Returning function values? about klara.jl HOT 6 CLOSED

Hi
I reread the code in SerialMC.jl and apparently we don't store the log posterior in the diagnostics field (even though the samplers apparently return them).

I'll see if I can add that... (no delay promises ! ).

from klara.jl.

I think it is better to store and return the absolute minimum in order to avoid excessive storage requirements for the output. It is easy to evaluate the log-posterior using the returned chain if desired. I will write up a simple example to clarify how the user can achieve this.

from klara.jl.

Ok, here is the example:

using MCMC

mymodel1 = model(v-> -dot(v,v), init=ones(3)) 
mychain = run(mymodel1, RWM(0.1), SerialMC(steps=1000, burnin=100))

# Compute the log-posterior for the first sample (i.e. first MCMC iteration)
mymodel1.eval(mychain.samples[1, :])

# Compute the log-posterior for all the samples
[mymodel1.eval(mychain.samples[i, :]) for i in 1:size(mychain.samples, 1)]

from klara.jl.

This is a good point, we should be parcimonious in the data stored especially for very long chains. On the other hand this would not significantly increase the memory requirements : we already store the parameter vector, gradient and diagnostics for each iteration, adding a single float will not change much. Secondly, recalculating the log-posterior after the run will take about the same time as the run itself so it is not viable for very long runs.

Perhaps we could add some kind of opting-out of the diagnostics produced ? (that the user could choose if allocations made at the run initiation fail).

from klara.jl.

Yes, maybe we can provide the user with the option of choosing what to return. At the same time, I think that if we decide to return the log-posterior, it should be a standalone field outside the diagnostics; we provide the gradient outside the diagnostics which is more secondary than the log-posterior.

from klara.jl.

@carloalbert : the current master should provide the return value of the function (in the field logtargets of the MCMCChain type).
The new version of MCMC is 0.0.6 (run Pkg.update())

from klara.jl.