README.md

Gibbs Prior Diagnostics

Code for the AISTATS 2022 paper titled "Discovering Inductive Bias with Gibbs Priors: A Diagnostic Tool for Approximate Bayesian Inference" by Luca Rendsburg, Agustinus Kristiadi, Philipp Hennig, and Ulrike von Luxburg.

Setting up:

1. Install CUDA >=11 and CuDNN >=8.2: https://gist.github.com/kmhofmann/cee7c0053da8cc09d62d74a6a4c1c5e4.
2. Install Miniconda: https://docs.conda.io/en/latest/miniconda.html.
3. Create a virtual environment: conda create --name {ENV_NAME} --file requirements.txt, for an arbitrary {ENV_NAME}.
4. Activate: conda activate {ENV_NAME}.

Reproducing the paper's results:

1. Choose a folder in experiments/ to reproduce the experiments for the log-normals model in Section 5.1 (experiments/log_normal), for the volatility model in Section 5.2 (experiments/volatility), for the baseline (experiments/baseline), or for convergence monitoring (experiments/convergence)
2. Run the run_%.ipynb file(s) to generate the data, which are saved in the corresponding folder in results/
3. Run the eval_%.ipynb file to create the figures used in the paper.

Citing the paper:

@inproceedings{rendsburg2022gibbs-prior,
  title={Discovering Inductive Bias with {G}ibbs Priors: A Diagnostic Tool for Approximate {B}ayesian Inference},
  author={Rendsburg, Luca and Kristiadi, Agustinus and Hennig, Philipp and von Luxburg, Ulrike},
  booktitle={AISTATS},
  year={2022}
}