DAGNN

A graph neural network tailored to directed acyclic graphs that outperforms conventional GNNs by leveraging the partial order as strong inductive bias besides other suitable architectural features.
Source code for ICLR 2021 paper https://openreview.net/forum?id=JbuYF437WB6.

Installation

• Tested with Python 3.7, PyTorch 1.5.0, and PyTorch Geometric 1.6.0
• Set up an Anaconda environment: ./setup.sh
  (set the CUDA variables in the script)
• Alternatively, install the above and the packages listed in requirements.txt

Overview

• /dvae
  Basically the experiment code and data from D-VAE. The DAGNN models in this directory contain additional methods to support the decoding loss.
• /ogb
  Basically the code from the Open Graph Benchmark (OGB). We just added additional data preprocessing steps necessary for DAGNN.
• /ogbg-code
  Experiment code for TOK and LP experiments over the ogbg-code2 data from OGB. The DAGNN model in this directory, can be considered as the basic implementation of DAGNN.
• /scripts
  Scripts for running the experiments.
• /src
  Basic utilities used in all experiments.

OGB Experiments

• Run ./scripts/ogb_tok.sh or ./scripts/ogb_lp.sh
• The ogbg-code2 data will be downloaded at first run, set the directory for storage in the scripts ($DATA). We updated the code since the ogbg-code dataset used in the paper is deprecated and not available anymore.
• By default, the script will run DAGNN over a 15% random subset of ogbg-code2. To change these settings, see the comments in the scripts.

DAG Scoring Experiments

• The two datasets are Neural Architectures (NA) and Bayesian Networks (BN) in /dvae/data. Since the latter is rather large, we suggest to start with the former.
• Run ./scripts/na_train.sh [DEVICE_ID] [MODEL] for training and ./scripts/na_eval.sh [DEVICE_ID] [MODEL] [EVALUATION_POINT] for evaluation. And similar for BN.
  [MODEL] can be one of: DAGNN, DAGNN_BN, SVAE_GraphRNN, DVAE_GCN, DVAE_DeepGMG, DVAE, DVAE_BN.
• For example ./scripts/na_train.sh 0 DAGNN, ./scripts/na_eval.sh 0 DAGNN 100.
• Note that the learning rates and epoch numbers in the scripts are the ones which we used for DAGNN. The experiment parameters for the D-VAE models and baselines can be found here.

Please leave an issue if you have any trouble running the code or suggestions for improvements.