Skip to contents

Getting Started with kindling

Source: vignettes/kindling.Rmd
kindling.Rmd

Introduction

kindling bridges the gap between torch and tidymodels, providing a streamlined interface for building, training, and tuning deep learning models. This vignette will guide you through the basic usage.

Installation

You can install kindling on CRAN:

install.packages('kindling')

Or install the development version from GitHub:

# install.packages("pak")
pak::pak("joshuamarie/kindling")
## devtools::install_github("joshuamarie/kindling") 
library(kindling)
#> 
#> Attaching package: 'kindling'
#> The following object is masked from 'package:base':
#> 
#>     args

Before using {kindling}

Before starting, you need to install LibTorch, the backend of PyTorch which also the backend of torch R package:

torch::install_torch()

Four Levels of Interaction

kindling offers flexibility through four levels of abstraction:

  1. Code Generation - Generate raw torch::nn_module code
  2. Direct Training - Train models with simple function calls
  3. tidymodels Integration - Use with parsnip, recipes, and workflows
  4. Hyperparameter Tuning - Optimize models with tune and dials

Level 1: Code Generation

Generate PyTorch-style module code:

ffnn_generator(
    nn_name = "MyNetwork",
    hd_neurons = c(64, 32),
    no_x = 10,
    no_y = 1,
    activations = 'relu'
)

Level 2: Direct Training

Train a model with one function call:

model = ffnn(
    Species ~ .,
    data = iris,
    hidden_neurons = c(10, 15, 7),
    activations = act_funs(relu, elu), # c("relu", "elu")
    loss = "cross_entropy",
    epochs = 100
)

predictions = predict(model, newdata = iris)

Level 3: tidymodels Integration

Work with neural networks like any other parsnip model:

box::use(
    parsnip[fit, augment],
    yardstick[metrics]
)

nn_spec = mlp_kindling(
    mode = "classification",
    hidden_neurons = c(10, 7),
    activations = act_funs(relu, softshrink = args(lambd = 0.5)),
    epochs = 100
)

nn_fit = fit(nn_spec, Species ~ ., data = iris)
augment(nn_fit, new_data = iris) |> 
    metrics(truth = Species, estimate = .pred_class)

Learn More