A lightweight framework for extra quick iteration on Bayesian model development using black box variational inference and mean-field gaussian assumptions.
More documentation to come soon!
This framework is an extension of the work done by David Duvenaud and Ryan Adams in "Black-Box Stochastic Variational Inference in Five Lines of Python". (See https://www.cs.toronto.edu/~duvenaud/papers/blackbox.pdf). The object oriented framework built here was also inspired in part by James Vucovic's (@jamesvuc) fork, in which he implemented a variety of really useful gradient estimates. Currently, this package uses the reparameterization gradient; in the future, I may include his other estimates as well.