attempting a run without normalising digits to between -1 and 1

notebooks/mnist.qmd

@@ -137,15 +137,16 @@ savefig(plt, joinpath(output_images_path, "surrogate_gone_wrong.png"))
 ```{julia}
 function pre_process(x; noise::Float32=0.03f0)
     ϵ = Float32.(randn(size(x)) * noise)
-    x = @.(2 * x - 1) .+ ϵ
+    # x = @.(2 * x - 1) 
+    x += ϵ
     return x
 end
 ```
 
 ```{julia}
 # Hyper:
-_retrain = false
-_regen = false
+_retrain = true
+_regen = true
 
 # Data:
 n_obs = 10000
@@ -179,7 +180,7 @@ _finaliser = x -> x                         # finaliser function
 
 ```{julia}
 # JEM parameters:
-𝒟x = Uniform(-1,1)
+𝒟x = Uniform(0,1)
 𝒟y = Categorical(ones(output_dim) ./ output_dim)
 sampler = ConditionalSampler(
     𝒟x, 𝒟y, 
@@ -306,7 +307,7 @@ model_performance = DataFrame()
 for (mod_name, mod) in model_dict
     # Test performance:
     test_data = load_mnist_test()
-    test_data.X = pre_process.(test_data.X, noise=0.0f0)
+    # test_data.X = pre_process.(test_data.X, noise=0.0f0)
     _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

paper/paper.tex

@@ -310,6 +310,7 @@ As noted by \citet{guidotti2022counterfactual}, these distance-based measures ar
   \item ECCCo is sensitive to optimizer (Adam works well), learning rate and distance metric (l1 currently only one that works)
   \item SGLD takes time 
   \item REVISE has benefit of lower dimensional space
+  \item For MNIST it seems that ECCCo is better at reducing pixel values than increasing them (better at erasing than writing)
 \end{itemize}
 
 \section{Discussion}