xybFight / GNARKD

PyTorch code for the GNARKD.

Distilling Autoregressive Models to Obtain High-Performance Non-Autoregressive Solvers for Vehicle Routing Problems with Faster Inference Speed

The PyTorch Implementation of AAAI 2024 -- "Distilling Autoregressive Models to Obtain High-Performance Non-Autoregressive Solvers for Vehicle Routing Problems with Faster Inference Speed"pdf.

This paper introduce a novel and generic method for solving VRPs named GNARKD to transform AR models into NAR ones to improve the inference speed while preserving essential knowledge.

How to Run

# 1. Training (for each teacher, e.g. POMO for TSP)
python -u GNARKD-POMO\TSP\Training.py

# Note that due to file size limitations, we removed the teacher's pre-training parameters, which you can download from the github link mentioned in the corresponding paper for successful training.


# 2. Testing (e.g., GNARKD-POMO for TSP)
python -u GNARKD-POMO\TSP\Test_file.py

The detail performance is as follows.

Acknowledgments

• We would like to thank the anonymous reviewers and (S)ACs of AAAI 2024 for their constructive comments and dedicated service to the community. The reviews are available here

• We also would like to thank the following repositories, which are baselines of our code:

  • https://github.com/wouterkool/attention-learn-to-route

  • https://github.com/yd-kwon/POMO

  • https://github.com/xbresson/TSP_Transformer

Citation

If you find our paper and code useful, please cite our paper:

@InProceedings{GNARKD2024, 
              author={Xiao, Yubin and Wang, Di and Li, Boyang and Wang, Mingzhao and Wu, Xuan and Zhou, Changliang and Zhou, You}, 
              title = {Distilling Autoregressive Models to Obtain High-Performance Non-autoregressive Solvers for Vehicle Routing Problems with Faster Inference Speed}, 
              volume={38}, 
              number={18}, 
              booktitle={Proceedings of the AAAI Conference on Artificial Intelligence}, 
              year={2024}, 
              pages={20274-20283},
              DOI={10.1609/aaai.v38i18.30008}, }

We purchased an additional page, but it still wasn't enough to fully show our work. Please refer to the Arxiv version for all content.