Skip to contents

This CallbackTuning integrates early stopping into the hyperparameter tuning of an XGBoost learner. Early stopping estimates the optimal number of trees (nrounds) for a given hyperparameter configuration. Since early stopping is performed in each resampling iteration, there are several optimal nrounds values. The callback writes the maximum value to the archive in the max_nrounds column. In the best hyperparameter configuration (instance$result_learner_param_vals), the value of nrounds is replaced by max_nrounds and early stopping is deactivated.

Details

Currently, the callback does not work with GraphLearners from the package mlr3pipelines. The callback is compatible with the AutoTuner. The final model is fitted with the best hyperparameter configuration and max_nrounds i.e. early stopping is not performed.

Resources

Examples

clbk("mlr3tuning.early_stopping")
#> <CallbackTuning:mlr3tuning.early_stopping>: Early Stopping Callback
#> * Active Stages: on_eval_before_archive, on_eval_after_benchmark,
#>   on_result, on_optimization_begin
# \donttest{
if (requireNamespace("mlr3learners") && requireNamespace("xgboost") ) {
  library(mlr3learners)

  # activate early stopping on the test set and set search space
  learner = lrn("classif.xgboost",
    eta = to_tune(1e-02, 1e-1, logscale = TRUE),
    early_stopping_rounds = 5,
    nrounds = 100,
    early_stopping_set = "test")

  # tune xgboost on the pima data set
  instance = tune(
    method = tnr("random_search"),
    task = tsk("pima"),
    learner = learner,
    resampling = rsmp("cv", folds = 3),
    measures = msr("classif.ce"),
    term_evals = 10,
    callbacks = clbk("mlr3tuning.early_stopping")
  )
}
#> Loading required namespace: mlr3learners
#> Loading required namespace: xgboost
# }