predict.train fails with single observation in new data for gbms

[jknowles]

I have found an odd bug on both the current CRAN and dev versions of caret which is related to zachmayer/caretEnsemble#171. The error is that predict.train(model, newdata, type = 'prob') does not work consistently for single case values of newdata for gbm models. Here's an MWE using the iris dataset:

library(caret)
gbmMod <- train(iris[, 1:2], iris[, 5],  method = "gbm", 
     trControl=trainControl(method="cv", number=2, 
                     savePredictions=TRUE, classProbs=TRUE))

head(predict(gbmMod, type = "prob"))

This nicely produces:

     setosa   versicolor    virginica
1 0.9997469 1.541028e-04 9.896410e-05
2 0.9940483 3.459335e-03 2.492320e-03
3 0.9999370 3.382596e-05 2.913693e-05
4 0.9999131 5.922540e-05 2.762628e-05
5 0.9998697 7.918245e-05 5.111574e-05
6 0.9992466 5.095148e-04 2.438444e-04

And if we pass in newdata, that works too: predict(gbmMod, type = "prob", newdata = iris[100:105, c(1:2)]):

 setosa versicolor virginica
1 3.347258e-03 0.77843367 0.2182191
2 7.632194e-04 0.24274868 0.7564881
3 3.165440e-04 0.57012163 0.4295618
4 2.054595e-05 0.06004269 0.9399368
5 8.409241e-05 0.68628701 0.3136289
6 6.959093e-05 0.14138125 0.8585492

But, if newdata only has 1 row: predict(gbmMod, type = "prob", newdata = iris[100, c(1:2)]):

Error in `[.data.frame`(out, , obsLevels, drop = FALSE) : 
  undefined columns selected

[topepo]

It is a bug. I've fixed the model file. Until the next release, you can source this file and use the new code via:

gbmMod <- train(iris[, 1:2], iris[, 5],  method = modelInfo, 
                trControl=trainControl(method="cv", number=2, 
                                       savePredictions=TRUE, classProbs=TRUE))

[jknowles]

Awesome thanks Max!