- the accompanying paper will be available soon ...
- Python 3.8 or later
- PyTorch 1.8 or later
- PyTorch geometric 1.6.3 or later
- Networkit
- Concorde TSP solver (optional)
(Commands are supposed to be run from the root environment)
- Use conda to create a new environemnt with
export ENV_NAME="hamgnn"
conda env create --file requirements.yaml --name $ENV_NAME
- Then add this package in develop mode
pip install -e .
- Start the training with
python terminal_scripts/train_main_model.py
This should log a wandb run and produce a checkpoint in MODEL_CHECKPOINTS directory. Training data is generated on the fly.
- Install Concorde TSP solver from [https://www.math.uwaterloo.ca/tsp/concorde.html]
- Update the following fields in
./config.cfgfile:- CONCORDE_SCRIPT_PATH - path to "concorde" executable created during concorde install
- Generate data for various experiments by running
python terminal_scripts/generate_data.py
After completion, this should create several dataset in the ./DATA folder.
- The model can be tested on generated data by running
python terminal_scripts/test_main_model.py
It prints the output in somewhat raw format, the percentage of HCP solved can be seen in perc_hamilton_found column. A nicer output format will be added soon...
- To generate all the figures presented in the paper, run
python terminal_scripts/generate_figures.py
Note, this will take a very long time since several different models need to be trained.
- A list of model that have been tried out during experimentation can be found in
hamgnn/models_list.py. All models listed there should theoretically be runnable with:
python terminal_scripts/train_model_variation.py <experiment_name>
where <experiment_name> corresponds to the name of the python variable to which the experiment has been assigned inside hamgnn/models_list.py file.
- The code in
hamgnn/legacy/is depricated and will be removed eventually