https://github.com/steindoringi/Variational_Integrator_Networks

https://github.com/steindoringi/Variational_Integrator_Networks

Last synced: about 2 months ago
JSON representation

• Host: GitHub
• URL: https://github.com/steindoringi/Variational_Integrator_Networks
• Owner: steindoringi
• Created: 2020-02-20T02:00:29.000Z (over 4 years ago)
• Default Branch: master
• Last Pushed: 2022-12-08T03:45:39.000Z (almost 2 years ago)
• Last Synced: 2024-07-04T02:14:11.791Z (3 months ago)
• Language: Python
• Size: 44.9 MB
• Stars: 18
• Watchers: 4
• Forks: 7
• Open Issues: 8
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

        
# Variational Integrator Networks

## Overview

### Link to paper

![Variational Integrator Networks for Physically Structured Embeddings](https://arxiv.org/abs/1910.09349)

### 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)