LoginSign Up
← Back to all sparks
T

tune

ANALYTICS
Velocity0.0

Hyperparameter tuning and model optimization for the tidymodels framework

github.com/tidymodels/tune ↗

tune extends tuning past the model itself to postprocessors, and adds a second parallel backend

hyperparameter-tuningtidymodelsparallelismpostprocessingbayesian-optimization
◆Current state
tune runs hyperparameter search for tidymodels. Version 2.0.0 rewrote tune_grid() to make postprocessing tunable alongside preprocessing and the model, changed the .config naming scheme to match, and added mirai as a parallel backend next to future. Version 2.1.0 followed with quantile regression support and a replacement Gaussian process engine.
◆Where it's heading
Two migrations run through this timeline. The tunable surface keeps widening - first censored regression as a mode, then postprocessors via tailor - so that a candidate is now a preprocessor, model and postprocessor triple rather than just a model. The parallel story has moved from foreach to future and now to mirai, each step deprecating the last. Neither is finished.
◆Prediction
Expect the foreach path to be removed outright, and the postprocessing surface to grow as tailor gains more steps; the GauPro switch will likely need follow-up as its behavior differs from the old engine.

◆Recent moves

  1. 3mo ago

    Quantile regression tuning; Bayesian search moves to GauPro

    Model tuning reaches quantile regression, and resampling estimates can be weighted by assessment set size. The breaking part is a forced dependency swap: the Gaussian process engine moves from GPfit to GauPro because the former is unmaintained, which changes how qualitative predictors are handled and how poor fits are recovered from.

    View source ↗
  2. 9mo ago

    Fixes int_pctl() with future parallelism on last_fit()

    A single bug fix for percentile intervals failing on last_fit() results when future parallelism was enabled. Cleanup after the 2.0.0 parallelism rework.

    View source ↗
  3. 11mo ago

    Postprocessors become tunable; mirai joins future as a backend

    ⚡ SPARK

    The release the rest of this timeline orbits: tune_grid() was rewritten so that postprocessing is a tunable stage, which is why the .config scheme had to change to name three stages instead of two. Adding mirai continues the parallel-backend migration that 1.2.0 started and 1.3.0 warned about.

    View source ↗
  4. 11mo ago

    Development snapshot re-enabling skipped tests

    A pre-release development tag whose notes are the squashed commit log - test re-enabling and code review fixups. It appears here because the repository tags development versions, not because anything shipped.

    View source ↗
  5. 1y ago

    Warns on foreach parallelism; space-filling grids by default

    Signals the next step in the parallel migration by warning when foreach is enabled without future, and states plainly that the following version will be pure future. Automatic grids switch to grid_space_filling(), which produces optimized designs rather than depending on random draws.

    View source ↗
  6. 2y ago

    Fixes parallel tuning errors under multisession plans

    Two parallelism bug fixes, one for bootstrap intervals under PSOCK clusters and one for an object-not-found error when tuning across everything. The release notes also correct a mismatched GitHub tag.

    View source ↗