A High Dimensional Model for Adversarial Training: Geometry and Trade-Offs

To reproduce the figures from the paper, please use define_experiment.ipynb in the experiments folder.

All experiments have a definition for the data-model and problem types considered. The choice of the sweep parameters has to be customised.

Once you've created an experiment.json file and you would like to run the experiment from the command line instead of the jupyter notebook, use:

mpiexec -n 5 python sweep/run_sweep.py --file experiment.json

Alternatively, in a cluster environment, it is possible to use a run.sh file. There's an example in experiments/run.sh.

The experiments on real data have been performed in the pca_experiments.ipynb notebook.

How to use

Install mpi using brew

brew install mpich

Install an environment for execution:

pip install uv
uv venv --python 3.11

source .venv/bin/activate
uv pip install -r pyproject.toml

Install this package

uv pip install .

Build brentq.c:

gcc -shared -o numerics/brentq.so numerics/brentq.c

How to contibute

Install mpi using brew

brew install mpich

Install an environment for development:

pip install uv
uv venv --python 3.11

source .venv/bin/activate
uv pip install -r pyproject.toml --extra dev

Install this package in editable mode

uv pip install -e .

Build brentq.c:

gcc -shared -o numerics/brentq.so numerics/brentq.c

Install a pre-commit hook for ruff

pre-commit install

Run the tests like this to see logging output

pytest -o log_cli=true -o log_cli_level=INFO

If you work with vs code, add this to your .vscode/settings.json file in your workspace for the notebooks to resolve your python path correctly:

"python.analysis.extraPaths": [
        "${workspaceFolder}"
    ]