fix cases where watchlist can be improperly passed as nothing (#29)

* fix cases where watchlist can be improperly passed as nothing

* bump version

* fix disaster caused by fixing XGBoost predict output

Project.toml

@@ -1,7 +1,7 @@
 name = "MLJXGBoostInterface"
 uuid = "54119dfa-1dab-4055-a167-80440f4f7a91"
 authors = ["Anthony D. Blaom <[email protected]>"]
-version = "0.3.3"
+version = "0.3.4"
 
 [deps]
 MLJModelInterface = "e80e1ace-859a-464e-9ed9-23947d8ae3ea"

src/MLJXGBoostInterface.jl

@@ -107,7 +107,7 @@ function kwargs(model, verbosity, obj)
     o = merge(o, (silent=(verbosity ≤ 0),))
     # watchlist is for log output, so override if it's default and verbosity ≤ 0
     wl = (verbosity ≤ 0 && isnothing(model.watchlist)) ? (;) : model.watchlist
-    o = merge(o, (watchlist=wl,))
+    isnothing(wl) || (o = merge(o, (watchlist=wl,)))
     merge(o, (objective=_fix_objective(obj),))
 end
 
@@ -174,16 +174,8 @@ function MMI.predict(model::XGBoostClassifier, fitresult, Xnew)
     nlevels = length(classes)
     npatterns = MMI.nrows(Xnew)
 
-    if nlevels == 2
-        true_class_probabilities = reshape(o, 1, npatterns)
-        false_class_probabilities = 1 .- true_class_probabilities
-        o = vcat(false_class_probabilities, true_class_probabilities)
-    end
-
-    prediction_probabilities = reshape(o, nlevels, npatterns)
-
-    # note we use adjoint of above:
-    MMI.UnivariateFinite(classes, prediction_probabilities')
+    # XGB can return a rank-1 array for binary classification
+    MMI.UnivariateFinite(classes, o, augment=ndims(o)==1)
 end