Merge pull request #898 from JuliaAI/loss-functions-0.9

Bump compat LossFunctions = "0.9" and address breakage

Project.toml

@@ -35,7 +35,7 @@ CategoricalDistributions = "0.1"
 ComputationalResources = "0.3"
 Distributions = "0.25.3"
 InvertedIndices = "1"
-LossFunctions = "0.5, 0.6, 0.7, 0.8"
+LossFunctions = "0.9"
 MLJModelInterface = "1.7"
 Missings = "0.4, 1"
 OrderedCollections = "1.1"

src/measures/loss_functions_interface.jl

@@ -44,7 +44,7 @@ err_wrap(n) = ArgumentError("Bad @wrap syntax: $n. ")
 
 # We define amacro to wrap a concrete `LossFunctions.SupervisedLoss`
 # type and define its constructor, and to define property access in
-# case of paramters; the macro also defined calling behaviour:
+# case of parameters; the macro also defines calling behaviour:
 macro wrap_loss(ex)
  ex.head == :call || throw(err_wrap(1))
  Loss_ex = ex.args[1]
@@ -130,7 +130,7 @@ MMI.prediction_type(::Type{<:DistanceLoss}) = :deterministic
 MMI.target_scitype(::Type{<:DistanceLoss}) = Union{Vec{Continuous},Vec{Count}}
 
 call(measure::DistanceLoss, yhat, y) =
- LossFunctions.value(getfield(measure, :loss), y, yhat)
+ LossFunctions.value(getfield(measure, :loss), yhat, y)
 
 function call(measure::DistanceLoss, yhat, y, w::AbstractArray)
  return w .* call(measure, yhat, y)
@@ -148,7 +148,7 @@ _scale(p) = 2p - 1
 function call(measure::MarginLoss, yhat, y)
  probs_of_observed = broadcast(pdf, yhat, y)
  return (LossFunctions.value).(getfield(measure, :loss),
- 1, _scale.(probs_of_observed))
+ _scale.(probs_of_observed), 1)
 end
 
 call(measure::MarginLoss, yhat, y, w::AbstractArray) =

test/measures/loss_functions_interface.jl

@@ -42,12 +42,9 @@ end
 
  for M_ex in MARGIN_LOSSES
  m = eval(:(MLJBase.$M_ex()))
- @test m(yhat, y) ≈ LossFunctions.value(getfield(m, :loss), ym, yhatm)
- @test MLJBase.Mean()(m(yhat, y, w)) ≈
- LossFunctions.value(getfield(m, :loss),
- ym,
- yhatm,
- WeightedSum(w))/N
+ @test m(yhat, y) ≈ LossFunctions.value(getfield(m, :loss), yhatm, ym)
+ @test m(yhat, y, w) ≈
+ w .* LossFunctions.value(getfield(m, :loss), yhatm, ym)
  end
 end
 
@@ -64,10 +61,8 @@ end
  m_ex = MLJBase.snakecase(M_ex)
  @test m == eval(:(MLJBase.$m_ex))
  @test m(yhat, y) ≈
- LossFunctions.value(getfield(m, :loss), y, yhat)
- @test mean(m(yhat ,y, w)) ≈
- LossFunctions.value(getfield(m, :loss), y, yhat,
- WeightedSum(w))/N
-
+ LossFunctions.value(getfield(m, :loss), yhat, y)
+ @test m(yhat ,y, w) ≈
+ w .* LossFunctions.value(getfield(m, :loss), yhat, y)
  end
 end