Time Quantization PipeOp

[pfistfl]

I am just gonna leave some code here until I have time to improve and create a PR.
The goal is to reduce the number of unique time points during training in order to make prediction faster.
Would guess this should perhaps live here instead of mlr3pipelines?

PipeOpTimeQuantiles = R6::R6Class("PipeOpTimeQuantiles",
  inherit = mlr3pipelines::PipeOpTaskPreproc,
  public = list(

    #' @description
    #'   Initialize a Multi-Calibration PipeOp.
    #'
    #' @param id [`character`] \cr
    #'   The `PipeOp`'s id. Defaults to "mcboost".
    #' @param param_vals [`list`] \cr
    #'   List of hyperparameters for the `PipeOp`.
    initialize = function(id = "timequantiles", param_vals = list()) {
      param_set = paradox::ParamSet$new(list(paradox::ParamInt$new("num_uniques", lower = 0L, upper = Inf, default =500L, tags = "train")))
      param_set$valus = list("num_uniques" = 500L)

      super$initialize(id, param_set = param_set, param_vals = param_vals)
    }
  ),
  private = list(
    .train_task = function(task) {
      times = task$times()
      qs = self$param_set$get_values()$num_uniques
      if (length(unique(times)) < qs) {
        newtimes = times
      } else {
        qt = round(unique(quantile(unique(times), seq(from = 0, to = 1, length.out = qs), type=2)))
      }
      # assign closest match
      new_target = data.table()
      set(new_target, j = task$target_names[[1]], value = qt[apply(abs(outer(qt, times, "-")), 2, which.min)])
      task$cbind(new_target)
    }
  )
  )
)