioner:
    • renaming PreConditioner -> Preconditioner
    • fit to TensorOperator PR #601
    • Bumped zarr dependency to v3 [PR #668](https://github...

v0.9.2

0.9.2 - 🏗 Bug fixes, logging improvement

Added

• Add progress bars to the computation of LazyChunkSequence and
  NestedLazyChunkSequence
  PR #567
• Add a device fixture for pytest, which depending on the availability and
  user input (pytest --with-cuda) resolves to cuda device
  PR #574

Fixed

• Fixed logging issue in decorator log_duration
  PR #567
• Fixed missing move of tensors to model device in EkfacInfluence
  implementation PR #570
• Missing move to device of preconditioner in CgInfluence implementation
  PR #572
• Raise a more specific error message, when a RunTimeError occurs in
  torch.linalg.eigh, so the user can check if it is related to a known
  issue
  PR #578
• Fix an edge case (empty train data) in the test
  test_classwise_scorer_accuracies_manual_derivation, which resulted
  in undefined behavior (np.nan to int conversion with different results
  depending on OS)
  PR #579

Changed

• Changed logging behavior of iterative methods LissaInfluence and
  CgInfluence to warn on not achieving desired tolerance within maxiter,
  add parameter warn_on_max_iteration to set the level for this information
  to logging.DEBUG
  PR #567

v0.9.1

0.9.1

Fixed

• FutureWarning for ParallelConfig constantly raised without actually
  instantiating the object
  PR #562
• Modify log level for implementations of TorchInfluenceFunctionModel
• Add duration logging to output of SequentialCalculator

v0.9.0

🆕 New methods, better docs and bugfixes 📚🐞

Added

• New method MSR Banzhaf with accompanying notebook, and new stopping
  criterion RankCorrelation PR #520
• New method: NystroemSketchInfluence PR #504
• New preconditioned block variant of conjugate gradient PR #507
• Improvements to documentation: fixes, links, text, example gallery, LFS and more PR #532, PR #543
• Glossary of data valuation and influence terms in the documentation PR #537
• Documentation about writing notes for new features, changes or deprecations PR #557

Fixed

• Bug in LissaInfluence, when not using CPU device PR #495
• Memory issue with CgInfluence and ArnoldiInfluence PR #498
• Raising specific error message with install instruction when trying to load pydvl.utils.cache.memcached without pymemcache installed. If pymemcache is available, all symbols from pydvl.utils.cache.memcached are available through pydvl.utils.cache PR #509

Changed

• Add property model_dtype to instances of type TorchInfluenceFunctionModel
• Bump versions of CI actions to avoid warnings PR #502
• Add Python Version 3.11 to supported versions PR #510
• Documentation improvements and cleanup PR #521, PR #522
• Simplified parallel backend configuration PR #549

New Contributors

• @jakobkruse1 made their first contribution in #510

Full Changelog: v0.8.1...v0.9.0

v0.8.1

🆕 New method and notebook, Games with exact shapley values, bug fixes and cleanup 🏗

Added

• Implement new method: EkfacInfluence #451
• New notebook to showcase ekfac for LLMs #483
• Implemented exact games in Castro et al. 2009 and 2017 #341

Fixed

• Bug in using DaskInfluenceCalcualator with TorchnumpyConverter for single dimensional arrays #485
• Fix implementations of to methods of TorchInfluenceFunctionModel implementations #487
• Fixed bug with checking for converged values in semivalues #341

Docs

• Add applications of data valuation section, display examples more prominently, make all sections visible in table of contents, use mkdocs material cards in the home page #492

New Contributors

• @opcode81 made their first contribution in #481
• @dependabot made their first contribution in #455

Full Changelog: v0.8.0...v0.8.1

v0.8.0

🆕 New interfaces, scaling computation, bug fixes and improvements 🎁

Added

• New cache backends: InMemoryCacheBackend and DiskCacheBackend PR #458
• New influence function interface InfluenceFunctionModel
• Data parallel computation with DaskInfluenceCalculator PR #26
• Sequential batch-wise computation and write to disk with SequentialInfluenceCalculator PR #377
• Adapt notebooks to new influence abstractions PR #430

Changed

• Refactor and simplify caching implementation PR #458
• Simplify display of computation progress PR #466
• Improve readme and explain better the examples PR #465
• Simplify and improve tests, add CodeCov code coverage PR #429
• Breaking Changes
  • Removed compute_influences and all related code.
    Replaced by new InfluenceFunctionModel interface. Removed modules:
    • influence.general
    • influence.inversion
    • influence.twice_differentiable
    • influence.torch.torch_differentiable

Fixed

• Import bug in README PR #457

Full Changelog: v0.7.1...v0.8.0

v0.7.1

🆕 New methods, bug fixes and improvements for local tests 🐞🧪

Added

• New method: Class-wise Shapley values PR #338
• New method: Data-OOB by @BastienZim PR #426, PR #431
• Added AntitheticPermutationSampler PR #439
• Faster semi-value computation with per-index check of stopping criteria (optional) PR #437

Changed

• No longer using docker within tests to start a memcached server PR #444
• Using pytest-xdist for faster local tests PR #440
• Improvements and fixes to notebooks PR #436
• Refactoring of parallel module. Old imports will stop working in v0.9.0 PR #421

Fixed

• Fix initialization of data_names in ValuationResult.zeros() PR #443

v0.7.0

📚🆕 Documentation and IF overhaul, new methods and bug fixes 💥🐞

This is our first β release! We have worked hard to deliver improvements across
the board, with a focus on documentation and usability. We have also reworked
the internals of the influence module, improved parallelism and handling of
randomness.

Added

• Implemented solving the Hessian equation via spectral low-rank approximation PR #365
• Enabled parallel computation for Leave-One-Out values PR #406
• Added more abbreviations to documentation PR #415
• Added seed to functions from pydvl.utils.numeric, pydvl.value.shapley and pydvl.value.semivalues. Introduced new type Seed and conversion function ensure_seed_sequence. PR #396

Changed

• Replaced sphinx with mkdocs for documentation. Major overhaul of documentation PR #352
• Made ray an optional dependency, relying on joblib as default parallel backend PR #408
• Decoupled ray.init from ParallelConfig PR #373
• Breaking Changes
  • Signature change: return information about Hessian inversion from compute_influence_factors PR #375
  • Major changes to IF interface and functionality. Foundation for a framework abstraction for IF computation. PR #278, PR #394
  • Renamed semivalues to compute_generic_semivalues PR #413
  • New joblib backend as default instead of ray. Simplify MapReduceJob. PR #355
  • Bump torch dependency for influence package to 2.0. PR #365

Fixed

• Fixes to parallel computation of generic semi-values: properly handle all samplers and stopping criteria, irrespective of parallel backend. PR #372
• Optimize memory usage in IF calculation PR #375
• Fix adding valuation results with overlapping indices and different lengths PR #370
• Fixed bugs in conjugate gradient and linear_solve PR #358
• Fix installation of dev requirements for Python 3.10 PR #382
• Improvements to IF documentation PR #371

New Contributors

• @schroedk made their first contribution in #378

Full Changelog: v0.6.1...v0.7.0

v0.6.1

🏗 Bug fixes and minor improvements

• Fix parsing keyword arguments of compute_semivalues dispatch function by @kosmitive in #333
• Create new RayExecutor class based on the concurrent.futures API, use the new class to fix an issue with Truncated Monte Carlo Shapley (TMCS) starting too many processes and dying, plus other small changes by @AnesBenmerzoug in #329
• Fix creation of GroupedDataset objects using the from_arrays and from_sklearn class methods by @AnesBenmerzoug in #334
• Fix release job not triggering on CI when a new tag is pushed by @AnesBenmerzoug in #331
• Added alias ApproShapley from Castro et al. 2009 for permutation Shapley by @mdbenito in #332

Full Changelog: v0.6.0...v0.6.1

v0.6.0

🆕 New algorithms, cleanup and bug fixes 🏗

• Fix/stopping checks by @mdbenito in #283
• Fix Monte Carlo Least Core error when n_iterations < len(dataset) by @AnesBenmerzoug in #281
• Hide parallel backend in tmcs main function by @mdbenito in #293
• Cosmetic changes to Dataset by @mdbenito in #290
• Refactor/nicer imports by @mdbenito in #284
• Fix StandardError stopping criterion by @mdbenito in #300
• Remove unpackable decorator, use asdict() by @mdbenito in #233
• Add burn-in param to AbsoluteStandardError by @mdbenito in #305
• Remove default non-negativity constraint on least core subsidy by @AnesBenmerzoug in #304
• Close #280: Add py.typed by @mdbenito in #307
• Minor docstring and cosmetic changes by @mdbenito in #317
• Allow passing additional kwargs to Dataset class' classmethods by @AnesBenmerzoug in #316
• Semi-values and samplers by @mdbenito in #319
• Remove bogus iter method. by @kosmitive in #326
• Improvements to ValuationResult by @mdbenito in #327

Full Changelog: v0.5.0...v0.6.0