TensorFlow Probability (TFP) is an open-source Python library for probabilistic reasoning and statistical analysis. TFP implements a suite of distributions and bijectors which are accurate, efficient, and differentiable. TFP also provides building blocks for modern inference algorithms like gradient-based Markov chain Monte Carlo and variational inference. We present a case-study of end-to-end Bayesian modeling - from writing down a generative model and reasoning about the prior predictive distribution to performing Hamiltonian Monte Carlo and diagnosing the quality of the fit. Along the way, we highlight TFP’s unique features. Specifically we cover the JointDistribution abstraction - a declarative representation of graphical models. We also showcase the performance benefits when fitting models using specialized hardware such as GPUs and TPUs.
Presenter: Colin Carroll