v0.2.0 Lightning 2.0, RegressionRoutine & SegmentationRoutine

🚀 TorchUncertainty 0.2.0 Released! 🚀

We're thrilled to unveil TorchUncertainty 0.2.0!

This update brings a complete overhaul reconstruction around our uncertainty-aware routines. Highlights include:

• Lightning 2.0: Support and a complete overhaul of the command-line interface.

• RegressionRoutine: Fully functional, now supporting probabilistic regression with PyTorch distributions.

• SegmentationRoutine: Introduces semantic segmentation support for datasets like Cityscapes and MUAD.

Stay tuned for even more (Monocular depth estimation!) in TorchUncertainty 0.2.1!

Breaking Changes

As we are still in pre-release, this version breaks a large part of the routine and CLI components of TorchUncertainty 0.1.6.

CLI

The behavior of the CLI has completely changed and is now based on the configuration files from Lightning 2.0. We provide a new page that explains how to leverage Baselines using the CLI for easy benchmarking.

Routines

Notably, there is no more distinction between ensemble and single routines to reduce code entropy: single routines are ensemble routines with 1 estimator. Furthermore, the routines' loss parameters now take an instantiated loss instead of a type, the optimization_procedure is renamed optim_recipe and is now a dictionary and not a callable. The ood_criterion and the calibration sets are now strings.

Metrics

The NegativeLogLikelihood metric is renamed CategoricalNLL.

Baselines

All baselines have been renamed to explicitly contain "Baselines" in their name.

Tutorials

We have rewritten and updated the tutorials should now be clearer. Send us feedback!

What's Changed

• ➖ Avoid using Argvcontext in tutorials by @o-laurent in #82
• 🚀 Upgrade to Lightning 2.0 by @alafage in #79
• 🚀 Update to Lightning 2.0, Add Segmentation, & Rework Regression by @o-laurent in #85

Full Changelog: v0.1.6...v0.2.0