google/uncertainty-baselines

Implements uncertainty quantification methods (ensembles, Bayesian neural networks, temperature scaling) alongside standard baselines, all built on TensorFlow/Keras with support for TPU acceleration and standardized evaluation metrics like calibration error and negative log-likelihood. Each baseline is self-contained and independently reproducible, enabling direct comparison across different uncertainty approaches on common benchmarks like CIFAR-10 and ImageNet without reimplementation overhead.

1,568 stars and 54 monthly downloads. Actively maintained with 1 commit in the last 30 days. Available on PyPI.