notebooks/mnist.qmd

@@ -7,7 +7,7 @@ eval(setup_notebooks)
 
 ```{julia}
 # Data:
-n_obs = 1000
+n_obs = 10000
 counterfactual_data = load_mnist(n_obs)
 X, y = CounterfactualExplanations.DataPreprocessing.unpack_data(counterfactual_data)
 X = table(permutedims(X))
@@ -22,15 +22,16 @@ First, let's create a couple of image classifier architectures:
 ```{julia}
 # Model parameters:
 epochs = 100
-batch_size = Int(round(n_obs/10))
+batch_size = minimum([Int(round(n_obs/10)), 100])
 n_hidden = 32
-activation = Flux.swish
+activation = Flux.relu
 builder = MLJFlux.@builder Flux.Chain(
     Dense(n_in, n_hidden),
     BatchNorm(n_hidden, activation),
     Dense(n_hidden, n_out),
     BatchNorm(n_out)
 )
+# builder = MLJFlux.Short(n_hidden=n_hidden, dropout=0.2, σ=activation)
 # builder = MLJFlux.MLP(
 #     hidden=(
 #         n_hidden,
@@ -39,7 +40,7 @@ builder = MLJFlux.@builder Flux.Chain(
 #     ), 
 #     σ=activation
 # )
-α = [1.0,1.0,1e-1]
+α = [1.0,1.0,5e-1]
 
 # Simple MLP:
 mlp = NeuralNetworkClassifier(
@@ -58,7 +59,11 @@ jem = JointEnergyClassifier(
     batch_size=batch_size,
     finaliser=x -> x,
     loss=Flux.Losses.logitcrossentropy,
-    jem_training_params=(α=α,verbosity=10,),
+    jem_training_params=(
+        α=α,verbosity=10,
+        use_gen_loss=false,
+        use_reg_loss=false,
+    ),
     sampling_steps=20,
     epochs=epochs,
 )
@@ -106,12 +111,12 @@ println("F1 score (test): $(round(f1,digits=3))")
 Random.seed!(1234)
 
 # Set up search:
-factual_label = 2
+factual_label = 9
 x = reshape(counterfactual_data.X[:,rand(findall(predict_label(M, counterfactual_data).==factual_label))],input_dim,1)
-target = 0
+target = 4
 factual = predict_label(M, counterfactual_data, x)[1]
 γ = 0.5
-T = 100
+T = 1
 
 # Generate counterfactual using generic generator:
 generator = GenericGenerator()