I’m looking for some direction or examples on implementing a factor graph + belief propagation model in Edward.
For example, I have a factor graph representation of sequential states (variables) that have Gaussian process factors between them, and on each variable I also have unary factors representing Gaussian priors. Then I run message-passing over the factor graph (to optimize all the model’s Gaussian params), converging on a set of values for each variable as a solution. How can this model + inference be implemented in Edward?
I’m aware of the graphical models examples (http://edwardlib.org/api/model-compositionality), but it’s still unclear how to implement my slightly more complex graph.
Thanks in advance!!