Gen Compose

Gen.jl provides an excelent ecosystem for probablistic programming, enabling a unified language for generative models and infernce procedures. However, using Gen to develop and analyze combinations of posteriors and inference procedures requires a considerable amount of boilerplait. Gen_Compose aims to standardize inference approximation without limiting the expressivity of Gen.

Design Philosophy

Gen_Compose revolves around InferenceChain{Q,P} which is parameterized by a combination of Q<:Query and P<:InferenceProcedure. A query denotes an approximation target (e.g., a conditional distribution / posterior over a generative model) or more formally as an estimand. A inference procedure serves as an estimator over the estimand. The result of this combination is the InferenceChain (the estimate).

julia> run_chain(pf, seq_query, nsteps)
PFChain{SequentialQuery, ParticleFilter}

The queries and procedures provided by Gen_Compose are not intended to be exhaustive, but instead provide a template for trivial extensability. For example, a more complex instance of AbstractParticleFilter or MCMC could be implemented by adding a new step! or initialize_procedure method.

Gen_Compose provides two basic forms of queries:

1. StaticQuery : an unfactorized conditional of the form $Pr(H \mid O)$
2. SequentialQuery : an time-factorized conditional of the form $Pr(H \mid O) \sym Pr(H_0) * \prod_t Pr(O_t \mid H_t) Pr(H_t \mid H_{t-1})$

Like procedures, new queries could be defined that explore different forms of factorization over $Pr(H \mid O)$. It's straightforward to extend a procedure to take advantage of such factorization by simply defining a method such as

step!(chain::PFChain{Q}) where {Q<:MyQuery}

If one requires a more extensive overhaul, then defining new subtypes of InferenceChain{Q, P} could look like

abstract type MyParticleFilter <: AbstractParticleFilter end
mutable struct MyPFChain{Q<:MyQuery, P<:MyParticleFilter} <: InferenceChain{Q, P}
    ...
end

Chain Logging

dlog = JLD2Logger(nsteps, "chain_log.jld2";overwrite = true)
chain = run_chain(proc, query, nsteps, dlog)