Create a stack model with the package mlr3

Question

I'm using the mlr3pipelines package to It define a pipeline object named "stack" which is used for stacking ensemble learning.

However, I'm unable to find an alternative to the 'po' function, it's not suitable for stacking models as I get this error:

Error: Element with key 'stack' not found in DictionaryPipeOp!

What is a good way to create stacking?

Here is the code whole code if it helps:

super_train <- function(data, n_models = 10, cpus = 32) {


  train <- as.data.table(data[[1]])
  test <- as.data.table(data[[2]])
        
  task <- TaskClassif$new(id = "my_task", backend = train, target = "CR")
  
  base_learners <- lapply(1:n_models, function(i) {
    lrn_lgbm <- lrn("classif.lightgbm", predict_type = "response", objective = "binary")
    lrn_lgbm$param_set$values$num_iterations <- 1000 # set a large number of trees
    lrn_lgbm$param_set$values$early_stopping_rounds <- 50 # early stopping
    lrn_lgbm$param_set$values$verbose <- -1 # suppress lightGBM's output
    lrn_lgbm
  })
  
  meta_learners <- lapply(1:n_models, function(i) {
    lrn_rf <- lrn("classif.randomForest", predict_type = "response", mtry = 3, ntree = 500, importance = "gini")
    return(lrn_rf)
  })
  
  base_resampling <- rsmp("cv", folds = 10)
  
  # set up stacking
  stack <- mlr3pipelines::po(.obj = "stack", base_learners = base_learners, metalearner = lrn("classif.gam"))
  
  # set up stacking resampling
  stack_resampling <- rsmp("cv", folds = 10)
  
  ps_lgbm <- ParamSet$new(params = list(
    makeDiscreteParam("booster", values = c("gbtree","gblinear")), 
    makeNumericParam("eta", lower = 0.01, upper = 0.06), 
    makeIntegerParam("max_depth", lower = 3L, upper = 10L), 
    makeNumericParam("min_child_weight", lower = 2L, upper = 8L), 
    makeNumericParam("subsample", lower = 0.5, upper = 0.5), 
    makeNumericParam("colsample_bytree", lower = 0.5, upper = 0.8)
  ))
  
  
  # set up hyperparameters tuning
  base_tuning <- tune_grid(
    base_learners,
    task,
    resampling = base_resampling,
    measure = msr("classif.auc"),
    control = ctrl_tuning(
      method = "random",
      repeats = 10,
      show_info = FALSE,
      save_pred = "none"
    ),
    par.set = ps_lgbm
  )
  
  base_models <- lapply(base_tuning$learners, function(lrn) {
    train(lrn, task)
  })
  
  # train the meta-learners
  meta_models <- lapply(meta_learners, function(lrn) {
    train(stack, task, stack_resampling)
  })
  
  best_meta_model <- meta_models[[which.max(sapply(meta_models, function(m) m$aggregated_result$score))]]
  
  best_base_models <- lapply(base_models, function(m) m$learner)
  final_model <- stack$train(best_base_models, best_meta_model, task)
        
  return(final_model)
}

score 1 · Answer 1 · answered Mar 27 '23 at 14:40

There are two types of dictionaries from which you can retrieve objects to define graphs:

PipeOps can be retrieved using the po() function. See an overview here
Predefined Graphs can be retrieved using the ppl() function. See an overview here

You seem to be looking for the graph that can be retrieved using ppl("stacking"). Its manual page can be found here.

Create a stack model with the package mlr3

1 Answers1