Merge be7a6f8 into 662654c

NAMESPACE

@@ -1,6 +1,7 @@
 # Generated by roxygen2: do not edit by hand
 
 S3method(print,ParConfig)
+S3method(trainLearner,HyperoptWrapper)
 export(downloadParConfig)
 export(downloadParConfigs)
 export(generateHyperControl)
@@ -21,6 +22,7 @@ export(getParConfigParVals)
 export(getTaskDictionary)
 export(hyperopt)
 export(makeHyperControl)
+export(makeHyperoptWrapper)
 export(makeParConfig)
 export(setHyperControlMeasures)
 export(setHyperControlMlrControl)
@@ -40,3 +42,4 @@ import(lhs)
 import(methods)
 import(mlr)
 import(stringi)
+importFrom(utils,getFromNamespace)

R/HyperoptWrapper.R

@@ -0,0 +1,58 @@
+#' @title Fuse learner with mlrHyperopt tuning.
+#'
+#' @description
+#' Fuses an mlr base learner with mlrHyperopt tuning.
+#' Creates a learner object, which can be used like any other learner object.
+#' If the train function is called on it, \code{\link{hyperopt}} is invoked to select an optimal set of hyperparameter values.
+#' Finally, a model is fitted on the complete training data with these optimal hyperparameters and returned.
+#'
+#' @template arg_learner
+#' @template arg_parconfig
+#' @template arg_hypercontrol
+#' @template arg_showinfo
+#' @return [\code{\link{Learner}}].
+#' @export
+#' @family tune
+#' @family wrapper
+#' @examples
+#' \donttest{
+#' task = makeClassifTask(data = iris, target = "Species")
+#' lrn = makeLearner("classif.svm")
+#' lrn = makeHyperoptWrapper(lrn)
+#' mod = train(lrn, task)
+#' print(getTuneResult(mod))
+#' # nested resampling for evaluation
+#' # we also extract tuned hyper pars in each iteration
+#' r = resample(lrn, task, cv3, extract = getTuneResult)
+#' getNestedTuneResultsX(r)
+#' }
+#' @importFrom utils getFromNamespace
+makeHyperoptWrapper = function(learner, par.config = NULL, hyper.control = NULL, show.info = getMlrOptions()$show.info) {
+  learner = checkLearner(learner)
+  id = stri_paste(learner$id, "hyperopt", sep = ".")
+  # more or less just an empty dummy control
+  makeTuneControl = getFromNamespace("makeTuneControl", "mlr")
+  makeOptWrapper = getFromNamespace("makeOptWrapper", "mlr")
+  control = makeTuneControl(same.resampling.instance = FALSE, cl = "TuneControlHyperopt")
+  x = makeOptWrapper(id = id, learner = learner, resampling = NULL, measures = NULL, par.set = NULL, bit.names = character(0L), bits.to.features = function(){}, control = control, show.info = show.info, learner.subclass = c("HyperoptWrapper", "TuneWrapper"), model.subclass = "TuneModel")
+  x$hyper.control = hyper.control
+  x$par.config = par.config
+  return(x)
+}
+
+#' @export
+trainLearner.HyperoptWrapper = function(.learner, .task, .subset = NULL,  ...) {
+  .task = subsetTask(.task, .subset)
+  or = hyperopt(task = .task, learner = .learner$next.learner, par.config = .learner$par.config, hyper.control = .learner$hyper.control)
+  lrn = or$learner
+  or$learner = NULL
+  if ("DownsampleWrapper" %in% class(.learner$next.learner) && !is.null(.learner$control$final.dw.perc) && !is.null(getHyperPars(lrn)$dw.perc) && getHyperPars(lrn)$dw.perc < 1) {
+    messagef("Train model on %f on data.", .learner$control$final.dw.perc)
+    lrn = setHyperPars(lrn, par.vals = list(dw.perc = .learner$control$final.dw.perc))
+  }
+  m = train(lrn, .task)
+  makeChainModel = getFromNamespace("makeChainModel", "mlr")
+  x = makeChainModel(next.model = m, cl = "TuneModel")
+  x$opt.result = or
+  return(x)
+}

R/hyperopt.R

@@ -11,9 +11,7 @@
 #'  If no learner is given the learner referenced in the \code{par.config} will be used, if available.
 #' @template arg_parconfig
 #' @template arg_hypercontrol
-#' @param show.info [\code{logical(1)}]\cr
-#'   Print verbose output on console?
-#'   Default is set via \code{\link{configureMlr}}.
+#' @template arg_showinfo
 #' @return [\code{\link[mlr]{TuneResult}}]
 #' @import mlr
 #' @examples

man-roxygen/arg_showinfo.R

@@ -0,0 +1,3 @@
+#' @param show.info [\code{logical(1)}]\cr
+#'   Print verbose output on console?
+#'   Default is set via \code{\link{configureMlr}}.

tests/testthat/test_HyperoptWrapper.R

@@ -0,0 +1,30 @@
+context("HyperoptWrapper")
+
+test_that("hyperoptWrapper works", {
+  mlr::configureMlr(show.info = FALSE, show.learner.output = FALSE)
+  lrn = makeLearner("classif.svm")
+  lrn2 = makeHyperoptWrapper(lrn)
+  task = iris.task
+  res = resample(learner = lrn2, task = task, resampling = cv2, extract = getTuneResult)
+  expect_class(res$extract[[1]], "TuneResult")
+  expect_data_frame(getNestedTuneResultsX(res))
+  expect_data_frame(getNestedTuneResultsOptPathDf(res))
+
+  # some random workflow
+  # triggers Random Search
+  hyper.control = makeHyperControl(
+    mlr.control = makeTuneControlRandom(maxit = 10),
+    resampling = makeResampleDesc("Holdout"),
+    measures = list(auc)
+  )
+  par.config = generateParConfig(lrn, sonar.task)
+  par.set = getParConfigParSet(par.config)
+  par.set = filterParams(par.set, ids = "cost")
+  par.config = setParConfigParSet(par.config, par.set)
+  par.config = setParConfigParVals(par.config, par.vals = list())
+  lrn2 = makeHyperoptWrapper(learner = lrn, par.config = par.config, hyper.control = hyper.control)
+  res = resample(learner = lrn2, task = sonar.task, resampling = cv2, extract = getTuneResult)
+  expect_class(res$extract[[1]], "TuneResult")
+  expect_data_frame(getNestedTuneResultsX(res))
+  expect_data_frame(getNestedTuneResultsOptPathDf(res))
+})