mlr3torch: A Deep Learning Framework in R based on mlr3 and torch

A team of researchers including Sebastian Fischer and Bernd Bischl has introduced mlr3torch, a new deep learning framework designed to bring the power of PyTorch to the R programming language within the established mlr3 machine learning ecosystem. The package, detailed in a recent arXiv preprint, acts as a bridge, allowing R users to define, train, and evaluate neural networks for both tabular data and generic tensors (like images) using a familiar, unified interface. Its core integration with torch provides the computational backend, while its design as an mlr3 "learner" means any torch model can be seamlessly plugged into mlr3's extensive toolkit for resampling, benchmarking, and preprocessing.

A key innovation of mlr3torch is its use of the graph-based workflow language from mlr3pipelines. This allows data scientists to construct an entire modeling pipeline—from data augmentation and preprocessing steps to the neural network architecture itself—as a single, definable graph. This approach significantly streamlines experimentation and reproducibility. The authors demonstrate the package's capabilities through practical use cases like hyperparameter tuning, model fine-tuning, and building architectures for multimodal data, positioning it as a robust solution for R users who need to incorporate modern deep learning techniques into their statistical and machine learning workflows without leaving their preferred environment.