Official PyTorch implementation of "Physics-aware Difference Graph Networks for Sparsely-Observed Dynamics".

Last update: Nov 20, 2022

Related tags

Deep Learning ICLR2020-PADGN

Overview

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}
}

Official PyTorch implementation of "Physics-aware Difference Graph Networks for Sparsely-Observed Dynamics".

Related tags

Overview

Physics-aware Difference Graph Networks for Sparsely-Observed Dynamics

Data

Environment

Docker (Recommended!)

Manual Installation

Run

Experiments in Section 3.1 "Approximation of Directional Derivatives"

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

Summary of Results

Reference

Owner

USC-Melady

Deep GPs built on top of TensorFlow/Keras and GPflow

Fake News Detection Using Machine Learning Methods

Code for "The Box Size Confidence Bias Harms Your Object Detector"

Code repository for the paper "Doubly-Trained Adversarial Data Augmentation for Neural Machine Translation" with instructions to reproduce the results.

Personalized Transfer of User Preferences for Cross-domain Recommendation (PTUPCDR)

Action Segmentation Evaluation

ContourletNet: A Generalized Rain Removal Architecture Using Multi-Direction Hierarchical Representation

Code and training data for our ECCV 2016 paper on Unsupervised Learning

AlphaBot2 Pi Core software for interfacing with the various components.

The PyTorch implementation of Directed Graph Contrastive Learning (DiGCL), NeurIPS-2021

GBIM(Gesture-Based Interaction map)

DGN pymarl - Implementation of DGN on Pymarl, which could be trained by VDN or QMIX

✅ How Robust are Fact Checking Systems on Colloquial Claims?. In NAACL-HLT, 2021.

OBBDetection is a oriented object detection library, which is based on MMdetection.

An easier way to build neural search on the cloud

Ivy is a templated deep learning framework which maximizes the portability of deep learning codebases.

Leaf: Multiple-Choice Question Generation

Code for the Interspeech 2021 paper "AST: Audio Spectrogram Transformer".

Code for STFT Transformer used in BirdCLEF 2021 competition.

[NeurIPS'20] Multiscale Deep Equilibrium Models