dmeoli / optiml / 28243370665

Builds Branch Commit Type Ran Committer Via Coverage
28243370665 master Bump version to 1.8 Release that ships the ML-CUP data inside the package (package_data) and makes the example notebooks runnable on Google Colab. push dmeoli github
84.3
28240491172 master Update ML_report notebook to the current API so it runs on Binder The notebook used the removed DualSVC/DualSVR classes and the renamed OneHotEncoder(sparse=) argument, which would break on Binder. Migrate to: - SVC(loss=hinge, dual=True, reg_int... push dmeoli github
84.39
28237978870 master Fix run-together docstring field lists and surface algorithm mechanics - add the missing blank line before the :param/:return field lists that were rendering run-together on the docs site (Quadratic and base function/jacobian/ hessian, the Ro... push dmeoli github
84.45
28237576398 master Auto-publish to PyPI on GitHub Release via Trusted Publishing Add a workflow that builds and uploads the package to PyPI whenever a GitHub Release is published, authenticating with PyPI Trusted Publishing (OIDC) so no API token needs to be stored... push dmeoli github
84.39
28236897742 master Demonstrate the generic lower bound in the constrained-methods notebook Add a short section showing that the box-constrained optimizers now also solve the generic lb <= x <= ub program (the L-BFGS section was already added to the line-search note... push dmeoli github
84.45
28236701695 master Bump version to 1.7 Release including L-BFGS, the generic lb <= x <= ub box-constrained optimizers, the certified cvxpy f_star, the SMO linear-kernel fix, the rigorous/portable test suite, the GitHub Actions CI, the GitHub Pages documentation wit... push dmeoli github
84.52
28231681366 master Support a generic lower bound lb <= x <= ub in the box-constrained optimizers Generalize ProjectedGradient, FrankWolfe, ActiveSet and InteriorPoint (and the BoxConstrainedQuadraticOptimizer base) from 0 <= x <= ub to a generic lb <= x <= ub via a... push dmeoli github
84.48
28231146333 master Surface the Interior Point KKT system in its class docstring Lift the slackened KKT system and the Newton-step reduction (H dx = w) from the minimize() implementation comments into the class docstring as rendered math, so the documentation site d... push dmeoli github
84.43
28231051902 master Add tests for previously-uncovered code - test the four weight initializers (shape, finiteness, reproducibility, bounds) - test the Rosenbrock/Ackley/Six-Hump-Camel benchmark functions (optimum, f_star, stationarity/finiteness of the gradient) ... push dmeoli github
84.33
28230539641 master Remove the last two stray assert_optimal calls in the L2 SVR test These two lines carried a trailing comment and so were missed by the previous score-only revert, leaving an undefined-name reference. push dmeoli github
82.0
28214388774 master Use the robust tolerance for all SVC convergence checks The multiclass one-vs-rest SVC subproblems on the tiny iris dataset produce optimality gaps that are sensitive to the BLAS/platform (a "smooth" L2 line-search gap of ~1e-9 locally became ~9e... push dmeoli github
82.28
28213300295 master Make L1 SVC stochastic test check only the score The stochastic optimizers converge too erratically on the nonsmooth multiclass hinge primal to meet a fixed optimality-gap tolerance across platforms/seeds (CI showed gaps up to 1.6e-2); their conv... push dmeoli github
68.0
28212963022 master Render docstring math as LaTeX for the documentation site Convert the plain-text formulas in the docstrings across the whole library to reStructuredText math (:math: / .. math::) so Sphinx + MathJax renders them properly on the documentation site... push dmeoli github
82.09
28212477813 master Fix modern numpy/scipy/sklearn/cvxpy compatibility so CI passes - np.sum(generator) is rejected by numpy 2.x: use the builtin sum() for the neural network regularization terms - OneHotEncoder(sparse=...) was renamed to sparse_output in scikit-l... push dmeoli github
79.19
1262 master fixed ld plot push Donato Meoli travis-ci-com pending completion