Discrete Graph Auto-Encoder

Welcome to the official repository of the Discrete Graph Auto-Encoder (DGAE)!

This repository contains the source code for the VQ-GAE, a powerful autoencoder for graph data that uses vector quantization to learn discrete representations.

Installation

Their are 4 dependencies to install. It is recommanded to follow this order.

1. rdkit rdkit.org/docs/Install.html

2. pytorch pytorch.org/get-started/locally/

3. pytorch-geometric pytorch-geometric.readthedocs.io/en/latest/install/installation.html

4. dig https://diveintographs.readthedocs.io/en/latest/intro/installation.html

Run the model

Once you have intalled the dependencies, you can run the DGAE by executing the main.py file.

The --work_type argument allows you to specify the type of work you want to do with the DGAE (train_autoencoder, train_prior, sample).

The --dataset argument specify the dataset, you want to use (ego, community, enzymes, qm9 or zinc).

The --model_folder, specify the folder name where the model parameters are stored. If you choose to train the prior or sample from a model, you will need to specify a folder using the --model_folder argument. This folder should contain the configuration for your model in a yaml file. ()

All the other configurations can be specify in a the yaml file as stored in the folder 'config'.

Checkpoints

We provide checkpoints for 3 datasets 'enzymes', 'qm9' and 'zinc': The data should be downloaded when running the model (be aware that the preprocessed (kekulized) zinc dataset is 5.4 GB in size).

• Enzymes:

  • Auto-encoder
  • Prior

• Qm9:

  • Auto-encoder
  • Prior

• Zinc:

  • Auto-encoder
  • Prior

Thank you for using the DGAE or using it as baseline.