mroavi/junctiontrees.jl

A metaprogramming-based implementation of the junction tree algorithm.

junctiontrees.jl's Issues

`asia` dataset is missing in examples folder

I am reading this documentation page, but I can not reproduce it without model data
https://mroavi.github.io/JunctionTrees.jl/stable/usage/

I would be great if the asia dataset can be in the examples folder.

Move file IO to another package

The file IO is used in another package. Moving uai files and its reading functions to another package will make the dependency smaller.
https://github.com/mroavi/TensorInference.jl

This design pattern is similar to GraphIO: https://github.com/JuliaGraphs/GraphIO.jl that loads graphs from disk.

Replace MetaGraphs.jl with a type-stable alternative

MetaGraphsNext.jl is an experimental, type-stable replacement for MetaGraphs.jl.

Replace MacroTools.jl in favor of MLStyle.jl

The OMEinsum boost optimization uses MLStyle.jl. Consider replacing the usage of MacroTools.jl's macros with MLStyle.jl.

TagBot trigger issue

This issue is used to trigger TagBot; feel free to unsubscribe.

If you haven't already, you should update your TagBot.yml to include issue comment triggers.
Please see this post on Discourse for instructions and more details.

If you'd like for me to do this for you, comment TagBot fix on this issue.
I'll open a PR within a few hours, please be patient!

Make OMEinsum the default backend

OMEinsum has a much better performance than the default backend in place.

Do not assume that observed values start from 1

Instead, assume that they start from 0, just like they are defined in the UAI evid file.
I probably made this assumption because it plays nicely with the redu function.

Replace eval with macros

The minimum fix should be like

using JunctionTrees

problem_number = "34"
problem_filename = joinpath("Promedus_" * problem_number)
problem_dir = joinpath(@__DIR__, "../examples/problems/Promedus/", problem_number)
uai_filepath = joinpath(problem_dir, problem_filename * ".uai")
uai_evid_filepath = joinpath(problem_dir, problem_filename * ".uai.evid")
uai_mar_filepath = joinpath(problem_dir, problem_filename * ".uai.MAR")
td_filepath = joinpath(problem_dir, problem_filename * ".tamaki.td")

reference_marginals = JunctionTrees.read_uai_mar_file(uai_mar_filepath)
obsvars, obsvals = JunctionTrees.read_uai_evid_file(uai_evid_filepath)

# ------------------------------------------------------------------------------
# Posterior marginals given evidence
# ------------------------------------------------------------------------------

macro inferencemodel(uai_filepath, uai_evid_filepath)
    compile_algo(
            eval(uai_filepath),
            uai_evid_filepath = eval(uai_evid_filepath),
        )
end

algo = @inferencemodel uai_filepath uai_evid_filepath

marginals = algo(obsvars, obsvals) |> x -> map(y -> y.vals, x)

Stack overflow during inference on julia#master due to long tuple types

During testing of this package (specifically, as part of the compile algorithm testset while compiling run_algo(Array{Int64, 1}, Array{Int64, 1})), inference is being invoked with very long tuple types:

Internal error: stack overflow in type inference of run_algo(Array{Int64, 1}, Array{Int64, 1}).
This might be caused by recursion over very long tuples or argument lists.

(rr) call jl_(types)
Tuple{typeof(Base.promote_typeof), Vararg{JunctionTrees.Factor{Float64, _A} where _A, 249}}

(the max number of args is much higher; this is just at some point during the backtrace.)

Invoking inference with such types is generally not OK, and here causes a stack overflow after recursing >10000 frames deep. I'm assuming somewhere a long list of values is being accidentally splat into a function invocation.

If you think this specific case should be supported, please reduce it to a minimal example and open an issue on the Julia repository. In the meantime, I'll assume this is a bug in the package, and will temporarily blacklist it in PkgEval.jl.