Resampling with store_models = TRUE is slow

[be-marc]

library(mlr3learners)

task = tsk("pima")
learner = lrn("classif.ranger", num.trees = 5000)
resampling = rsmp("cv", folds = 10)

system.time(resample(task, learner, resampling))
 
#    user  system elapsed 
#  25.037   1.413  24.330 

system.time(resample(task, learner, resampling, store_models = TRUE))

#    user  system elapsed 
#  27.560   1.701  27.138 

Saving the models take 3 seconds longer. While tuning with store_models it took almost 8 seconds longer. The ranger models get quite large ~60MB. The 3 seconds get lost when creating this data.table:

mlr3/R/resample.R

Lines 121 to 131 in 2c8734a

	data = data.table(
	task = list(task),
	learner = grid$learner,
	learner_state = map(res, "learner_state"),
	resampling = list(resampling),
	iteration = seq_len(n),
	prediction = map(res, "prediction"),
	uhash = UUIDgenerate(),
	param_values = map(res, "param_values"),
	learner_hash = map_chr(res, "learner_hash")
	)

Profiling of this part