How to specify a Turing model to have the 2 vectors as input for a distribution that returns a tuple

Specific Domains Probabilistic programming
, , ,
16

Thanks. I looked through the doc strings, but still have some questions about what it means to implement a model as @model. However, before I get too far into the details, I wanted to ask you about the code below. I adapted Dominique’s example for a Wald model (basically a reparameterized InverseGaussian), which is actually a subtype of UnivariateContinuousDistribution:

julia> dist = Wald(ν = 2.0, α=1.0, θ=.3); rand(dist)
0.5196099061252006

If that was the problem (or sole problem), I would expect the code below to work. For some reason it does not. I want to bring that up to make sure we are addressing the correct issue. In addition, the same thing happens if I use InverseGaussian directly instead of Wald.

Wald

using Turing
using SequentialSamplingModels
using Random
using LinearAlgebra


# Generate data with different drifts for two conditions A vs. B
Random.seed!(254)
rts = rand(Wald(ν=2.0, α=0.8, θ=0.3), 10)

@model function model_wald(rts)
    min_rt = minimum(rts)

    ν ~ truncated(Normal(2, 1), 0.0, Inf)
    α ~ truncated(Normal(0.8, 0.4), 0.0, Inf)
    θ ~ Uniform(0.0, min_rt)

    # Likelihood
    rts ~ Wald(; ν, α, θ)
    return (; rts, ν, α, θ)
end

chain = sample(model_wald(rts), Prior(), 100)

predictions = predict(model_wald(rts), chain)

Inverse Gaussian

using Turing
using Distributions
using Random
using LinearAlgebra


# Generate data with different drifts for two conditions A vs. B
Random.seed!(254)
rts = rand(InverseGaussian(1.0, 0.8,), 10)

@model function model_IG(rts)

    ν ~ truncated(Normal(2, 1), 0.0, Inf)
    α ~ truncated(Normal(0.8, 0.4), 0.0, Inf)

    # Likelihood
    rts ~ InverseGaussian(ν, α)
    return (; rts, ν, α)
end

chain = sample(model_IG(rts), Prior(), 100)

predictions = predict(model_IG(rts), chain)