q_out must be one-hot encoded

[mateusjoffily]

Hi,

I recently updated Julia to version 1.12 and reinstalled RxInfer. I'm now getting an error that didn't exist with the previous version. The minimal example code is:

using RxInfer

@model function model_urn(y,a)
    x ~ Dirichlet(a)
    y ~ Categorical(x)
end

# Y  = [1.0, 0.0]                  # observation
Y  = [0.7, 0.3]                  # prior beliefs about future observation

result = infer(
    model = model_urn(a=[1.0, 1.0]),
    data  = (y = Y, )
)

a_p = result.posteriors[:x].alpha

and the error is

ArgumentError: q_out must be one-hot encoded. Got: [0.7, 0.3]

Any suggestions on the best way to fix this? Thank you in advance for any help.

Mateus

[bvdmitri]

The output of a Categorical distribution is a category, so Y must be either [1.0, 0.0] or [0.0, 1.0]. Since Y is used in data it does not represent prior belief but instead an actual observation that must be one-hot encoded

[mateusjoffily]

Thank you for your help. I was expecting the previous model to be equivalent to the one below:

using RxInfer

@model function model_urn(y,a,y0)
    x ~ Dirichlet(a)
    y ~ Categorical(x)
    y ~ Categorical(y0)
end

# Y  = [1.0, 0.0]                  # observation
Y  = [0.7, 0.3]                  # prior beliefs about future observation

result = infer(
    model = model_urn(a=[1.0, 1.0], y0=Y),
    data  = (y = UnfactorizedData(missing), )
)

a_p = result.posteriors[:x].alpha

However, I was also expecting the updated concentration vector a_p to change given the prior y0.

I’m probably missing something - could you please help me understand why a_p isn't updated in this implementation?

[mateusjoffily]

As a follow-up to my previous question, would it be possible to implement the following model without including y? The output of this model (a_p) is actually what I was looking for with the first model, using belief propagation instead of variational message passing. Thank you.

using RxInfer

@model function model_urn(y,a,y0)
    x ~ Dirichlet(a)
    y ~ Categorical(x)
    y1 ~ Categorical(x)
    y1 ~ Categorical(y0)
end

constraints = @constraints begin
    q(x,y,y1) = q(x)q(y,y1)
end

ireturnvars = (
    x = KeepLast(),
)

# Y  = [1.0, 0.0]                  # observation
Y  = [0.7, 0.3]                  # prior beliefs about future observation

init = @initialization begin
    q(x) = Dirichlet([1.0, 1.0])
end

result = infer(
    model          = model_urn(a=[1.0, 1.0], y0=Y),
    data           = (y = UnfactorizedData(missing),), 
    iterations     = 10, 
    initialization = init, 
    returnvars     = ireturnvars, 
    constraints    = constraints
)

a_p = result.posteriors[:x].alpha

display(a_p)

2-element Vector{Float64}:
 1.8080707825928837
 1.1919292174071163