CNN_Transformer

Code for paper: A Lightweight CNN-Transformer Model for Learning Traveling Salesman Problems

Runs (Examples)

(Updated 04-30-2024) Now you can use Docker to train & test our model withour installing all the dependencies by running: docker run -it --gpus all --rm fmsjung/lit-cnn-trnsf-tsp

To train the proposed model on TSP20, run the following:

$ python train.py --exp_name Experiment-Name -n 20 --bsz 512 --gpu_id 0 --embedding conv --nb_neighbors 10 --kernel_size 11 --segm_len 5 --batchnorm

To test the trained model on TSP20 random dataset, run the following:

$ python test.py -n 20 --embedding conv --nb_neighbors 10 --kernel_size 11 --gpu_id 0 --ckpt_file Experiment-Name-on-Train.pkl --exp_tag Result-File-Name --segm_len 5 --greedy --beamsearch --batchnorm

OR you can instead run shell scripts for train and test respectively. You may modify the .sh files uploaded on this repository if necessary.

$ sh train_sh/train-sh-filename.sh
$ sh test_sh/test-sh-filename.sh

You can use test_tsplib.ipynb to test the trained model on the TSPLIB dataset

The pretrained model checkpoints are uploaded in the checkpoint directory

Requirements

• pytorch 1.12.1
• numpy 1.22.3
• pandas 1.4.3
• python 3.10.5
• matplotlib 3.7.1

Acknowledgement

The codes here are based on the work of Xavier Bresson's TSP Transformer

Cite

@article{jung2024lightweight,
  title={A lightweight CNN-transformer model for learning traveling salesman problems},
  author={Jung, Minseop and Lee, Jaeseung and Kim, Jibum},
  journal={Applied Intelligence},
  pages={1--12},
  year={2024},
  publisher={Springer}
}