MCMC chains are inappropriately shown as tables

[nilshg]

No video on this one, you just need one line of code (plus the relevant package):

The code is:

using TuringGLM
sample(turing_model(@formula(y ~ x1 + x2), (y = rand(10), x1 = rand(10), x2 = rand(10))), NUTS(), 2_000)

When running this in the REPL you get:

Which is the desired output as users are interested in the summary statistics by parameter, rather than the individual iterations of the MCMC chain.

The Pluto display used to be the same as in the REPL, so there must have been some recent change to how chains get displayed/when the table display kicks in.

[nilshg]

julia> Tables.istable(chain)
true

where chain is the result of the sample call above - so I guess the answer is yes.

[pankgeorg]

julia> Tables.istable(chain)
true

where chain is the result of the sample call above - so I guess the answer is yes.

Context: In a now-deleted comment, I asked if the chains are Table.jl like objects. I understood a bit later that the chains are "overloaded" to include both the iteration data and the metadata about the run, and what you want to see is the metadata and that's why I removed my comment.

The action here is to check whether we can make an HTML show method for these objects, and to make sure that Pluto will prefer these to the table viewer. I'll check the latter and suggest the former, probably with a PR, when I find some time (I do double duties as a dad lately so my contributions are a bit more limited 😮‍💨; but I'll get to it! friends say Turing is way to important not to pay appropriate attention!)

[nilshg]

Thanks - I can relate, if you look at my repos you see how much time I get to push forward the things I'd like to see in Julia, in large part due to my kids (not complaining!)

And probably agreed that Turing is one of these showcase packages with a large enough userbase that you probably want to provide a smooth experience in Pluto for its users.

[fonsp]

I'm busy moving but I will respond soonish!

We want to find a general solution to this (something like JuliaLang/julia#47270 or JuliaPluto/AbstractPlutoDingetjes.jl#10) but until then it might be easiest to add a special case for this package, just like Symbolics.jl:

Pluto.jl/src/runner/PlutoRunner.jl

Lines 1500 to 1505 in 994ba02

	Integration(
	id = Base.PkgId(UUID("0c5d862f-8b57-4792-8d23-62f2024744c7"), "Symbolics"),
	code = quote
	pluto_showable(::MIME"application/vnd.pluto.tree+object", ::Symbolics.Arr) = false
	end,
	),

[Batmaev]

You can still see the summary statistics, just call summarize(..):

I personally prefer displaying the chain as a table of individual iterations because it provides a better understanding of what is inside the object and what you can do with it