This repository provides the codes for all experiments shown in the paper Compositional Generalization from First Principles.

• use Python 3.9
• run pip install -r requirements.txt
• if desired, a Dockerfile is provided. You can create the container with docker build .

• all experiments where run on a single NVIDIA-RTX-2080Ti
• config files for all experiments can be found in cfgs/
• to start an experiment, run main.py --config path/to/experiment/config.yml

If you find the insights from the paper or our code base useful, please cite

@misc{wiedemer2023compositional,
      title={Compositional Generalization from First Principles}, 
      author={Thaddäus Wiedemer and Prasanna Mayilvahanan and Matthias Bethge and Wieland Brendel},
      year={2023},
      eprint={2307.05596},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}