We implement a Bayesian Neural Network on Koopman operator with

• Tensorflow 1.8
• Edward 1.3.5

in Python 2 environment.

At the time when I was writing the code, there is a shift from Python 2 to 3 and shifting of Edward to Tensorflow.Probability. Thus, this code for now is only for informative purpose. I will come back for an upgrade in the future into Python 3 environment and make everything in PyTorch.

It will be helpful if you want to know

• how is everything exactly coded to enforce stability constraint of Koopman operator?
• how is SVD-augmented autoencoder built?
• how is Koopman operator is built recurrently? how is the loss function constructed?

@article{pan2020physics,
title={Physics-informed probabilistic learning of linear embeddings of nonlinear dynamics with guaranteed stability},
author={Pan, Shaowu and Duraisamy, Karthik},
journal={SIAM Journal on Applied Dynamical Systems},
volume={19},
number={1},
pages={480--509},
year={2020},
publisher={SIAM}
}