PyTorch implementation of Interaction Networks (Battaglia et al., 2016).
Interaction Networks for Learning about Objects, Relations and Physics https://arxiv.org/abs/1612.00222
"Reasoning about objects, relations, and physics is central to human intelligence, and a key goal of artificial intelligence. Here we introduce the interaction network, a model which can reason about how objects in complex systems interact, supporting dynamical predictions, as well as inferences about the abstract properties of the system."
python physics.py
Automatically generates motion data for a three-particle system with arbitrary orbit, together with a video showing a sample of the data saved as "test.mp4"
python model.py
Trains the Interaction Network and plot loss for each epoch