Installation

poetry install

Launching Experiments

The experiments rely on Weights & Biases (wandb) for experiment tracking.

Launching goes in two steps. First, generate the bash scripts.

argmapper sweep exp/config/tabular_dp_experiments.yml --out exp/config/tabular_dp_experiments.runfile
argmapper sweep exp/config/cifar10_dp_experiments.yml --out exp/config/tabular_dp_experiments.runfile

Next, launch the commands in the generated bash scripts in any way you prefer. You can use:

argmapper launch exp/config/tabular_dp_experiments.runfile
argmapper launch exp/config/cifar10_dp_experiments.runfile

Keep in mind you need to have wandb set up and running so that the experiment metadata are saved.

Analyzing Experiments

Analyses of the experiments are done using Jupyter notebooks in the notebooks directory. They are saved as Python scripts in git. Use jupytext <notebook>.py --to ipynb to regenerate the notebook *.ipynb. Other than the analysis of experiments, the notebooks directory contains standalone experiments with synthetic data and theoretical expressions.