A number of imperative Probabilistic Programming Languages (PPLs) have been recently proposed, but the imperative style choice makes it very hard to deduce the dependence structure between the latent variables, which can also change from iteration to iteration. We propose a new declarative style PPL, Bean Machine, and demonstrate that in this new language, the dynamic dependence structure is readily available. Although we are not the first to propose a declarative PPL or to observe the advantages of knowing the dependence structure, we take the idea further by showing other inference techniques that become feasible or easier in this style. We show that it is very easy for users to program inference by composition (combining different inference techniques for different parts of the model), customization (providing a custom hand-written inference method for specific variables), and blocking (specifying blocks of random variables that should be sampled together) in a declarative language. A number of empirical results are provided where we backup these claims modulo the runtime inefficiencies of unvectorized Python. As a fringe benefit, we note that it is very easy to translate statistical models written in mathematical notation into our language.
