mlr3tuning

Package website: release | dev

tic CRAN Status StackOverflow Mattermost CodeFactor

This package provides hyperparameter tuning for mlr3. It offers various tuning methods e.g. grid search, random search and generalized simulated annealing and iterated racing and different termination criteria can be set and combined. AutoTuner provides a convenient way to perform nested resampling in combination with mlr3. The package is build on bbotk which provides a common framework for optimization.

Resources

Installation

Install the last release from CRAN:

install.packages("mlr3tuning")

Install the development version from GitHub:

remotes::install_github("mlr-org/mlr3tuning")

Example

Basic hyperparameter tuning

library("mlr3tuning")

# retrieve task
task = tsk("pima")

# load learner and set search space
learner = lrn("classif.rpart", cp = to_tune(1e-04, 1e-1, logscale = TRUE))

# hyperparameter tuning on the pima indians diabetes data set
instance = tune(
  method = "random_search",
  task = task,
  learner = learner,
  resampling = rsmp("cv", folds = 3),
  measure = msr("classif.ce"),
  term_evals = 10,
  batch_size = 5
)

# best performing hyperparameter configuration
instance$result
##           cp learner_param_vals  x_domain classif.ce
## 1: -2.774656          <list[2]> <list[1]>  0.2617188
# all evaluated hyperparameter configuration
as.data.table(instance$archive)
##            cp classif.ce  x_domain_cp runtime_learners           timestamp batch_nr      resample_result
##  1: -6.355096  0.2786458 0.0017378683            0.024 2021-09-02 13:45:24        1 <ResampleResult[20]>
##  2: -5.937751  0.2799479 0.0026379549            0.025 2021-09-02 13:45:24        1 <ResampleResult[20]>
##  3: -4.280177  0.2734375 0.0138402055            0.023 2021-09-02 13:45:24        1 <ResampleResult[20]>
##  4: -7.539553  0.2786458 0.0005316351            0.053 2021-09-02 13:45:24        1 <ResampleResult[20]>
##  5: -7.140496  0.2786458 0.0007923588            0.027 2021-09-02 13:45:24        1 <ResampleResult[20]>
##  6: -9.114735  0.2786458 0.0001100325            0.021 2021-09-02 13:45:25        2 <ResampleResult[20]>
##  7: -5.401998  0.2695312 0.0045075636            0.022 2021-09-02 13:45:25        2 <ResampleResult[20]>
##  8: -4.703298  0.2773438 0.0090653290            0.022 2021-09-02 13:45:25        2 <ResampleResult[20]>
##  9: -2.774656  0.2617188 0.0623709163            0.022 2021-09-02 13:45:25        2 <ResampleResult[20]>
## 10: -4.441283  0.2734375 0.0117808090            0.024 2021-09-02 13:45:25        2 <ResampleResult[20]>
# fit final model on complete data set
learner$param_set$values = instance$result_learner_param_vals
learner$train(task)

Automatic tuning

# retrieve task
task = tsk("pima")

# construct auto tuner
at = auto_tuner(
  method = "random_search",
  learner = lrn("classif.rpart", cp = to_tune(1e-04, 1e-1, logscale = TRUE)),
  resampling = rsmp("cv", folds = 3),
  measure = msr("classif.ce"),
  term_evals = 10,
  batch_size = 5
)

# train/test split
train_set = sample(task$nrow, 0.8 * task$nrow)
test_set = setdiff(seq_len(task$nrow), train_set)

# tune hyperparameters and fit final model on the complete data set in one go
at$train(task, row_ids = train_set)

# best performing hyperparameter configuration
at$tuning_result
##           cp learner_param_vals  x_domain classif.ce
## 1: -4.159136          <list[2]> <list[1]>  0.2590228
# all evaluated hyperparameter configuration
as.data.table(at$archive)
##            cp classif.ce  x_domain_cp runtime_learners           timestamp batch_nr      resample_result
##  1: -6.469604  0.2671051 0.0015498397            0.022 2021-09-02 13:45:26        1 <ResampleResult[20]>
##  2: -5.975552  0.2671051 0.0025400997            0.022 2021-09-02 13:45:26        1 <ResampleResult[20]>
##  3: -6.304555  0.2671051 0.0018279594            0.022 2021-09-02 13:45:26        1 <ResampleResult[20]>
##  4: -2.885359  0.2703730 0.0558347447            0.022 2021-09-02 13:45:26        1 <ResampleResult[20]>
##  5: -5.878631  0.2671051 0.0027986148            0.038 2021-09-02 13:45:26        1 <ResampleResult[20]>
##  6: -5.316384  0.2671051 0.0049104786            0.022 2021-09-02 13:45:26        2 <ResampleResult[20]>
##  7: -4.159136  0.2590228 0.0156210471            0.021 2021-09-02 13:45:26        2 <ResampleResult[20]>
##  8: -9.206399  0.2671051 0.0001003949            0.024 2021-09-02 13:45:26        2 <ResampleResult[20]>
##  9: -4.869814  0.2785350 0.0076747945            0.023 2021-09-02 13:45:26        2 <ResampleResult[20]>
## 10: -6.649454  0.2671051 0.0012947286            0.022 2021-09-02 13:45:26        2 <ResampleResult[20]>
# predict new data with model trained on the complete data set and optimized hyperparameters
at$predict(task, row_ids = test_set)
## <PredictionClassif> for 154 observations:
##     row_ids truth response
##           3   pos      pos
##           6   neg      neg
##          11   neg      neg
## ---                       
##         756   pos      pos
##         758   pos      pos
##         768   neg      neg

Nested resampling

# retrieve task
task = tsk("pima")

# load learner and set search space
learner = lrn("classif.rpart", cp = to_tune(1e-04, 1e-1, logscale = TRUE))

# nested resampling
rr = tune_nested(
  method = "random_search",
  task =  task,
  learner = learner,
  inner_resampling = rsmp("holdout"),
  outer_resampling = rsmp("cv", folds = 3),
  measure = msr("classif.ce"),
  term_evals = 10,
  batch_size = 5
)

# aggregated performance of all outer resampling iterations
rr$aggregate()
## classif.ce 
##  0.2473958
# performance scores of the outer resampling
rr$score()
##                 task task_id         learner          learner_id         resampling resampling_id iteration              prediction classif.ce
## 1: <TaskClassif[47]>    pima <AutoTuner[40]> classif.rpart.tuned <ResamplingCV[19]>            cv         1 <PredictionClassif[19]>  0.2187500
## 2: <TaskClassif[47]>    pima <AutoTuner[40]> classif.rpart.tuned <ResamplingCV[19]>            cv         2 <PredictionClassif[19]>  0.2500000
## 3: <TaskClassif[47]>    pima <AutoTuner[40]> classif.rpart.tuned <ResamplingCV[19]>            cv         3 <PredictionClassif[19]>  0.2734375
# inner resampling results
extract_inner_tuning_results(rr)
##    iteration        cp classif.ce learner_param_vals  x_domain task_id          learner_id resampling_id
## 1:         1 -2.768620  0.2573099          <list[2]> <list[1]>    pima classif.rpart.tuned            cv
## 2:         2 -3.880799  0.2046784          <list[2]> <list[1]>    pima classif.rpart.tuned            cv
## 3:         3 -8.862942  0.2748538          <list[2]> <list[1]>    pima classif.rpart.tuned            cv
# inner resampling archives
extract_inner_tuning_archives(rr)
##     iteration        cp classif.ce  x_domain_cp runtime_learners           timestamp batch_nr      resample_result task_id          learner_id resampling_id
##  1:         1 -4.539772  0.2748538 0.0106758449            0.008 2021-09-02 13:45:27        1 <ResampleResult[20]>    pima classif.rpart.tuned            cv
##  2:         1 -7.559936  0.2631579 0.0005209086            0.007 2021-09-02 13:45:27        1 <ResampleResult[20]>    pima classif.rpart.tuned            cv
##  3:         1 -8.648543  0.2631579 0.0001753822            0.007 2021-09-02 13:45:27        1 <ResampleResult[20]>    pima classif.rpart.tuned            cv
##  4:         1 -6.297959  0.2631579 0.0018400560            0.008 2021-09-02 13:45:27        1 <ResampleResult[20]>    pima classif.rpart.tuned            cv
##  5:         1 -8.947182  0.2631579 0.0001301033            0.007 2021-09-02 13:45:27        1 <ResampleResult[20]>    pima classif.rpart.tuned            cv
##  6:         1 -8.067483  0.2631579 0.0003135715            0.008 2021-09-02 13:45:27        2 <ResampleResult[20]>    pima classif.rpart.tuned            cv
##  7:         1 -8.350241  0.2631579 0.0002363396            0.007 2021-09-02 13:45:27        2 <ResampleResult[20]>    pima classif.rpart.tuned            cv
##  8:         1 -5.913481  0.2631579 0.0027027622            0.007 2021-09-02 13:45:27        2 <ResampleResult[20]>    pima classif.rpart.tuned            cv
##  9:         1 -8.752513  0.2631579 0.0001580636            0.008 2021-09-02 13:45:27        2 <ResampleResult[20]>    pima classif.rpart.tuned            cv
## 10:         1 -2.768620  0.2573099 0.0627485302            0.008 2021-09-02 13:45:27        2 <ResampleResult[20]>    pima classif.rpart.tuned            cv
## 11:         2 -3.286175  0.2105263 0.0373966171            0.007 2021-09-02 13:45:27        1 <ResampleResult[20]>    pima classif.rpart.tuned            cv
## 12:         2 -4.124071  0.2456140 0.0161785202            0.007 2021-09-02 13:45:27        1 <ResampleResult[20]>    pima classif.rpart.tuned            cv
## 13:         2 -2.385855  0.2105263 0.0920102553            0.007 2021-09-02 13:45:27        1 <ResampleResult[20]>    pima classif.rpart.tuned            cv
## 14:         2 -3.880799  0.2046784 0.0206343371            0.007 2021-09-02 13:45:27        1 <ResampleResult[20]>    pima classif.rpart.tuned            cv
## 15:         2 -4.328644  0.2456140 0.0131854101            0.007 2021-09-02 13:45:27        1 <ResampleResult[20]>    pima classif.rpart.tuned            cv
## 16:         2 -4.394274  0.2456140 0.0123478454            0.007 2021-09-02 13:45:28        2 <ResampleResult[20]>    pima classif.rpart.tuned            cv
## 17:         2 -5.922306  0.2690058 0.0026790151            0.007 2021-09-02 13:45:28        2 <ResampleResult[20]>    pima classif.rpart.tuned            cv
## 18:         2 -7.331236  0.2690058 0.0006547636            0.007 2021-09-02 13:45:28        2 <ResampleResult[20]>    pima classif.rpart.tuned            cv
## 19:         2 -4.992721  0.2690058 0.0067871745            0.007 2021-09-02 13:45:28        2 <ResampleResult[20]>    pima classif.rpart.tuned            cv
## 20:         2 -3.362562  0.2105263 0.0346463879            0.008 2021-09-02 13:45:28        2 <ResampleResult[20]>    pima classif.rpart.tuned            cv
## 21:         3 -8.862942  0.2748538 0.0001415380            0.007 2021-09-02 13:45:28        1 <ResampleResult[20]>    pima classif.rpart.tuned            cv
## 22:         3 -7.347079  0.2748538 0.0006444723            0.007 2021-09-02 13:45:28        1 <ResampleResult[20]>    pima classif.rpart.tuned            cv
## 23:         3 -7.067612  0.2748538 0.0008522663            0.007 2021-09-02 13:45:28        1 <ResampleResult[20]>    pima classif.rpart.tuned            cv
## 24:         3 -8.044845  0.2748538 0.0003207512            0.006 2021-09-02 13:45:28        1 <ResampleResult[20]>    pima classif.rpart.tuned            cv
## 25:         3 -6.697964  0.2748538 0.0012334207            0.007 2021-09-02 13:45:28        1 <ResampleResult[20]>    pima classif.rpart.tuned            cv
## 26:         3 -6.530743  0.2748538 0.0014579225            0.007 2021-09-02 13:45:28        2 <ResampleResult[20]>    pima classif.rpart.tuned            cv
## 27:         3 -4.267553  0.2982456 0.0140160370            0.007 2021-09-02 13:45:28        2 <ResampleResult[20]>    pima classif.rpart.tuned            cv
## 28:         3 -6.352161  0.2748538 0.0017429764            0.007 2021-09-02 13:45:28        2 <ResampleResult[20]>    pima classif.rpart.tuned            cv
## 29:         3 -6.222573  0.2748538 0.0019841342            0.007 2021-09-02 13:45:28        2 <ResampleResult[20]>    pima classif.rpart.tuned            cv
## 30:         3 -3.570023  0.2865497 0.0281551963            0.007 2021-09-02 13:45:28        2 <ResampleResult[20]>    pima classif.rpart.tuned            cv
##     iteration        cp classif.ce  x_domain_cp runtime_learners           timestamp batch_nr      resample_result task_id          learner_id resampling_id