Lanczos Network

This is the PyTorch implementation of Lanczos Network as described in the following ICLR 2019 paper:

@inproceedings{liao2019lanczos,
  title={LanczosNet: Multi-Scale Deep Graph Convolutional Networks},
  author={Liao, Renjie and Zhao, Zhizhen and Urtasun, Raquel and Zemel, Richard},
  booktitle={ICLR},
  year={2019}
}

Visualization

Benchmark

We also provide our own implementation of 9 recent graph neural networks on the QM8 benchmark:

• graph convolution networks for fingerprint (GCN-FP)
• gated graph neural networks (GGNN)
• diffusion convolutional neural networks (DCNN)
• Chebyshev networks (ChebyNet)
• graph convolutional networks (GCN)
• message passing neural networks (MPNN)
• graph sample and aggregate (GraphSAGE)
• graph partition neural networks (GPNN)
• graph attention networks (GAT)

You should be able to reproduce the following results of weighted mean absolute error (MAE x 1.0e-3):

Note:

• Above results are averaged over 3 runs with random seeds {1234, 5678, 9012}

Setup

To set up experiments, we need to download the preprocessed QM8 data and build our customized operators by running the following scripts:

./setup.sh

Note:

• We also provide the script dataset/get_qm8_data.py to preprocess the raw QM8 data which requires the installation of DeepChem. It will produce a different train/dev/test split than what we used in the paper due to the randomness of DeepChem. Therefore, we suggest using our preprocessed data for a fair comparison.

Dependencies

Python 3, PyTorch(1.0)

Other dependencies can be installed via

pip install -r requirements.txt

Run Demos

Train

• To run the training of experiment X where X is one of {qm8_lanczos_net, qm8_ada_lanczos_net, ...}:

  python run_exp.py -c config/X.yaml

Note:

• Please check the folder config for a full list of configuration yaml files.
• Most hyperparameters in the configuration yaml file are self-explanatory.

Test

• After training, you can specify the test_model field of the configuration yaml file with the path of your best model snapshot, e.g.,

  test_model: exp/qm8_lanczos_net/LanczosNet_chemistry_2018-Oct-02-11-55-54_25460/model_snapshot_best.pth

• To run the test of experiments X:

  python run_exp.py -c config/X.yaml -t

Run on General Graph Datasets

I provide an example code for a synthetic graph regression problem, i.e., given multiple graphs, each of which is accompanied with node embeddings, it is required to predict a real-valued graph embedding vector per graph.

• To generate the synthetic dataset:

  cd dataset

  PYTHONPATH=../ python get_graph_data.py

• To run the training:

  python run_exp.py -c config/graph_lanczos_net.yaml

Note:

• Please read dataset/get_graph_data.py for more information on how to adapt it to your own graph datasets.
• I only add support to LanczosNet by slightly modifying the learnable node embedding to the input node embedding in model/lanczos_net_general.py. It should be straightforward for you to add support to other models if you are interested.

Cite

Please cite our paper if you use this code in your research work.

Questions/Bugs

Please submit a Github issue or contact [email protected] if you have any questions or find any bugs.