Something not right with serialisation of python (sk-learn) models

using MLJ

model = @load RandomForestClassifier pkg=ScikitLearn

data = load_iris()
y, X = unpack(data, ==(:target), n->true, rng=1234)

mach = machine(model, X, y) |> fit!
MLJ.save("junk.jlso", mach)

## NEW REPL SESSION

using MLJ

model = @load RandomForestClassifier pkg=ScikitLearn

mach_predict_only = machine("junk.jlso") 

Sometimes the last expression hangs and if I interrupt Julia crashes. Sometimes it works fine. Same is true for JLSO.load("junk.jlso"). Anyone able to recreate?

(junk) pkg> st
Status `~/Dropbox/Julia7/MLJ/MLJ/sandbox/junk/Project.toml`
  [9da8a3cd] JLSO v2.3.2
  [add582a8] MLJ v0.14.1
  [5ae90465] MLJScikitLearnInterface v0.1.6

julia> versioninfo()
Julia Version 1.5.1
Commit 697e782ab8 (2020-08-25 20:08 UTC)
Platform Info:
  OS: macOS (x86_64-apple-darwin19.5.0)
  CPU: Intel(R) Core(TM) i7-8850H CPU @ 2.60GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-9.0.1 (ORCJIT, skylake)
Environment:
  JULIA_LTS_PATH = /Applications/Julia-1.0.app/Contents/Resources/julia/bin/julia
  JULIA_PATH = /Applications/Julia-1.5.app/Contents/Resources/julia/bin/julia
  JULIA_NUM_THREADS = 5

It may be that the python objects bound to fitresult are not reliably persistent and we need to do implement custom MMI.save and MMI.restore along the lines of XGBoost.

Assuming ScikitLearn.jl exposes the python serialization, this ought to be easy to do.

@OkonSamuel Thoughts?

Start by reproducing the provided MLJ.save and machine("junk.jlso") sequence in a fresh Julia REPL using the listed Julia, MLJ, JLSO, and ScikitLearnInterface versions. Inspect the ScikitLearn interface's serialization behavior and compare it with the MMI.save and MMI.restore implementation linked for XGBoost. Done means loading the saved model completes reliably without hanging or crashing.