Beware: this is nothing but a proof of concept. Specifically, due to unfortate choices in
IRTracker, handling traces of models with data sets not particularly small leads to exploding
type inference time, making the method unusable.
The first thing this library does is to slice out dependency graphs of the random variables in a DynamicPPL model:
@model function test0(x)
λ ~ Gamma(2.0, inv(3.0))
m ~ Normal(0, sqrt(1 / λ))
x ~ Normal(m, sqrt(1 / λ))
end
graph0 = trackdependencies(test0(1.4))will result in
⟨2⟩ = 1.4
⟨4⟩ = λ ~ Gamma(2.0, 0.3333333333333333) → 0.5181638066817373
⟨5⟩ = /(1, ⟨4⟩) → 1.9298916425751296
⟨6⟩ = sqrt(⟨5⟩) → 1.3892053997070157
⟨8⟩ = m ~ Normal(0, ⟨6⟩) → 0.7621682353026793
⟨9⟩ = /(1, ⟨4⟩) → 1.9298916425751296
⟨10⟩ = sqrt(⟨9⟩) → 1.3892053997070157
⟨12⟩ = x ⩪ Normal(⟨8⟩, ⟨10⟩) ← ⟨2⟩
Notation: references to values are written in ⟨angle brackets⟩. The names of random variables in
tilde statements are annotated with names (and possibly indices). Unobserved random variables
(“assumptions”) are notated by a simple ~, observed ones use ⩪. Equality signs denote
assignment of intermediate deterministic values, as in normal SSA-form IR. The results of
deterministic or assumed values come after the →; ← is used for values that are observed.
Numbers of references are quite contiguously numbered, but some numbers may be missing, since they get removed by the slicing process.
The result of trackdependencies is a Graph, which is essentially an ordered dictionary from
References to statements, that can be either of Assumption, Observation, Call, or
Constant. For debugging, you can also use integers to index the graph:
julia> graph0[4]
Gamma(2.0, 0.3333333333333333) → 1.0351608689025245
You can also write the graph to a GraphViz dot file:
julia> AutoGibbs.savedot("/tmp/graph.dot", graph0)
shell> dot /tmp/graph.dot -Tpdf > /tmp/graph.pdf
But don’t expect any magic to make it look good:
