workflow_idea.R

library(recipes)
library(parsnip)
library(workflows)

spline_cars <-
  recipe(mpg ~ ., data = mtcars) %>% 
  step_ns(disp, deg_free = 10)

bayes_lm <-
  linear_reg() %>% 
  set_engine("stan")

spline_cars_prepped <- prep(spline_cars, mtcars)
bayes_lm_fit <- fit(bayes_lm, mpg ~ ., data = juice(spline_cars_prepped))

car_wflow <-
  mtcars %>% 
  workflow() %>%
  add_initial_split(prop = 3/4) %>%
  add_vfold_cv(v = 10) %>% 
  add_recipe(
    recipe(mpg ~ .) %>%
    step_ns(disp, deg_free = 10)
  ) %>% 
  add_model(bayes_lm)

car_wflow %>%
  fit()

car_wflow %>%
  update_step(
    .rcp %>%
    step_arrange(cyl) %>%
    step_ns(disp, deg_free = 10)
  ) %>%
  fit()



rcp <- function (formula, data = NULL, ...) {
    f_funcs <- recipes:::fun_calls(formula)

    if (any(f_funcs == "-")) {
        rlang::abort("`-` is not allowed in a recipe formula. Use `step_rm()` instead.")
    }

    ## Extract all terms from formula
    all_terms <- as.character(attr(terms(formula), "variables"))
    all_terms <- setdiff(all_terms, "list")

    # Create empty df with column names
    data <- rep(list(numeric()), length(all_terms))
    names(data) <- all_terms
    data <- structure(data, class = "data.frame")

    # Continue with normal `recipe`
    args <- recipes:::form2args(formula, data)
    obj <- recipes:::recipe.data.frame(x = args$x,
                                       formula = NULL,
                                       vars = args$vars,
                                       roles = args$roles)
    
    class(obj) <- c("no_data", class(obj))
    obj
}

prep.no_data <- function(x,
                         training = NULL,
                         fresh = FALSE,
                         verbose = FALSE,
                         retain = TRUE,
                         strings_as_factors = TRUE,
                         ...) {

  if (nrow(x$template) == 0) {
    rlang::abort("No data has been specified in the recipe.")
  }
  class(x) <- setdiff(class(x), "no_data")
  
  ## Because we haven't really checked that all vars
  ## are in the data (since we hijacked `rcp` with an empty df)
  ## proceed with the normal recipe checks for raising an error
  # in case the supplied df is not compatible with the recipe
  args <- recipes:::form2args(recipe2formula(x), x$template)
  obj <- recipes:::recipe.data.frame(x = args$x,
                                     formula = NULL,
                                     vars = args$vars,
                                     roles = args$roles)


  ## The only extra thing to add is all the steps the user supplied
  ## earlier.
  obj$steps <- x$steps
  
  prep(obj,
       training = training,
       fresh = fresh,
       verbose = verbose, 
       retain = retain,
       strings_as_factors = strings_as_factors,
       ...)
}

my_rcp <-
  rcp(mpg ~ cyl + am + drat) %>%
  step_log(mpg, base = 10) %>%
  step_log(drat)

my_rcp$template <- mtcars
prep(my_rcp)

recipe2formula <- function(x) {
  x <- summary(x)
  x_vars <- x$variable[x$role == "predictor"]
  x_vars <- x_vars[!is.na(x_vars)]
  if (length(x_vars) == 0) 
    x_vars <- 1
  y_vars <- x$variable[x$role == "outcome"]
  y_vars <- y_vars[!is.na(y_vars)]
  if (length(y_vars) == 0) 
    y_vars <- ""
  x_vars <- paste0(x_vars, collapse = "+")
  y_vars <- paste0(y_vars, collapse = "+")
  as.formula(paste(y_vars, x_vars, sep = "~"))
}