Latent Program Network (LPN)

Code for the paper Searching Latent Program Spaces, submitted to the ARC Prize 2024 competition.

Authors and maintainers: Clément Bonnet & Matthew V Macfarlane.

Links

• 📄 Paper on arXiv
• 🚀 Interactive Demo

Overview

The LPN is an architecture for inductive program synthesis that builds in test-time adaption by learning a latent space that can be used for search.

Installation

Install JAX using the official documentation here. Then, install the required packages using the following commands:

git clone https://github.com/clement-bonnet/lpn
cd lpn
export PYTHONPATH=${PYTHONPATH}:${PWD}
pip install -U -r requirements.txt

Add your secrets to the environment variables (HuggingFace token and WandB API key):

export HF_TOKEN=...
export WANDB_API_KEY=...

Repository Structure

src/
├── configs/        # Configuration files (use with hydra)
├── datasets/       # Data processing utilities (re-arc)
├── models/         # Neural network architectures (lpn)
└── train.py        # Main training script

Usage

To train a model, run the following command (replace the config name with the desired configuration):

python src/train.py --config-name pattern_2d

License

This project is licensed under the open-source Apache 2.0 License. See the LICENSE file for more details.

Citation

If you use this code in your research, please cite:

@misc{bonnet2024searchinglatentprogramspaces,
      title={Searching Latent Program Spaces}, 
      author={Clément Bonnet and Matthew V Macfarlane},
      year={2024},
      eprint={2411.08706},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2411.08706}, 
}