properly resolve obj alias objective

[nipnipj]

Error message Error in init(env): For early stopping, valids must have at least one element is produced when trying to train a multi-class model with

lgb_model <-
  boost_tree(
      stop_iter = 1000
  ) %>%
  set_engine(   
     "lightgbm",
      validation = 0.2,
      objective = 'multiclass',
      metric="none",
      eval = 'auc_mu',
      num_class = n_classes
  )

[simonpcouch]

Thanks for the issue. Training for a multi-class task with early stopping indeed works fine:

library(bonsai)
#> Loading required package: parsnip
library(modeldata)

data(penguins)

lgb_model <-
  boost_tree("classification", stop_iter = 1000) %>%
  set_engine("lightgbm", validation = 0.2, eval = 'auc_mu')

fit(
  lgb_model,
  formula = species ~ flipper_length_mm + island, 
  data = penguins
)
#> parsnip model object
#> 
#> <lgb.Booster>
#>   Public:
#>     add_valid: function (data, name) 
#>     best_iter: 14
#>     best_score: 0.973090277777778
#>     current_iter: function () 
#>     dump_model: function (num_iteration = NULL, feature_importance_type = 0L) 
#>     eval: function (data, name, feval = NULL) 
#>     eval_train: function (feval = NULL) 
#>     eval_valid: function (feval = NULL) 
#>     finalize: function () 
#>     initialize: function (params = list(), train_set = NULL, modelfile = NULL, 
#>     lower_bound: function () 
#>     params: list
#>     predict: function (data, start_iteration = NULL, num_iteration = NULL, 
#>     raw: NA
#>     record_evals: list
#>     reset_parameter: function (params, ...) 
#>     rollback_one_iter: function () 
#>     save: function () 
#>     save_model: function (filename, num_iteration = NULL, feature_importance_type = 0L) 
#>     save_model_to_string: function (num_iteration = NULL, feature_importance_type = 0L) 
#>     set_train_data_name: function (name) 
#>     to_predictor: function () 
#>     update: function (train_set = NULL, fobj = NULL) 
#>     upper_bound: function () 
#>   Private:
#>     eval_names: auc_mu
#>     get_eval_info: function () 
#>     handle: lgb.Booster.handle
#>     higher_better_inner_eval: TRUE
#>     init_predictor: NULL
#>     inner_eval: function (data_name, data_idx, feval = NULL) 
#>     inner_predict: function (idx) 
#>     is_predicted_cur_iter: list
#>     name_train_set: training
#>     name_valid_sets: list
#>     num_class: 3
#>     num_dataset: 2
#>     predict_buffer: list
#>     set_objective_to_none: FALSE
#>     train_set: lgb.Dataset, R6
#>     train_set_version: 1
#>     valid_sets: list

^{Created on 2023-04-24 with reprex v2.0.2}

bonsai handles setting the obj argument to lgb.train, and will warn if it's passed:

library(bonsai)
#> Loading required package: parsnip
library(modeldata)

data(penguins)

lgb_model <-
  boost_tree("classification", stop_iter = 1000) %>%
  # note `object = .`, since R partial-matches `obj` to `object`
  set_engine(object = ., "lightgbm", obj = "multiclass", validation = 0.2, eval = 'auc_mu')

fit(
  lgb_model,
  formula = species ~ flipper_length_mm + island, 
  data = penguins
)
#> Warning: The following argument(s) are guarded by bonsai and will not be passed
#> to `lgb.train`: obj
#> parsnip model object
#> ... [clipped]

^{Created on 2023-04-24 with reprex v2.0.2}

but fails to do so when that argument is passed by its alias objective. Will leave this open to make sure we handle that alias properly in the next release.

[nipnipj]

Your suggestion worked. I'm getting another error, not sure if it is related, tho.

data(penguins, package= "modeldata")
data <- penguins
rec <- data %>% 
  recipe(species ~ flipper_length_mm + island) 

lgbm_model <-
  boost_tree(
    stop_iter = 100,
    trees = 1000, 
    learn_rate =  0.07, 
  ) %>%
  set_engine(
    object = .,
    "lightgbm",
    obj = 'multiclass',
    metric="none",
    eval = 'multi_logloss',
    validation = 0.2
 ) %>% 
  set_mode("classification)

learner_lgbm <- workflow() %>% 
  add_recipe(rec) %>% 
  add_model(lgbm_model)

cv_folds <- data %>%
  vfold_cv(v = 2, repeats = 1, strata  = species) 

lg <- fit_resamples(
  object = learner_lgbm,
  resamples = cv_folds,
  metrics = metric_set(roc_auc)
)

Then → A | warning: The following argument(s) are guarded by bonsai and will not be passed to lgb.train: obj There were issues with some computations A: x2 error appears.

[simonpcouch]

See:

bonsai handles setting the obj argument to lgb.train, and will warn if it's passed

No need to pass that argument!👍