Torch and TorchVision R Improvements

This repository contains my submission for the R-GSOC project "torchvision in R improvements". The project aims to enhance the R interface for torch and torchvision, making computer vision models and datasets more accessible to R users.

Background

torch is a powerful machine learning package for R, based on the LibTorch C++ library. torchvision provides additional functionality specifically for computer vision tasks, including pre-trained models, datasets, and image transformations. However, the R implementation of torchvision is currently missing several features compared to its Python counterpart. This project aims to address some of these limitations.

Tasks Completed

Easy Task: Gaussian Linear Model

For the easy task, I installed torch and created a simple Gaussian Linear Model using the instructions from the distributions vignette. The implementation demonstrates:

• Creating and visualizing synthetic data
• Building a custom neural network model with torch
• Training the model using autograd and optimization
• Comparing results with R's built-in glm function
• Visualizing the fitted model against the true relationship

The implementation is available in easy.Rmd with its rendered output as a PDF.

Medium Task: Custom Dataset for Spam Classification

The medium task involved adapting the "Loading data" vignette to work with the spam dataset from "Elements of Statistical Learning". This implementation showcases:

• Creating a custom dataset class for email classification
• Implementing helper functions for data download and preprocessing
• Building a dataloader for efficient batch processing
• Creating a neural network for binary classification
• Training and evaluating the model on spam detection

The implementation is available in medium.Rmd with its rendered output as a PDF.

Hard Task: Pull Request for Spam Dataset

For the hard task, I forked the torch repository and created a PR that adds a proper data loader for the spam dataset, complete with tests and documentation. This contribution makes the spam dataset easily accessible to other R users via the torch package.

PR link: mlverse/torch#1294

Repository Structure

.
├── README.md               # This file
├── easy.Rmd                # Source for the easy task (Gaussian Linear Model)
├── easy.pdf                # Rendered output of easy task
├── medium.Rmd              # Source for the medium task (Spam Dataset)
└── medium.pdf              # Rendered output of medium task

How to Run

Requirements

• R (>= 4.0.0)
• torch package
• knitr and rmarkdown for rendering the Rmd files

Installation

install.packages("torch")
install.packages(c("knitr", "rmarkdown"))

Rendering the Documents

To render the Rmd files:

rmarkdown::render("easy.Rmd")
rmarkdown::render("medium.Rmd")