Error in varImp[, "%IncMSE"] on RFE rfFuncs regression subset fit

[floriandeboissieu]

varImp on the final regression fit of an RFE with rfFuncs gives an error. It comes from rfFuncs$fit.
It can be fixed modifying rfFuncs$fit, see example below.

library(caret)
library(randomForest)
set.seed(123)

# Simulate data
x <- matrix(rnorm(1e3), ncol = 10)  # Feature matrix
colnames(x) <- paste0('x', seq_len(10))
y_cont <- runif(1e2)                            # Continuous response
ctrl <- rfeControl(functions = rfFuncs, method = 'cv'
fit_cont <- rfe(x, y_cont, sizes = seq_len(3), rfeControl = ctrl)

varImp(fit_cont$fit)

# adding importance computation for selected model
rfFuncs$fit <- function (x, y, first, last, ...) 
{
  loadNamespace("randomForest")
  randomForest::randomForest(x, y, importance = (first|last), ...)
}

ctrl <- rfeControl(functions = rfFuncs, method = 'cv')
fit_cont <- rfe(x, y_cont, sizes = seq_len(3), rfeControl = ctrl)
varImp(fit_cont$fit)

[topepo]

Looks good. Thanks