torch-uncertainty

Why we included this project

For models that must report how confident they are, not just what they predict, torch-uncertainty gives you one place to start. It wraps deep ensembles, MC-dropout, variational Bayesian networks and evidential classification into ready-to-use layers, losses, metrics and Lightning training routines, so classification, regression and segmentation tasks can be compared side by side without reimplementing each method. The evaluation side is automated too: calibration scoring, selective classification, out-of-distribution detection and distribution-shift metrics come prebuilt, and post-processing like temperature scaling and conformal prediction are plain Python methods. Several methods are official implementations of papers from major machine learning and vision conferences, and the package featured in tutorials at WACV, HAICON and ECCV 2024.