'How to set the parameters grids correctly when tuning the workflowset with tidymodels?

I try to use tidymodels to tune the workflow with recipe and model parameters. When tuning a single workflow there is no problem. But when tuning a workflowsets with several workflows it always fails. Here is my codes:

# read the training data
train <- read_csv("../../train.csv")
train <- train %>% 
    mutate(
      id = row_number(),
      across(where(is.double), as.integer),
      across(where(is.character), as.factor),
      r_yn = fct_relevel(r_yn, "yes")) %>% 
  select(id, r_yn, everything())

# setting the recipes

# no precess
rec_no <- recipe(r_yn ~ ., data = train) %>%
  update_role(id, new_role = "ID")

# downsample: tuning the under_ratio
rec_ds_tune <- rec_no %>% 
  step_downsample(r_yn, under_ratio = tune(), skip = TRUE, seed = 100) %>%
  step_nzv(all_predictors(), freq_cut = 100)

# setting the models

# randomforest
spec_rf_tune <- rand_forest(trees = 100, mtry = tune(), min_n = tune()) %>%
  set_engine("ranger", seed = 100) %>%
  set_mode("classification")

# xgboost
spec_xgb_tune <- boost_tree(trees = 100, mtry = tune(), tree_depth = tune(), learn_rate = tune(), min_n = tune()) %>% 
   set_engine("xgboost") %>% 
   set_mode("classification")

# setting the workflowsets
wf_tune_list <- workflow_set(
  preproc = list(no = rec_no, ds = rec_ds_tune),
  models = list(rf = spec_rf_tune, xgb = spec_xgb_tune),
  cross = TRUE)

# finalize the parameters, I'm not sure it is correct or not
rf_params <- spec_rf_tune %>% parameters() %>% update(mtry = mtry(c(1, 15)))
xgb_params <- spec_xgb_tune %>% parameters() %>% update(mtry = mtry(c(1, 15)))
ds_params <- rec_ds_tune %>% parameters() %>% update(under_ratio = under_ratio(c(1, 5)))

wf_tune_list_finalize <- wf_tune_list %>% 
  option_add(param = ds_params, id = c("ds_rf", "ds_xgb")) %>% 
  option_add(param = rf_params, id = c("no_rf", "ds_rf")) %>% 
  option_add(param = xgb_params, id = c("no_xgb", "ds_xgb"))

I check the option in wf_tune_list_finalize it shows:

> wf_tune_list_finalize$option
[[1]]
a list of options with names:  'param'

[[2]]
a list of options with names:  'param'

[[3]]
a list of options with names:  'param'

[[4]]
a list of options with names:  'param'

Then I tune this workflowset:

# tuning the workflowset
cl <- makeCluster(detectCores())
registerDoParallel(cl)
wf_tune_race <- wf_tune_list_finalize %>%
  workflow_map(fn = "tune_race_anova",
               seed = 100,
               resamples = cv_5,
               grid = 3,
               metrics = metric_auc,
               control = control_race(parallel_over = "everything"), 
               verbose = TRUE)
stopCluster(cl)

The verbose messages shows that there is something wrong with my parameters in the workflow ds_rf and ds_xgb:

i 1 of 4 tuning:     no_rf
i Creating pre-processing data to finalize unknown parameter: mtry
�� 1 of 4 tuning:     no_rf (1m 44.4s)
i 2 of 4 tuning:     no_xgb
i Creating pre-processing data to finalize unknown parameter: mtry
�� 2 of 4 tuning:     no_xgb (28.9s)
i 3 of 4 tuning:     ds_rf
x 3 of 4 tuning:     ds_rf failed with: Some tuning parameters require finalization but there are recipe parameters that require tuning. Please use `parameters()` to finalize the parameter ranges.
i 4 of 4 tuning:     ds_xgb
x 4 of 4 tuning:     ds_xgb failed with: Some tuning parameters require finalization but there are recipe parameters that require tuning. Please use `parameters()` to finalize the parameter ranges.

The result is:

> wf_tune_race
# A workflow set/tibble: 4 x 4
  wflow_id info             option      result        
  <chr>    <list>           <list>      <list>        
1 no_rf    <tibble [1 x 4]> <wrkflw__ > <race[+]>     
2 no_xgb   <tibble [1 x 4]> <wrkflw__ > <race[+]>     
3 ds_rf    <tibble [1 x 4]> <wrkflw__ > <try-errr [1]>
4 ds_xgb   <tibble [1 x 4]> <wrkflw__ > <try-errr [1]>

What's more, although the no_rf and no_xgb have tuning results, I find that the range of mtry in these two workflows is not the range I set above, that means the parameters range setting step is totally fail. I have followed the tutorials from https://www.tmwr.org/workflow-sets.html and https://workflowsets.tidymodels.org/ but still have no ideas.

So how to set both the recipe and model parameters correctly when tuning workflowsets?

The train.csv in my code is here: https://github.com/liuyifeikim/Some-data



Solution 1:[1]

I have modified the parameter setting step, and the tuning result is correct now:

# setting the parameters on each workflow seperately
no_rf_params <- wf_set_tune_list %>% 
  extract_workflow("no_rf") %>% 
  parameters() %>% 
  update(mtry = mtry(c(1, 15)))

no_xgb_params <- wf_set_tune_list %>% 
  extract_workflow("no_xgb") %>% 
  parameters() %>% 
  update(mtry = mtry(c(1, 15)))

ds_rf_params <- wf_set_tune_list %>% 
  extract_workflow("ds_rf") %>% 
  parameters() %>% 
  update(mtry = mtry(c(1, 15)), under_ratio = under_ratio(c(1, 5)))

ds_xgb_params <- wf_set_tune_list %>% 
  extract_workflow("ds_xgb") %>% 
  parameters() %>% 
  update(mtry = mtry(c(1, 15)), under_ratio = under_ratio(c(1, 5)))

# update the workflowset
wf_set_tune_list_finalize <- wf_set_tune_list %>% 
  option_add(param_info = no_rf_params, id = "no_rf") %>%
  option_add(param_info = no_xgb_params, id = "no_xgb") %>% 
  option_add(param_info = ds_rf_params, id = "ds_rf") %>% 
  option_add(param_info = ds_xgb_params, id = "ds_xgb")

The rest remains the same. I think there may be some efficient ways to set the parameters.

Sources

This article follows the attribution requirements of Stack Overflow and is licensed under CC BY-SA 3.0.

Source: Stack Overflow

Solution Source
Solution 1 Kim.L