The LearnerRegrAuto is an automated machine learning (AutoML) system for regression tasks. It combines preprocessing, a switch between multiple learners, and hyperparameter tuning to find the best model for the given task.
Value
Object of class R6::R6Class and LearnerRegrAuto.
Debugging
Set options(bbotk.debug = TRUE) to run the tuning in the main session.
Set encapsulate_learner = FALSE to remove encapsulation of the learner.
Set encapsulate_mbo = FALSE to catch no errors in mbo.
Parameters
- learner_timeout
(
integer(1))
Timeout for training and predicting with a single learner.- n_threads
(
integer(1))
Number of threads used for training a single learner.- n_cpu
(named
integer())
Number of CPUs a single training of a learner uses, named by learner id, e.g.c(xgboost = 1). Overrides the default of the learner. Must be at least 1 because the worker always runs on the CPU. Currently informational only; the number of threads is controlled byn_threads.- n_gpu
(named
integer())
Number of GPUs a single training of a learner uses, named by learner id, e.g.c(xgboost = 1). Overrides the default of the learner. Can only be 0 or 1 for now. The torch learners, TabPFN, TabFM, and fastai default to 1; all other learners default to 0. Only effective when"cuda"is part ofdevices; otherwise every learner is trained on the CPU. When the requirements are mixed and the daemons of the mirai compute profiles"mlr3automl_cpu"and"mlr3automl_gpu"are set up withrush::rush_plan(), the search space is partitioned into a cpu and a gpu subspace which are tuned with mlr3mbo::TunerADBOSubspaces. The workers of a profile only ever propose and evaluate points of the subspace of that profile, so the number of workers per subspace is the number of workers of its profile. Otherwise the cpu and gpu learners are tuned in a single search space with mlr3mbo::TunerAsyncMbo.- memory_limit
(
integer(1))
Memory limit for training a single learner in MB. The limit is shared across the parallel workers, i.e. divided by the number of workers.- devices
(
character())
Devices to use for model training. Possible values are"cpu"and"cuda". If"cuda", learners with an_gpurequirement of 1 are trained on a GPU, while the remaining learners stay on the CPU.- large_data_size
(
integer(1))
Threshold for the data set size (number of rows times number of columns) above which large-data rules apply. Beyond this threshold the number of parallel workers is reduced to a quarter, rounded up, and each remaining worker is given proportionally more threads and memory. When the workers are distributed over mirai compute profiles, the number of workers of every profile is reduced, but every profile keeps at least one worker. The"mlr3automl_gpu"profile is exempt because its number of workers is fixed by the number of GPUs. It keeps its workers, threads, and memory limit so that the gpu learners do not claim the CPU cores and the memory that are freed on the cpu profiles.- small_data_size
(
integer(1))
Threshold value for the data set size (rows) from which special rules apply.- small_data_resampling
(mlr3::Resampling)
Resampling strategy to use for model training on small data sets.- initial_design_default
(
logical(1))
Whether to use the default design of the learner.- initial_design_set
(
integer(1))
Number of points to use for the initial design set.- initial_design_size
(
integer(1))
Size of the random, sobol or lhs initial design.- initial_design_type
(
character(1))
Type of the initial design used for mbo. Possible values are"lhs","sobol","random"."lhs"uses a Latin Hypercube Sampling design."sobol"uses a Sobol sequence design."random"uses a random design.- initial_design_fraction
(
numeric(1))
Fraction of the budget to use for the initial design. When the search space is partitioned into a cpu and a gpu subspace, the remaining points of both designs are dropped, because every compute profile has its own queue.- resampling
(mlr3::Resampling)
Resampling strategy used for tuning.- terminator
(bbotk::Terminator)
Terminator criterion for tuning.- measure
(mlr3::Measure)
Measure used for tuning.- callbacks
(mlr3tuning::CallbackAsyncTuning)
Callbacks used for tuning.- store_benchmark_result
(
logical(1))
Whether to store the benchmark result.- store_models
(
logical(1))
Whether to store the models.- encapsulate_learner
(
logical(1))
Whether to encapsulate the learner. Change toFALSEto debug.- encapsulate_mbo
(
logical(1))
Whether to encapsulate the tuning. Change toFALSEto debug.- check_learners
(
logical(1))
Whether to check if the learners are compatible with the task. Change toFALSEto debug.
Python learners
Python learners like TabPFN, TabFM, and fastai run via reticulate and
therefore need a Python installation with their required packages.
There are two ways to provide it:
Do nothing and let
reticulate::py_require()install the required packages into an ephemeral virtual environment automatically.Point the
RETICULATE_PYTHONenvironment variable to a Python installation that has the required packages installed.
We recommend option 2 when running on many workers,
as it avoids the overhead of downloading and installing the packages on each worker.
Use install_python_learners() to create a conda environment with the required packages and set
RETICULATE_PYTHON to the returned Python binary.
The TabPFN learner additionally requires the TABPFN_TOKEN environment variable to download the model weights.
Super classes
mlr3::Learner -> LearnerAuto -> LearnerRegrAuto
Methods
LearnerRegrAuto$new()
Creates a new instance of this R6 class.
Usage
LearnerRegrAuto$new(id = "regr.auto", learner_ids, rush = NULL)Arguments
id(
character(1))
Identifier for the new instance.learner_ids(
character())
Learner that should be used.rushrush::Rush
Rush instance.
Examples
packages = c(
"mlr3extralearners", "catboost", "ranger", "callr", "mlr3torch",
"glmnet", "kknn", "MASS", "lightgbm", "e1071", "xgboost"
)
if (mlr3misc::require_namespaces(packages, quietly = TRUE)) {
learner = lrn("regr.auto")
learner
}