R/botCallWrapper.R

In ja-thomas/OMLbots: Bot to run experiments on OpenML

#' @title runBot
#' Main function to run the bot. This evaluates batch.size configurations. 
#' Task, learner and the actual configuration are defined by the corresponding functions. 
#' @param batch.size number of configurations to evaluate in parallel
#' @param sample.learner.fun function to sample a learner from tunePair
#' @param sample.task.fun sample a OML task
#' @param sample.configuration given a lrn sample a configuration
#' @param min.resources minimal resources to start benchmark (list with elements walltime and memory)
#' @param max.resources maximum resources allowed for each evaluation (list with elements walltime and memory)
#' @param lrn.ps.sets of available learners with matching parameter sets
#' @param upload should the run be uploaded to OpenML
#' @param path where should the registry be created, defaults to tempdir()
#' @export
runBot = function(batch.size, sample.learner.fun = sampleRandomLearner, 
  sample.task.fun = sampleSimpleTask, sample.configuration.fun = sampleRandomConfiguration, 
  min.resources = NULL, max.resources = NULL, lrn.ps.sets = simple.lrn.par.set, upload = FALSE,
  path = NA, extra.tag = "botV1") {
  
  task = sample.task.fun()
  task$task = getOMLTask(task$id)
  messagef("Selected OML task: %s (id %s)", task$name, task$id)
  
  lrn = sample.learner.fun(lrn.ps.sets)
  messagef("Selected learner: %s", lrn$learner$short.name)
  
  par = sample.configuration.fun(batch.size, lrn$param.set)
  attr(par, "additional.tags") = c(attr(par, "additional.tags"), paste0("sciBenchV", packageDescription("rscimark")$Version))
  attr(par, "additional.tags") = c(attr(par, "additional.tags"), extra.tag)
  print("Selected configurations:")
  print(par)
  
  if (getLearnerPackages(lrn$learner) == "ranger" & extra.tag != "referenceV1") {
    #Avoid fails by adjusting mtry
    p = ncol(task$task$input$data.set$data) - 1
    par$mtry = ceiling(p * par$mtry)
    n = nrow(task$task$input$data.set$data)
    par$min.node.size = round(2^(log(n, 2) * par$min.node.size))
  }
  
  if (getLearnerPackages(lrn$learner) == "xgboost") {
    #Convert factors to numeric
    target = task$task$input$data.set$target.features
    cols = which(colnames(task$task$input$data.set$data) != target)
    task$task$input$data.set$data = data.frame(sapply(dummy.data.frame(task$task$input$data.set$data[,cols], sep = "_._"), as.numeric), 
      task$task$input$data.set$data[,target,drop = F])
    colnames(task$task$input$data.set$data) = make.names(colnames(task$task$input$data.set$data))
  }
  
  evalConfigurations(lrn$learner, task = task, par, min.resources, max.resources, upload = upload, path = path)
}