Physics-aware Difference Graph Networks for Sparsely-Observed Dynamics

This repository is the official PyTorch implementation of "Physics-aware Difference Graph Networks for Sparsely-Observed Dynamics".

Sungyong Seo*, Chuizheng Meng*, Yan Liu, Physics-aware Difference Graph Networks for Sparsely-Observed Dynamics, ICLR 2020.

Data

Download the requried data.zip from Google Drive. Then,

cd /path/to/the/root/of/project
mkdir data
mv /path/to/data.zip ./data/
cd data
unzip data.zip

Environment

Docker (Recommended!)

First follow the official documents of Docker and nvidia-docker to install docker with CUDA support.

Use the following commands to build a docker image containing all necessary packages:

cd docker
bash build_docker.sh

This script will also copy the jupyter_notebook_config.py, which is the configuration file of Jupyter Notebook, into the docker image. The default password for Jupyter Notebook is 12345.

Use the following script to create a container from the built image:

bash rundocker-melady.sh

If the project directory is not under your home directory, modify rundocker-melady.sh to change the file mapping.

Manual Installation

# install python packages
pip install pyyaml tensorboardX geopy networkx tqdm
conda install pytorch==1.1.0 torchvision==0.2.2 cudatoolkit=9.0 -c pytorch
conda install -y matplotlib scipy pandas jupyter scikit-learn geopandas
conda install -y -c conda-forge jupyterlab igl meshplot

# install pytorch_geometric
export PATH=/usr/local/cuda/bin:$PATH
export CPATH=/usr/local/cuda/include:$CPATH
export LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH
pip install --verbose --no-cache-dir torch-scatter==1.2.0
pip install --verbose --no-cache-dir torch-sparse==0.4.0
pip install --verbose --no-cache-dir torch-cluster==1.3.0
pip install --verbose --no-cache-dir torch-spline-conv==1.1.0
pip install torch-geometric==1.1.2

# specify numpy==1.16.2 to avoid loading error (>=1.16.3 may require allow_pickle=True in np.load)
pip install -I numpy==1.16.2 

Run

Experiments in Section 3.1 "Approximation of Directional Derivatives"

See the Jupyter Notebook approx-gradient/synthetic-gradient-approximation.ipynb for details.

Experiments in Section 3.2 "Graph Signal Prediction" and Section 4 "Prediction: Graph Signals on Land-based Weather Stations"

cd scripts
python train.py --extconf /path/to/exp/config/file --mode train --device cuda:0

Examples:

• PA-DGN, Graph Signal Prediction of Synthetic Data

cd scripts
python train.py --extconf ../confs/iclrexps/irregular_varicoef_diff_conv_eqn_4nn_42_250sample/GraphPDE_GN_sum_notshared_4nn/conf.yaml --mode train --device cuda:0

• PA-DGN, Prediction of Graph Signals on Land-based Weather Stations

cd scripts
python train.py --extconf ../confs/iclrexps/noaa_pt_states_withloc/GraphPDE_GN_RGN_16_notshared_4nn/conf.yaml --mode train --device cuda:0

• PA-DGN, Sea Surface Temperature (SST) Prediction

cd scripts
python train.py --extconf ../confs/iclrexps/sst-daily_4nn_42_250sample/GraphPDE_GN_sum_notshared_4nn/conf.yaml --mode train --device cuda:0

Summary of Results

You can use results/print_results.ipynb to print tables of experiment results, including the mean value and the standard error of mean absolution error (MAE) of prediction tasks.

Reference

@inproceedings{seo*2020physicsaware,
title={Physics-aware Difference Graph Networks for Sparsely-Observed Dynamics},
author={Sungyong Seo* and Chuizheng Meng* and Yan Liu},
booktitle={International Conference on Learning Representations},
year={2020},
url={https://openreview.net/forum?id=r1gelyrtwH}
}