NeuralNets_GraphColoringSolving
Solving a dicision version of Graph Coloring Problem (GCP) by DeepLearning approach
Creating venv
pip install -r requirements.txt
Also you can install cuda version of torch and run training on GPU (though it doesn't give much speed up)
Data
To create synthetic data (hard to color graph instances) for training
python dataset_generator.py --samples number_of_samples --nmin min_number_of_vertices --nmax max_number_of_vertcies --root path_to_dataset --test is_test
Also you can convert dataset from https://machine-reasoning-ufrgs.github.io/GNN-GCP/ through running:
python convert_gnn_gcp_dataset.py path_to_downloaded_dataset
Model
Currently available Reccurent Graph Neural Network model similar to https://arxiv.org/abs/1903.04598 with various attention mechanisms and lack of layer normalization.
Training and evaluating
python --mode train_or_test --config path_to_your_config --data path_to_dataset --print_freq print_freq --log_dir dir_to_save_checkpoints_and_log --save_freq save_freq --test_freq test_freq
Attention aggegators
You can change version of attention aggregators through setting such option in config:
{
"attention": true,
"attention_version": "pairwise_0"
}
There are available "pairwise_0", "pairwise_1" and "pairwise_2" aggergators (named in ascending of computational complexity) Basic graph attention aggregator related paper (https://arxiv.org/abs/1710.10903)
Current results
| Config | Data generation model | Data vertices number | Data chrom number | Acc |
|---|---|---|---|---|
| Basic | This | 35-45 | 3-8 | 87.3 |
| Pairwise_0 | This | 35-45 | 3-8 | 87.3 |
Experiments on GNN_GCP (https://machine-reasoning-ufrgs.github.io/GNN-GCP/) haven't gone well (more likely there is bug somewhere)