Fitting tabnet with tidymodels


In this vignette we show how to create a TabNet model using the tidymodels interface.

We are going to use the lending_club dataset available in the modeldata package.

First let’s split our dataset into training and testing so we can later access performance of our model:

data("lending_club", package = "modeldata")
split <- initial_split(lending_club, strata = Class)
train <- training(split)
test  <- testing(split)

We now define our pre-processing steps. Note that tabnet handles categorical variables, so we don’t need to do any kind of transformation to them. Normalizing the numeric variables is a good idea though.

rec <- recipe(Class ~ ., train) %>%

Next, we define our model. We are going to train for 50 epochs with a batch size of 128. There are other hyperparameters but, we are going to use the defaults.

mod <- tabnet(epochs = 50, batch_size = 128) %>%
  set_engine("torch", verbose = TRUE) %>%

We also define our workflow object:

wf <- workflow() %>%
  add_model(mod) %>%

We can now define our cross-validation strategy:

folds <- vfold_cv(train, v = 5)

And finally, fit the model:

fit_rs <- wf %>%

After a few minutes we can get the results:

# A tibble: 2 x 5
  .metric  .estimator  mean     n  std_err
  <chr>    <chr>      <dbl> <int>    <dbl>
1 accuracy binary     0.946     5 0.000713
2 roc_auc  binary     0.732     5 0.00539 

And finally, we can verify the results in our test set:

model <- wf %>% fit(train)
test %>% 
    predict(model, test, type = "prob")
  ) %>% 
  roc_auc(Class, .pred_bad)
# A tibble: 1 x 3
  .metric .estimator .estimate
  <chr>   <chr>          <dbl>
1 roc_auc binary         0.710