Posterior samples as vector of named tuples

Specific Domains Probabilistic programming
1

Let’s say I have a simple Turing model

@model function model()
    C ~ LKJ(3, 1.0)
    x ~ MvNormal(zeros(3), C)
end

and obtained some posterior samples via MCMC

chain = sample(model(), NUTS(), 100)

or variational inference

q = vi(model(), ADVI(10, 1_000))
samples = rand(q, 100)

Is there an (easy) way to transform such samples into a vector of named tuples, i.e., [(C = Matrix, x = Vector), (C = Matrix, x = Vector), ...] corresponding to the sampled random variables of the Turing model?
Could not find it in the documentation (hope to have just missed it).

2

What are you trying to accomplish that requires the vector of named tuples?

3

Basically, I wanted to condition the model on some of its parameters using the posterior mean or a couple of samples as values.
Have found a way to do it though, e.g., DynamicPPL.value_iterator_from_chain provides samples in a suitable format.

1 Like
4

Here is something that might be along those lines: API · DynamicPPL