Trying to adapt the initial coin-flip example on RxInfer to model a weighted die

[bsyiem]

Hi all,

I’m completely new to Julia and I apologize if this is a naive question, but I have been attempting to model a weighted die using RxInfer. I’ve followed the same structure as the coin_toss example provide in the RxInfer getting-started page but can’t seem to get it to work.

The issue seems to arise when I actually try to use the “infer” function, which throws the error: ERROR: probvec(::PointMass{ <: Real }) is not defined

I would really appreciate any help with this. I’ve attached the minimal amount of code to reproduce this below:

using RxInfer, Distributions, Random
using Cairo, GraphPlot, Plots #for graph visuals

#initial parameters
rng = MersenneTwister(20)

n_10 = 10
die_weightings = [0.05, 0.05, 0.05, 0.283, 0.283, 0.284]
distribution = Categorical(die_weightings)
dataset = rand(rng, distribution, n_10)

@model function die_model(y, α)
  p ~ Dirichlet(α)
  y .~ Categorical(p) # 6-sided die
end

result = infer(
  model = die_model(α = ones(6)),
  data = (y = dataset,)
)

Thanks!

P.S: Conditioning and visualizing the model seems to work as intended even with deferred data:

conditioned_with_deffered_data = die_model() | (
y = [ 4, 1, 6 ],
α = RxInfer.DeferredDataHandler(),
)

model_with_deffered_data = RxInfer.create_model(conditioned_with_deffered_data)
GraphPlot.gplot(RxInfer.getmodel(model_with_deffered_data))

Thank you.

[bvdmitri]

You need to convert the observations (which are y = dataset in your example) to the one-hot encoding for the Categorical node. So, for example, 5 becomes [ 0.0, 0.0, 0.0, 0.0, 1.0, 0.0 ] since the die has 6 categories in total, the one hot-encoding also should be of length 6.

[bsyiem]

Much appreciated, I was wondering what I was doing wrong for a few days.

Thank you!