Code for recreating the experiments in the paper "Integral Probability Metrics PAC-Bayes Bounds ", R. Amit, B. Epstein, S. Moran, and R. Meir, NeurIPS 2022.

Pre-requisites

Tested on Python 3.10.4, conda environment with:

• matplotlib 3.5.2
• numpy 1.23.2
• pandas 1.4.3
• pytorch 1.12.1
• Jinja2 3.0.3

How to re-create results

• python main.py --sigma_P 0.01 --sigma_Q 0.001 --optim_objective klpb_risk_bound
• python main.py --sigma_P 0.0001 --sigma_Q 0.001 --optim_objective klpb_risk_bound
• python main.py --sigma_P 0. --sigma_Q 0. --optim_objective wpb_risk_bound