Particle Gibbs does not return log probability

[colehurwitz]

Would it be possible to return the log probability when using PG?

[mohamed82008]

If you have usage questions, please post a minimal working example. These questions are also more suited to the Julia slack (https://julialang.slack.com) channel #turing.

[cpfiffer]

@colehurwitz31 You should be able to extract the log probability from the chain for each sample after you've run it using chain[:lp], with chain being the variable containing the value returned from sample. Does that work for you?

[xukai92]

@mohamed82008 @cpfiffer Sorry for the confusing - I was helping @colehurwitz31 with Turing and found this issue so I asked him to create an issue on this.

Yes PG indeed doesn't report log-joint in chain correct - all :lp in the chain are 0s.

using Turing

@model gdemo(x, y) = begin
    s ~ InverseGamma(10,3)
    m ~ Normal(0,sqrt(s))
    x ~ Normal(m, sqrt(s))
    y ~ Normal(m, sqrt(s))
    return s, m
end

x, y = 1.5, 2.0

chn = sample(gdemo(x, y), PG(20, 500))
chn[:lp] # => all 0s

[xukai92]

Because PG doesn't track the log-joint, there is no way to extract it directly - is my understanding correct? @yebai

I guess a simple solution is to evaluate the log-joint after each MCMC step, though it adds some more computation (ofc we should have an interface to make it optional). Otherwise we should provide some utility functions to re-evaluate the log-joint for each sample in the MCMC chain.

[yebai]

Because PG doesn't track the log-joint, there is no way to extract it directly - is my understanding correct? @yebai

That's right, the weight associated with each particle is reset to 1 after each resampling step. I think this feature has been brought up in the past, see related issues: https://github.com/TuringLang/Turing.jl/issues/493 #426. It's actually really simple to support this feature using API from https://github.com/TuringLang/Turing.jl/issues/634

logp(model, vi)

However, if we accumulate log weights and store them somewhere, we can avoid an extra call to runmodel (or logp(model, vi) using the new API from https://github.com/TuringLang/Turing.jl/issues/634).

[xukai92]

However, if we accumulate log weights and store them somewhere, ...

So it seems to me we'd add another field to the ParticleContainer right?

[yebai]

Likely out of date due to the effort of separating particle MCMC and SMC samplers into https://github.com/TuringLang/AdvancedPS.jl