README.md

Deep Symbolic Regression for Recurrent Sequences

This repository contains code for the paper Deep Symbolic Regression for Recurrent Sequences. An interactive demonstration of the paper may be found here.

The code is based on the repository Deep Learning for Symbolic Mathematics. Most of the code specific to recurrent sequences lies in the folder src/envs.

Run the model

To launch a small Transformer on the CPU, run: python train.py --cpu True

Main arguments

--float_sequences		# if True, run the float model, otherwise run the integer model
--output_numeric		# if True, run the numeric model, otherwise run the symbolic model
--use_sympy 			# whether to use sympy simplification

The arguments specific to the generator can be found in src/envs/recurrence.py.

Multinode training

Distributed training is available via Slurm and submitit:

pip install submitit

To launch a run on 2 nodes with 8 GPU each, use the submit.py script.

Dependencies

• Python 3
• NumPy
• SymPy
• PyTorch (tested on version 1.3)

Citation

If you want to reuse this material, please considering citing the following:

@article{d2022deep,
  title={Deep symbolic regression for recurrent sequences},
  author={d'Ascoli, St{\'e}phane and Kamienny, Pierre-Alexandre and Lample, Guillaume and Charton, Fran{\c{c}}ois},
  journal={arXiv preprint arXiv:2201.04600},
  year={2022}
}

License

The majority of this repository is released under the Apache 2.0 license as found in the LICENSE file.