prototype code to update parsnip and recipe objects with final parameter values

finalize.R

finalize_model <- function(x, param) {
  if (!inherits(x, "model_spec")) {
    stop("`x` should be a parsnip model specification.")
  }
  # check for nrow > 1 and tibble/list
  pset <- parameters(x)
  if (tibble::is_tibble(param)) {
    param <- as.list(param)
  }
  
  param <- param[names(param) %in% pset$id]
  
  # make this a function:
  discordant <- dplyr::filter(pset, id != name & id %in% names(param))
  if (nrow(discordant) > 0) {
    for (i in 1:nrow(discordant)) {
      names(param)[ names(param) == discordant$id[i] ] <- discordant$name[i]
    }
  }
  # end
  rlang::exec(update, object = x, !!!param)
}


finalize_recipe <- function(x, param) {
  if (!inherits(x, "recipe")) {
    stop("`x` should be a recipe.")
  }
  # check for nrow > 1 and tibble/list
  pset <- 
    dials::parameters(x) %>% 
    dplyr::filter(id %in% names(param) & source == "recipe")

  if (tibble::is_tibble(param)) {
    param <- as.list(param)
  }
  
  param <- param[names(param) %in% pset$id]
  param <- param[pset$id]
  pset <- split(pset, pset$component_id)
  for (i in seq_along(pset)) {
    x <- complete_steps(param[[i]], pset[[i]], x)
  }
  x
}

complete_steps <- function(param, pset, object) {
  # find the corresponding step in the recipe
  step_ids <- purrr::map_chr(object$steps, ~ .x$id)
  step_index <- which(pset$component_id == step_ids)
  tmp <- object$steps[[step_index]]
  
  # if the id is not the argument name:
  list_para <- list(param)
  names(param) <- pset$name
  
  tmp <- rlang::exec(update, object = tmp, !!!param)
  object$steps[[step_index]] <- tmp
  object
}

# ------------------------------------------------------------------------------

library(tidymodels)
library(tune)
library(mlbench)

data(PimaIndiansDiabetes)

# ------------------------------------------------------------------------------

set.seed(151)
pima_rs <- vfold_cv(PimaIndiansDiabetes, repeats = 1)

pca_rec <- 
  recipe(diabetes ~ ., data = PimaIndiansDiabetes) %>% 
  step_pca(all_predictors(), num_comp = tune("# components"))

tree_mod <-
  decision_tree(cost_complexity = tune("cp"), min_n = tune()) %>%
  set_mode("classification") %>%
  set_engine("rpart")

roc_vals <- metric_set(roc_auc)

set.seed(3625)
pima_res <-
  tune_grid(pca_rec,
            tree_mod,
            resamples = pima_rs,
            metrics = roc_vals)

final_param <- select_best(pima_res)

final_rec <- finalize_recipe(pca_rec, final_param)
final_mod <- finalize_model(tree_mod, final_param)

Or it could be built into a workflow as you suggest in the vignette - then the final solution would be even more elegant. It could be something along these lines:

pca_rec <- 
  recipe(diabetes ~ ., data = PimaIndiansDiabetes) %>% 
  step_pca(all_predictors(), num_comp = tune("# components"))

tree_mod <-
  decision_tree(cost_complexity = tune("cp"), min_n = tune()) %>%
  set_mode("classification") %>%
  set_engine("rpart")

wrkflow <- 
    workflow() %>%
    add_model(pca_rec) %>% 
    add_recipe(tree_mod)

roc_vals <- metric_set(roc_auc)

set.seed(3625)
pima_res <-
  tune_grid(wrkflow,
                   resamples = pima_rs,
                   metrics = roc_vals)

final_param <- select_best(pima_res)

wrkflow_fin <- wrkflow %>%
         update(select_best(pima_res)) %>% ### could be directly done here 
         fit(diabetes ~ ., data = PimaIndiansDiabetes)

I'm writing those methods now. I might make it into the update() methods but that would require some minor changes to parsnip.

Any progress here? It seems to me select_best() misses step-tune-data. Am I wrong?

This is already implemented in tune; look at those functions instead.

R 4.x, Ubuntu 16.x LTS

My fault, thank you.