correct clamping clamp

jaak-s · Jan 27, 2015 · ea944e2 · ea944e2
1 parent 6b69614
commit ea944e2
Show file tree

Hide file tree

Showing 2 changed files with 17 additions and 13 deletions.
diff --git a/src/BMRF.jl b/src/BMRF.jl
@@ -20,7 +20,7 @@ function BMRF(data::RelationData;
   roc_avg  = 0.0
   rmse_avg = 0.0
 
-  local probe_rat_all
+  local probe_rat_all, clamped_rat_all
 
   ## Gibbs sampling loop
   for i in 1 : burnin + psamples
@@ -57,14 +57,7 @@ function BMRF(data::RelationData;
     end
 
     # clamping maybe needed for MovieLens data
-    local probe_rat
-    if isempty(clamp)
-      probe_rat = pred(rel.test_vec, data.entities[2].model.sample, data.entities[1].model.sample, rel.mean_rating)
-    else
-      probe_rat = pred(rel.test_vec, data.entities[2].model.sample, data.entities[1].model.sample, rel.mean_rating)
-      probe_rat[ probe_rat .< clamp[1] ] = clamp[1]
-      probe_rat[ probe_rat .> clamp[2] ] = clamp[2]
-    end
+    probe_rat = pred(rel.test_vec, data.entities[2].model.sample, data.entities[1].model.sample, rel.mean_rating)
 
     if i > burnin
       probe_rat_all = (counter_prob*probe_rat_all + probe_rat)/(counter_prob+1)
@@ -80,9 +73,13 @@ function BMRF(data::RelationData;
     correct  = (rel.test_label .== (probe_rat_all .< class_cut) )
     err_avg  = mean(correct)
     err      = mean(rel.test_label .== (probe_rat .< class_cut))
-    rmse_avg = haveTest ? sqrt(mean( (rel.test_vec[:,3] - probe_rat_all) .^ 2 )) : NaN
-    rmse     = haveTest ? sqrt(mean( (rel.test_vec[:,3] - probe_rat) .^ 2 ))     : NaN
-    roc_avg  = haveTest ? AUC_ROC(rel.test_label, -vec(probe_rat_all))           : NaN
+
+    clamped_rat     = isempty(clamp) ?probe_rat     :makeClamped(probe_rat, clamp)
+    clamped_rat_all = isempty(clamp) ?probe_rat_all :makeClamped(probe_rat_all, clamp)
+
+    rmse_avg = haveTest ? sqrt(mean( (rel.test_vec[:,3] - clamped_rat_all) .^ 2 )) : NaN
+    rmse     = haveTest ? sqrt(mean( (rel.test_vec[:,3] - clamped_rat) .^ 2 ))     : NaN
+    roc_avg  = haveTest ? AUC_ROC(rel.test_label, -vec(probe_rat_all))             : NaN
     verbose && @printf("Iteration %d:\t avgAcc %6.4f Acc %6.4f | avgRMSE %6.4f | avgROC %6.4f | FU(%6.2f) FM(%6.2f) Fb(%6.2f) [%2.0fs]\n", i, err_avg, err, rmse_avg, roc_avg, vecnorm(data.entities[1].model.sample), vecnorm(data.entities[2].model.sample), vecnorm(data.entities[1].model.beta), time1 - time0)
   end
 
@@ -91,6 +88,6 @@ function BMRF(data::RelationData;
   result["RMSE"]        = rmse_avg
   result["accuracy"]    = err_avg
   result["ROC"]         = roc_avg
-  result["predictions"] = probe_rat_all
+  result["predictions"] = clamped_rat_all
   return result
 end
diff --git a/src/sampling.jl b/src/sampling.jl
@@ -2,6 +2,13 @@ function pred(probe_vec, sample_m, sample_u, mean_rating)
   sum(sample_m[probe_vec[:,2],:].*sample_u[probe_vec[:,1],:],2) + mean_rating
 end
 
+function makeClamped(x, clamp::Vector{Float64})
+  x2 = copy(x)
+  x2[x2 .< clamp[1]] = clamp[1]
+  x2[x2 .> clamp[2]] = clamp[2]
+  return x2
+end
+
 function ConditionalNormalWishart(U::Matrix{Float64}, mu::Vector{Float64}, kappa::Real, T::Matrix{Float64}, nu::Real)
   N = size(U, 1)
   Ū = mean(U,1)