Provably Powerful Graph Networks

This repository holds the code for the paper: https://arxiv.org/abs/1905.11136. If you want to learn more about it, check out these two blog posts explaining invariant graph networks and their power: http://irregulardeep.org. A torch version of this algorithm can be found in https://github.com/hadarser/ProvablyPowerfulGraphNetworks_torch.

Data

Before running the code the data should be downloaded using the following commands:

cd ProvablyPowerfulGraphNetworks
python utils/get_data.py

This script will download all the data from Dropbox links.

Code

Prerequisites

python3

TensorFlow gpu 1.9.0.

Additional modules: numpy, pandas, matplotlib, tqdm, easydict

Running the tests

The folder main_scripts contains scripts that run different experiments:

1. To run 10-fold cross-validation with our chosen hyper-parameters, run the main_10fold_experiment.py script. You can choose the dataset in 10fold_config.json or using the command line option. These hyper-parameters refer to version 1 from the paper. example: to run 10-fold cross-validation experiment:

python main_scripts/main_10fold_experiment.py --config=configs/10fold_config.json --dataset_name=NCI1

2. To run the QM9 experiment with our hyper-parameters, run the main_qm9_experiment.py script:

python main_scripts/main_qm9_experiment.py --config=configs/qm9_config.json

In the paper, we have two models for the QM9 task: the first predicts a all the outputs quantity at once, while the other predicts a single chosen output quantity. You can switch between these models by changing the following line in the configs/qm9_config.json:

  "target_param": false,

to the chosen target (range 0-11).

Running other vesions from the paper

Version 2

in configs/10fold_config.json, change:

    "new_suffix": false

in utils/config.py change:

LEARNING_RATES = {'COLLAB': 0.0001, 'IMDBBINARY': 0.00005, 'IMDBMULTI': 0.0001, 'MUTAG': 0.0001, 'NCI1':0.0001, 'NCI109':0.0001, 'PROTEINS': 0.001, 'PTC': 0.0001}
DECAY_RATES = {'COLLAB': 0.5, 'IMDBBINARY': 0.5, 'IMDBMULTI': 0.75, 'MUTAG': 1.0, 'NCI1':0.75, 'NCI109':0.75, 'PROTEINS': 0.5, 'PTC': 1.0}
CHOSEN_EPOCH = {'COLLAB': 150, 'IMDBBINARY': 100, 'IMDBMULTI': 150, 'MUTAG': 500, 'NCI1': 200, 'NCI109': 250, 'PROTEINS': 100, 'PTC': 400}

Version 3

in configs/10fold_config.json, change:

    "block_features": [256,256,256],
    "depth_of_mlp": 3,
    "new_suffix": true

in utils/config.py change:

LEARNING_RATES = {'COLLAB': 0.0005, 'IMDBBINARY': 0.00001, 'IMDBMULTI': 0.0001, 'MUTAG': 0.0005, 'NCI1':0.0005, 'NCI109':0.0001, 'PROTEINS': 0.0005, 'PTC': 0.001}
DECAY_RATES = {'COLLAB': 0.5, 'IMDBBINARY': 0.75, 'IMDBMULTI': 1.0, 'MUTAG': 0.5, 'NCI1':0.75, 'NCI109': 1.0, 'PROTEINS': 0.75, 'PTC': 0.5}
CHOSEN_EPOCH = {'COLLAB': 85, 'IMDBBINARY': 100, 'IMDBMULTI': 150, 'MUTAG': 150, 'NCI1': 100, 'NCI109': 300, 'PROTEINS': 100, 'PTC': 200}

Note: The script mentioned in the data section above will download a processed version of QM9 which is needed for our main code. We also share our processing code (requires pytorch), which is is based on the pytorch-geometric package. see: https://pytorch-geometric.readthedocs.io/en/latest/_modules/torch_geometric/datasets/qm9.html#QM9:

cd ProvablyPowerfulGraphNetworks
python utils/get_qm9_data.py