rank_results() bug: Warning: Unknown or uninitialised column: `result`. AND(!?)✖ Column `metric` is not found.

[PathosEthosLogos]

suppressPackageStartupMessages(suppressWarnings({
  library(tidytable)
  library(tidymodels)

  conflicted::conflict_prefer_all(winner = "tidytable",
                                  quiet = TRUE)
}))

# Create training data with a bit of preprocessing included
df_train = attenu |> 
  select(-station)

# Set up model formula
rec_formula = df_train |> 
  recipe(dist ~ .) |> 
  prep()

# Set up cross validation for time series
resamples = df_train |> 
  sliding_index(event,
                lookback = 3)

# Set hyperparameter tuning algorithm specification
spec_enet = linear_reg(penalty = tune(), mixture = tune()) |> 
  set_engine('glmnet')

# For performance metrics
metrics = metric_set(huber_loss, rmse, mae, smape)

# Put the model steps together
wf_enet = workflow(preprocessor = rec_formula,
                   spec = spec_enet)

# Run the models and time it
system.time({
  tuned_model = wf_enet |> 
    tune_grid(grid = 5,
              resamples = resamples,
              metrics = metrics,
              control = control_resamples(save_pred = T))
})
#>    user  system elapsed 
#>   42.41    0.61   43.03

# Pick the best model
best_models = rank_results(tuned_model,
                           rank_metric = huber_loss,
                           select_best = T)
#> Warning: Unknown or uninitialised column: `result`.
#> Error in `dplyr::group_by()`:
#> ! Must group by variables found in `.data`.
#> ✖ Column `metric` is not found.
#> Backtrace:
#>     ▆
#>  1. ├─workflowsets::rank_results(...)
#>  2. │ └─workflowsets:::pick_metric(x, rank_metric)
#>  3. │   └─workflowsets:::collate_metrics(x)
#>  4. │     └─metrics %>% dplyr::group_by(metric) %>% ...
#>  5. ├─dplyr::summarize(...)
#>  6. ├─dplyr::group_by(., metric)
#>  7. └─dplyr:::group_by.data.frame(., metric)
#>  8.   └─dplyr::group_by_prepare(.data, ..., .add = .add, error_call = current_env())
#>  9.     └─rlang::abort(bullets, call = error_call)

^{Created on 2023-12-26 with reprex v2.0.2}

[PathosEthosLogos]

I was digging into it a bit. There are two places where rank_results() calls for the column result after running tune_grid().

First, it runs into problem at collate_metrics() in source misc.R.

Then it calls for the column twice again in rank_results()

types <- x %>% full_join(wflow_info, by = "wflow_id") %>% 
    mutate(is_race = map_lgl(**result**, ~inherits(.x,
                                               "tune_race")), num_rs = map_int(**result**x, get_num_resamples)) %>% 
    select(wflow_id, is_race, num_rs)
  ranked <- full_join(results, types, by = "wflow_id") %>% 
    filter(.metric == metric)

It seems that tune_grid() should be creating the column result but it does not. Was there some new update that changed this? It has been working.

[simonpcouch]

Thanks for the issue @PathosEthosLogos!

tune_grid() doesn't create a column called result and hasn't before; workflow_map() creates that column, and its entries are the outputs of each call to tune_grid().

The issue you're seeing here is that rank_results() is a function defined for workflow sets and doesn't know what to do with a tuning result like tuned_model. In workflowsets, we should better check inputs and error more informatively there.

If you want output similar to rank_results() output for tuning results, you could try select_best() or collect_metrics()!

[simonpcouch]

Related to #131.