Calling a scikit-learn model

[fabiensatalia]

I want to use CounterfactualExplanations.jl to generate counterfactual explanations with a model trained with scikit-learn in Python. After some reading, my current idea is to write a Julia wrapper (using PyCall) to call that Python model from Julia, and declare this wrapper as a subtype of AbstractFittedModel. Is this a smart way to go about this? Am I missing a simpler way to do this?

[pat-alt]

That would be a very welcome contribution actually. And yes, that's the right approach 👍 Just keep in mind that depending on what type of model you want to use, you may not be able to use gradient-based generators. Interested to see what you come up with, if you are able to share. Thanks!

[fabiensatalia]

I did implement the things mentioned above but when applied to a toy workflow inspired from the doc, it is currently not working. I am unclear on what the error message means. The cause could be what you mentioned in your previous reply, i.e. that gradient-based generators won't work with a random forest model.

What I did can be found here: https://github.com/fabiensatalia/SklearnBackend.jl

Here is the error message I get when running test.jl:

ERROR: Compiling Tuple{typeof(convert), Type{PyCall.PyAny}, PyCall.PyObject}: ArgumentError: array must be non-empty
Stacktrace:
  [1] macro expansion
    @ ~/.julia/packages/Zygote/YYT6v/src/compiler/interface2.jl:101 [inlined]
  [2] _pullback(::Zygote.Context{true}, ::typeof(convert), ::Type{PyCall.PyAny}, ::PyCall.PyObject)
    @ Zygote ~/.julia/packages/Zygote/YYT6v/src/compiler/interface2.jl:101
  [3] _pullback
    @ ~/.julia/packages/PyCall/ilqDX/src/PyCall.jl:318 [inlined]
  [4] _pullback(::Zygote.Context{true}, ::typeof(getproperty), ::PyCall.PyObject, ::Symbol)
    @ Zygote ~/.julia/packages/Zygote/YYT6v/src/compiler/interface2.jl:0
  [5] macro expansion
    @ ~/.julia/packages/Zygote/YYT6v/src/lib/literal_getproperty.jl:93 [inlined]
  [6] _pullback
    @ ~/.julia/packages/Zygote/YYT6v/src/lib/literal_getproperty.jl:117 [inlined]
  [7] _pullback
    @ ~/projects/CounterFactuals_contribu/julia/SklearnBackend/src/SklearnBackend.jl:36 [inlined]
  [8] _pullback(::Zygote.Context{true}, ::typeof(probs), ::SklearnModel, ::Matrix{Float64})
    @ Zygote ~/.julia/packages/Zygote/YYT6v/src/compiler/interface2.jl:0
  [9] _pullback
    @ ~/projects/CounterFactuals_contribu/julia/SklearnBackend/src/SklearnBackend.jl:40 [inlined]
 [10] _pullback(::Zygote.Context{true}, ::typeof(logits), ::SklearnModel, ::Matrix{Float64})
    @ Zygote ~/.julia/packages/Zygote/YYT6v/src/compiler/interface2.jl:0
 [11] _pullback
    @ ~/.julia/packages/CounterfactualExplanations/ATUIK/src/objectives/loss_functions.jl:23 [inlined]
 [12] _pullback(::Zygote.Context{true}, ::CounterfactualExplanations.Objectives.var"##logitcrossentropy#12", ::Base.Pairs{Symbol, Union{}, Tuple{}, NamedTuple{(), Tuple{}}}, ::typeof(Flux.Losses.logitcrossentropy), ::CounterfactualExplanation)
    @ Zygote ~/.julia/packages/Zygote/YYT6v/src/compiler/interface2.jl:0
 [13] _pullback
    @ ~/.julia/packages/CounterfactualExplanations/ATUIK/src/objectives/loss_functions.jl:22 [inlined]
 [14] _pullback(ctx::Zygote.Context{true}, f::typeof(Flux.Losses.logitcrossentropy), args::CounterfactualExplanation)
    @ Zygote ~/.julia/packages/Zygote/YYT6v/src/compiler/interface2.jl:0
 [15] _pullback
    @ ~/.julia/packages/CounterfactualExplanations/ATUIK/src/generators/loss.jl:18 [inlined]
 [16] _pullback(::Zygote.Context{true}, ::typeof(CounterfactualExplanations.Generators.ℓ), ::CounterfactualExplanations.Generators.GradientBasedGenerator, ::Nothing, ::CounterfactualExplanation)
    @ Zygote ~/.julia/packages/Zygote/YYT6v/src/compiler/interface2.jl:0
 [17] _pullback
    @ ~/.julia/packages/CounterfactualExplanations/ATUIK/src/generators/loss.jl:7 [inlined]
 [18] _pullback(::Zygote.Context{true}, ::typeof(CounterfactualExplanations.Generators.ℓ), ::CounterfactualExplanations.Generators.GradientBasedGenerator, ::CounterfactualExplanation)
    @ Zygote ~/.julia/packages/Zygote/YYT6v/src/compiler/interface2.jl:0
 [19] _pullback
    @ ~/.julia/packages/CounterfactualExplanations/ATUIK/src/generators/gradient_based/loss.jl:13 [inlined]
 [20] _pullback(::Zygote.Context{true}, ::CounterfactualExplanations.Generators.var"#15#16"{CounterfactualExplanations.Generators.GradientBasedGenerator, CounterfactualExplanation})
    @ Zygote ~/.julia/packages/Zygote/YYT6v/src/compiler/interface2.jl:0
 [21] pullback(f::Function, ps::Zygote.Params{Zygote.Buffer{Any, Vector{Any}}})
    @ Zygote ~/.julia/packages/Zygote/YYT6v/src/compiler/interface.jl:414
 [22] gradient(f::Function, args::Zygote.Params{Zygote.Buffer{Any, Vector{Any}}})
    @ Zygote ~/.julia/packages/Zygote/YYT6v/src/compiler/interface.jl:96
 [23] ∂ℓ(generator::CounterfactualExplanations.Generators.GradientBasedGenerator, M::SklearnModel, ce::CounterfactualExplanation)
    @ CounterfactualExplanations.Generators ~/.julia/packages/CounterfactualExplanations/ATUIK/src/generators/gradient_based/loss.jl:13
 [24] ∇(generator::CounterfactualExplanations.Generators.GradientBasedGenerator, M::SklearnModel, ce::CounterfactualExplanation)
    @ CounterfactualExplanations.Generators ~/.julia/packages/CounterfactualExplanations/ATUIK/src/generators/gradient_based/loss.jl:46
 [25] propose_state(generator::CounterfactualExplanations.Generators.GradientBasedGenerator, ce::CounterfactualExplanation)
    @ CounterfactualExplanations.Generators ~/.julia/packages/CounterfactualExplanations/ATUIK/src/generators/gradient_based/utils.jl:9
 [26] generate_perturbations(generator::CounterfactualExplanations.Generators.GradientBasedGenerator, ce::CounterfactualExplanation)
    @ CounterfactualExplanations.Generators ~/.julia/packages/CounterfactualExplanations/ATUIK/src/generators/gradient_based/utils.jl:24
 [27] update!(ce::CounterfactualExplanation)
    @ CounterfactualExplanations ~/.julia/packages/CounterfactualExplanations/ATUIK/src/counterfactuals/search.jl:9
 [28] generate_counterfactual(x::Matrix{Float64}, target::Int64, data::CounterfactualData, M::SklearnModel, generator::CounterfactualExplanations.Generators.GradientBasedGenerator; num_counterfactuals::Int64, initialization::Symbol, generative_model_params::NamedTuple{(), Tuple{}}, max_iter::Int64, decision_threshold::Float64, gradient_tol::Float64, min_success_rate::Float64, converge_when::Symbol, timeout::Nothing, invalidation_rate::Float64, learning_rate::Float64, variance::Float64)
    @ CounterfactualExplanations ~/.julia/packages/CounterfactualExplanations/ATUIK/src/generate_counterfactual/core.jl:86
 [29] generate_counterfactual(x::Matrix{Float64}, target::Int64, data::CounterfactualData, M::SklearnModel, generator::CounterfactualExplanations.Generators.GradientBasedGenerator)
    @ CounterfactualExplanations ~/.julia/packages/CounterfactualExplanations/ATUIK/src/generate_counterfactual/core.jl:43
 [30] top-level scope
    @ REPL[49]:1

[fabiensatalia]

Update: I have the same error when using a sklearn.linear_model.LogisticRegression classifier.

[pat-alt]

Thanks for looking into this. I've transferred this to an issue here. Maybe we can add support for sklearn there, although likely limited (reasons discussed in issue).

[fabiensatalia]

Thanks, the link does not seem to work on my computer, is it working on yours? I would like to follow it and see if there's any way I can contribute.

[pat-alt]

My bad, the repo was set to private, because it's still in very early development. Should be accessible now.