R/getCaretParamSet.R

In berndbischl/mlr: Machine Learning in R

#' @title Get tuning parameters from a learner of the caret R-package.
#'
#' @description
#' Constructs a grid of tuning parameters from a learner of the `caret`
#' R-package. These values are then converted into a list of non-tunable
#' parameters (`par.vals`) and a tunable
#' [ParamHelpers::ParamSet] (`par.set`), which can be used by
#' [tuneParams] for tuning the learner. Numerical parameters will
#' either be specified by their lower and upper bounds or they will be
#' discretized into specific values.
#'
#' @param learner (`character(1)`)\cr
#'   The name of the learner from `caret`
#'   (cf. <https://topepo.github.io/caret/available-models.html>). Note that the
#'   names in `caret` often differ from the ones in `mlr`.
#' @param length (`integer(1)`)\cr
#'   A length / precision parameter which is used by `caret` for
#'   generating the grid of tuning parameters. `caret` generates either as
#'   many values per tuning parameter / dimension as defined by `length`
#'   or only a single value (in case of non-tunable `par.vals`).
#' @param task ([Task])\cr
#'   Learning task, which might be requested for creating the tuning grid.
#' @param discretize (`logical(1)`)\cr
#'   Should the numerical parameters be discretized? Alternatively, they will
#'   be defined by their lower and upper bounds. The default is `TRUE`.
#' @return (`list(2)`). A list of parameters:
#' \itemize{
#'   \item{`par.vals`} contains a list of all constant tuning parameters
#'   \item{`par.set`} is a [ParamHelpers::ParamSet], containing all the configurable
#'   tuning parameters
#' }
#' @export
#' @examples
#' if (requireNamespace("caret") && requireNamespace("mlbench")) {
#'   library(caret)
#'   classifTask = makeClassifTask(data = iris, target = "Species")
#'
#'   # (1) classification (random forest) with discretized parameters
#'   getCaretParamSet("rf", length = 9L, task = classifTask, discretize = TRUE)
#'
#'   # (2) regression (gradient boosting machine) without discretized parameters
#'   library(mlbench)
#'   data(BostonHousing)
#'   regrTask = makeRegrTask(data = BostonHousing, target = "medv")
#'   getCaretParamSet("gbm", length = 9L, task = regrTask, discretize = FALSE)
#' }
getCaretParamSet = function(learner, length = 3L, task, discretize = TRUE) {

  td = getTaskData(task, target.extra = TRUE)
  caret.grid = caret::getModelInfo(learner)[[learner]]$grid(
    x = td$data, y = td$target, len = length)

  # transfer caret parameters into mlr parameters
  params = lapply(colnames(caret.grid), function(i) {
    par.vals = sort(unique(caret.grid[, i]))
    cl = class(par.vals)
    if (cl == "factor") {
      if (all(levels(par.vals) %in% c("TRUE", "FALSE"))) {
        par.vals = as.logical(as.character(par.vals))
        cl = "logical"
      } else {
        par.vals = as.character(par.vals)
        cl = "character"
      }
    }
    if (discretize) {
      cl = "character"
    }
    switch(cl,
      character = makeDiscreteParam(id = i, values = par.vals),
      logical = makeLogicalParam(id = i),
      numeric = makeNumericParam(id = i, lower = min(par.vals), upper = max(par.vals)),
      integer = makeIntegerParam(id = i, lower = min(par.vals), upper = max(par.vals))
    )
  })
  names(params) = colnames(caret.grid)

  # are the parameters configurable or are the values unique?
  is.tunable = vlapply(params, function(x) {
    (!is.null(x$values) && length(x$values) > 1) |
      (!is.null(x$lower) && !is.null(x$upper) && (x$upper > x$lower))
  })

  # define par.vals (if existing)
  if (all(is.tunable)) {
    par.vals = NULL
  } else {
    par.vals = lapply(caret.grid[!is.tunable], function(x) {
      if (is.factor(x)) {
        x = as.character(x)
      }
      return(x[1L])
    })
    # convert integerish variables into integer
    par.vals[vlapply(par.vals, testIntegerish)] =
      lapply(par.vals[vlapply(par.vals, testIntegerish)], as.integer)
  }

  return(list(par.vals = par.vals,
    par.set = do.call(makeParamSet, params[is.tunable])))
}