This folder contains the code used for the experiments in the paper "Cold Posterior through PAC-Bayes". 

The folder "utils" contains the main functions which are used in common by all datasets. This includes functions that train multiple MAP estimates and estimate the Alquier and Catoni bounds.

The folder "laplace_package_extensions" contains an extension of the Laplace package such that it can support our experiments. In particular the Isotropic Laplace is defined, and we scale quantities to much our paper. 

The folder "fashionmnist" contains example scripts of how the utils + laplace_package_extensions resources are used by the fashionmnist dataset.

The env.yml file describes the conda environment used in the experiments.

---------------------------------------------------------------------------

The code is meant as an example of the one we used and we didn't test that the parts included are functional by themselves. In particular the dataloaders functions are not included, though they can be easily replicated and are the standard ones used in practice.