Add shorter non-display printing of VarInfo

[phipsgabler]

Currently, a VarInfo prints to a huge blob showing all kinds of stuff. Which is useful on its own in the REPL, but not as part of another printed structure.

I moved that version to show(::IO, ::MIME"text/plain", ::UntypedVarInfo), and added a dummy replacement

Base.show(io::IO, vi::UntypedVarInfo) = print(io, "VarInfo of ", vi.metadata.vns)

Please suggest what things should actually printed in the short version.

[devmotion]

Maybe just print the number of variables and possibly the log probability in the short version?

[codecov]

Codecov Report

Merging #66 into master will decrease coverage by 2.99%.
The diff coverage is 0.00%.

@@            Coverage Diff             @@
##           master      #66      +/-   ##
==========================================
- Coverage   79.42%   76.43%   -3.00%     
==========================================
  Files          13       13              
  Lines         841      853      +12     
==========================================
- Hits          668      652      -16     
- Misses        173      201      +28     

Impacted Files	Coverage Δ
src/varinfo.jl	84.42% <0.00%> (-3.85%)	⬇️
src/compiler.jl	79.86% <0.00%> (-9.40%)	⬇️
src/utils.jl	53.06% <0.00%> (-3.55%)	⬇️
src/varname.jl	76.47% <0.00%> (+0.96%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 9fbe09a...6903e3e. Read the comment docs.

[phipsgabler]

VarInfo (2 variables, logp = -3.304)?

[devmotion]

Yes, something like this. I guess that should be the most important information, shouldn't it? But maybe it's still to verbose?

[phipsgabler]

For me everything that short is fine -- it shouldn't clutter printing of nested structures.

I'll leave this open some time to wait for feedback.

[yebai]

VarInfo (2 variables, logp = -3.304)?

Maybe add:

• model dimensionality
• whether variables are in unconstrained space or not
• observed variable names (just the symbol is fine, we don't need indexing here).

[devmotion]

Hmm I'm not sure, the intention here is to have a really short condensed way of printing a VarInfo that will be used, e.g., when printing arrays of VarInfo objects. If an expression returns a single VarInfo object, the REPL will use the longer version implemented by show(io, ::MIME"plain/text", ::VarInfo) to display it.

[phipsgabler]

I concur with David, although I think that adding the variables names is not a bad idea.

[yebai]

I see David's point, but model dimensionality is something valuable and can be packed into a few words.

[phipsgabler]

@yebai What exactly do you mean by that? The sum of the sizes of all variables? Is this currently used anywhere in the VarInfo code?

[yebai]

Yes, exactly.

[phipsgabler]

What about this:

VarInfo (λ (size 1), m (size 1); logp: -3.612)

where size = length(md.ranges[md.idcs[vn]])?

[yebai]

Looks good, maybe reformat the dimensionality message

VarInfo (λ: scalar, m: scalar, x: (3,4); logp: -3.612)

Also, this might gets messy when we have a lot of parameters. That's where a single model dimensionality might be less verbose.

[phipsgabler]

Ok, I tried to extract the size:

function _vname_info(vn, vi)
    md = vi.metadata
    s = size(md.vals[md.ranges[md.idcs[vn]]])
    if s == ()
        return "$(getsym(vn))"
    else
        return "$(getsym(vn)) $s"
    end
end

But since the values are only given through ranges, there seems to be no way to distinguish scalars from size 1 vectors? E.g. for a gdemo-like model:

VarInfo (λ (1,), m (1,); logp: -3.612)

Or is there?

[yebai]

But since the values are only given through ranges, there seems to be no way to distinguish scalars from size 1 vectors?
VarInfo (λ (1,), m (1,); logp: -3.612)

That works.

[devmotion]

But since the values are only given through ranges, there seems to be no way to distinguish scalars from size 1 vectors?

No, AFAIK one can only distinguish between scalars and vectors with one element by checking if the corresponding distribution is uni- or multivariate. IMO that's one of the inconveniences caused by vectorizing everything in VarInfo, and leads to some heuristics (and ugly code) in the implementation of the ESS and MH sampler in Turing.

[phipsgabler]

Ok, based on this feedback I think it's best to show just the overall dimension of the model, plus a limited number of symbols of variable names (given by a constant _MAX_VARS_SHOWN = 4). The latter makes the VarInfo more easily to recognize, but prevents clutter for large models.

The current way of implementating this does not distinguish between one @varname x of size 3 and [@varname(x[1]), @varname(x[2]), @varname(x[3])] of 3 scalars, see test1 and test2 below.

julia> @model function test3()
           x ~ Bernoulli()
           y ~ Bernoulli()
           z ~ Bernoulli()
           v ~ Bernoulli()
           w ~ MvNormal(zeros(3), 1.2)
       end
ModelGen{var"###generator#628",(),(),Tuple{}}(##generator#628, NamedTuple())

julia> vi = VarInfo(); test3()(vi); show(vi)
VarInfo (5 variables (w, y, v, z, ...), dimension 7; logp: -11.837)

julia> @model function test2()
           s ~ Gamma(1.0, 1.0)
           x = zeros(3)
           for i in 1:3
               x[i] ~ Normal(0.0, s)
           end
       end
ModelGen{var"###generator#644",(),(),Tuple{}}(##generator#644, NamedTuple())

julia> vi = VarInfo(); test2()(vi); show(vi)
VarInfo (2 variables (s, x), dimension 4; logp: 3.142)

julia> @model function test1()
           s ~ Gamma(1.0, 1.0)
           x ~ MvNormal(zeros(3), s)
       end
ModelGen{var"###generator#660",(),(),Tuple{}}(##generator#660, NamedTuple())

julia> vi = VarInfo(); test1()(vi); show(vi)
VarInfo (2 variables (s, x), dimension 4; logp: -3.651)

[yebai]

This looks nice. To distinguish parameters and data, maybe group them and annotate using a semicolon, e.g.

VarInfo (6 variables (w, y, v, z, ...; d), dimension 7; logp: -11.837)

where w,y,v,z... are model parameters, and d is observed data.

[phipsgabler]

But the observed data is not contained in the VarInfo at all, so where would I get it from -- or am I mistaken?

[devmotion]

No, I think you're right, it's not available (and hence probably should not be printed?).

[yebai]

But the observed data is not contained in the VarInfo at all, so where would I get it from -- or am I mistaken?

Hmm, I didn't release this - sorry for the noise. We might want to store the observed variable information in VarInfo in the future, but it is out of this PR's scope.