Models and ConditionalModels

A Model in Soss

Model Combinators

Building Inference Algorithms

Inference Primitives

At its core, Soss is about source code generation. Instances of this are referred to as inference primitives, or simply "primitives". As a general rule, new primitives are rarely needed. A wide variety of inference algorithms can be built using what's provided.

To easily find all available inference primitives, enter Soss.source<TAB> at a REPL. Currently this returns this result:

julia> Soss.source
sourceLogdensity         sourceRand            sourceXform
sourceParticles      sourceWeightedSample

The general pattern is that a primitive sourceFoo specifies how code is generated for an inference function foo.

For more details on inference primitives, see the Internals section.

Inference Functions

An inference function is a function that takes a ConditionalModel as an argument, and calls at least one inference primitive (not necessarily directly). The wrapper around each primitive is a special case of this, but most inference functions work at a higher level of abstraction.

There's some variability , but is often of the form

foo(d::ConditionalModel, data::NamedTuple)

For example, advancedHMC uses TuringLang/AdvancedHMC.jl , which needs a logdensity and its gradient.

Most inference algorithms can be expressed in terms of inference primitives.

Chain Combinators

Internals

Models

struct Model{A,B}
    args  :: Vector{Symbol}
    vals  :: NamedTuple
    dists :: NamedTuple
    retn  :: Union{Nothing, Symbol, Expr}
end

function sourceWeightedSample(_data)
    function(_m::Model)

        _datakeys = getntkeys(_data)
        proc(_m, st :: Assign)     = :($(st.x) = $(st.rhs))
        proc(_m, st :: Return)     = nothing
        proc(_m, st :: LineNumber) = nothing

        function proc(_m, st :: Sample)
            st.x ∈ _datakeys && return :(_ℓ += logdensity_def($(st.rhs), $(st.x)))
            return :($(st.x) = rand($(st.rhs)))
        end

        vals = map(x -> Expr(:(=), x,x),variables(_m))

        wrap(kernel) = @q begin
            _ℓ = 0.0
            $kernel

            return (_ℓ, $(Expr(:tuple, vals...)))
        end

        buildSource(_m, proc, wrap) |> flatten
    end
end