← Back to all sparks
F

finetune

ANALYTICS
Velocity0.0

Additional model tuning strategies including racing and simulated annealing

finetune tracks tune's evolving contracts more than it advances racing itself

tidymodelshyperparameter-tuningracingsimulated-annealingsurvival-analysis
Current state
finetune provides the racing and simulated-annealing alternatives to grid search in tidymodels. The core algorithms have been stable since 1.0.x; what has changed is everything around them — censored regression support arriving with a tune release, weighted resampling estimates preserved through racing, and a breaking move to named-only optional arguments.
Where it's heading
This is a package operating downstream of tune, adopting whatever the shared resampling machinery grows next rather than proposing new search strategies. The 1.3.0 weighting work is a clear example: tune changed how resampling estimates are computed, and finetune's job was to not lose the weights during racing. Error messages and input checks are the steady internal theme.
Prediction
Expect the next release to absorb whatever tune changes about metric collection or resampling weights; nothing in the entries points to a new search algorithm.

Recent moves

  1. 3mo ago

    Racing preserves tune's assessment-set weighting

    Keeps finetune correct now that tune can weight resampling estimates by assessment-set size — without this, racing would compare configurations on differently-weighted numbers. Also fixes NULL-result logging in simulated annealing.

    View source ↗
  2. 1y ago

    Maintenance release; magrittr pipe replaced with base pipe

    A CRAN-required maintenance release whose only substantive change is internal pipe style.

    View source ↗
  3. 2y ago

    Censored regression models can be raced and annealed

    Survival models become full citizens of finetune's search strategies, arriving with the tune release that enabled them rather than as independent work. Optional arguments also become named-only, which breaks positional calls to collect_metrics() and show_best().

    View source ↗
  4. 3y ago

    Keep-up release for tune and dplyr; .config alignment fixed

    Mostly compatibility churn, with one real fix so .config values in predictions line up with those in metrics.

    View source ↗
  5. 3y ago

    Racing results filter to fully resampled configurations

    select_best() and show_best() stop reporting configurations that racing eliminated early, and a complete argument makes that filtering explicit in collect_*(). Corrects a real source of misleading racing output.

    View source ↗
  6. 3y ago

    Informative error when resamples are too few for racing

    A 1.0.0 release whose notes list only an error message and an argument-passing fix. The version number carries more weight than the contents.

    View source ↗