Fix MLJ Serialization, add test for single class classifiers, fix single class classifiers predict

Project.toml

@@ -1,7 +1,7 @@
 name = "CatBoost"
 uuid = "e2e10f9a-a85d-4fa9-b6b2-639a32100a12"
 authors = ["Beacon Biosignals, Inc."]
-version = "0.3.4"
+version = "0.3.5"
 
 [deps]
 MLJModelInterface = "e80e1ace-859a-464e-9ed9-23947d8ae3ea"

src/mlj_catboostclassifier.jl

@@ -100,10 +100,11 @@ MMI.reports_feature_importances(::Type{<:CatBoostClassifier}) = true
 function MMI.predict(mlj_model::CatBoostClassifier, fitresult, X_pool)
     if fitresult[1] === nothing
         # Always predict the single class
-        n = nrow(X_pool)
+        n = pyconvert(Int, X_pool.shape[0])
         classes = [fitresult.single_class]
         probs = ones(n, 1)
-        return MMI.UnivariateFinite(classes, probs; pool=fitresult.y_first)
+        pool = MMI.categorical([fitresult.y_first])
+        return MMI.UnivariateFinite(classes, probs; pool=pool)
     end
 
     model, y_first = fitresult
@@ -116,8 +117,8 @@ end
 function MMI.predict_mode(mlj_model::CatBoostClassifier, fitresult, X_pool)
     if fitresult[1] === nothing
         # Return probability 1 for the single class
-        n = nrow(X_pool)
-        return hcat(ones(n), zeros(n))
+        n = pyconvert(Int, X_pool.shape[0])
+        return fill(fitresult.y_first, n)
     end
 
     model, y_first = fitresult

src/mlj_serialization.jl

@@ -1,24 +1,41 @@
 # Taken from https://github.com/JuliaAI/MLJXGBoostInterface.jl
 # It is likely also not the optimal method for serializing models, but it works
+
 """
-    _persistent(booster)
+    _persistent(model::CatBoostModels, fitresult)
 
 Private method.
 
 Return a persistent (ie, Julia-serializable) representation of the
-CatBoost.jl model `booster`.
+CatBoost.jl model `fitresult`.
 
-Restore the model with [`booster`](@ref)
+Restore the model with [`fitresult`](@ref)
 """
-function _persistent(booster)
+function _persistent(::CatBoostRegressor, fitresult)
     ctb_file, io = mktemp()
     close(io)
 
-    booster.save_model(ctb_file)
+    fitresult.save_model(ctb_file)
     persistent_booster = read(ctb_file)
     rm(ctb_file)
     return persistent_booster
 end
+function _persistent(::CatBoostClassifier, fitresult)
+    model, y_first = fitresult
+    if model === nothing
+        # Case 1: Single unique class
+        return (nothing, fitresult.single_class, y_first)
+    else
+        # Case 2: Multiple unique classes
+        ctb_file, io = mktemp()
+        close(io)
+
+        model.save_model(ctb_file)
+        persistent_booster = read(ctb_file)
+        rm(ctb_file)
+        return (persistent_booster, y_first)
+    end
+end
 
 """
     _booster(persistent)
@@ -28,24 +45,45 @@
 Return the CatBoost.jl model which has `persistent` as its persistent
 (Julia-serializable) representation. See [`persistent`](@ref) method.
 """
-function _booster(persistent)
+function _booster(::CatBoostRegressor, persistent)
+    ctb_file, io = mktemp()
+    write(io, persistent)
+    close(io)
+
+    booster = catboost.CatBoostRegressor().load_model(ctb_file)
+
+    rm(ctb_file)
+
+    return booster
+end
+function _booster(::CatBoostClassifier, persistent)
     ctb_file, io = mktemp()
     write(io, persistent)
     close(io)
 
-    booster = catboost.CatBoost().load_model(ctb_file)
+    booster = catboost.CatBoostClassifier().load_model(ctb_file)
 
     rm(ctb_file)
 
     return booster
 end
 
-function MMI.save(::CatBoostModels, fr; kw...)
-    (booster, a_target_element) = fr
-    return (_persistent(booster), a_target_element)
+function MMI.save(model::CatBoostModels, fitresult; kwargs...)
+    return _persistent(model, fitresult)
+end
+
+function MMI.restore(model::CatBoostRegressor, serializable_fitresult)
+    return _booster(model, serializable_fitresult)
 end
 
-function MMI.restore(::CatBoostModels, fr)
-    (persistent, a_target_element) = fr
-    return (_booster(persistent), a_target_element)
+function MMI.restore(model::CatBoostClassifier, serializable_fitresult)
+    if serializable_fitresult[1] === nothing
+        # Case 1: Single unique class
+        return (model=nothing, single_class=serializable_fitresult[2],
+                y_first=serializable_fitresult[3])
+    else
+        # Case 2: Multiple unique classes
+        persistent_booster, y_first = serializable_fitresult
+        return (_booster(model, persistent_booster), y_first)
+    end
 end

test/mlj_interface.jl

@@ -22,7 +22,24 @@
         preds = MLJBase.predict(mach, X)
         probs = MLJBase.predict_mode(mach, X)
 
-        serializable_fitresult = MLJBase.save(mach, mach.fitresult)
+        serializable_fitresult = MLJBase.save(mach, mach)
+        restored_fitresult = MLJBase.restore(mach, serializable_fitresult)
+    end
+
+    @testset "CatBoostClassifier - single class" begin
+        X = (; a=[1, 4, 5, 6], b=[4, 5, 6, 7])
+        y = [0, 0, 0, 0]
+
+        # MLJ Interface
+        model = CatBoostClassifier(; iterations=5)
+        mach = machine(model, X, y)
+        MLJBase.fit!(mach)
+        preds = MLJBase.predict(mach, X)
+        println(preds)
+        probs = MLJBase.predict_mode(mach, X)
+        println(probs)
+
+        serializable_fitresult = MLJBase.save(mach, mach)
         restored_fitresult = MLJBase.restore(mach, serializable_fitresult)
     end
 
@@ -36,7 +53,7 @@
         MLJBase.fit!(mach)
         preds = MLJBase.predict(mach, X)
 
-        serializable_fitresult = MLJBase.save(mach, mach.fitresult)
+        serializable_fitresult = MLJBase.save(mach, mach)
         restored_fitresult = MLJBase.restore(mach, serializable_fitresult)
     end