brulee_logistic_reg() expected 'mixture' to be a double.

Machine Learning and Modeling
,
1

I tried using tidymodels and grid to tune brulee logistic regression and it gave me error that expeced "mixture" tuning to be a double. But if I do hyper_log_grid$mixture %>% class() its output is 'numeric'. Numeric implies double right? Man, I even added %>% as.double() at the input.

Here's my code:

wf_brulee <-
  workflow() %>% 
  add_recipe(X_rec) %>% 
  add_model(brulee_spec)

grid_ctrl <-
   control_grid(
      save_pred = TRUE,
      parallel_over = "everything",
      save_workflow = TRUE
   )

hyper_log_grid <- grid_latin_hypercube(
  penalty(),
  mixture(range = as.double(c(0.0,1.0))),
  original = T,
  size = 10
)

grid_brulee_hyper <- 
  wf_brulee %>% 
  tune_grid(
    resamples = X_folds,
    grid = hyper_log_grid,
    control = grid_ctrl
  )

grid_brulee_hyper %>% 
  show_best(metric = "roc_auc")

Here's the error:

collect_notes(x = grid_brulee_hyper) %>% 
  pull(note)

[1] "\033[1m\033[33mError\033[39m in \033[38;5;254m`check_double()`\033[39m:\033[22m\n\033[33m!\033[39m brulee_logistic_reg() expected 'mixture' to be a double."
  [2] "\033[1m\033[33mError\033[39m in \033[38;5;254m`check_double()`\033[39m:\033[22m\n\033[33m!\033[39m brulee_logistic_reg() expected 'mixture' to be a double."
  [3] "\033[1m\033[33mError\033[39m in \033[38;5;254m`check_double()`\033[39m:\033[22m\n\033[33m!\033[39m brulee_logistic_reg() expected 'mixture' to be a double."
  [4] "\033[1m\033[33mError\033[39m in \033[38;5;254m`check_double()`\033[39m:\033[22m\n\033[33m!\033[39m brulee_logistic_reg() expected 'mixture' to be a double."
2

Thanks for the post!

It's difficult to troubleshoot here without a reproducible example, though a couple things that come to mind:

  • We've just recently put out new releases of parsnip and brulee—do you have parsnip 1.0.2 and brulee 0.2.0 installed?
  • Does this work as expected with allow_par = FALSE in control_grid()?
  • Does the error persist if you set grid = 10 as opposed to specifying the latin hypercube design yourself?

If this doesn't do the trick, here's some code to get you started with a reproducible example:

library(tidymodels)
library(brulee)

head(two_class_dat)

brulee_spec <- 
  logistic_reg(penalty = tune(), mixture = tune()) %>%
  set_engine("brulee")

X_rec <- recipe(Class ~ ., data = two_class_dat)

X_folds <- bootstraps(two_class_dat, 10)

wf_brulee <-
  workflow() %>% 
  add_recipe(X_rec) %>% 
  add_model(brulee_spec)

grid_ctrl <-
  control_grid(
    save_pred = TRUE,
    parallel_over = "everything",
    save_workflow = TRUE
  )

hyper_log_grid <- grid_latin_hypercube(
  penalty(),
  mixture(range = as.double(c(0.0,1.0))),
  original = T,
  size = 10
)

grid_brulee_hyper <- 
  wf_brulee %>% 
  tune_grid(
    resamples = X_folds,
    grid = hyper_log_grid,
    control = grid_ctrl
  )

grid_brulee_hyper %>% 
  show_best(metric = "roc_auc")
3

Hi simoncouch, thanks for the reply.

I didn't have the chance to apply the allow_par = FALSE, but I vaguely remember that I was still getting error with grid = 10. But don't take my words for it, since I decided to ditch the brulee logistic regression for C.50 decision tree.