Turing resamples variable that is passed as variable to condition on when querying

[mgmverburg]

When running the below code, one observes that x is being resampled when performing the result1 query, and that a different value for x is therefore printed instead of "1.0" as the value should be (since that is what it is conditioned on).

using Turing, Distributions

@model function gdemo(x, y)
	println("START")
    s ~ InverseGamma(2, 3)
    m ~ Normal(0, sqrt(s))
    x ~ filldist(Normal(m, sqrt(s)), length(y))
    @show x
    for i in 1:length(y)
        @show x[i]
        y[i] ~ Normal(x[i], sqrt(s))
    end
end

model_gdemo = gdemo([1.0, 0.0], [1.5, 0.0])
c2 = sample(model_gdemo, NUTS(0.65), 100)

result1 = prob"y = [1.5] | chain=c2, model = model_gdemo, x = [1.0]"

[devmotion]

This is caused by the branch in

DynamicPPL.jl/src/prob_macro.jl

Line 190 in d480908

if isdefined(right, :chain)

: If chain is specified, then only the values in the chain are considered but not any separately specified random variables it seems (BTW this reveals another bug,

DynamicPPL.jl/src/prob_macro.jl

Line 197 in d480908

model(vi, SampleFromPrior(), ctx)

does not resample the model so all x values are identical).

I guess here it would actually make sense to construct a new model with gdemo([1.0], [1.5]) (it is also possible to change some values in an existing model, there is just no convenient API yet IIRC) and then use the loglikelihood function to compute the log-likelihood for each sample of s in the chain - it is possible to do this with individual VarInfo objects but we do not have a function that evaluates loglikelihood for a chain automatically.

[devmotion]

You can e.g. manually compute

julia> vi = Turing.VarInfo(gdemo(missing, [1.5]));

julia> map(c2[:s], c2[:m]) do s, m
        loglikelihood((y = [1.5],), (x = [1.0], s = s, m = m), model_gdemo, vi)
       end

In this call you can also use nothing instead of vi but then every call of loglikelihood will sample from the prior of the model first to populate a VarInfo object before evaluating the provided values.

[mgmverburg]

Just to be clear, is the suggestion an alternative/workaround to the bug or?

Currently trying to figure out why I get a stackoverflowerror on some of my other models with the VarInfo part. Trying to reduce it to a MWE but have not succeeded with that yet.

[devmotion]

Yes, you could use this instead of the prob macro (actually it returns the logarithm of what you wanted to compute above, so maybe you want to apply exp). But I'm not sure if I understand your question?

[mohdibntarek]

I have a fix for this. I will make a PR soon.

[devmotion]

The fix is merged and a new version of DynamicPPL will be available soon: JuliaRegistries/General#25122 Thanks @mohamed82008!