rezaqorbani / f-CMI

Information-theoretic generalization bounds for black-box algorithms

Replicating the experiments

1. Open .bashrc or .zshrc (depending which shell you use) and add a line “export $DATA_DIR={path-of-the-data-dir}.
2. Generate the commands for the experiment using scripts/fcmi_scripts.py script.
3. Parse the result of the experiment using the scripts/fcmi_parse_results.py script.
4. Use the notebooks/fcmi-plots.ipynb to generate plots from the parsed results.

Requirements

• Basic libraries such as numpy, scipy, tqdm, matplotlib, and seaborn.
• We used Pytorch 1.7.0, but higher versions should work too.