Diffusion Probabilistic Models for Graph Structured predictions

This is an official implementation of our paper "Diffusion Probabilistic Models for Graph Structured predictions."

Repository Overview

We provide the PyTorch implementation for DPM-GSP framework here. The repository is organised as follows:

|-- DPM-GSP-{fully-supervised, semi-supervised, reasoning} # DPM-GSP for supervised node classification, semi-supervised node classification, and reasoning tasks
    |-- config/ # configurations (Hyperparameters used in the experiments are specified in the Appendix C of our paper.)
    |-- parsers/ # the argument parser
    |-- models/ # model definition
    |-- method_series/ # training method
    |-- data/ # dataset
    |-- logs_train/ # training logs
    |-- utils/ # data process and others
    |-- main.py # the training code

Setting up the environment

You can set up the environment by following commands.

conda create -n DPM-GSP python=3.10
conda activate DPM-GSP
conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 cudatoolkit=11.3 -c pytorch
pip install tqdm pyyaml easydict torch-sparse torch-scatter==2.0.9

You also need to install torch-geometric package. Each experiment requires a different version.

DPM-GSP for fully-supervised and reasoning

pip install torch-geometric==1.7.1

DPM-GSP-semi-supervised

pip install torch-geometric==2.1.0

Running

At each directory, you can use the following command to obtain the results of our paper.