Bug in predict.train

[jknowles]

I think this is causing our issue over on zachmayer/caretEnsemble#180

Here's a simple minimal reproducible example:

library(caret)
myGBM <- train(x=iris[, 1:2], y = iris[, 5], 
tuneLength = 1, method = "gbm", 
trControl = trainControl(method = "cv", number =2, 
savePredictions = "final", classProbs = FALSE), verbose = FALSE)

predict(myGBM) # results in error
predict(myGBM, newdata = NULL) # results in error

Here is the error message:

Error in object$var.levels[[i]] : subscript out of bounds
Called from: predict.gbm(modelFit, newdata, type = "response", 
n.trees = modelFit$tuneValue$n.trees)

This behavior is inconsistent, as the following works:

myRF <- train(x=iris[, 1:2], y = iris[, 5], 
tuneLength = 1, method = "rf", 
trControl = trainControl(method = "cv", number =2, 
savePredictions = "final", classProbs = FALSE), verbose = FALSE)

predict(myRF) # results in no error
predict(myRF, newdata = NULL) # results in no error

@zachmayer and @topepo -- any advice about investigating whether this is just limited to the predict.gbm implementation? How might we investigate it further?

[topepo]

In the last version, I changed predict.train to avoid using extractPrediction and go straight to predictionFunction. In doing so I didn't account for the use of predict(obj) with no newdata.

It's generally not a good idea to do that but I will fix it since it was working before I "fixed" the other issue =]

[jknowles]

Thanks Max. If it's not a good idea to do predict(modelobj) without newdata, perhaps in the case where our users do not specify newdata in caretEnsemble we should modify the predict call to get the fitted values from the train objects instead of doing using the predict function.

What do you think @zachmayer ?

[zachmayer]

@jknowles How about if it's not a good idea to call predict.train without newdata, we disable calling predict.caretEnsemble without new data?

[topepo]

Do these changes fix the issue? The examples that you've shown don't show errors now.