Printing empty VarInfo

[phipsgabler]

Line

DynamicPPL.jl/src/varinfo.jl

Line 968 in f74e039

print(io, "), dimension ", sum(prod(size(md.vals[md.ranges[md.idcs[vn]]])) for vn in vns))

sums over the sizes of the varnames, which fails in case there are none. (You only notice this in very weird settings, but IRTracker is one of those.)

I'll replace the sum by reduce(+, ...; init=0) some time soon.

[devmotion]

I'm a bit confused, what exactly do we mean with dimension here? Can't we just write

 print(io, "), dimension ", length(md.vals))

? Or at least

 print(io, "), dimension ", sum(length(md.vals[md.ranges[md.idcs[vn]]]) for vn in vns)) 

(possibly with the reduce fix)?

[phipsgabler]

Hong used that term to refer to the "dimension" of the variable space, and I just kept it. What you propose I'd call the rank, but yes, it's a bit ambiguous. What about "degrees of freedom"? Or just "total size"?

[devmotion]

I'm a bit confused because md.vals is always a vector (see

DynamicPPL.jl/src/varinfo.jl

Line 59 in f74e039

vals :: TVal # AbstractVector{<:Real}

), so it seems a bit weird to compute prod(size(md.vals[md.ranges[md.idcs[vn]]])) since that should just be equal to length(md.ranges[md.idcs[vn]]), and in turn if you sum everything up you should just get length(md.vals) it seems?

[phipsgabler]

That all sounds very reasonable. I really wrote this line without thinking about the semantic meaning of those values 🤷

Now, what do you suggest for the wording of "dimension"?

[devmotion]

It's the total number of scalar samples in the model, so I guess it's reasonable to call it "dimension". Something like "length" does not imply any relation to multidimensionality or the (original) shape of the variables, but is less clear IMO. I think we should maybe just simplify the implementation for now, and maybe someone has a better suggestion in the corresponding PR.

[devmotion]

Fixed in #125.