tuning a model with a single glmnet penalty value

[topepo]

library(tidymodels)

set.seed(1)
rs <- vfold_cv(two_class_dat)

# error 1: At least one penalty value is required for glmnet.

thresh_tlr <- tailor() |> adjust_probability_threshold(threshold = tune())

glmn_spec <- 
  logistic_reg(penalty = 0.01) |> 
  set_engine("glmnet")

glmn_wflow <- workflow(Class ~ ., glmn_spec, thresh_tlr)

set.seed(2)
res <- 
  glmn_wflow |> 
  tune_grid(resamples = rs)
#> Error in `dplyr::mutate()`:
#> ℹ In argument: `model_stage = purrr::map(...)`.
#> Caused by error in `purrr::map()`:
#> ℹ In index: 1.
#> Caused by error in `no_penalty()`:
#> ! At least one penalty value is required for glmnet.

# This does not help

glmn_spec <- 
  logistic_reg(penalty = 0.01) |> 
  set_engine("glmnet", path_values = 10^seq(-4:0))

glmn_wflow <- workflow(Class ~ ., glmn_spec, thresh_tlr)

set.seed(2)
res <- 
  glmn_wflow |> 
  tune_grid(resamples = rs)
#> Error in `dplyr::mutate()`:
#> ℹ In argument: `model_stage = purrr::map(...)`.
#> Caused by error in `purrr::map()`:
#> ℹ In index: 1.
#> Caused by error in `no_penalty()`:
#> ! At least one penalty value is required for glmnet.

^{Created on 2026-01-12 with reprex v2.1.1}

This appears to be a min_grid() issue. Even though the penalty is not tuned, it is checking to see if there is a penalty column in the grid.

[topepo]

This does work though:

library(tidymodels)

set.seed(1)
rs <- vfold_cv(two_class_dat)

thresh_tlr <- tailor() |> adjust_probability_threshold(threshold = tune())

cart_spec <- 
  decision_tree() |> 
  set_mode("classification")

cart_wflow <- workflow(Class ~ ., cart_spec, thresh_tlr)

set.seed(2)
res <- 
  cart_wflow |> 
  tune_grid(resamples = rs)
res
#> # Tuning results
#> # 10-fold cross-validation 
#> # A tibble: 10 × 4
#>    splits           id     .metrics          .notes          
#>    <list>           <chr>  <list>            <list>          
#>  1 <split [711/80]> Fold01 <tibble [30 × 5]> <tibble [0 × 4]>
#>  2 <split [712/79]> Fold02 <tibble [30 × 5]> <tibble [0 × 4]>
#>  3 <split [712/79]> Fold03 <tibble [30 × 5]> <tibble [0 × 4]>
#>  4 <split [712/79]> Fold04 <tibble [30 × 5]> <tibble [0 × 4]>
#>  5 <split [712/79]> Fold05 <tibble [30 × 5]> <tibble [0 × 4]>
#>  6 <split [712/79]> Fold06 <tibble [30 × 5]> <tibble [0 × 4]>
#>  7 <split [712/79]> Fold07 <tibble [30 × 5]> <tibble [0 × 4]>
#>  8 <split [712/79]> Fold08 <tibble [30 × 5]> <tibble [0 × 4]>
#>  9 <split [712/79]> Fold09 <tibble [30 × 5]> <tibble [0 × 4]>
#> 10 <split [712/79]> Fold10 <tibble [30 × 5]> <tibble [0 × 4]>

^{Created on 2026-02-02 with reprex v2.1.1}

[topepo]

I think the issue is related to how we S3 dispatch to min_grid().

• In the glmnet case, we dispatch to min_grid.logistic_reg(), and it expects there to be a model parameter, so it fails.

• For the CART model, we dispatch to min_grid.model_spec() which just returns the grid with an empty placeholder:

> min_grid(extract_spec_parsnip(wflow), model_grid) |>
+     dplyr::select(dplyr::all_of(model_param))
# A tibble: 1 × 0

and then it will move on to the predict and post stages.

I think that there are two options:

1. Add an escape check in each min_grid() method to detect if there are no model parameters. This sounds tedious but there is effectively only one. As of now, the ones listed here just call fit_max_value(). However, that isn't great since outside contributors may not do that.
2. Circumvent proper S3 dispatch here and force the usage of min_grid.model_spec() when there are no model parameters. Again, this is also unhelpful for outside contributors.

I found four instances of min_grid() methods outside of tune: false.alarm, offsetreg, tabnet, and tidyclust.

I'll make changes for the second approach and test with reveps. false.alarm isn't on CRAN so I'll test locally.