Based on question/conversation on slack (with @devmotion, @mohamed82008):

A bit of a different take on TuringLang/DynamicPPL.jl#94

I would like to specify which variables are tracked by Turing. In the above issue I believe users are trying to track additional variables (which are transforms of others), but sometimes I also want to specify that I do not want to track a specific variable (in my example below for example n might be very large, and I don't want to track all of theta).

@model function tpl(y, n, m, ::Type{TV}=Vector{Float64}) where {TV}
    a = TV.(undef, m)
    b = TV.(undef, m)
    for i in 1:m
        a[i] ~ LogNormal(0.0, 0.5)
        b[i] ~ Normal(0.0, 3.0)
    end
    theta = TV(undef, n)
    for p in 1:n
        theta[p] ~ Normal(0,1)
        for i in 1:m
            prob = invlogit(a[i] * (theta[p] - b[i]))
            y[i,p] ~ Bernoulli(prob)
        end
    end
    return a, b
end;

Users in the slack have suggested using the variables returned in the model macro (a and b in my example) as the variables which are tracked. I think this is a great solution (and intuitively this is what I thought the return would do when I first read through the tutorial.