Univariate Finite 2 (#77)

Author: tlienart

Project.toml

@@ -1,7 +1,7 @@
 name = "MLJLinearModels"
 uuid = "6ee0df7b-362f-4a72-a706-9e79364fb692"
 authors = ["Thibaut Lienart <[email protected]>"]
-version = "0.4.0"
+version = "0.5.0"
 
 [deps]
 DocStringExtensions = "ffbed154-4ef7-542d-bbb7-c09d3a79fcae"
@@ -16,7 +16,7 @@ Parameters = "d96e819e-fc66-5662-9728-84c9c7592b0a"
 DocStringExtensions = "^0.8"
 IterativeSolvers = "^0.8"
 LinearMaps = "^2.6"
-MLJModelInterface = "^0.1,^0.2"
+MLJModelInterface = "^0.3"
 Optim = "^0.20,^0.21"
 Parameters = "^0.12"
 julia = "^1"

src/mlj/classifiers.jl

@@ -25,8 +25,6 @@ the strength of the L2 (resp. L1) regularisation components.
 * `fit_intercept` (Bool): whether to fit an intercept (Default: `true`)
 * `penalize_intercept` (Bool): whether to penalize intercept (Default: `false`)
 * `solver` (Solver): type of solver to use, default if `nothing`.
-* `multi_class` (Bool): whether it's a binary or multi class classification
-                        problem. This is usually set automatically.
 """
 @with_kw_noshow mutable struct LogisticClassifier <: MMI.Probabilistic
     lambda::Real             = 1.0
@@ -35,17 +33,15 @@ the strength of the L2 (resp. L1) regularisation components.
     fit_intercept::Bool      = true
     penalize_intercept::Bool = false
     solver::Option{Solver}   = nothing
-    multi_class::Bool        = false
-    nclasses::Int            = 2
 end
 
-glr(m::LogisticClassifier) =
+glr(m::LogisticClassifier, nclasses::Integer) =
     LogisticRegression(m.lambda, m.gamma;
                        penalty=Symbol(m.penalty),
-                       multi_class=m.multi_class,
+                       multi_class=(nclasses > 2),
                        fit_intercept=m.fit_intercept,
                        penalize_intercept=m.penalize_intercept,
-                       nclasses=m.nclasses)
+                       nclasses=nclasses)
 
 descr(::Type{LogisticClassifier}) = "Classifier corresponding to the loss function ``L(y, Xθ) + λ|θ|₂²/2 + γ|θ|₁`` where `L` is the logistic loss."
 
@@ -66,15 +62,14 @@ to `true` by default. The other parameters are the same.
     fit_intercept::Bool      = true
     penalize_intercept::Bool = false
     solver::Option{Solver}   = nothing
-    nclasses::Int            = 2       # leave to 2, cf LogisticRegression
 end
 
-glr(m::MultinomialClassifier) =
+glr(m::MultinomialClassifier, nclasses::Integer) =
     MultinomialRegression(m.lambda, m.gamma;
                           penalty=Symbol(m.penalty),
                           fit_intercept=m.fit_intercept,
                           penalize_intercept=m.penalize_intercept,
-                          nclasses=m.nclasses)
+                          nclasses=nclasses)
 
 descr(::Type{MultinomialClassifier}) =
     "Classifier corresponding to the loss function " *

src/mlj/interface.jl

@@ -54,36 +54,31 @@ function MMI.fit(m::Union{CLF_MODELS...}, verb::Int, X, y)
         yplain[yplain .== 1] .= -1
         yplain[yplain .== 2] .= 1
         # force the binary case
-        m.multi_class = false
-        m.nclasses    = 0
-    else # > 2
-        m.nclasses = nclasses
+        nclasses    = 0
     end
     # NOTE: here the number of classes is either 0 or > 2
-    clf = glr(m)
+    clf = glr(m, nclasses)
     solver = m.solver === nothing ? _solver(clf, size(Xmatrix)) : m.solver
     # get the parameters
     θ = fit(clf, Xmatrix, yplain, solver=solver)
     # return
-    return (θ, features, classes), nothing, NamedTuple{}()
+    return (θ, features, classes, nclasses), nothing, NamedTuple{}()
 end
 
-function MMI.predict(m::Union{CLF_MODELS...}, (θ, features, classes), Xnew)
+function MMI.predict(m::Union{CLF_MODELS...}, (θ, features, classes, c), Xnew)
     Xmatrix = MMI.matrix(Xnew)
-    c = m.nclasses
     preds = apply_X(Xmatrix, θ, c)
     if c > 2 # multiclass
         preds .= softmax(preds)
     else # binary (necessarily c==0)
         preds  .= sigmoid.(preds)
         preds   = hcat(1.0 .- preds, preds) # scores for -1 and 1
-        return [MMI.UnivariateFinite(classes, preds[i, :]) for i in 1:size(Xmatrix,1)]
+        return MMI.UnivariateFinite(classes, preds)
     end
-    return [MMI.UnivariateFinite(classes, preds[i, :]) for i in 1:size(Xmatrix,1)]
+    return MMI.UnivariateFinite(classes, preds)
 end
 
-function MMI.fitted_params(m::Union{CLF_MODELS...}, (θ, features, classes))
-    c = m.nclasses
+function MMI.fitted_params(m::Union{CLF_MODELS...}, (θ, features, classes, c))
     # helper function to assemble the results
     _fitted_params(coefs, features, intercept) =
         (classes=classes, coefs=coef_vec(coefs, features), intercept=intercept)

test/interface/fitpredict.jl

@@ -63,7 +63,7 @@ end
 @testset "String-Symbol" begin
     model = LogisticClassifier(penalty="l1")
     @test model.penalty == "l1"
-    gr = MLJLinearModels.glr(model)
+    gr = MLJLinearModels.glr(model, 2)
     @test gr isa GLR
     @test gr.penalty isa ScaledPenalty{L1Penalty}
 end