如何调优tidymodels中recipe步骤的自定义参数？

• 语法位置错误：你把digits = tune("digits")写在了step_mutate()的顶级参数位置，step_mutate()会将其识别为你要新增一个名为digits的列，赋值为tune()返回的调用对象，这就是你报错「digits必须为向量而非调用对象」的直接原因。
• 调优规则不符：tune()仅能标记recipe step本身公开的可调节参数，无法识别嵌套在step内部表达式里的普通函数（如round()）参数，所以即使你把digits = tune("digits")放到round()括号内，tune_grid()也检测不到待调参数。

你可以通过自定义recipe step的方式将round的digits参数公开为可调节参数，可运行的完整示例如下：

library(tidyverse)
library(tidymodels)

# 自定义step_round：指定列四舍五入，digits设为可调参数
step_round <- function(
  recipe, 
  ..., 
  role = NA, 
  trained = FALSE, 
  columns = NULL, 
  digits = 0, 
  skip = FALSE,
  id = rand_id("round")
) {
  add_step(
    recipe,
    step_round_new(
      terms = enquos(...),
      role = role,
      trained = trained,
      columns = columns,
      digits = digits,
      skip = skip,
      id = id
    )
  )
}

step_round_new <- function(terms, role, trained, columns, digits, skip, id) {
  step(
    subclass = "round",
    terms = terms,
    role = role,
    trained = trained,
    columns = columns,
    digits = digits,
    skip = skip,
    id = id
  )
}

# 定义step预处理逻辑
prep.step_round <- function(x, training, info = NULL, ...) {
  col_names <- recipes_eval_select(x$terms, training, info)
  
  step_round_new(
    terms = x$terms,
    role = x$role,
    trained = TRUE,
    columns = col_names,
    digits = x$digits,
    skip = x$skip,
    id = x$id
  )
}

# 定义step执行逻辑
bake.step_round <- function(object, new_data, ...) {
  col_names <- object$columns
  for (col in col_names) {
    new_data[[col]] <- round(new_data[[col]], digits = object$digits)
  }
  new_data
}

# 注册可调参数
tunable.step_round <- function(x, ...) {
  tibble::tibble(
    name = "digits",
    call_info = list(list(pkg = "dials", fun = "digits")),
    source = "recipe",
    component = "step_round",
    component_id = x$id
  )
}

# 调整后的recipe，使用自定义step标记digits为可调参数
glmnet_recipe <- 
  recipe(formula = Species ~ Sepal.Length + Petal.Length, data = iris) %>% 
  step_round(Sepal.Length, digits = tune("digits")) %>% 
  step_zv(all_predictors()) %>% 
  step_normalize(all_predictors(), -all_nominal()) 

glmnet_spec <- 
  multinom_reg(penalty = 0, mixture = 0) %>% 
  set_mode("classification") %>% 
  set_engine("glmnet") 

glmnet_workflow <- 
  workflow() %>% 
  add_recipe(glmnet_recipe) %>% 
  add_model(glmnet_spec) 

glmnet_grid <- expand_grid(digits = c(0, 1))

# 运行调优
glmnet_tune <- 
  tune_grid(glmnet_workflow, resamples = vfold_cv(iris), grid = glmnet_grid)

运行上述代码即可正常识别到待调的digits参数，无报错完成调优。

内容的提问来源于stack exchange，提问作者AlbertRapp