I am currently interested in performing variational inference on large-scale real-valued graphical models (dynamic Bayesian networks for instance, with tens of thousands of nodes). While I am impressed by the width and depth of the Pyro library, I wonder if it is the right tool for the job.
Indeed, the graphical models that are discussed in the tutorials (mixtures, HMMs, neural networks) have simple “layered” structures that work well with vectorized
plate statements. For general graphs, I feel like the only way is to
pyro.sample each variable by hand from its parents in an endless
for loop, thus missing out on the benefits of the Pytorch backend.
Does this problem ring a bell for anyone?
Thanks in advance