Way to Declare Data

[ParadaCarleton]

Mentioned on the call today with Hong and Cameron -- perhaps we should add a wrapper type for data that throws an error if any variable is sampled from a distribution instead of being added to the log-likelihood. (Possibly exempting missing values.) Cameron suggested an @data convenience macro where users can list all their data variables and have them automatically converted by Turing.

This could avoid some of the errors we've been seeing on Discourse where users accidentally generate a sample instead of adding it to the log-likelihood.

[torfjelde]

IMO we should just do what's already done in TuringLang/DynamicPPL.jl#275.

The reason why I'm in favour of that instead of something like a wrapper-type will require changing the way we're performing checks, otherwise we'll run into method ambiguities real quickly. In contrast, the @observe approach is completely non-intrusive.

[ParadaCarleton]

IMO we should just do what's already done in TuringLang/DynamicPPL.jl#275.

The reason why I'm in favour of that instead of something like a wrapper-type will require changing the way we're performing checks, otherwise we'll run into method ambiguities real quickly. In contrast, the @observe approach is completely non-intrusive.

Hmm; Hong and Cameron suggested the type approach as a way to avoid having to deal with metaprogramming here, but maybe using types isn't worth it. My original suggestion was actually something much closer to the @observe you mentioned, with the exception that it would declare a block of variables to be data at the start, rather than adding @observe to each individual line. If we do go with the @observe approach, maybe we should have an @data macro that adds @observe to each line involving a data variable.

[yebai]

I think we can now make use of the nice callable object support (TuringLang/DynamicPPL.jl#398) to declare variables as data, e.g.

struct MyModel
    a::Int
    x::Float64
end

is_observation(m::MyModel, s::Symbol) = s == :x ? true : false 

@model function (f::MyModel)(x)
    m ~ Normal(f.a, 1)
    # we can modify the `~` implementation to call `is_observation` internally. 
    return m.x ~ Normal(m, 1) 
end

[yebai]

We likely will promote the syntax in #1736 (comment) as a solution.