Variational Integrator Networks

Overview

Link to paper

Variational Integrator Networks for Physically Structured Embeddings

Run experiment

python run_exp.py root_dir model_name system_name observations num_train_traj num_train_steps seed
e.g.
python run_exp.py experiments VIN_VV pendulum pixels 1 60 1

Dependencies

  • tensorflow 2.1
  • tensorflow_probability
  • gin-config
  • see requirements.txt

Example: Ideal Pendulum, Noisy Observations

Setup

  • Train on 15s of observations (150 datapoints)
  • Test on noisless initial state, forecast for 10s

Recurrent Residual Network (Left) / Variational Integrator Network (Right)

Example: Ideal Pendulum, Pixel Observations

Setup

  • Train on 6s of 28x28 pixel observations (60 datapoints)
  • Infer latent initial state from 1s of data
  • Forecast for 10s, reconstruct latent path

Recurrent ResNet (Left) / VIN (Middle) / VIN on SO(2) Manifold (Right)