patch params argument with xgboost engine in boost_tree()

[simonpcouch]

Closes #774, closes #459. Related to #411.

The goal of this PR is to ensure that folks can pass arguments that live in the param argument to xgb.train. The new docs section is probably the best place to start for big picture here. :)

Some notes-to-self that helped me keep track of arguments:

• xgb.train routes arguments from its dots to the params argument. For simplicity of our own argument routing, this PR proposes we pass non-main params arguments to the dots rather than params.
• xgb_train previously took objective as a main argument. This argument is eventually routed to the params argument in xgb.train, so again for simplicity of our machinery, I deleted the objective argument to xgb_train so that it will be passed through dots. This change isn’t user-facing and doesn’t change the way the argument is passed in practice.
• We’d prefer that users pass elements of the params argument directly to set_engine rather than as part of the params list so that tune machinery works “out-of-the-box.” This PR now raises a warning when users supply a non-empty params argument (though now correctly handles patching the params argument with main boost_tree arguments).

An additional unit test that arguments passed via ... to xgb_train can indeed be tuned; will PR to extratests after this is merged:

# with this devel parsnip:
library(tidymodels)

ctrl$verbosity <- 0L
#> Error in ctrl$verbosity <- 0L: object 'ctrl' not found

# define base model spec
spec_base <-
  boost_tree() %>%
  set_mode("regression") %>%
  set_engine("xgboost", eval_metric = tune())

res <-
  tune_grid(
    spec_base,
    preprocessor = mpg ~ .,
    resamples = vfold_cv(mtcars, v = 6),
    param_info =
      extract_parameter_set_dials(spec_base) %>%
      update(eval_metric = new_qual_param(
        type = "character",
        values = c("rmse", "logloss"),
        label = c(eval_metric = "Evaluation Metric")
      )),
    control = control_grid(save_workflow = TRUE)
  )

res_eval_values <-
  collect_metrics(res) %>%
  pull(eval_metric)

all(c("logloss", "rmse") %in% res_eval_values)
#> [1] TRUE

^{Created on 2022-08-15 by the reprex package (v2.0.1)}

[topepo]

I'm ready to merge this but it breaks a package. Can you investigate and, if needed, add a PR (or issue) to https://github.com/Harrison4192/autostats/issues?

autostats

• Version: 0.3.0
• GitHub: https://github.com/Harrison4192/autostats
• Source code: https://github.com/cran/autostats
• Date/Publication: 2022-02-09 08:50:02 UTC
• Number of recursive dependencies: 231

Run revdep_details(, "autostats") for more info

Newly broken

• checking examples ... ERROR

  Running examples in ‘autostats-Ex.R’ failed
  The error most likely occurred in:
  
  > ### Name: get_params
  > ### Title: get params
  > ### Aliases: get_params get_params.xgb.Booster
  > 
  > ### ** Examples
  > 
  > 
  ...
  > iris_dummies %>%
  +   tidy_formula(target = Petal.Length) -> p_form
  > 
  > iris_dummies %>%
  +   tidy_xgboost(p_form, mtry = .5, trees = 5L, loss_reduction = 2, sample_size = .7) -> xgb
  Warning: `early_stop` was reduced to 4.
  Warning: `early_stop` was reduced to 4.
  Error in if (objective == "multi:softmax") { : argument is of length zero
  Calls: %>% ... withCallingHandlers -> %>% -> tidy_predict -> tidy_predict.xgb.Booster
  Execution halted
  

[simonpcouch]

Yup, sure thing.

[simonpcouch]

Submitted a fix!👍

[github-actions]

This pull request has been automatically locked. If you believe you have found a related problem, please file a new issue (with a reprex: https://reprex.tidyverse.org) and link to this issue.