MLJ implementations of KernelPerceptronClassifier, PegasosClassifier, PerceptronClassifier broken #79
Description
Context: JuliaAI/MLJ.jl#1182 (comment)
using MLJ
using MLJTestIntegration
A = @load KernelPerceptronClassifier
X, y = MLJTestIntegration.datasets(A)[1];
mach = machine(A(), X, y)
MLJ.fit!(mach, rows=1:4);
MLJ.predict(mach, X)
# ERROR: xtrain, ytrain or α have a length not compatible with the number of classes in this model.
# Stacktrace:
# [1] error(s::String)
# @ Base ./error.jl:35
# [2] predict(x::Matrix{…}, xtrain::Vector{…}, ytrain::Vector{…}, α::Vector{…}, classes::Vector{…}; K::typeof(radial_kernel))
# @ BetaML.Perceptron ~/.julia/packages/BetaML/GVo1h/src/Perceptron/Perceptron_kernel.jl:268
# [3] predict
# @ ~/.julia/packages/BetaML/GVo1h/src/Perceptron/Perceptron_kernel.jl:259 [inlined]
# [4] predict(model::BetaML.Bmlj.KernelPerceptronClassifier, fitresult::Tuple{…}, Xnew::@NamedTuple{…})
# @ BetaML.Bmlj ~/.julia/packages/BetaML/GVo1h/src/Bmlj/Perceptron_mlj.jl:285
# [5] predict(mach::Machine{…}, Xraw::@NamedTuple{…})
# @ MLJBase ~/.julia/packages/MLJBase/F8Zzu/src/operations.jl:135
# [6] top-level scope
# @ REPL[55]:1
# Some type information was truncated. Use `show(err)` to see complete types.This looks like only levels explicitly seen in training (and not the complete categorical pool) are being to recorded in the train. I see that there was some unwrapping added in the last release, which I suppose is the cause.
If I recall correctly, the recent patch was to anticipate changes to CategoricalArrays.jl (which MLJ has not addressed). Did the baby get thrown out with the bathwater?