AD3: A Multi-Scale Asynchronous Divide-Diffuse-Decode Diffusion Framework for TSP Solving [Scalable Data Science]

Code for the paper AD3: A Multi-Scale Asynchronous Divide-Diffuse-Decode Diffusion Framework for TSP Solving [Scalable Data Science]

Dependencies

AD3 is built in Python 3.8.13 and Pytorch 1.11.0. And Please use the following command to install the requirements:

pip install -r requirements.txt

ckpt

The models for reproduce is epoch_32.ckpt, placed in the "storage/ckpt/" directory by default.

Evaluate

We provide both single-GPU and multi-GPU configurations in AD3_evaluate_S.py and AD3_evaluate_P.py. We are currently integrating them.

- `evaluate_single_gpu.py`: single-GPU evaluation script
- `evaluate_multi_gpu.py`: multi-GPU evaluation script

The results are recorded in "storage/logs" directory by default.

Usage

Run from project root with relative paths only:

python evaluate_single_gpu.py --help
python evaluate_multi_gpu.py --help

Then provide your own relative paths for checkpoints/data if needed.