Merge pull request #880 from JuliaAI/dev

For a 0.21.5 release

Project.toml

@@ -1,7 +1,7 @@
 name = "MLJBase"
 uuid = "a7f614a8-145f-11e9-1d2a-a57a1082229d"
 authors = ["Anthony D. Blaom <[email protected]>"]
-version = "0.21.4"
+version = "0.21.5"
 
 [deps]
 CategoricalArrays = "324d7699-5711-5eae-9e2f-1d82baa6b597"

src/composition/learning_networks/signatures.jl

@@ -17,7 +17,7 @@ single machine.
 function machines_given_model(node::AbstractNode)
     ret = LittleDict{Symbol,Any}()
     for mach in machines(node)
-        model = mach.model 
+        model = mach.model
         model isa Symbol || continue
         if !haskey(ret, model)
             ret[model] = Any[mach,]
@@ -245,6 +245,17 @@ See also [`MLJBase.Signature`](@ref).
 """
 operations(signature::Signature) = keys(operation_nodes(signature))
 
+"""
+    glb(signature::Signature)
+
+**Private method.**
+
+Return the greatest lower bound of all operation nodes, report nodes and fitted parameter
+nodes associated with `signature`.
+
+See also [`MLJBase.Signature`](@ref).
+
+"""
 function glb(signature::Signature)
     grab(f) = values(f(signature)) |> collect
     nodes = vcat(

src/composition/models/stacking.jl

@@ -34,10 +34,28 @@ const ERR_BAD_METALEARNER = ArgumentError(
     "of $(Union{Deterministic, Probabilistic})"
 )
 
+ERR_BAD_BASEMODEL(model) = ArgumentError(
+    "The base model $model is not supported as it appears to "*
+    "be a classifier predicting point values. Ordinarily, the "*
+    "the model must either be a "*
+    "probabilistic classifier (output of `predict` is a vector of  "*
+    "`UnivariateFinite`) "*
+    "or a regressor (`target_scitype(model) <: "*
+    "AbstractVector{<:Union{Continuous,Missing}}`). "
+)
+
 const ERR_NO_METALEARNER = ArgumentError(
     "No metalearner specified. Use Stack(metalearner=..., model1=..., model2=...)"
 )
 
+# checks `model` is either a probabilistic classifier or a regressor:
+function check_valid_basemodel(model)
+    problem = prediction_type(model) === :deterministic &&
+        target_scitype(model) <: AbstractVector{<:Union{Finite,Missing}}
+    problem && throw(ERR_BAD_BASEMODEL(model))
+    return nothing
+end
+
 mutable struct DeterministicStack{
     modelnames,
     inp_scitype,
@@ -179,8 +197,13 @@ function MMI.clean!(stack::Stack{modelnames, inp_scitype, tg_scitype}) where {
     tg_scitype
 }
 
-    # We only carry checks and don't try to correct the arguments here
+    # We only carry out checks and don't try to correct the arguments here
     message = ""
+
+    # check basemodels:
+    basemodels = map(name -> getproperty(stack, name), modelnames)
+    check_valid_basemodel.(basemodels)
+
     # Checking target_scitype and input_scitype have not been changed from the original
     # stack:
     glb_inp_scitype, glb_tg_scitype =

src/machines.jl

@@ -965,7 +965,7 @@ A machine returned by `serializable` is characterized by the property
 
     # This machine can now be serialized
     smach = serializable(mach)
-    JLSO.save("machine.jlso", machine => smach)
+    JLSO.save("machine.jlso", :machine => smach)
 
     # Deserialize and restore learned parameters to useable form:
     loaded_mach = JLSO.load("machine.jlso")[:machine]

src/show.jl

@@ -1,5 +1,3 @@
-## REGISTERING LABELS OF OBJECTS DURING ASSIGNMENT
-
 """
     color_on()
 
@@ -20,6 +18,8 @@ macro colon(p)
     Expr(:quote, p)
 end
 
+## REGISTERING LABELS OF OBJECTS DURING ASSIGNMENT
+
 """
     @constant x = value
 
@@ -70,8 +70,9 @@ end
 
 # long version of showing a named tuple:
 Base.show(stream::IO, ::MIME"text/plain", t::NamedTuple) = fancy_nt(stream, t)
-fancy_nt(t) = fancy_nt(stdout, t)
-fancy_nt(stream::IO, t) = fancy_nt(stream, t, 0)
+fancy_nt(t) = fancy_nt(stdout, t) # is this used?
+fancy_nt(stream, t::NamedTuple{(), Tuple{}}) = print(stream, "NamedTuple()")
+fancy_nt(stream, t) = fancy_nt(stream, t, 0)
 fancy_nt(stream, t, n) = show(stream, t)
 function fancy_nt(stream, t::NamedTuple, n)
     print(stream, "(")

test/composition/models/stacking.jl

@@ -158,6 +158,13 @@ end
 end
 
 @testset "Stack constructor valid argument checks" begin
+
+    # baselearner cannot be a deterministic classifier:
+    @test_throws(
+    MLJBase.ERR_BAD_BASEMODEL(DeterministicConstantClassifier()),
+    Stack(metalearner=ConstantClassifier(), mymodel=DeterministicConstantClassifier()),
+    )
+
     # metalearner should be `Deterministic` or `Probablisitic`:
     @test_throws(
         MLJBase.ERR_BAD_METALEARNER,