Issue in the dot product within a NormalMeanVariance distribution

[jesper-jpg]

Hi, I just tried to create a model in which I use the dot product to calculate \mu, the mean prediction. For clarification, i'll provide the code below:

@model function polynomial_regression(y, X, D, σ2)

# Prior distribution
β ~ MvNormalMeanCovariance(zeros(D + 1), 0.1 * diagm(ones(D + 1)))
# Define empty z
z = Vector{Vector{Float64}}(undef, length(X))

# Likelihood
for i in eachindex(X)

    # Polynomial
    z[i] = [X[i]^j for j in 0:D]
    
    # prediction
    μ = dot(β, z[i])
    # μ = β ⋅ z
    y[i] ~ NormalMeanVariance(μ, σ2[i])

end

end

Initialize variables

y = df[!,:y]
X = df[!,:x]
D = 3
σ2 = df[!,:σ2]
results = infer(
model = polynomial_regression(D = D, σ2 = σ2),
data = (y = y, X = X),
)

This code gives me the following error: StackOverflowError: Stacktrace: [1] dot(x::GraphPPL.NodeLabel, y::Float64) (repeats 79984 times) @ LinearAlgebra C:\Users\jespe.julia\juliaup\julia-1.10.7+0.x64.w64.mingw32\share\julia\stdlib\v1.10\LinearAlgebra\src\generic.jl:861

The error occurs in the dot function and it seems like it enters an infinite loop, creating a huge model or it tries to recursively resolve something adding TODO's in the stack until it overflows.

However, when I switch from:
μ = dot(β, z[i]) y[i] ~ NormalMeanVariance(μ, σ2[i])

To:
y[i] ~ NormalMeanVariance(dot(β, z[i]), σ2[i])

The function/model seems to work fine. Could this be an error in the package? Thanks for looking into it.
Kind regards,
Jesper

[bvdmitri]

You should use := for deterministic relationships in the @model definition (ref1, ref2).

So in your example you should write

μ := dot(β, z[i])

Can you confirm that this fixes issue for you?

[jesper-jpg]

That works! Thank you for the clarification and my apologies for the inconvenience.

[bvdmitri]

@jesper-jpg Not a problem ,feel free to open other issues! I used your feedback here #391