```{julia}
include("$(pwd())/notebooks/setup.jl")
eval(setup_notebooks)
```

# Real-World Data

```{julia}
# Hyper:
_retrain = true

# Data:
test_size = 0.2
n_obs = Int(10000 / (1.0 - test_size))
counterfactual_data, test_data = train_test_split(load_california_housing(n_obs); test_size=test_size)
X, y = CounterfactualExplanations.DataPreprocessing.unpack_data(counterfactual_data)
X = table(permutedims(X))
labels = counterfactual_data.output_encoder.labels
input_dim, n_obs = size(counterfactual_data.X)
output_dim = length(unique(labels))
```

First, let's create a couple of image classifier architectures:

```{julia}
# Model parameters:
epochs = 100
batch_size = minimum([Int(round(n_obs/10)), 128])
n_hidden = 128
activation = Flux.relu
builder = MLJFlux.@builder Flux.Chain(
    Dense(n_in, n_hidden, activation),
    Dense(n_hidden, n_hidden, activation),
    Dense(n_hidden, n_out),
)
n_ens = 5                                   # number of models in ensemble
_loss = Flux.Losses.logitcrossentropy       # loss function
_finaliser = x -> x                         # finaliser function
```

```{julia}
# JEM parameters:
𝒟x = Normal()
𝒟y = Categorical(ones(output_dim) ./ output_dim)
sampler = ConditionalSampler(
    𝒟x, 𝒟y, 
    input_size=(input_dim,), 
    batch_size=10,
)
α = [1.0,1.0,1e-1]      # penalty strengths
```

```{julia}
# Simple MLP:
mlp = NeuralNetworkClassifier(
    builder=builder, 
    epochs=epochs,
    batch_size=batch_size,
    finaliser=_finaliser,
    loss=_loss,
)

# Deep Ensemble:
mlp_ens = EnsembleModel(model=mlp, n=n_ens)

# Joint Energy Model:
jem = JointEnergyClassifier(
    sampler;
    builder=builder,
    epochs=epochs,
    batch_size=batch_size,
    finaliser=_finaliser,
    loss=_loss,
    jem_training_params=(
        α=α,verbosity=10,
    ),
    sampling_steps=30,
)

# JEM with adversarial training:
jem_adv = deepcopy(jem)
# jem_adv.adv_training = true

# Deep Ensemble of Joint Energy Models:
jem_ens = EnsembleModel(model=jem, n=n_ens)

# Deep Ensemble of Joint Energy Models with adversarial training:
# jem_ens_plus = EnsembleModel(model=jem_adv, n=n_ens)

# Dictionary of models:
models = Dict(
    "MLP" => mlp,
    "MLP Ensemble" => mlp_ens,
    "JEM" => jem,
    "JEM Ensemble" => jem_ens,
    # "JEM Ensemble+" => jem_ens_plus,
)
```


```{julia}
# Train models:
function _train(model, X=X, y=labels; cov=.95, method=:simple_inductive, mod_name="model")
    conf_model = conformal_model(model; method=method, coverage=cov)
    mach = machine(conf_model, X, y)
    @info "Begin training $mod_name."
    fit!(mach)
    @info "Finished training $mod_name."
    M = ECCCo.ConformalModel(mach.model, mach.fitresult)
    return M
end
if _retrain
    model_dict = Dict(mod_name => _train(mod; mod_name=mod_name) for (mod_name, mod) in models)
    Serialization.serialize(joinpath(output_path,"cal_housing_models.jls"), model_dict)
else
    model_dict = Serialization.deserialize(joinpath(output_path,"cal_housing_models.jls"))
end
```

```{julia}
# Evaluate models:

measure = Dict(
    :f1score => multiclass_f1score, 
    :acc => accuracy, 
    :precision => multiclass_precision
)
model_performance = DataFrame()
for (mod_name, mod) in model_dict
    # Test performance:
    _perf = CounterfactualExplanations.Models.model_evaluation(mod, test_data, measure=collect(values(measure)))
    _perf = DataFrame([[p] for p in _perf], collect(keys(measure)))
    _perf.mod_name .= mod_name
    model_performance = vcat(model_performance, _perf)
end
Serialization.serialize(joinpath(output_path,"cal_housing_model_performance.jls"), model_performance)
CSV.write(joinpath(output_path, "cal_housing_model_performance.csv"), model_performance)
model_performance
```

## Benchmark

```{julia}
λ₁ = 0.25
λ₂ = 0.75
λ₃ = 0.75
Λ = [λ₁, λ₂, λ₃]

# Benchmark generators:
generator_dict = Dict(
    "Wachter" => WachterGenerator(),
    "REVISE" => REVISEGenerator(),
    "Schut" => GreedyGenerator(),
    "ECCCo" => ECCCoGenerator(λ=Λ),
)
```

```{julia}
# Measures:
measures = [
    CounterfactualExplanations.distance,
    ECCCo.distance_from_energy,
    ECCCo.distance_from_targets,
    CounterfactualExplanations.Evaluation.validity,
    CounterfactualExplanations.Evaluation.redundancy,
    ECCCo.set_size_penalty
]

bmks = []
for target in sort(unique(labels))
    for factual in sort(unique(labels))
        if factual == target
            continue
        end
        bmk = benchmark(
            counterfactual_data; 
            models=model_dict, 
            generators=generator_dict, 
            measure=measures,
            suppress_training=true, dataname="California Housing",
            n_individuals=10,
            target=target, factual=factual,
            initialization=:identity,
            converge_when=:generator_conditions,
        )
        push!(bmks, bmk)
    end
end
bmk = reduce(vcat, bmks)
CSV.write(joinpath(output_path, "cal_housing_benchmark.csv"), bmk())
```

```{julia}
df = @chain bmk() begin
    @mutate(variable = ifelse.(variable .== "distance_from_energy", "Non-Conformity", variable))
    @mutate(variable = ifelse.(variable .== "distance_from_targets", "Implausibility", variable))
    @mutate(variable = ifelse.(variable .== "distance", "Cost", variable))
    @mutate(variable = ifelse.(variable .== "redundancy", "Redundancy", variable))
    @mutate(variable = ifelse.(variable .== "Validity", "Validity", variable))
end
plt = AlgebraOfGraphics.data(df) * visual(BoxPlot) * 
    mapping(:generator, :value, row=:variable, col=:model, color=:generator)
plt = draw(
    plt, axis=(xlabel="", xticksvisible=false, xticklabelsvisible=false, width=150, height=120), 
    facet=(; linkyaxes=:none)
)   
display(plt)
save(joinpath(output_images_path, "cal_housing_benchmark.png"), plt, px_per_unit=5)
```