Global Loss Function Capabilities

[Waswas321123]

I used to be able to define a global loss function (specifically Spearman rank correlation), but now element-wise loss makes that impossible. Is there a workaround for disabling the element wise requirement in Python? Or at least a good approximation for rank correlation?

[MilesCranmer]

Sorry I am not sure I follow. Are you okay to use loss_function? This let's you write an arbitrary loss.

[Waswas321123]

When I try precomputing data rankings and then fitting a model using
loss_function = """ begin using Statistics function loss_function(prediction, target) return 1 - cor(prediction, target) end end """
I run into the very long Julia error

juliacall.JuliaError: MethodError: no method matching loss_function(::Node{Float32}, ::SymbolicRegression.CoreModule.DatasetModule.BasicDataset{Float32, Float32, Matrix{Float32}, Vector{Float32}, Nothing, @NamedTuple{}, Nothing, Nothing, Nothing, Nothing}, ::Options{SymbolicRegression.CoreModule.OptionsStructModule.ComplexityMapping{Int64, Int64}, OperatorEnum, Node, Expression, @NamedTuple{}, MutationWeights, false, false, nothing, Nothing, 5}) The function loss_function exists, but no method is defined for this combination of argument types.

Apologies for the very long printout. Seems no matter how I've tried it I'm coming up with error

[MilesCranmer]

Try

from pysr import jl
jl.seval("using Pkg: Pkg")
jl.Pkg.add("Statistics")
jl.seval("using Statistics: cor")

loss_function = """
function eval_loss(tree, dataset::Dataset{T,L}, options)::L where {T,L}
    prediction, flag = eval_tree_array(tree, dataset.X, options)
    if !flag
        return L(Inf)
    end
    return 1 - cor(prediction, dataset.y)
end
"""

Basically elementwise_loss and loss_function are two different things. See https://ai.damtp.cam.ac.uk/pysr/api/