Naive passing of Julia class_weight dictionary does not work

[ablaom]

At least it doesn't work for RandomForestClassifier:

using MLJ 
Forest  = @load RandomForestClassifier pkg=MLJScikitLearnInterface
w = Dict("setosa"=>0.2, "versicolor"=>0.7, "virginica"=>0.1)
forest = Forest(class_weight=w)

julia> mach = machine(forest, X, y) |> fit!
[ Info: Training machine(RandomForestClassifier(n_estimators = 100, …), …).
┌ Error: Problem fitting the machine machine(RandomForestClassifier(n_estimators = 100, …), …). 
└ @ MLJBase ~/.julia/packages/MLJBase/fEiP2/src/machines.jl:682
[ Info: Running type checks... 
[ Info: Type checks okay. 
ERROR: Python: InvalidParameterError: The 'class_weight' parameter of RandomForestClassifier must be a str among {'balanced_subsample', 'balanced'}, an instance of 'dict', an instance of 'list' or None. Got Julia:
Dict{String, Float64} with 3 entries:
  "virginica"  => 0.1
  "setosa"     => 0.2
  "versicolor" => 0.7 instead.
Python stacktrace:
 [1] validate_parameter_constraints
   @ sklearn.utils._param_validation 

Currently the struct has class_weight::Any. I wonder if it makes any difference if this is Union{Nothing, AbstractDict}.

@tylerjthomas9 Is there an easy fix for this? Workaround?

[tylerjthomas9]

PythonCall.jl doesn't automatically convert Dicts to Python. You can manually do it with pydict. This weekend I will investigate a better way than manually converting them to Python dictionaries. However, the class

using MLJ 
using PythonCall
X, y = @load_iris
Forest  = @load RandomForestClassifier pkg=MLJScikitLearnInterface
w = Dict(1=>0.2, 2=>0.7, 3=>0.1)
forest = Forest(class_weight=pydict(w))
mach = machine(forest, X, y) |> fit!

One thing to note, it is that we convert the y to Int using yplain = MMI.int(y). We may want to change this so we don't have to use unnamed integers as the classes in the dictionary.

[ablaom]

@tylerjthomas9 Thanks for the quick response and the workaround.

[ablaom]

Closed by #64