Update to LossFunctions.jl v0.10

[juliohm]

We are gradually cleaning up the LossFunctions.jl package, simplifying the interface and making the code more general. Can you please review the latest changes and bump the version here again?

The main breaking changes from v0.9 to v0.10 are:

1. The value(loss, yhat, y) function is replaced by the functor loss(yhat, y)
2. The AggMode submode is gone in favor of sum(loss, yhats, ys, [weights]) and mean(loss, yhats, ys, [weights])

cc: @ablaom @OkonSamuel

[ablaom]

@juliohm It seems that calling behaviour of loss functions has also changed:

In 0.9:

julia> loss = DWDMarginLoss(1.0)
DWDMarginLoss{Float64}(1.0)

julia> yhat = [-0.5, 0.5, 0.2]
3-element Vector{Float64}:
 -0.5
  0.5
  0.2

julia> y = [-1, 1 , 1]
3-element Vector{Int64}:
 -1
  1
  1

julia> loss(yhat, y)
3-element Vector{Float64}:
 0.5
 0.5
 0.8

In 0.10:

julia> loss(yhat, y)
ERROR: MethodError: no method matching (::DWDMarginLoss{Float64})(::Vector{Float64}, ::Vector{Int64})
Stacktrace:
 [1] top-level scope
   @ REPL[7]:1

Now that value is gone, what is the preferred way to apply losses to vectors? Am I expected to just broadcast?

[juliohm]

Am I expected to just broadcast?

Exactly. We let Julia do the work as it is more idiomatic and can fuse with other outer calls.