Skip to content


Switch branches/tags

Name already in use

A tag already exists with the provided branch name. Many Git commands accept both tag and branch names, so creating this branch may cause unexpected behavior. Are you sure you want to create this branch?

Latest commit

* Add scheduler reduce_on_plateau to tabnet supervised training

* Add scheduler reduce_on_plateau to tabnet pretraining

* Update documentation for parameter `lr_scheduler`

* Add unit test for when lr_scheduler is set to reduce_on_plateau

* Fix using correct lr scheduler function at unit test

* Fix that step with loss is requested when function requires it

* Update news and description

Git stats


Failed to load latest commit information.


R build status Lifecycle: experimental CRAN status Discord

An R implementation of: TabNet: Attentive Interpretable Tabular Learning. The code in this repository is an R port of dreamquark-ai/tabnet PyTorch’s implementation using the torch package.


You can install the released version from CRAN with:


The development version can be installed from GitHub with:

# install.packages("remotes")

Basic Binary Classification Example

Here we show a binary classification example of the attrition dataset, using a recipe for dataset input specification.


data("attrition", package = "modeldata")
test_idx <-, size = 0.2 * nrow(attrition))

train <- attrition[-test_idx,]
test <- attrition[test_idx,]

rec <- recipe(Attrition ~ ., data = train) %>% 
  step_normalize(all_numeric(), -all_outcomes())

fit <- tabnet_fit(rec, train, epochs = 30, valid_split=0.1, learn_rate = 5e-3)

The plots gives you an immediate insight about model overfitting, and if any, the available model checkpoints available before the overfitting

Keep in mind that regression as well as multi-class classification are also available, and that you can specify dataset through data.frame and formula as well. You will find them in the package vignettes.

Model performance results

As the standard method predict() is used, you can rely on your usual metric functions for model performance results. Here we use {yardstick} :

metrics <- metric_set(accuracy, precision, recall)
cbind(test, predict(fit, test)) %>% 
  metrics(Attrition, estimate = .pred_class)
#> # A tibble: 3 × 3
#>   .metric   .estimator .estimate
#>   <chr>     <chr>          <dbl>
#> 1 accuracy  binary         0.837
#> 2 precision binary         0.837
#> 3 recall    binary         1
cbind(test, predict(fit, test, type = "prob")) %>% 
  roc_auc(Attrition, .pred_No)
#> # A tibble: 1 × 3
#>   .metric .estimator .estimate
#>   <chr>   <chr>          <dbl>
#> 1 roc_auc binary         0.554

Explain model on test-set with attention map

TabNet has intrinsic explainability feature through the visualization of attention map, either aggregated:

explain <- tabnet_explain(fit, test)

or at each layer through the type = "steps" option:

autoplot(explain, type = "steps")

Self-supervised pretraining

For cases when a consistent part of your dataset has no outcome, TabNet offers a self-supervised training step allowing to model to capture predictors intrinsic features and predictors interactions, upfront the supervised task.

pretrain <- tabnet_pretrain(rec, train, epochs = 50, valid_split=0.1, learn_rate = 1e-2)

The exemple here is a toy example as the train dataset does actually contain outcomes. The vignette on Unsupervised training and fine-tuning will gives you the complete correct workflow step-by-step.

Missing data in predictors

{tabnet} leverage the masking mechanism to deal with missing data, so you don’t have to remove the entries in your dataset with some missing values in the predictors variables.