High-Level Modeling Functions with 'torch'

Provides high-level modeling functions to define and train models using the 'torch' R package. Models include linear, logistic, and multinomial regression as well as multilayer perceptrons.

brulee

The R brulee package contains several basic modeling functions that use the torch package infrastructure, such as:

neural networks
linear regression
logistic regression
multinomial regression
residual networks (ResNet)
regularization learning networks (RLN)
AutoInt
Self-Attention and Inter-sample Attention Transformer (Saint)
Chronos2 foundational model for forecasting
Transformer-based foundation model TabICL

Chronos2 and TabICL are pretrained models, requiring a one-time download of about 500MB and 400MB, respectively.

Installation

You can install the released version of brulee from CRAN with:

install.packages("brulee")

And the development version from GitHub with:

# install.packages("pak")
pak::pak("tidymodels/brulee")

Example

brulee has formula, x/y, and recipe user interfaces for each function. For example:

library(brulee)
library(recipes)
library(yardstick)

data(bivariate, package = "modeldata")
set.seed(20)
nn_log_biv <- brulee_mlp(Class ~ log(A) + log(B), data = bivariate_train, 
                         hidden_units = 3)

# We use the tidymodels semantics to always return a tibble when predicting
predict(nn_log_biv, bivariate_test, type = "prob") 
#> # A tibble: 710 × 2
#>    .pred_One .pred_Two
#>        <dbl>     <dbl>
#>  1     0.675     0.325
#>  2     0.673     0.327
#>  3     0.679     0.321
#>  4     0.688     0.312
#>  5     0.685     0.315
#>  6     0.679     0.321
#>  7     0.674     0.326
#>  8     0.681     0.319
#>  9     0.697     0.303
#> 10     0.675     0.325
#> # ℹ 700 more rows

A recipe can also be used if the data require some sort of preprocessing (e.g., indicator variables, transformations, or standardization):

library(recipes)

rec <- 
  recipe(Class ~ ., data = bivariate_train) |>  
  step_YeoJohnson(all_numeric_predictors()) |> 
  step_normalize(all_numeric_predictors())

set.seed(20)
nn_rec_biv <- brulee_mlp(rec, data = bivariate_train, 
                         epochs = 150, hidden_units = 3)

# A little better
predict(nn_rec_biv, bivariate_test, type = "prob") |> 
  bind_cols(bivariate_test) |> 
  roc_auc(Class, .pred_One)
#> # A tibble: 1 × 3
#>   .metric .estimator .estimate
#>   <chr>   <chr>          <dbl>
#> 1 roc_auc binary         0.866