update

Author: BatyLeo

scripts/main.jl

@@ -4,9 +4,9 @@ using MLUtils
 using Statistics
 using Plots
 
-res = fyl_train_model(ArgmaxBenchmark(); epochs=10_000)
-plot(res.validation_loss[100:end]; label="Validation Loss")
-plot!(res.training_loss[100:end]; label="Training Loss")
+res = fyl_train_model(StochasticVehicleSchedulingBenchmark(); epochs=100)
+plot(res.validation_loss; label="Validation Loss")
+plot!(res.training_loss; label="Training Loss")
 
 baty_train_model(DynamicVehicleSchedulingBenchmark(; two_dimensional_features=false))
 DAgger_train_model(DynamicVehicleSchedulingBenchmark(; two_dimensional_features=false))

src/dagger.jl

@@ -67,9 +67,7 @@ end
 
 function DAgger_train_model(b::AbstractStochasticBenchmark{true}; kwargs...)
     dataset = generate_dataset(b, 30)
-    train_instances, validation_instances, test_instances = dataset[1:10],
-    dataset[11:20],
-    dataset[21:30]
+    train_instances, validation_instances, _ = splitobs(dataset; at=(0.3, 0.3, 0.4))
     train_environments = generate_environments(b, train_instances; seed=0)
     validation_environments = generate_environments(b, validation_instances)
     model = generate_statistical_model(b)

src/fyl.jl

@@ -3,6 +3,8 @@
 # TODO: Implement validation loss as a metric callback
 # TODO: batch training option
 # TODO: parallelize loss computation on validation set
+# TODO: have supervised learning training method, where fyl_train calls it, therefore we can easily test new supervised losses if needed
+# TODO: easier way to define and provide metrics
 
 function fyl_train_model!(
     model,
@@ -13,7 +15,7 @@ function fyl_train_model!(
     maximizer_kwargs=(sample -> (; instance=sample.info)),
     metrics_callbacks::NamedTuple=NamedTuple(),
 )
-    perturbed = PerturbedAdditive(maximizer; nb_samples=50, ε=1.0, threaded=true, seed=0)
+    perturbed = PerturbedAdditive(maximizer; nb_samples=50, ε=0.0, threaded=true, seed=0)
     loss = FenchelYoungLoss(perturbed)
 
     optimizer = Adam()
@@ -86,7 +88,7 @@ function fyl_train_model!(
 end
 
 function fyl_train_model(b::AbstractBenchmark; kwargs...)
-    dataset = generate_dataset(b, 100)
+    dataset = generate_dataset(b, 20)
     train_dataset, validation_dataset, _ = splitobs(dataset; at=(0.3, 0.3, 0.4))
     model = generate_statistical_model(b)
     maximizer = generate_maximizer(b)