Lighter: Configuration-Driven Deep Learning
Ibrahim Hadžić, Suraj Pai, Keno K. Bressem, Borek Foldyna, Hugo J.W.L. Aerts · The Journal of Open Source Software · 2025
Lighter is a configuration-driven deep learning (DL) framework that separates experimental setup from code implementation.Models, datasets, and other components are defined through structured configuration files (configs).Configs serve as snapshots of the experiments, enhancing reproducibility while eliminating unstructured and repetitive scripts.Lighter uses (i) PyTorch Lightning (Falcon & The PyTorch Lightning team, 2019) to implement a task-agnostic DL logic, and (ii) MONAI Bundle configuration (Cardoso et al., 2022) to manage experiments using YAML configs.