Neural Relational Inference (NRI)

Graph Neural Network for interacting systems

Given a time series data of nodes, the NRI model predicts the future node states and underlying relashionship between the nodes as edges.

This is a reproduction work of the neural relational inference (NRI) in Chainer. The original implementation by the authors is found here: ethanfetaya/NRI.

Please refer for details to the paper:
Neural relational inference for interacting systems.
Thomas Kipf*, Ethan Fetaya*, Kuan-Chieh Wang, Max Welling, Richard Zemel.
https://arxiv.org/abs/1802.04687 (*: equal contribution)

Dataset

Particle Physics Simulation Dataset

cd data
python generate_dataset.py

Training

Particle Physics Simulation Dataset

python train.py --gpu 0

Visualize results

python utils/visualize_results.py \
--args-file results/2019-01-22_10-20-25_0/args.json \
--encoder-snapshot results/2019-01-22_10-20-25_0/encoder_epoch-500.npz \
--decoder-snapshot results/2019-01-22_10-20-25_0/decoder_epoch-500.npz \
--gpu 0

Quantitative evaluation

Accuracy (in %) of unsupervised interaction recovery

Mean squared error (MSE) in predicting future states for simulations with 5 nodes