Where/how can I apply last_fit() after tuning? Then how do I apply extract_workflow() to new data without y values?

[pathos]

Below is my code so far from sample time series data.
My questions are:

1. How do I apply last_fit() from the result of tune_grid()?
2. How do I apply this best model to extra new data (say, df_test, or some imaginary df_test2, or even df_train, or any of the CV folds) with only yearr, monthh, etc., but no y to predict y?

#rm(list=ls())
# Libs
suppressPackageStartupMessages(library(readr))
suppressPackageStartupMessages(library(tidymodels))
suppressPackageStartupMessages(library(glmnet))
suppressPackageStartupMessages(library(lubridate))
suppressPackageStartupMessages(library(stacks))
suppressPackageStartupMessages(library(recipes))
suppressPackageStartupMessages(library(workflows))
suppressPackageStartupMessages(library(tune))
suppressPackageStartupMessages(library(yardstick))

# Data
df = read_csv('https://github.com/andrew-couch/Tidy-Tuesday/raw/master/Season%202/Data/example_retail_sales.csv', show_col_types = FALSE)

# Dates feature engineering
dff = df |>
  mutate(yearr = year(ds),
         monthh = month(ds),
         quarterr = quarter(ds),
         semesterr = semester(ds),
         ydayy = yday(ds)) |>
  select(-ds)

# Sample
df_split = initial_time_split(dff)
df_train = training(df_split)
df_test = testing(df_split)

folds = vfold_cv(df_train)

# Elastic net
spec_enet = linear_reg(mode = 'regression', penalty = tune(), mixture = tune()) |>
  set_engine('glmnet')

# Recipes
rec_df = df_train |>
  recipe(y ~ .)

# IGNORED
holidays <- c("AllSouls", "AshWednesday", "ChristmasEve", "Easter", 
              "ChristmasDay", "GoodFriday", "NewYearsDay", "PalmSunday")

rec_generic = rec_df |>
  #step_dummy(all_nominal()) |>
  #step_zv(all_numeric(), all_outcomes()) |>
  #step_normalize(all_numeric(), -all_outcomes()) |>
  #update_role(datee, new_role = 'ID') #|> # CHANGE
  #step_zv(all_predictors())
  step_holiday(ds, holidays = holidays) |> # NEED DATEE
  step_rm(ds)
# IGNORED

# Controls
metric = metric_set(rmse)

#grid_ctrl = control_stack_grid()
#res_ctrl = control_stack_resamples()

# Workflow
wf_enet = workflow() |>
  add_model(spec_enet) |>
  add_recipe(rec_df)

enet_grid = tibble(penalty = 10^seq(-4, -1, length.out = 30),
                   mixture = seq(from = 0, to = 1, by = (1/29)))

# Tune enet hyperparameters
tune_enet = wf_enet |>
  tune_grid(folds,
            grid = enet_grid,
            control = control_grid(save_pred = TRUE),
            metrics = metric_set(rmse))

enet_best = tune_enet |>
  show_best(metric = "rmse")

#autoplot(tune_enet)

tune_enet |>
  collect_predictions()
#> # A tibble: 6,570 x 7
#>    id       .pred  .row penalty mixture      y .config              
#>    <chr>    <dbl> <int>   <dbl>   <dbl>  <dbl> <chr>                
#>  1 Fold01 197658.    14  0.0001       0 150087 Preprocessor1_Model01
#>  2 Fold01 201480.    27  0.0001       0 192319 Preprocessor1_Model01
#>  3 Fold01 232054.    35  0.0001       0 202520 Preprocessor1_Model01
#>  4 Fold01 205302.    40  0.0001       0 196039 Preprocessor1_Model01
#>  5 Fold01 197658.    50  0.0001       0 192380 Preprocessor1_Model01
#>  6 Fold01 228232.    58  0.0001       0 222663 Preprocessor1_Model01
#>  7 Fold01 224410.    69  0.0001       0 222807 Preprocessor1_Model01
#>  8 Fold01 193837.    73  0.0001       0 207853 Preprocessor1_Model01
#>  9 Fold01 209123.    77  0.0001       0 245027 Preprocessor1_Model01
#> 10 Fold01 197658.    86  0.0001       0 220650 Preprocessor1_Model01
#> # ... with 6,560 more rows

^{Created on 2022-01-10 by the reprex package (v2.0.1)}

[hannah]

last_fit() emulates the process where, after determining the best model, the final fit on the entire training set is needed and is then evaluated on the test set.

So to use last_fit() you need to determine the best model and fill in the placeholders for the hyper-parameter (which you set with tune()) with actual values. You can extract those with select_best() and then "fill in" with finalize_workflow().

If you want to use the fitted workflow to predict on any data set, you can extract it with extract_workflow().

# select best values for the tuning parameter
best_hyperparameters <- select_best(tune_enet, metric = "rmse")

# finalize the workflow with those parameter values
final_wflow <- wf_enet %>%
  finalize_workflow(best_hyperparameters)

# `last_fit()` for fit on training, predict on test, and report performance
lf <- last_fit(final_wflow, df_split)
lf

# extract fitted workflow to predict on any data set (here df_test)
lf %>%
  extract_workflow() %>%
  predict(df_test)