cvpreds.R

# Extract cross-validated predicted values (in order of original rows)
h2o.cvpreds <- function(object, single_col = TRUE) {
  
  # TO DO: Check that object is an H2OModel
  # TO DO: Check that keep_cross_validation_predictions = TRUE in the model
  # TO DO: Need to add support for returning a multiclass prediction and binary (full frame: predict, p0, p1)
  # TO DO: Remove family variable and just check class(object) directly
  
  # Need to extract family from model object
  if (class(object) == "H2OBinomialModel") family <- "binomial"
  if (class(object) == "H2OMulticlassModel") family <- "multinomial"
  if (class(object) == "H2ORegressionModel") family <- "gaussian"
  
  # return the frame_id of the resulting 1-col Hdf of cvpreds for learner l
  V <- object@allparameters$nfolds
  
  if (single_col) {
    if (family %in% c("bernoulli", "binomial")) {
      predlist <- sapply(1:V, function(v) h2o.getFrame(object@model$cross_validation_predictions[[v]]$name)[,3], simplify = FALSE)
    } else {
      predlist <- sapply(1:V, function(v) h2o.getFrame(object@model$cross_validation_predictions[[v]]$name)$predict, simplify = FALSE)
    }
    cvpred_sparse <- h2o.cbind(predlist)  #N x V H2OFrame with rows that are all zeros, except corresponding to the v^th fold if that rows is associated with v
    cvpreds <- apply(cvpred_sparse, 1, sum)    
  } else {
    stop("single_col = FALSE not yet implemented")
  }
  return(cvpreds)
}