Caret in R: Set number of cores for allowParallel?

5
votes

I am using R's caret package, and in the training function (train) I use the allowParallel Parameter, which works. However, it uses all of the cores, and since the training runs on my local PC I would rather leave one core for myself to be able to work while training models. Is there any way to do this?

From what I've gathered it seems that different model types might use different parallelization packages. I am working on windows, so I guess it's not using doMC (where I know how to set the number of cores...)

2
rparallel-processingr-caret
Try this: github.com/lgreski/datasciencectacontent/blob/master/markdown/… – tushaR
I tried it, but this seems to be specific for randomforest only. With method="ranger" (which I use) it just uses all 8 cores... – Thomas

2 Answers

9
votes

So after more research, I found a way to use the number of cores I want: train has the option to directly specify the number of cores to use with num.threads = 7 (for 7 out of 8 cores)

rf_model<-train(Target~., data = df_tree_train, method = "ranger",
                trControl = trainControl(method = "oob"
                                       , verboseIter  = TRUE
                                       , allowParallel = TRUE
                                       , classProbs = TRUE
                )
                , verbose = T
                , tuneGrid = tuneGrid
                , num.trees = 50
                , num.threads = 7  # <- This one
)
2
votes

I'm surprised that:

library("doParallel")
registerParallel(parallel::detectCores() - 1)

does not do it. Maybe there is recursive parallelism that does not acknowledge the above. You could try with the doFuture package:

library("doFuture")
registerDoFuture()
plan(multiprocess, workers = availableCores() - 1)

which should protected against unwanted nested parallelism.