NAP

NAP (Neural Acquisition Process)

This repository is the official implementation of End-to-End Meta-Bayesian Optimisation with Transformer Neural Processes. The code provided in this repo allows the user to train, validate and test NAP on `HPO-B and Antigen experiments.

Setup

Setup a virtualenv/conda/miniconda environment with at least python3.8 and use the requirements.txt to install the dependencies.

# Example with virtualenv
sudo apt-get install python3.8-venv  # for Ubuntu 18.04 LTS
python3.8 -m venv nap_env
. nap_env/bin/activate
pip install --upgrade pip
pip install -r requirements.txt

Training

To train NAP on HPO:

PYTHONPATH=. python scripts/nap/train_nap_hpo.py
# if it complains about the number of opened files, first run
ulimit -Sn 10000

Testing

Adjust the paths inside scripts/nap/test_nap_hpo.py and run the script.

PYTHONPATH=. python scripts/nap/test_nap_hpo.py

Results

Cite us

@misc{maraval2023endtoend,
      title={End-to-End Meta-Bayesian Optimisation with Transformer Neural Processes}, 
      author={Alexandre Maraval and Matthieu Zimmer and Antoine Grosnit and Haitham Bou Ammar},
      year={2023},
      eprint={2305.15930},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}

Contributors

Alexandre Max Maraval, Matthieu Zimmer, Antoine Grosnit, Haitham Bou Ammar

License

• https://github.com/metabo-iclr2020/MetaBO is under GNU APGL-3.0 License.