LG-ODE is an overall framework for learning continuous multi-agent system dynamics from irregularly-sampled partial observations considering graph structure.

You can see our Neurips 2020 paper “Learning Continuous System Dynamics from Irregularly-Sampled Partial Observations” for more details.

This implementation of LG-ODE is based on Pytorch Geometric API.

Generate simulated datasets (spring, charged particles) by running:

cd data
python generate_dataset.py 

This generates the springs dataset, use --simulation charged for charged particles.

As simulated data is too large, we provide a toy-data from spring dataset and can be found under data/example_data

Motion dataset can be downloaded CMU MoCap

This implementation is based on pytorch_geometric. To run the code, you need the following dependencies:

• Python 3.6.10

• Pytorch 1.4.0

• pytorch_geometric 1.4.3

  • torch-cluster==1.5.3
  • torch-scatter==2.0.4
  • torch-sparse==0.6.1

• torchdiffeq 0.0.1

• numpy 1.16.1

Execute the following scripts to train on the sampled data from spring system:

python run_models.py

There are some key options of this scrips:

• --sample-percent-train: This is the observed percentage in your training data.

• --sample-percent-test: This is the observed percentage in your testing data.

• --solver : This is for choosing your ODE Solver.

• --extrap: Set True to run in the extrapolation mode, otherwise run in the interpolation mode.

The details of other optional hyperparameters can be found in run_models.py.

Please consider citing the following paper when using our code for your application.

@inproceedings{LG-ODE,
  title={Learning Continuous System Dynamics from Irregularly-Sampled Partial Observations},
  author={Zijie Huang and Yizhou Sun and Wei Wang},
  booktitle={Advances in Neural Information Processing Systems},
  year={2020}
}