fix bug in TuneWrapper for param in used in predict function, issue 2472

R/TuneWrapper.R

@@ -74,8 +74,16 @@ trainLearner.TuneWrapper = function(.learner, .task, .subset = NULL,  ...) {
 
 #' @export
 predictLearner.TuneWrapper = function(.learner, .model, .newdata, ...) {
-  lrn = setHyperPars(.learner$next.learner, par.vals = .model$learner.model$opt.result$x)
-  predictLearner(lrn, .model$learner.model$next.model, .newdata, ...)
+  # setHyperPars not used because par.vals are not acessed anymore
+  arglist = list(.learner = .learner$next.learner, .model = .model$learner.model$next.model, .newdata = .newdata)
+  arglist = insert(arglist, list(...))
+
+  # get x from opt result and only select those that are used for predition
+  opt.x = .model$learner.model$opt.result$x
+  ps = getParamSet(.learner)
+  ns = Filter(function(x) ps$pars[[x]]$when %in% c("both", "predict"), getParamIds(ps))
+  arglist = insert(arglist, opt.x[ns])
+  do.call(predictLearner, arglist)
 }
 
 #' @export

tests/testthat/test_base_TuneWrapper.R

@@ -120,3 +120,34 @@ test_that("TuneWrapper with glmnet (#958)", {
   expect_error(pred, NA)
 })
 
+test_that("TuneWrapper respects train parameters (#2472)", {
+
+  # make task with only 0 as y
+  tsk = makeRegrTask("dummy", data = data.frame(y = rep(0L, 100), x = rep(1L, 100)), target = "y")
+
+  ps = makeParamSet(
+    makeNumericLearnerParam("p1", when = "train", lower = 0, upper = 10),
+    makeNumericLearnerParam("p2", when = "predict", lower = 0, upper = 10),
+    makeNumericLearnerParam("p3", when = "both", lower = 0, upper = 10)
+  )
+
+  lrn = makeLearner("regr.__mlrmocklearners__4", predict.type = "response", p1 = 10, p2 = 10, p3 = 10)
+  # prediction of this learner is always
+  # train_part = p1 + p3
+  # y = train_part + p2 + p3
+  # therefore p1 = p2 = p3 = 0 is the optimal setting
+  # we set params to bad values p1 = p2 = p3 = 10, meaning |y_hat-y| would be 40
+
+  lrn2 = makeTuneWrapper(lrn, resampling = makeResampleDesc("Holdout"),
+    par.set = ps,
+    control = makeTuneControlGrid(resolution = 2L))
+  mod = train(lrn2, tsk)
+  # we expect that the optimal parameters are found by the grid search.
+  expect_equal(mod$learner.model$opt.result$x, list(p1 = 0, p2 = 0, p3 = 0))
+  expect_true(mod$learner.model$opt.result$y == 0)
+  pred = predict(mod, tsk)
+  # we expect that the optimal parameter are also applied for prediction and therefore y_hat = p1+p2+p3+p3 should be 0
+  expect_true(all(pred$data$response == 0))
+})
+
+