Equivariant Graph Mechanics Networks with Constraints (ICLR 2022)

Wenbing Huang*, Jiaqi Han*, Yu Rong, Tingyang Xu, Fuchun Sun, Junzhou Huang

[OpenReview] [Paper] [Poster]

Graph Mechanics Networks (GMNs) are novel graph neural networks particularly powerful for modeling the dynamics of constrained systems. GMNs are equivariant to translations, rotations, and reflections. The flowchart of our model is provided in the figure below. Please refer to our paper for more details.

Dependencies

python==3.8
torch==1.8.0
scikit-learn==0.24.2
networkx==2.5.1

A more detailed Python environment is depicted in requirements.yml.

Data Preparation

1. Simulation Dataset

Our simulation scripts are placed under spatial_graph/n_body_system/dataset.

To generate datasets containing multiple isolated particles, sticks, and hinges, use the following command under the path spatial_graph/n_body_system/dataset:

python -u generate_dataset.py --num-train 5000 --seed 43 --n_isolated 3 --n_stick 2 --n_hinge 1 --n_workers 50

where the arguments n_isolated, n_stick, and n_hinge indicate the number of isolated particles, sticks, and hinges, respectively. The argument n_workers refers to the number of parallel threads for parallel data generation. For other potential arguments, please refer to generate_dataset.py.

By default, the generated data will be placed in a new folder named data under the current path.

Note: On our CPU machine with 50 parallel workers, the entire data generation process takes from 10 minutes to 1~2 hours, depending on the complexity of the particle system (number of particles, sticks, and hinges).

2. MD17

The MD17 dataset can be downloaded from MD17. Download the dataset and place the files under spatial_graph/MD17. The splits are also provided in the MD17 folder.

3. Motion Capture

The raw data were obtained via CMU Motion Capture Database. The preprocessed dataset as well as the splits are provided in spatial_graph/motion folder.

Training and Evaluation

1. Simulation Dataset

Under the root path, simply use

python -u spatial_graph/main.py --config_by_file

where the --config_by_file option enables loading the hyper-parameters from the config files in configs folder.

To run experiments in different scenarios, simply change the hyper-parameters in the config files. For instance, one may change the n_isolated, n_stick, and n_hinge options to evaluate the model under various combinations of isolated particles, sticks, and hinges.

2. MD17

python -u spatial_graph/main_md17.py --config_by_file

3. Motion Capture

python -u spatial_graph/main_motion.py --config_by_file

Visualizations

We provide a toy script for visualization in visualization/vis.ipynb.

We exhibit some visualizations of the predictions as below (Left: initial position, Middle: GMN, Right: EGNN):

Simulation dataset

Motion Capture

Here we also provide some dynamic cases tested with a batch of continuous input frames (Blue: prediction, Red: ground truth). The toy script is deferred to visualization/vis_cont.ipynb.

Citation

If you find our work helpful, please cite as:

@inproceedings{
huang2022equivariant,
title={Equivariant Graph Mechanics Networks with Constraints},
author={Wenbing Huang and Jiaqi Han and Yu Rong and Tingyang Xu and Fuchun Sun and Junzhou Huang},
booktitle={International Conference on Learning Representations},
year={2022},
url={https://openreview.net/forum?id=SHbhHHfePhP}
}

New Update

We update the code of our GMN-OC, an enhanced version of GMN for molecular data, to assets/gmnoc folder. GMN-OC contributes to our winning the first place of The Open Catalyst Challenge 2022.

Typo: Delta frame should be 3000 instead of 5000 in the paper (on Page 16). The configuration described in the code (https://github.com/hanjq17/GMN/blob/main/configs/simple_config_md17.json) is correct.

Contact

If you have any questions, feel free to reach us at:

Jiaqi Han: [email protected]