Sampling hangs on integer data

[anhi]

Hi,

I'm trying to fit a Weibull distrbution to input data that happen to be exact integers, but sampling doesn't terminate. Here is a reduced example:

using Turing, Distributions

@model weibull(x) = begin
    μ ~ Uniform(1, 10)
    σ ~ Uniform(1, 10)

    x ~ Weibull.(σ, μ)
end

w = Weibull(2, 4);
data = rand(w, 1000);
rounded = round.(data)
rounded_noisy = rounded + 0.01 .* rand(length(rounded));

# this works
chain = sample(weibull(data), NUTS(0.65), 1000, progress=true)
# this works, too
chain =  sample(weibull(rounded_noisy), NUTS(0.65), 1000, progress=true)
# this doesn't
chain =  sample(weibull(rounded), NUTS(0.65), 1000, progress=true)```
`

[luiarthur]

I think the behavior is due to rounded containing zeros. Note that pdf(w, 0) == 0. Thus, when zeros are present in x, the likelihood, and thus the joint posterior density, will always evaluate to 0. So all samples of (mu, sigma) will be rejected.

If rounded does not contain any zeros (but only positive integers), it also runs fine.

[anhi]

Ah, that makes perfect sense! You are absolutely right, so I'll close this issue.