Mlj.jl: Problems deserialising pipeline models because of auto-generated type name

[ablaom]

From an issue raised on slack:

I am trying to load a saved model (MLJ.save("hottree2.jlso", mtm, compression=:none) )from https://alan-turing-institute.github.io/DataScienceTutorials.jl/isl/lab-8/ but get the following error
julia> mach2 = machine("hottree2.jlso")
ERROR: MethodError: no method matching Machine(::Array{UInt8,1})
Closest candidates are:
Machine(::M, ::AbstractNode...) where M<:MLJModelInterface.Model at /home/jade/.julia/packages/MLJBase/cJmIS/src/machines.jl:19
Stacktrace:
[1] machine(::String; kwargs::Base.Iterators.Pairs{Union{},Union{},Tuple{},NamedTuple{(),Tuple{}}}) at /home/jade/.julia/packages/MLJBase/cJmIS/src/machines.jl:676
[2] machine(::String) at /home/jade/.julia/packages/MLJBase/cJmIS/src/machines.jl:674
And my response there:

Interesting issue. I’m assuming you executed the same code that defined your pipeline model before attempting to retrieve the saved machine. In that case I expect I can guess the problem - the auto-generated model type name for the pipeline is different. A workaround is to explicitly give your pipeline mode type a name by specifying @pipeline …… name=MyPipe . Let me know if that works for you.

The workaround resolves the issue.

---

I personally find the automatic generation of a type name very convenient and am guessing the best solution more generally is to flag the problem in documentation. Any other ideas?

[OkonSamuel]

@ablaom I agree documenting this and also putting this in the doc-string of MLJ.save
But there is another issue that occurs when de-serializing user defined models (including Composites e.g Pipelines) even in the REPL. (well not really an issue if it's documented)
The following code shows this

# enter the following in a REPL session
using MLJ
import RDatasets: dataset

import DataFrames: DataFrame, select, Not
@load DecisionTreeClassifier pkg=DecisionTree

carseats = dataset("ISLR", "Carseats")

High = ifelse.(carseats.Sales .<= 8, "No", "Yes") |> categorical;
carseats[!, :High] = High;

X = select(carseats, Not([:Sales, :High]))
y = carseats.High;

HotTreeClf = @pipeline(OneHotEncoder(),
                       DecisionTreeClassifier(), name = MyPipe) #note that i named my Pipeline

mdl = HotTreeClf
mach = machine(mdl, X, y)
fit!(mach);

MLJ.save("hottree.jlso", mach)
machine("hottree.jlso") # this works

# Now restart the REPL and enter the follwoing code
using MLJ
julia> machine("hottree.jlso")
[warn | JLSO]: UndefVarError: DecisionTree_ not defined
[warn | JLSO]: UndefVarError: DecisionTree_ not defined
[warn | JLSO]: UndefVarError: MyPipe not defined
ERROR: MethodError: no method matching Machine(::Array{UInt8,1})
....

julia> julia> @load DecisionTreeClassifier pkg=DecisionTree
DecisionTreeClassifier(
    max_depth = -1,
    min_samples_leaf = 1,
    min_samples_split = 2,
    min_purity_increase = 0.0,
    n_subfeatures = 0,
    post_prune = false,
    merge_purity_threshold = 1.0,
    pdf_smoothing = 0.0,
    display_depth = 5) @548

julia> machine("hottree.jlso")
[warn | JLSO]: UndefVarError: MyPipe not defined
ERROR: MethodError: no method matching Machine(::Array{UInt8,1})
...

# I had to redefine the MyPipe for this to work
@pipeline(OneHotEncoder(),
                              DecisionTreeClassifier(), name = MyPipe)

julia> machine("hottree.jlso")
Machine{MyPipe} @856 trained 1 time.
  args:

I think we need to document this. i.e for user defined models including Pipelines, the implementation of the models type has to be loaded or defined for deserialization to work.

[ablaom]

I think we need to document this. i.e for user defined models including Pipelines, the implementation of the models type has to be loaded or defined for deserialization to work.

I think this is pretty well understood but agree newbies will rub against it, and it should be documented.

[ablaom]

Regarding the original post here, one solution is to make all pipeline model a single type, along the lines suggested in https://github.com/alan-turing-institute/MLJ.jl/issues/594, but this is some work and needs to be carefully thought through.

[drcxcruz]

Hello, let me know if I could help in documenting it. Please let us know of any suggestions. Thanks