zhusuan

Why we included this project

ZhuSuan targets developers doing Bayesian deep learning who want TensorFlow's familiar graph and gradient machinery plus the ability to express uncertainty in their models. It layers probabilistic building blocks onto the same stack, so you can define stochastic latent variables and train generative models such as variational autoencoders without hand-writing the inference math. The library offers several inference routes: variational inference with a choice of gradient estimators, importance sampling, Hamiltonian Monte Carlo with parallel chains, and a family of stochastic-gradient MCMC samplers. The bundled examples show how a probabilistic model gets assembled and trained on a standard deep-learning stack, which makes a practical starting point if you are new to the area.