/LG-ODE

Modified version of LG-ODE model

Primary LanguagePython

Modified version of LG-ODE implementation

This repository contains a modified version of the LG-ODE implementation, used in our paper Temporal Graph Neural Networks for Irregular Data. Please see our main repository for the paper for more information. We are grateful to the original LG-ODE authors for making their code openly available.

The modifications made to the implementation are mainly:

  • Ability to load our datasets
  • The use of a validation set during training
  • Addition of a function to separately evaluate the model on test data

Using the LG-ODE model with datasets from TGNN4I

To train and evaluate the LG-ODE model with our datasets:

  • Convert the dataset to a suitable format for LG-ODE using the script convert_to_lgode.py found in the main repository.
  • Place the converted dataset (the full folder) in the data directory in this repository.
  • Train the LG-ODE model using run_models.py.
  • To evaluate a trained model on the test set, use run_models.py with the flags --test 1 and --load set to the name of the file to load model parameters from. The evaluation script will then create a .pt-file with predictions.
  • Use the script eval_saved_predictions.py in the main repository to compute the final test errors.

(Original README) LG-ODE

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.

Data Generation

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

Setup

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

Usage

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.

Citation

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}
}