Pretty printing of Turing models

[DominiqueMakowski]

I am trying to use TuringGLM to better learn Turing, and so I wrote the simplest LM and would like to re-write that in Turing.

using RDatasets
using TuringGLM

data = RDatasets.dataset("datasets", "mtcars")

fm = @formula(MPG ~ WT)
model = turing_model(fm, data)
model

DynamicPPL.Model{TuringGLM.var"#normal_model#19"{Int64,Int64,CustomPrior},(:y, :X, :predictors, :μ_X, :σ_X, :prior, :residual),(:predictors, :μ_X, :σ_X, :prior, :residual),(),Tuple{Vector{Float64},Matrix{Float64},Int64,Int64,Int64,CustomPrior,Float64},Tuple{Int64,Int64,Int64,CustomPrior,Float64},DynamicPPL.DefaultContext}(TuringGLM.var"#normal_model#19"{Int64,Int64,CustomPrior}(0, 1, CustomPrior(TDist{Float64}(ν=3.0), LocationScale{Float64,Continuous,TDist{Float64}}(
                μ:19.2σ:5.411498097545447ρ:TDist{Float64}(ν=3.0)
            ), nothing), Core.Box(TuringGLM.var"#normal_model#18#20"(Core.Box(TuringGLM.var"#normal_model#19"{Int64,Int64,CustomPrior}())))), (y=[21.0, 21.0, 22.8, 21.4, 18.7, 18.1, 14.3, 24.4, 22.8, 19.2 … 15.2, 13.3, 19.2, 27.3, 26.0, 30.4, 15.8, 19.7, 15.0, 21.4], X=[2.62; 2.875…3.57; 2.78;;], predictors=1, μ_X=0, σ_X=1, prior=CustomPrior(TDist{Float64}(ν=3.0), LocationScale{Float64,Continuous,TDist{Float64}}( #= circular reference @-4 =#
                μ:19.2σ:5.411498097545447ρ:TDist{Float64}(ν=3.0)
            ), nothing), residual=6.026948052089105), (predictors=1, μ_X=0, σ_X=1, prior=CustomPrior(TDist{Float64}(ν=3.0), LocationScale{Float64,Continuous,TDist{Float64}}(
                μ:19.2σ:5.411498097545447ρ:TDist{Float64}(ν=3.0)
            ), nothing), residual=6.026948052089105), DynamicPPL.DefaultContext())

Unfortunately, I can't get my head around this output to format in a standard Turing syntax. Is there a way to "pretty-print" or format the Turing model call? Thanks!

[storopoli]

This is how Turing.jl prints the model: as a Julia struct.
There is nothing special about how TuringGLM handles the model.
We just parse the formula as a Turing model and delegates that to Turing.

@torfjelde is there a way to get the underlying model of an instantiated Turing model?

[torfjelde]

is there a way to get the underlying model of an instantiated Turing model?

What do you mean by this? As in, what information are hoping to extract? As you say, it's just a struct and so you can access the arguments it's been instantiated with, etc.

As for the pretty-printing, I agree we should do something. I'll make an issue over at DPPL.

[storopoli]

What do you mean by this?

The underlying macro call.
Like:

@model function demo(x, y)
  # Assumptions
  σ2 ~ InverseGamma(2, 3)
  σ = sqrt(σ2)
  μ ~ Normal(0, σ)

  # Observations
  x ~ Normal(μ, σ)
  y ~ Normal(μ, σ)
end

[torfjelde]

That's not possible, but why would you want that? Sorry, maybe I'm not fully understanding.

…

On Fri, May 26, 2023, 9:24 AM Jose Storopoli ***@***.***> wrote: What do you mean by this? The underlying macro call. Like: @model function demo(x, y) # Assumptions σ2 ~ InverseGamma(2, 3) σ = sqrt(σ2) μ ~ Normal(0, σ) # Observations x ~ Normal(μ, σ) y ~ Normal(μ, σ)end — Reply to this email directly, view it on GitHub <#91 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ACUPZZAKFIFSIRCL44TBI6DXIBSEJANCNFSM6AAAAAAYH4KPGI> . You are receiving this because you were mentioned.Message ID: ***@***.***>

[storopoli]

They want to have the model macro when they create a model using turing_model and @formula.

[torfjelde]

Gotcha. That would have to be provided by TuringGLM if so. A macro generates code, so once you apply a macro, e.g. @model, you don't know what it used to look like.

[storopoli]

Yeah, but I don't want any scaffolding on Turing from TuringGLM. That would need a TuringGLM custom type to hold that as a field (maybe a String).

[DominiqueMakowski]

My rationale for this was two-fold:

• Educational: I know than many people learned for instance Stan by fitting models in brms and then inspecting the Stan code, and so I think that many new Julia transitionners might start with TuringGLM, but might be interested in seeing what the model looks like in raw Turing
• Clarity: When we print the model now it is not super clear what priors/model have been used which can impede easy and transparent reporting. Having some pretty printing like "these are the priors used:" - even if not exactly presented like a Turing model - would be useful in that regards :)

[torfjelde]

I don't have the time to pursue this right now, but you could do something like this:

struct TuringGLMModel{M,F,P}
    model::M
    formula::F
    priors::P
    # other stuff
    # ...
end

and then just display this in a different way than what you would for typical Turing.jl models.