Feature request: allow `initial` in workflow_map() to take as input an entire `workflow set` object

[MatthieuStigler]

A great feature of the tune_bayes() and tune_sim_anneal() is the initial argument that can be the output from a previous tune_grid() call.

Unfortunately, this feature becomes quite cumbersome for workflow_set objects? This stack post shows that passing workflow_map(initial =tune_grid) won't work, and the suggested answer is a rather verbose and manual recursive call:

workflow_set() %>% 
 ## add for one model
  option_add(
    id = "recipe_lasso",
    initial = tune_grid_res %>% extract_workflow_set_result("recipe_lasso")
  ) %>% 
  ## repeat for every model....
  option_add(..) %>% 
  option_add(..) %>% 
  workflow_map()

this is also slightly counter-intuitive, as the option_add() is added to the initial workflow_set, not to the workflow_map() call?

Would it be possible to have an explicit argument initial in workflow_map() that would possibly take as input a previous workflow_set and.... map it to the underlying modelling function?

Thanks!

Below is an example, together with a manual solution function option_add_initial() which is quite... untidy!

library(tidymodels)

# Load and prepare data
ames_data <- ames[,sapply(ames, class) %in% c("integer", "numeric")]

# Define a recipe
recipe <- recipe(Sale_Price ~ ., data = ames_data) %>%
  step_normalize(all_predictors())

# Define models
lasso_model <- linear_reg(penalty = tune(), mixture = 1) %>%
  set_engine("glmnet")

rf_model <- rand_forest(min_n = tune(), trees = 100) %>%
  set_engine("ranger") %>%
  set_mode("regression")

# Create workflows
lasso_wf <- workflow() %>%
  add_model(lasso_model) %>%
  add_recipe(recipe)

rf_wf <- workflow() %>%
  add_model(rf_model) %>%
  add_recipe(recipe)

cross_val <- vfold_cv(ames_data, v = 5)

tune_grid_res <- 
  workflow_set(
    preproc = list(recipe),
    models = list(lasso = lasso_model, rf = rf_model)) %>%
  workflow_map("tune_grid", resamples = cross_val, grid = 10)


option_add_initial <- function(workflow_set, workflow_with_grid_results){

  res <- workflow_set
  for(i in 1:nrow(workflow_with_grid_results)){
    id_i <- workflow_with_grid_results$wflow_id[[i]]
    res <- res %>% 
      option_add(
        id = id_i,
        initial = workflow_with_grid_results %>% extract_workflow_set_result(id_i)
      ) 
  }
  res
}

tune_bayes_res <- workflow_set(
  preproc = list(recipe),
  models = list(lasso = lasso_model, rf = rf_model)
  ) %>% 
  option_add_initial(workflow_with_grid_results = tune_grid_res) %>% 
  workflow_map("tune_bayes", resamples = cross_val)
#> ! All of the rmse values were identical. The Gaussian process model cannot be
#>   fit to the data. Try expanding the range of the tuning parameters.
#> → A | error:   Infinite values of the Deviance Function, 
#>                            unable to find optimum parameters 
#> 
#> There were issues with some computations   A: x1
#> ✖ Optimization stopped prematurely; returning current results.
#> There were issues with some computations   A: x1There were issues with some computations   A: x1

tune_bayes_res
#> # A workflow set/tibble: 2 × 4
#>   wflow_id     info             option    result   
#>   <chr>        <list>           <list>    <list>   
#> 1 recipe_lasso <tibble [1 × 4]> <opts[2]> <tune[+]>
#> 2 recipe_rf    <tibble [1 × 4]> <opts[2]> <tune[+]>

^{Created on 2025-03-21 with reprex v2.1.1}