Skip to content

GraphPKU/DisGNN

Repository files navigation

Is Distance Matrix Enough for Geometric Deep Learning?

Pytorch implementation of the paper Is Distance Matrix Enough for Geometric Deep Learning?

Dependencies

python=3.8

pytorch-lightning=2.0.1.post0

pytorch=2.0.0+cu118

torch_geometric=2.3.0

Train and Test $k$-DisGNN

To train and test k-DisGNN, run

python scripts/kDis_script.py [--model <model_name>] [--ds <dataset_name>] [--dname <data_name>] [--devices <device_ids>] [--data_dir <data_dir>] [--version <version>] [--resume] [--only_test] [--ckpt <checkpoint_path>] [--merge <merge hparam list>] [--use_wandb] [--proj_name <project_name>]

Arguments description:

  • --model: ["2EDis", "3EDis", "2FDis"], corresponds to the three models. By default, it is "2FDis".
  • --ds: ["md17", "qm9"], corresponds to the two datasets. By default, it is "md17".
  • --dname: In MD17, choices are the name of molecules, which can be found in md17; In QM9, choices are numbers in [0, 18], corresponding to the targets in qm9.
  • --devices: The number of the GPU devices. For example, if you want to use GPU 0 and 1, please specify it as 0 1. If no devices are specified, the code will run on the GPU with minimum memory usage. If it's set to -1, the code will run on CPU.
  • --data_dir: The directory of the dataset. By default, it is "~/datasets/MD17".
  • --version: The version of the log. By default, it is "NO_VERSION".
  • --resume: Whether to resume the training from the latest checkpoint. By default, it is False.
  • --only_test: Whether to only test the model. By default, it is False.
  • --ckpt: The path of the checkpoint. By default, it is None.
  • --merge: The hparams to merge. For example, if you want to modify the hparam model_config.block_num to 4, then use --merge model_config.block_num 4.
  • --use_wandb: Whether to use wandb to log the training process. By default, it is False.
  • --proj_name: The name of the project in wandb. By default, it is None.

Hyperparameters are specified in the dictionary hparams.

For example, to run 2FDis on revised MD17 (revised ethanol) using gpu 0:

python scripts/kDis_script.py --model 2FDis --ds md17 --dname 'revised ethanol' --devices 0 --data_dir <your_data_dir> --version test 

Run Counterexamples Verification Programs

To verify the counterexamples in Appendix A.1 and A.2 (the first two kinds), run

python counterexamples/verify_ce.py [--reg <regular_polyhedron>] [--verbose] [--visualize]

Arguments description:

  • --reg: ["12", "20", "cube_8", "cube_cube"], corresponds to the counterexamples sampled from different regular polyhedrons (or the combination of regular polyhedrons).
  • --verbose: store_true. Whether to show the detailed 1-WL-E process.
  • --visualize: store_true. Whether to visualize the counterexamples.

To verify the counterexamples in Appendix A.3 (the families), run

python counterexamples/verify_families.py [--reg <regular_polyhedron>] [--verbose] [--num_of_layers <num_of_layers>]

Arguments description:

  • --reg: ["12", "20"], corresponds to the counterexamples sampled from different regular polyhedrons.
  • --verbose: store_true. Whether to show the detailed 1-WL-E process.
  • --num_of_layers: The number of layers of the counterexamples. By default, it is None, which means that the number of layers is randomly sampled from [1, 10].

To verify the Lemma A.3 (Stable states), run

python counterexamples/verify_stable_state.py [--reg <regular_polyhedron>] [--verbose]

Arguments description:

  • --reg: ["12", "20"], corresponds to the counterexamples sampled from different regular polyhedrons.
  • --verbose: store_true. Whether to show the detailed 1-WL-E process.

Cite

If you use the code please cite our paper.

@article{li2023distance,
  title={Is Distance Matrix Enough for Geometric Deep Learning?},
  author={Li, Zian and Wang, Xiyuan and Huang, Yinan and Zhang, Muhan},
  journal={arXiv preprint arXiv:2302.05743},
  year={2023}
}

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published