Our paper, pytorch-ood: A Unified PyTorch Library for Out-of-Distribution Detection, has been published in the Journal of Open Source Software (JOSS). It describes the substantially extended and matured version of the library first introduced in our CVPR 2022 workshop paper.
A copy (CC BY 4.0) of the PDF is available here. The archived software release is on Zenodo.
Summary §
Machine learning models, particularly deep neural networks, can produce unreliable predictions when they encounter data that differs from what they were trained on. Detecting such out-of-distribution (OOD) inputs is particularly important in safety-critical applications. Despite rapid progress in the field, practical experimentation remains fragmented. Implementations are often tied to individual papers, interfaces differ across methods, and evaluation setups can be difficult to reproduce.
pytorch-ood is a Python library that provides a unified framework for OOD detection in the PyTorch deep learning ecosystem. It encompasses a broad range of detection methods, training objectives, datasets, pretrained models, and evaluation tools under a consistent interface. This allows users to apply different methods, compare them under shared conditions, and integrate them into existing workflows with minimal additional code. The library is designed to support flexible experimentation rather than a single fixed benchmark. It is accompanied by extensive documentation and tests, enabling reliable reuse and facilitating reproducible research.
Installation §
pip install pytorch-ood
Citation
@article{kirchheim2026pytorchood,
title = {{pytorch-ood}: A Unified {PyTorch} Library for Out-of-Distribution Detection},
author = {Kirchheim, Konstantin and Langer, Theo and Ortmeier, Frank},
journal = {Journal of Open Source Software},
publisher = {The Open Journal},
year = {2026},
volume = {11},
number = {126},
pages = {10806},
doi = {10.21105/joss.10806}
}DOI: 10.21105/joss.10806