Parallel Cross-Validation Example in R

gistfile1.r

# Parallel Cross-Validation Example

library(foreach)    # install.packages('foreach')
library(caret)      # install.packages('caret', dependencies = c("Depends", "Suggests"))
library(doParallel) # install.packages('doParallel')
registerDoParallel(makeCluster(4)) # Use 4 cores for parallel CV

data <- read.csv(…) # Assuming this is your dataset 

cv <- caret::createFolds(nrow(data), k=10, list=T) # Create 10 folds

# 'dopar' here would run this on multiple threads (change to just 'do' for synchronous runs)
results <- foreach(fold = cv) %dopar% {
  
  # Get the fold data where 'fold' is nothing more than a list of indexes for test observations in the data
  data.train <- data[-fold] # Get the opposite of the test observations to train on
  data.test <- data[fold]
  
  # Fit the model and make predictions
  fit <- glm(response ~ var1 + var2, data=data.train, family='binomial')
  y.pred <- predict(fit, newdata=data.test)
  y.true <- data.test$response
  
  # Return 2x2 table of predictions vs actual values as well as the fit model (so you could check coefficients)
  list(accuracy=table(y.pred, y.true, dnn=c('predicted', 'actual')),  model=fit)
}

# Results is a list, so you could process it to extract the accuracies like this:

library(dplyr) # install.packages('dplyr')
foreach(fold.result=results, fold.num=icount(), .combine=rbind) %do%{
  as.data.frame(fold.result$accuracy) %>% 
  mutate(actual = as.character(actual), predicted = as.character(predicted)) %>%
  group_by(Correct=actual == predicted) %>%  # Determine whether or not the predictions were correct
  summarise(Count=sum(Freq)) %>%  # Count the number of such predictions
  mutate(Fold=fold.num)  # Add the fold number to this result
}

### Returns something like:
Correct   Count   Fold
TRUE       123       1
FALSE      329       1
TRUE       923       2
…

So totally 10 folds in each of 4 cores will operate parallely, right. Totally 40?

@rahultam: each core will handle a fold thus minimizing the running time.