[R] Using doMC to run parallel SVM grid search?

Brian Feeny bfeeny at mac.com
Wed Nov 21 21:25:22 CET 2012


Has anyone used doMC to speed up an SVM grid search?  I am considering doing like so:

library(doMC)
registerDoMC()
foreach (i=0:3) %dopar% {
	tuned_part1 <- tune.svm(label~., data = trainset, gamma = 10^(-10:-6), cost = 10^(-1:1))
	tuned_part2 <- tune.svm(label~., data = trainset, gamma = 10^(-5:0),     cost = 10^(-1:1))
	tuned_part3 <- tune.svm(label~., data = trainset, gamma = 10^(1:-5),     cost = 10^(-1:1))
	tuned_part4 <- tune.svm(label~., data = trainset, gamma = 10^(5:10),    cost = 10^(-1:1))
}


I have a Quad Core processor, so if I understand correctly the above could split that up across the cores.

My goal would be a coarse grid search, not sure if the above parameters are good for that, it just seemed like 
some good starting points.

I would just manually look at each of the resulting files, although it would be cool if it resulted in an instance variable
being set of the best values. 

Has anyone used doMC for something like this?  Is there a better library to potentially use than doMC for doing 
something like splitting up an SVM grid search over multiple cores?

Brian




More information about the R-help mailing list