Contexts and acclogp!

[treigerm]

If one wants to use the acclogp! function to manually increment the log probability manually then this prevents the use of PriorContext or LikelihoodContext to compute the prior or likelihood, respectively. This is due to the fact that acclogp! does not dispatch on the current context so it always increments lp.

MWE:

using Turing

@model function model(y)
    x ~ Normal(0, 1)
    y ~ Normal(x, 1)
    Turing.acclogp!(_varinfo, 1.0)
end

xval = 1.0
yval = 3.0

m = model(yval)
vi = Turing.VarInfo(m)
vi[@varname(x)] = [xval;]

m(vi, Turing.SampleFromPrior(), Turing.PriorContext())
lp = Turing.getlogp(vi)

@show lp
@show logpdf(Normal(0, 1), xval)

Results in:

lp = -0.41893853320467267
logpdf(Normal(0, 1), xval) = -1.4189385332046727

Is this behaviour intended? Is there a known way to work around this?

[devmotion]

I guess the most common (only?) non-internal use case for acclogp! has been to reject samples by adding a acclogp!(_varinfo, -Inf); return statement somewhere in the model definition. In an upcoming release users won't have to deal with acclogp! and _varinfo anymore in this case and can just use @addlogprob -Inf (see #143).

There is an easy fix for your problem though. Similar to _varinfo, there exists an internal variable _context for the context as well. So you could just check

if _context isa DefaultContext
    acclogp!(_varinfo, 1.0)
end

instead.

[treigerm]

Yes that would solve my problem, thanks!

I have used acclogp! to manually marginalise out the responsibility parameter in a mixture model (as in the Stan User's Guide) before so that's another possible use case. In that case I would want acclogp! to be included in the prior term but I guess that can also be done by manually checking the context.

Would it be possible to maybe add a note to the new @addlogprob macro saying that this term will always be added to the log probability? I more or less accidentally stumbled upon this behaviour and would have almost missed it, so adding a note could possible save someone else some time.

[devmotion]

Fixed by TuringLang/Turing.jl#1391.