I mean, it should be.. you are learning a probability distribution over the nodes and then performing some kind of inference at eval time.