alstn12088/kse527-offline-co

Project

This is official PyTorch code for KSE527 final project held in 2022 sprining somestar in KAIST.

Source Code Reference

This code is originally implemented based on Attention Model , which is source code of the paper Attention, Learn to Solve Routing Problems! which has been accepted at ICLR 2019, cite as follows:

@inproceedings{
    kool2018attention,
    title={Attention, Learn to Solve Routing Problems!},
    author={Wouter Kool and Herke van Hoof and Max Welling},
    booktitle={International Conference on Learning Representations},
    year={2019},
    url={https://openreview.net/forum?id=ByxBFsRqYm},
}

Our work revised learning scheme of Attention Model (AM), improving sample efficiency in offline CO. The most of code configuration is same, except:

• Modified "net/attention_model.py" for state forcing with guiding state, and log-likelihood caluation of guiding action.
• Modified "train.py" to import offline data, to perform label augmentation and to perform self-supervised learning (psuedo label based behavior cloning).

We remarked "KSE527" to revised part in the source code .

How to Use

Unzip data

We provide pre-collected labeled data (from Concorde) which is pre-processed for training in "data/offline_labels.pkl". We provide pre-defined problem set which is corresponding to the labeled data. We provide benchmark test data in "data/tsp20_test_seed1234.pkl"

Evaluation with pretrained model

python eval.py --dataset_path data/tsp20_test_seed1234.pkl --model pretrained ours/tsp20/epoch-99.pt

python eval.py --dataset_path data/tsp100_test_seed1234.pkl --model pretrained ours/tsp100/epoch-99.pt

Training the Model

Training AM with our method (AM + Data augmentation + Self-supervised Learning)

python run.py --graph_size 20 --training_model ssl 

Training AM with data augmentation(AM + Data augmentation)

python run.py --graph_size 20 --training_model da

Training AM with naive offline behavior cloning (baseline)

python run.py --graph_size 20 --training_model bc

#### GPU

Only a single GPU is available in this code. 


### Other usage

```bash
python run.py -h
python eval.py -h

Dependencies

• Python>=3.8
• NumPy
• SciPy
• PyTorch>=1.1
• tqdm
• tensorboard_logger
• Matplotlib