Important!

LLLA, among other Laplace-approximation flavors are available in the easy-to-use laplace-torch library: https://github.com/AlexImmer/Laplace.

Please use that library instead of this code repository!

Last-layer Laplace approximation (LLLA)

Code examples of the last-layer Laplace approximation method used in "Being Bayesian, Even Just a Bit, Fixes Overconfidence in ReLU Networks" paper (ICML 2020). See bnn_laplace.ipynb and bnn_laplace_multiclass.ipynb. Codes for reproducing the paper's results are in the paper directory.

Requirements

• PyTorch
• Scikit-learn
• Matplotlib
• Seaborn
• Numpy
• BackPack

Credits

• https://github.com/AlexMeinke/certified-certain-uncertainty
• https://github.com/f-dangel/backpack
• https://github.com/wjmaddox/swa_gaussian