more work on mnist

notebooks/mnist.qmd

@@ -320,7 +320,7 @@ model_performance
 ```{julia}
 # ECCCo:
 λ=[0.5,0.1,0.5]
-temp=0.1
+temp=0.5
 η=0.01
 
 # Generate counterfactuals using ECCCo generator:

paper/paper.tex

@@ -272,7 +272,7 @@ As noted by \citet{guidotti2022counterfactual}, these distance-based measures ar
   \item Coverage and temperature impacts CCE in somewhat unpredictable ways
   \item It seems that models that are not explicitly trained for generative task, still learn it implictly
   \item Batch size seems to impact quality of generated samples (at inference, but not so much during JEM training)
-  \item ECCCo is sensitive to optimizer (Adam works well), learning rate and distance metric (l1 works well)
+  \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
 \end{itemize}

src/generator.jl

@@ -28,7 +28,7 @@ function ECCCoGenerator(;
     function _set_size_penalty(ce::AbstractCounterfactualExplanation)
         return ECCCo.set_size_penalty(ce; κ=κ, temp=temp)
     end
-    _penalties = [Objectives.distance_l1, _set_size_penalty, ECCCo.distance_from_energy]
+    _penalties = [Objectives.distance_l2, _set_size_penalty, ECCCo.distance_from_energy]
     λ = λ isa AbstractFloat ? [0.0, λ, λ] : λ
     return Generator(; penalty=_penalties, λ=λ, opt=opt, kwargs...)
 end