Want rxinfer equivalent of the following turing model

[prantik1998]

I am trying to port the following model from turing to rxinfer

@model function laplacian_model(A,y)
m,n = size(A)
mew ~ filldist(Laplace(10,1),n)
y ~ MvNormal(A*mew,[0.5 0;0 0.5])
end

I was not able to use laplce with rxinfer and also was not able to use cvi for non conjugate priors.

Any pointers on how to get this working in rxinfer.

[albertpod]

Hi @prantik1998. I think the best place for this question will be in Discussions.
Converting it.

[albertpod]

RxInfer.jl operates on factor graphs, which means that all of your Distributions that go after ~ should be available as nodes in ReactiveMP.jl. Unfortunately, there's no Laplace node available, but you can create your own following this manual. The absence of the nodes from the exponential family should not be a problem in future `RxInfer.jl releases.

Besides, there's no filldist option in RxInfer.jl, so the current solution will be somewhat ugly. I will show how I would do it in RxInfer. I will leave the implementation of the Laplace node to you and work with Gamma instead. While trying to give you a solution, I've realized that splatting is not available in RxInfer.jl just yet.

So here's a bit of an ugly sketch for the problem:

using RxInfer, StableRNGs, Optimisers

@model function laplacian_model(n)
    mew = randomvar(n)
    A   = constvar([0.5 0.0; 0.0 0.5])

    y   = datavar(Vector{Float64})

    for i in 1:n
        mew[i] ~ Gamma(α=10.0, β=1.0)
    end
    dew ~ mul(A, mew[1], mew[2]) # auxilary variable for multiplying A w vector of Gammas
    y   ~ MvNormal(μ=dew, Λ=[0.5 0;0 0.5])
end

@meta function cvi_meta(rng, nr_samples, nr_iterations, optimizer)
    mul() -> CVI(rng, nr_samples, nr_iterations, optimizer)
end;

n = 2

function mul(A, x, y)
    A*[x,y]
end

result = infer(data = (y = randn(2), ), model = laplacian_model(n), meta=cvi_meta(StableRNG(42), 100, 200, Optimisers.Descent(0.01)), iterations=10, free_energy=true, returnvars=KeepLast())

[bvdmitri]

It seems that you use an outdated version of RxInfer. The latest version is v2.16.1. You can try to create a new temporary environment with the ] activate --temp command in Julia and add the latest version with ] add RxInfer command.

[prantik1998]

@albertpod @bvdmitri Thank you for your help it worked

[prantik1998]

Hey i am facing an issue. If i change the prior from gamma to normal i am not able to run the inference. The error is as follows:-

Failed to compute node bound free energy component. The result is NaN. Use diagnostic_checks field in BetheFreeEnergy constructor or free_energy_diagnostics keyword argument in the inference function to suppress this error.

[albertpod]

Set free_energy_diagnostics=nothing, or turn off free_energy computations by setting it to false, i.e., free_energy=false.
Additionally, if you use Normal as a prior, I recommend switching to Unscented or Linearization approximation methods instead of CVI.

[prantik1998]

ok thanks