Skip to content

changelog.md

Latest commit

 

History

History
151 lines (115 loc) · 16.2 KB

changelog.md

File metadata and controls

151 lines (115 loc) · 16.2 KB
Raw

Changelog

Language: English
Last updated: 2026-08-09
This page: Changelog
Switch: Chinese

2026-08-08

PR #122 — Panel Tier-1 diagnostics Stage B

  • Added public structured PanelTestResult and PanelFitStatistics outputs plus standardized fit_statistics_ on the maintained panel estimators. The new fit statistics use parameter-based within/between/overall R², an explicitly defined adjusted R², and a classical homoskedastic model F statistic where the estimator has a residual-OLS fit space.
  • Kept Stage-A coefficient inference and legacy R²/df behavior unchanged. In particular, PanelOLS continues to expose its historical public residual df and BSE/t/p/CI, while Stage-B diagnostics use a separate standard fixed-effect nuisance-rank df; the classical Hausman calculation consumes only a diagnostic small covariance rescaled to that standard denominator.
  • Added the classical pooling F test for fixed effects, the one-way entity error-components Breusch-Pagan LM test including the Baltagi-Li unbalanced-panel formula, and the classical one-way entity FE-vs-RE Hausman test. Inapplicable econometric cases return structured reasons; singular positive-semidefinite Hausman covariance differences use a documented generalized-inverse/rank extension, while materially indefinite differences are rejected.
  • Added optional entity_ids to PooledOLS.fit() and FamaMacBeth.fit() solely for Stage-B within/between fit statistics and the panel BP-LM path. Pooled HAC sorting now carries entity diagnostic metadata through the same stable permutation as X/y. Formula missing-row filtering aligns observation-level side arrays before diagnostics are formed.
  • Added analytic/fitted regressions, maintained Python 3.9 + Torch 2.0 CPU parity, and an executable linearmodels==7.0 definition-alignment job. FirstDifference external comparison is restricted to panels where both implementations use the same transformed sample; Stage B does not silently redefine the Stage-A adjacent-observed-row differencing contract for internal time gaps.
  • Added dev/benchmarks/validate_panel_stage_b_gpu.py as the exact-head physical correctness/provenance gate. The previously accepted P100 artifacts at numerical implementation a57efcea29b0e87ecb89865c5a6902d5773812c6 remain immutable historical evidence: CuPy and Torch each passed all 17 estimator cases with requested-backend provenance and no fallback, while the focused disconnected two-way FE artifact validated the df=1 inference boundary to machine precision. The four Hausman parameterizations per backend in that run were all correctly structured applicable=false cases, so they validate applicability/reason parity but do not physically exercise an applicable Hausman statistic/p-value/df path.
  • The reopened physical gate is now closed on exact clean measurement head 2701aa9feb3796c33c94e6480fcb78c80c6a809c: Tesla P100 CuPy and Torch each passed all 17 estimator cases and all five Hausman diagnostics with requested/executed backend identity and no CPU fallback. The dedicated 48-observation nonzero-effect fixture is applicable=true on both backends with df=1; its Hausman statistic differs from NumPy by at most 1.10e-13 and p-value by at most 2.19e-14. The promoted 44-row canonical validation source preserves statistic/pvalue/df for that branch, while the older 42-row a57efcea source remains historical audit evidence. No timing or speedup claim is made.

Related: Issue #93 and pull request #122.

PR #121 — CuPy inverse-quantile LUT correctness

  • Corrected the CuPy LUT cache tuple order used by betaincinv() and gammaincinv(). The LUT builders already stored (x_grid, y_grid), but the cached values were unpacked in reverse, so inverse lookup searched the wrong axis and could collapse quantiles to clipped boundary values.
  • The original failure was exposed by Panel Stage A physical validation: on Tesla P100, the CuPy t.isf(0.025, 45) path produced an almost-zero critical value and zero-width confidence intervals. The corrected two-line numerical implementation now returns 2.014103388876289, within 4.04e-09 absolute error of the SciPy reference.
  • Added maintained regression coverage for raw inverse-beta/inverse-gamma cache reuse; public CuPy Student-t, Beta, F, Gamma, and chi-square PPF/ISF plus round trips; Student-t LUT/native-fallback boundaries at df 1, 10, 45, 60, and 80; module-level distribution proxies; legacy inverse-quantile aliases; and representative Panel inference consumers.
  • Expanded physical validation on exact numerical head f768b312d05f47debdb8fa13ae4da09b27d00239 used Tesla P100, Python 3.9.16, and CuPy 13.6.0 with a clean working tree. Maximum PPF/ISF absolute errors were 4.04e-09 for Student-t, 5.49e-13/1.68e-13 for Beta, 4.51e-12 for F, 1.45e-08 for Gamma, and 2.90e-08 for chi-square, all well inside the maintained inverse-quantile accuracy contract.
  • The reconstructed Panel CI and actual shared Panel inference consumer now match their references within 3.42e-10 and 1.96e-10, respectively; the formerly zero-width intervals are non-degenerate and agree with the Torch/reference result.

Related: Issue #120 and pull request #121.

2026-08-07

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

  • Added an internal BasePanelModel, PanelIndexInfo, PanelTestResult, and PanelFitStatistics substrate for Issue #93. Stage A establishes shared lifecycle and panel-structure contracts only; Hausman/pooling-F/Breusch-Pagan LM tests and expanded fit statistics remain Stage B work.
  • Centralized dispatch for the covariance definitions that already existed in the panel estimators. Nonrobust scaling, HC1 corrections, one-/two-way clustering, HAC behavior, rank/df conventions, and unsupported-name errors remain estimator-equivalent to the pre-refactor implementation. No HC0/HC2/HC3 or Driscoll-Kraay support is added in this stage.
  • Migrated PanelOLS, RandomEffects, PooledOLS, BetweenOLS, FirstDifferenceOLS, and FamaMacBeth to statistically neutral shared lifecycle helpers where valid. Fixed-effect recovery/prediction, Swamy-Arora variance components and quasi-demeaning, and Fama-MacBeth beta-series covariance remain model-specific.
  • Preserved the existing formula, missing-row alignment, intercept/effect-token behavior, prediction output contracts, summary schemas/printing behavior, balanced/unbalanced semantics, residual-df definitions, and strict explicit-device/no-fallback rules.
  • Added a pre-refactor golden suite before any panel source migration and kept it active after the refactor. Dedicated Python 3.9 + Torch 2.0 CPU CI now also executes the shared panel metadata/covariance/inference regression tests so optional Torch coverage cannot silently skip.

Stage B diagnostics and Stage C covariance expansion remain pending under Issue #93; this Stage-A refactor does not advertise them as public capabilities.

PR #116 — Torch LogisticRegressionCV strict-CUDA repair

  • Fixed the maintained Torch strict-CUDA LogisticRegressionCV failure in the batched GPU IRLS path. Mixed-precision CV now allocates parameters and ridge diagonals in the active working dtype, and coefficient/intercept paths remain backend-native through validation scoring.
  • Added regression coverage for float32 and float64 CV, weighted/unweighted execution, intercept/no-intercept paths, and the full selector consumer. A dedicated Python 3.9 + Torch 2.0 CPU CI job prevents the optional Torch regression suite from silently skipping.
  • Physical validation ran on exact numerical implementation head e6e4846b06604ed53e65fc9afd9054bd5777098f using Tesla P100-SXM2-16GB, Python 3.9.16, PyTorch 2.0.0+cu117 / CUDA 11.7, and CuPy 13.6.0. All four focused Torch CUDA cases selected the same C=0.2 as the CPU reference; the largest mean-loss difference was below 6.2e-8, with the float64 path agreeing to machine precision.
  • The canonical six-family rerun recorded all 18 statgpu NumPy/CuPy/Torch backend rows as successful, with zero failed candidates/folds and converged final refits. LogisticRegressionCV selected C=0.1 on NumPy, CuPy, Torch, and sklearn; the Torch/NumPy validation-loss difference was below 4.7e-8.
  • The historical pre-fix P100 source remains immutable and registered. The post-fix exact-head source is registered separately from results/pr116_p100/cv_benchmark_pr116_p100.json; focused_validation.json remains validation-only evidence rather than dashboard timing data.

Related: Issue #112 and pull request #116.

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 validation, transactional refits, convergence reporting, integer hard predictions, single-column response handling, and finite decision thresholds.
  • Unified fitted logistic likelihood diagnostics across NumPy, CuPy, and Torch with the registered stable LogisticLoss objective. Likelihood, AIC, BIC, pseudo-R², and convergence remain available independently of covariance inference.
  • Kept confusion-matrix metrics available for one-class targets while retaining explicit class-support errors for ROC-AUC and average precision.
  • Kept analytic weights device-native on CuPy/Torch fits and corrected weighted IRLS curvature, likelihood, dispersion, and sandwich-inference semantics.
  • Standardized GLM analytic-weight behavior across fitting, line search, diagnostics, and covariance. Globally rescaling analytic weights does not change fitted parameters or reported diagnostics.
  • Added backend-native response-domain, real-valued, finite, shape, and length validation for scalar GLMs, including penalized and CV entry points.
  • Aligned formula sample weights after Patsy row filtering and corrected weighted Gaussian FISTA centering.

Cross-validation, inference, and estimator contracts

  • Made RidgeCV, ElasticNetCV, and LogisticRegressionCV 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 standalone ElasticNet and final-refit ElasticNetCV inference across NumPy, CuPy, and Torch. Fold models remain estimation-only.
  • Corrected ElasticNet/Ridge scaling documentation: under the shared average-loss convention, ElasticNet(alpha, l1_ratio=0) matches Ridge(alpha).
  • Made public 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 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; non-smooth proximal-Newton requests delegate visibly to backend-native FISTA until a Hessian-metric proximal solver exists.
  • Narrowed Armijo, linear-solve, CV-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, and ADMM solvers to the preprocessed design backend, device, and dtype.
  • Completed ADMM's legitimate Cholesky fallback and hardened L-BFGS-B directions, backend-native bounds, and NaN-bound validation.
  • Added a centralized, observable Torch compile policy: eager remains the default for unset, auto, and disable; default and reduce-overhead are explicit opt-ins. 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 require a particular order.

Documentation 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 the ownership boundary between maintained pytest coverage and manual physical-GPU diagnostics.
  • Bumped package metadata to 0.2.4 and added .github/releases/v0.2.4.md as the authoritative GitHub Release body.

Validation

  • The final PR #87 implementation head passed 2239 tests with 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. Exact release-head hosted gates must pass before tag v0.2.4 is created.

Related: Issue #45, Issue #81, Issue #82, Issue #83, and pull request #87.

0.2.3 — 2026-08-04

Survival analysis

  • 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 objectives, gradients, information matrices, and baseline estimation. Exact tied-event partitions use backend-native dynamic programming.
  • Extended CoxPHCV held-out partial likelihood to all supported tie methods, delayed entry, start-stop rows, strata, and subject-grouped folds.
  • Hardened Cox inference, centered risk-set numerics, log-domain baseline prediction, 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 the Torch Efron value, gradient, and Hessian paths native.

Cross-validation and grouped penalties

  • Requested CoxPHCV two-stage and successive-halving controls now execute one explicit exhaustive full-precision candidate pass, preserving deterministic selection while avoiding repeated complete-grid fitting.
  • One-shot CoxPHCV.cv_splits iterators are reusable across repeated fit, scikit-learn clone, parameter reconstruction, and pickle.
  • Public Group Lasso and Adaptive Group Lasso use the generic loss-gradient and exact group-proximal path consistently across supported backends.

Validation and packaging

  • Hosted workflow #960 passed on 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 as schema-3 evidence. It records 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.
  • The package version is now 0.2.3. Release-package validation checks version consistency, builds the pure-Python wheel and sdist, runs twine check, validates artifact contents, and smoke-installs both distributions in clean environments.

Earlier history

Detailed entries through 2026-08-03 are retained in the archived changelog.