fix

3 files changed, 80 insertions, 70 deletions

diff --git a/training/dtrain/dtrain.cc b/training/dtrain/dtrain.cc
index 441e2cd7..0a27a068 100644
--- a/training/dtrain/dtrain.cc
+++ b/training/dtrain/dtrain.cc
@@ -414,6 +414,12 @@ main(int argc, char** argv)
 
       score_t kbest_loss_first, kbest_loss_last = 0.0;
 
+      for (vector<pair<ScoredHyp,ScoredHyp> >::iterator it = pairs.begin();
+           it != pairs.end(); it++) {
+        score_t model_diff = it->first.model - it->second.model;
+        kbest_loss_first += max(0.0, -1.0 * model_diff);
+      }
+
       for (int ki=0; ki < repeat; ki++) {
 
       score_t kbest_loss = 0.0; // test-k-best
@@ -520,21 +526,22 @@ main(int argc, char** argv)
         }
       }
 
-      if (ki==0) kbest_loss_first = kbest_loss;
       if (ki==repeat-1) { // done
         kbest_loss_last = kbest_loss;
-        score_t best_score = -1.;
-        score_t best_model = -std::numeric_limits<score_t>::max();
-        unsigned best_idx;
-        for (unsigned i=0; i < samples->size(); i++) {
-          score_t s = lambdas.dot((*samples)[i].f);
-          if (s > best_model) {
-            best_idx = i;
-            best_model = s;
+        if (repeat > 1) {
+          score_t best_score = -1.;
+          score_t best_model = -std::numeric_limits<score_t>::max();
+          unsigned best_idx;
+          for (unsigned i=0; i < samples->size(); i++) {
+            score_t s = lambdas.dot((*samples)[i].f);
+            if (s > best_model) {
+              best_idx = i;
+              best_model = s;
+            }
           }
+          score_sum += (*samples)[best_idx].score;
+          model_sum += best_model;
         }
-        score_sum += (*samples)[best_idx].score;
-        model_sum += best_model;
       }
     } // repeat
 
@@ -588,15 +595,14 @@ main(int argc, char** argv)
     cerr << _p << " (" << model_diff << ")" << endl;
     cerr << "           avg # pairs: ";
     cerr << _np << npairs/(float)in_sz << endl;
-    cerr << "     avg # margin viol: ";
-    cerr << margin_violations/(float)in_sz << endl;
     cerr << "        avg # rank err: ";
     cerr << rank_errors/(float)in_sz;
     if (faster_perceptron) cerr << " (meaningless)";
     cerr << endl;
+    cerr << "     avg # margin viol: ";
+    cerr << margin_violations/(float)in_sz << endl;
     if (batch) cerr << "            batch loss: " << batch_loss << endl;
-    if (repeat > 1) cerr << "       k-best loss imp: " << ((float)kbest_loss_improve/in_sz)*100 << "%" << endl;
-
+    cerr << "       k-best loss imp: " << ((float)kbest_loss_improve/in_sz)*100 << "%" << endl;
     cerr << "    non0 feature count: " <<  nonz << endl;
     cerr << "           avg list sz: " << list_sz/(float)in_sz << endl;
     cerr << "           avg f count: " << f_count/(float)list_sz << endl;
diff --git a/training/dtrain/examples/standard/dtrain.ini b/training/dtrain/examples/standard/dtrain.ini
index ef022469..fc83f08e 100644
--- a/training/dtrain/examples/standard/dtrain.ini
+++ b/training/dtrain/examples/standard/dtrain.ini
@@ -15,7 +15,7 @@ epochs=3                 # run over input 3 times
 k=100                    # use 100best lists
 N=4                      # optimize (approx) BLEU4
 scorer=fixed_stupid_bleu # use 'stupid' BLEU+1
-learning_rate=0.0001     # learning rate, don't care if gamma=0 (perceptron) and loss_margin=0 (not margin perceptron)
+learning_rate=0.1        # learning rate, don't care if gamma=0 (perceptron) and loss_margin=0 (not margin perceptron)
 gamma=0                  # use SVM reg
 sample_from=kbest        # use kbest lists (as opposed to forest)
 filter=uniq              # only unique entries in kbest (surface form)
diff --git a/training/dtrain/examples/standard/expected-output b/training/dtrain/examples/standard/expected-output
index a35bbe6f..75f47337 100644
--- a/training/dtrain/examples/standard/expected-output
+++ b/training/dtrain/examples/standard/expected-output
@@ -4,17 +4,18 @@ Reading ./nc-wmt11.en.srilm.gz
 ----5---10---15---20---25---30---35---40---45---50---55---60---65---70---75---80---85---90---95--100
 ****************************************************************************************************
   Example feature: Shape_S00000_T00000
-Seeding random number sequence to 4049211323
+Seeding random number sequence to 3751911392
 
 dtrain
 Parameters:
                        k 100
                        N 4
                        T 3
+                   batch 0
                   scorer 'fixed_stupid_bleu'
              sample from 'kbest'
                   filter 'uniq'
-           learning rate 1
+           learning rate 0.1
                    gamma 0
              loss margin 0
        faster perceptron 1
@@ -25,9 +26,9 @@ Parameters:
                   l1 reg 0 'none'
                     pclr no
                max pairs 4294967295
+                  repeat 1
                 cdec cfg './cdec.ini'
-                   input './nc-wmt11.de.gz'
-                    refs './nc-wmt11.en.gz'
+                   input './nc-wmt11.gz'
                   output '-'
               stop_after 10
 (a dot represents 10 inputs)
@@ -35,25 +36,26 @@ Iteration #1 of 3.
  . 10
 Stopping after 10 input sentences.
 WEIGHTS
-              Glue = -1100
-       WordPenalty = -82.082
-     LanguageModel = -3199.1
- LanguageModel_OOV = -192
-     PhraseModel_0 = +3128.2
-     PhraseModel_1 = -1610.2
-     PhraseModel_2 = -4336.5
-     PhraseModel_3 = +2910.3
-     PhraseModel_4 = +2523.2
-     PhraseModel_5 = +506
-     PhraseModel_6 = +1467
-       PassThrough = -387
+              Glue = -110
+       WordPenalty = -8.2082
+     LanguageModel = -319.91
+ LanguageModel_OOV = -19.2
+     PhraseModel_0 = +312.82
+     PhraseModel_1 = -161.02
+     PhraseModel_2 = -433.65
+     PhraseModel_3 = +291.03
+     PhraseModel_4 = +252.32
+     PhraseModel_5 = +50.6
+     PhraseModel_6 = +146.7
+       PassThrough = -38.7
         ---
        1best avg score: 0.16966 (+0.16966)
- 1best avg model score: 2.9874e+05 (+2.9874e+05)
-           avg # pairs: 906.3 (meaningless)
-        avg # rank err: 906.3
+ 1best avg model score: 29874 (+29874)
+           avg # pairs: 906.3
+        avg # rank err: 0 (meaningless)
      avg # margin viol: 0
-    non0 feature count: 825
+       k-best loss imp: 100%
+    non0 feature count: 832
            avg list sz: 91.3
            avg f count: 139.77
 (time 0.35 min, 2.1 s/S)
@@ -61,25 +63,26 @@ WEIGHTS
 Iteration #2 of 3.
  . 10
 WEIGHTS
-              Glue = -1221
-       WordPenalty = +836.89
-     LanguageModel = +2332.3
- LanguageModel_OOV = -1451
-     PhraseModel_0 = +1507.2
-     PhraseModel_1 = -2728.4
-     PhraseModel_2 = -4183.6
-     PhraseModel_3 = +1816.3
-     PhraseModel_4 = -2894.7
-     PhraseModel_5 = +1403
-     PhraseModel_6 = +35
-       PassThrough = -1097
+              Glue = -122.1
+       WordPenalty = +83.689
+     LanguageModel = +233.23
+ LanguageModel_OOV = -145.1
+     PhraseModel_0 = +150.72
+     PhraseModel_1 = -272.84
+     PhraseModel_2 = -418.36
+     PhraseModel_3 = +181.63
+     PhraseModel_4 = -289.47
+     PhraseModel_5 = +140.3
+     PhraseModel_6 = +3.5
+       PassThrough = -109.7
         ---
        1best avg score: 0.17399 (+0.004325)
- 1best avg model score: 49369 (-2.4937e+05)
-           avg # pairs: 662.4 (meaningless)
-        avg # rank err: 662.4
+ 1best avg model score: 4936.9 (-24937)
+           avg # pairs: 662.4
+        avg # rank err: 0 (meaningless)
      avg # margin viol: 0
-    non0 feature count: 1235
+       k-best loss imp: 100%
+    non0 feature count: 1240
            avg list sz: 91.3
            avg f count: 125.11
 (time 0.27 min, 1.6 s/S)
@@ -87,32 +90,33 @@ WEIGHTS
 Iteration #3 of 3.
  . 10
 WEIGHTS
-              Glue = -1574
-       WordPenalty = -17.372
-     LanguageModel = +6861.8
- LanguageModel_OOV = -3997
-     PhraseModel_0 = -398.76
-     PhraseModel_1 = -3419.6
-     PhraseModel_2 = -3186.7
-     PhraseModel_3 = +1050.8
-     PhraseModel_4 = -2902.7
-     PhraseModel_5 = -486
-     PhraseModel_6 = -436
-       PassThrough = -2985
+              Glue = -157.4
+       WordPenalty = -1.7372
+     LanguageModel = +686.18
+ LanguageModel_OOV = -399.7
+     PhraseModel_0 = -39.876
+     PhraseModel_1 = -341.96
+     PhraseModel_2 = -318.67
+     PhraseModel_3 = +105.08
+     PhraseModel_4 = -290.27
+     PhraseModel_5 = -48.6
+     PhraseModel_6 = -43.6
+       PassThrough = -298.5
         ---
        1best avg score: 0.30742 (+0.13343)
- 1best avg model score: -1.5393e+05 (-2.0329e+05)
-           avg # pairs: 623.8 (meaningless)
-        avg # rank err: 623.8
+ 1best avg model score: -15393 (-20329)
+           avg # pairs: 623.8
+        avg # rank err: 0 (meaningless)
      avg # margin viol: 0
-    non0 feature count: 1770
+       k-best loss imp: 100%
+    non0 feature count: 1776
            avg list sz: 91.3
            avg f count: 118.58
-(time 0.25 min, 1.5 s/S)
+(time 0.28 min, 1.7 s/S)
 
 Writing weights file to '-' ...
 done
 
 ---
 Best iteration: 3 [SCORE 'fixed_stupid_bleu'=0.30742].
-This took 0.86667 min.
+This took 0.9 min.