Inf is treated as variable rather than data

[xukai92]

In the MWE below, Inf is treated as a variable rather than the literal Inf.

using Turing

@model mwe() = begin
    m ~ Normal(0, 1)
    Inf ~ Normal(m, 1)
end

sample(mwe(), HMC(0.1, 2), 100)

[xukai92]

Reported by @jonasmac16.

[mohamed82008]

This is an interesting one, but may I ask how is supporting Inf as data useful?

[xukai92]

It's for something like below

res = solve(...)
if res.success
    data ~ dist
else
    Inf ~ dist
end

[mohamed82008]

Can we achieve the same goal by defining the Impossible distribution? rand(::Impossible) = NaN and logpdf(::Impossible, ::Any) = -Inf.

res = solve(...)
if res.success
    data ~ dist
else
    data ~ Impossible()
end

Supporting Inf on the LHS is also possible.

[xukai92]

Yes there are maily work aournds for this. Like the one you post or just pass in a data call data_inf which is Inf. But it's might be cool for the user to do so easily.

[yebai]

Maybe standardize the compiler API for LHS with concrete values, by e.g. introducing a function operate on Julia expressions (which can be called by the @model macro)? This would allow the user to overload ~ for specific use cases.

[mohamed82008]

Sounds good.

[devmotion]

As discussed in #57 (comment), I think the more natural way to write the example above would be

res = solve(...)
if !res.success
	@logpdf() = -Inf
	return
end

data ~ dist

So can this issue be closed, @xukai92, or are there other use cases that require a specific fix for Inf?

[xukai92]

It makes sense. The issue was originally rasied by a user in slack. Let's close if for now.