useR! 2024

mlr3torch is a high level deep learning framework for the mlr3 ecosystem designed to easily build, train, and evaluate neural networks in a few lines of code. It leverages the torch package, which is an R interface to the LibTorch C++ library. On the one hand, the package comes with predefined and easy-to-use neural network architectures, both for classification and regression. On the other hand, it defines a language that allows to easily define custom, fully parameterized neural networks. Because the package is integrated into the mlr3 ecosystem, these neural networks can be easily benchmarked, tuned, or combined with other machine learning workflows such as preprocessing or stacking. While mlr3’s focus is tabular data, mlr3torch extends this to other modalities such as images or text, by defining a new data type: the 'lazy_tensor' . This type can be treated similarly to standard vectors and can, e.g., be preprocessed without requiring the data to be stored in-memory. This presentation will give an overview of mlr3torch's features and demonstrate its application in both research and practical machine learning scenarios. https://github.com/mlr-org/mlr3torch