You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Added a pure Rust port of LIBLINEAR 2.50 with no C code, no unsafe code and no optional feature
Added the 12 solvers 0 to 7, 11 to 13 and 21: logistic regression, support vector classification, multiclass classification, support vector regression and one-class detection
Added the types Problem, Parameter, Model, Feature, SolverType and Error, the train call, Model::predict and Model::decision_values, and read_problem for case files
Added errors that name the failed check and the correction, and index and overflow checks that return an error instead of a panic
Added exact-bit results against the unchanged LIBLINEAR 2.50 source on 158 saved cases, built with g++ at -O0 and -O2, and a script that checks the saved reference
Added checks that the exp and log results, the random stream, the sort order and the BLAS sums match the C library by bits
Added bit-identical results across runs and thread counts on one CPU type
Added mutation checks that fail when a log, a sum order, the random generator, a sort, a shrinking rule or a threshold changes
Added fuzz targets for the case reader, training and prediction, proof harnesses for index bounds, and Miri runs, started by hand in one workflow
Added CONTRACT.md with units, shapes, error rules and a source map for every solver, and docs/PARITY.md with the result for each solver
Added train and predict timings for five solvers; the crate makes no speed claim