Skip to content

CHANGELOG.md

Latest commit

 

History

History
108 lines (79 loc) · 11.9 KB

CHANGELOG.md

File metadata and controls

108 lines (79 loc) · 11.9 KB
Raw

Changelog

All notable changes to statgpu are documented here, organized by release and date.

2026-08-08

PR #122 — Panel Tier-1 diagnostics Stage B

  • Added structured Panel fit_statistics_ with parameter-based within/between/overall R², adjusted R², and classical model F statistics while preserving Stage-A inference and legacy df/R² attributes.
  • Added classical pooling F, one-way entity Breusch-Pagan LM (including Baltagi-Li unbalanced panels), and one-way classical FE-vs-RE Hausman with explicit applicability diagnostics.
  • Added NumPy/CuPy/Torch coverage, formula-row alignment, maintained Torch 2.0 CPU tests, executable linearmodels 7.0 definition alignment, and an exact-head physical GPU acceptance runner that rechecks coefficient inference as well as Stage-B diagnostics.

PR #121 — CuPy inverse-quantile LUT correctness

  • Fixed CuPy betaincinv and gammaincinv LUT cache tuple ordering so inverse quantiles no longer collapse to boundary values and downstream confidence intervals retain their correct width.
  • Added maintained regression coverage for public CuPy Student-t, Beta, F, Gamma, and chi-square PPF/ISF paths, LUT/native-fallback boundaries, legacy inverse-quantile aliases, and Panel inference consumers.
  • Validated the unchanged two-line numerical fix on Tesla P100 with Python 3.9.16 and CuPy 13.6.0: the original t_{0.975,45} failure now agrees with SciPy within 4.04e-09, all expanded inverse-distribution checks pass, and the formerly zero-width Panel intervals match the reference.

2026-08-07

PR #119 — Panel Tier-1 shared framework Stage A

  • Added an internal BasePanelModel, shared panel index/balance metadata, and structured diagnostic/fit-stat result substrate for the later Tier-1 diagnostics stages without adding new public diagnostics.
  • Centralized the existing residual-based panel OLS covariance dispatch while preserving each estimator's current nonrobust, HC1, clustered, and HAC normalization/df conventions; Fama-MacBeth keeps its distinct beta-series covariance.
  • Migrated PanelOLS, RandomEffects, PooledOLS, BetweenOLS, FirstDifferenceOLS, and FamaMacBeth to the shared lifecycle where statistically valid, while preserving formula behavior, prediction/summary contracts, fixed-effect recovery, Swamy-Arora GLS, and backend-specific output semantics.
  • Added pre-refactor golden regression coverage plus maintained Torch 2.0 CPU coverage for shared panel metadata/covariance/inference paths. Stage B diagnostics and Stage C covariance expansion remain pending under Issue #93.

PR #116 — Torch LogisticRegressionCV strict-CUDA repair

  • Fixed the mixed-precision Torch strict-CUDA LogisticRegressionCV failure by allocating batched IRLS parameters and ridge diagonals in the active CV working dtype and keeping candidate path outputs backend-native through validation scoring.
  • Added maintained regression coverage for float32/float64 CV, weighted and unweighted fitting, intercept/no-intercept paths, and the full CV selector, plus a Python 3.9 + Torch 2.0 CPU CI gate so optional-Torch coverage cannot silently skip.
  • Validated the unchanged numerical implementation head e6e4846b06604ed53e65fc9afd9054bd5777098f on Tesla P100 with PyTorch 2.0.0+cu117/CUDA 11.7 and CuPy 13.6.0: all 18 statgpu canonical CV backend runs succeeded without CPU fallback, including LogisticRegressionCV on NumPy, CuPy, and Torch.
  • Retained the historical pre-fix P100 failure source unchanged and registered the exact-head post-fix source under results/pr116_p100/; focused physical validation evidence is retained separately from dashboard timing data.

0.2.4 — 2026-08-06

Logistic regression and GLM correctness

  • Corrected arbitrary-link Binomial IRLS Fisher weights, working responses, line-search objectives, backend-native warm starts, and quadratic-penalty validation.
  • Hardened direct LogisticRegression response/control validation, transactional refits, convergence reporting, integer prediction dtype, single-column response handling, and finite decision thresholds.
  • Unified fitted logistic likelihood diagnostics across NumPy, CuPy, and Torch with the registered numerically stable LogisticLoss objective; likelihood, AIC, BIC, pseudo-R², and convergence remain independent of covariance inference.
  • Kept confusion-matrix and hard classification metrics available for one-class targets while preserving explicit class-support requirements for ROC-AUC and average precision.
  • Kept CuPy/Torch analytic weights device-native and corrected weighted IRLS curvature, likelihood, dispersion, and sandwich-inference semantics.
  • Standardized GLM analytic-weight behavior across ridge scaling, line search, pseudo-loglikelihood, information criteria, dispersion, and covariance; global weight rescaling leaves estimates and diagnostics unchanged.
  • Added backend-native response-domain, finite-value, real-valued, shape, and length validation for scalar GLMs, including penalized and cross-validated entry points.
  • Aligned formula sample weights only after Patsy missing-row filtering and corrected weighted Gaussian FISTA centering.

Cross-validation, inference, and estimator contracts

  • Made RidgeCV, ElasticNetCV, and LogisticRegressionCV fits failure-safe: stale state is cleared before fitting and selected parameters are published only after the final full-data refit succeeds.
  • Preserved explicit Torch/CuPy requests and pinned device="auto" final refits to the backend selected during cross-validation.
  • Updated Logistic and Elastic Net default regularization grids to incorporate analytic weights and satisfy integer-weight row-replication equivalence.
  • Preserved declared validation losses and analytic weights in penalized CV; programming, shape, CUDA OOM, and device errors are no longer converted into candidate NaN values or unrelated MSE fallback.
  • Completed the standalone ElasticNet and final-refit ElasticNetCV inference contract across NumPy, CuPy, and Torch.
  • Corrected public ElasticNet/Ridge scaling documentation: under the shared average-loss convention, ElasticNet(alpha, l1_ratio=0) matches Ridge(alpha).
  • Made public estimator finite-input guards, cloning, sklearn tags, nested set_params, and fitted-state invalidation transactional, including legacy scikit-learn clone identity checks.

Solver and backend safety

  • Corrected the executable loss/penalty/solver matrix so Newton, L-BFGS, and L-BFGS-B reject unsupported non-smooth penalties rather than optimizing only the smooth component.
  • Removed the incorrect Euclidean-prox Newton shortcut. Smooth L2/no-penalty objectives retain Newton updates; non-smooth proximal-Newton requests delegate visibly to backend-native FISTA until a Hessian-metric proximal solver exists.
  • Narrowed Armijo, linear-solve, alpha-grid, and inference fallbacks to recognized numeric or rank failures; CUDA OOM, device, index, contract, and unrelated runtime failures propagate.
  • Normalized warm starts for FISTA, Newton-family, L-BFGS-family, L-BFGS-B, and ADMM solvers to the preprocessed design backend, device, and dtype.
  • Completed ADMM's legitimate Cholesky fallback and hardened L-BFGS-B feasible directions, backend-native bounds, and NaN-bound validation.
  • Added centralized, observable Torch compilation policy: eager remains the default for unset, auto, and disable; default and reduce-overhead are explicit opt-ins, and only the known CUDA Graph output-lifecycle failure becomes a permanent eager fallback.
  • Removed the package-initialization cycle between statgpu.glm_core and the Cox loss export by lazily exposing CoxPartialLikelihoodLoss; fresh-interpreter imports no longer depend on importing LogisticRegression first.

Documentation, testing, and release preparation

  • Reconciled the English and Chinese LogisticRegression, ElasticNet, cross-validation, solver-algorithm, and solver/penalty documentation with the maintained implementation.
  • Removed unsupported universal GPU speedup, backend-threshold, and coefficient-tolerance claims; performance guidance now requires workload-specific benchmarking.
  • Documented ownership boundaries between maintained pytest coverage and manual physical-GPU diagnostics.
  • Bumped package metadata to 0.2.4 and added the authoritative GitHub Release document at .github/releases/v0.2.4.md.

Validation

  • The final PR #87 implementation head passed the complete CPU suite with 2239 passed and 719 skipped, static and documentation contracts, Python 3.9–3.12 regression jobs, scikit-learn 1.2.2/1.3.2/latest compatibility, and release-package validation.
  • Physical NVIDIA validation passed on the unchanged numerical implementation: RTX 4090 with PyTorch 2.8.0+cu128 passed the selected compile/CUDA Graph matrix 9/9 and runtime assertions; Tesla P100 with CuPy 13.6.0 passed the corresponding runtime assertions.
  • The focused release PR changes version metadata and release-facing documentation only; all exact release-head hosted gates must pass before creating tag v0.2.4.

0.2.3 — 2026-08-04

Added

  • Completed CoxPH Phase 1 with Breslow, Efron, and Exact ties; delayed-entry and (start, stop] counting-process data; shared-coefficient stratification; subject identifiers; and Surv(start, stop, event) formula input.
  • Added shared NumPy, CuPy, and Torch-CUDA risk-set primitives for Cox objectives, gradients, information matrices, and baseline estimation, including backend-native dynamic programming for Exact ties.
  • Extended CoxPHCV held-out partial likelihood to all supported tie methods, delayed entry, start-stop rows, strata, and subject-grouped folds.

Changed

  • Hardened Cox inference, numerical stability, formula NA alignment, singular-information handling, CV cache identity, fold eligibility, selected-penalty refitting, and failed-fit state resets.
  • Hardened L1, L2, Elastic Net, SCAD, and MCP penalized Cox estimation; removed the unidentified intercept; corrected Cox-specific warm starts; and made Torch Efron value, gradient, and Hessian paths native.
  • Standardized public Group Lasso and Adaptive Group Lasso behavior through the generic loss-gradient and exact group-proximal path across supported backends.
  • Made requested CoxPHCV two-stage and successive-halving controls execute one explicit exhaustive full-precision candidate pass, avoiding repeated complete-grid fitting while preserving deterministic selection semantics.
  • Made one-shot CoxPHCV.cv_splits iterators reusable across repeated fit, scikit-learn clone, parameter reconstruction, and pickle.

Validation

  • Hosted workflow #960 passed on the final reviewed head f05a44ad363b46612e956e137e2f00d040765acb: documentation, static, full CPU, and Python 3.9–3.12 regression jobs all passed; the complete CPU suite reported 1881 passed and 662 skipped.
  • The final exact-head physical-GPU promotion artifact is published at https://gist.github.com/TheHiddenObserver/afdcad86a243e68a918d852b92e984a4. It records schema 3, 134/134 passing checks, zero child and nested return codes, empty gate-failure arrays, clean source state before and after execution, and SHA-256 bd4058450def691dd29e9d78853534016c6da70c33192a97dc312d95cbe5d76d.
  • Added release-package validation that checks version consistency, builds the pure-Python wheel and sdist, runs twine check, validates artifact contents, clean-installs the sdist on Ubuntu, and clean-installs the same wheel on Ubuntu, Windows, and macOS.

Packaging and release publication

  • Bumped the package version to 0.2.3 in pyproject.toml and statgpu/__init__.py.
  • The official wheel remains a universal py3-none-any artifact built with STATGPU_NO_EXT=1; optional Cython sources remain available in the sdist.
  • Added the authoritative GitHub Release document at .github/releases/v0.2.3.md, a release-note completeness gate, and tag automation that publishes that file as the GitHub Release body after the PyPI job succeeds.

Earlier history

Entries through 2026-07-27 are retained in CHANGELOG-history-through-2026-07-27.md.