Turing.jl: Question about custom inference parameterized by distribution types

[gajomi]

I have a noob question on how something might be done in Turing.jl, or any other extant PPL for that matter. Basically, I am wondering if I can implement bespoke inferences (for example, taking a particular expectation value) for hierarchical distributions that might only be appropriate for special cases of the hierarchical composition (as opposed to a general inference method like Gibbs sampling). I will try to clarify my question by way of an example.

Suppose I have a compound distribution involving some positive univariate distribution X of type T and a Poisson distribution Y = Poisson(X). I want to be able to estimate the mean of Y. In vanilla Julia I might define a compound distribution type (for univariate distributions that accept a single parameter)

struct CompoundDistribution{S,T}
   X{S}
   Y{T}
end

which would allow me to write a generic method that exploits the linear relationship between the mean of a Poisson distribution and its parameter

mean{S}(Z::CompoundDistribution{S,Poisson}) = mean(Z.X)

Is this sort of optimization something that Turing.jl supports or might in the future?

[trappmartin]

Turing is currently not exploiting special cases like conjugate priors. However, you can use a manually defined distribution, such as your CompoundDistribution, within Turing to exploit properties of special cases. You can have a look at http://turing.ml/latest/advanced.html for further details on how to define your CompoundDistribution so that it can be used in Turing.

I hope this answered your question.

[gajomi]

Yep. Thanks!