This code is a basic implementation in tensorflow, of the paper "Variational Inference for Nonlinear Dynamics", accepted for the Time Series Workshop at NIPS 2017. It represents a sequential variational autoencoder that is able to infer nonlinear dynamics in the latent space. The training algorithm makes use of a novel, two-step technique for optimization based on the Fixed Point Iteration method for finding fixed points of iterative equations.
| Original | Inferred |
|---|---|
 |
 |
The code is written in Python 3.5. You will need the bleeding edge versions of the following packages:
- tensorflow
- seaborn
In addition, up-to-date versions of numpy, scipy and matplotlib are expected.
Firing python runner.py works right off the bat. The code will find a two dimensional encoding and dynamical system describing the provided Gaussian data. A figure is provided with the original dynamical system and simulated trajectories that can be compared with the resulting fit. The hyperparameter plot2D, default-set to True, will produce these path+dynamics plots automatically for 2D latent spaces.