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.
# 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.
-
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:
If you find our paper and code useful, please cite our paper:
@misc{Xiao2023,
title={Distilling Autoregressive Models to Obtain High-Performance Non-Autoregressive Solvers for Vehicle Routing Problems with Faster Inference Speed},
author={Yubin Xiao and Di Wang and Boyang Li and Mingzhao Wang and Xuan Wu and Changliang Zhou and You Zhou},
year={2023},
eprint={2312.12469},
archivePrefix={arXiv},
}
Or after the publication of the AAAI24 paper: To be continued.