loss margin cfg, XYX improved, smooth bleu variant

2 files changed, 74 insertions, 72 deletions

diff --git a/dtrain/test/example/dtrain.ini b/dtrain/test/example/dtrain.ini
index e43d6b34..c8ac7c3f 100644
--- a/dtrain/test/example/dtrain.ini
+++ b/dtrain/test/example/dtrain.ini
@@ -5,7 +5,7 @@ decoder_config=test/example/cdec.ini # config for cdec
 # weights for these features will be printed on each iteration
 print_weights=Glue WordPenalty LanguageModel LanguageModel_OOV PhraseModel_0 PhraseModel_1 PhraseModel_2 PhraseModel_3 PhraseModel_4 PhraseModel_5 PhraseModel_6 PassThrough
 tmp=/tmp
-stop_after=20 # stop epoch after 20 inputs
+stop_after=10 # stop epoch after 10 inputs
 
 # interesting stuff
 epochs=3                # run over input 3 times
@@ -19,3 +19,4 @@ filter=uniq             # only unique entries in kbest (surface form)
 pair_sampling=XYX
 hi_lo=0.1               # 10 vs 80 vs 10 and 80 vs 10 here
 pair_threshold=0        # minimum distance in BLEU (this will still only use pairs with diff > 0)
+loss_margin=0
diff --git a/dtrain/test/example/expected-output b/dtrain/test/example/expected-output
index 08733dd4..25d2c069 100644
--- a/dtrain/test/example/expected-output
+++ b/dtrain/test/example/expected-output
@@ -15,7 +15,7 @@ State is 0 bytes for feature RuleNgramFeatures
 feature: RuleShape (no config parameters)
   Example feature: Shape_S00000_T00000
 State is 0 bytes for feature RuleShape
-Seeding random number sequence to 380245307
+Seeding random number sequence to 1072059181
 
 dtrain
 Parameters:
@@ -27,6 +27,7 @@ Parameters:
                   filter 'uniq'
            learning rate 0.0001
                    gamma 0
+             loss margin 0
                    pairs 'XYX'
                    hi lo 0.1
           pair threshold 0
@@ -35,90 +36,90 @@ Parameters:
                 cdec cfg 'test/example/cdec.ini'
                    input 'test/example/nc-wmt11.1k.gz'
                   output '-'
-              stop_after 20
+              stop_after 10
 (a dot represents 10 inputs)
 Iteration #1 of 3.
- .. 20
-Stopping after 20 input sentences.
+ . 10
+Stopping after 10 input sentences.
 WEIGHTS
-              Glue = -0.1015
-       WordPenalty = -0.0152
-     LanguageModel = +0.21493
- LanguageModel_OOV = -0.3257
-     PhraseModel_0 = -0.050844
-     PhraseModel_1 = +0.25074
-     PhraseModel_2 = +0.27944
-     PhraseModel_3 = -0.038384
-     PhraseModel_4 = -0.12041
-     PhraseModel_5 = +0.1047
-     PhraseModel_6 = -0.1289
-       PassThrough = -0.3094
+              Glue = -0.0293
+       WordPenalty = +0.049075
+     LanguageModel = +0.24345
+ LanguageModel_OOV = -0.2029
+     PhraseModel_0 = +0.0084102
+     PhraseModel_1 = +0.021729
+     PhraseModel_2 = +0.014922
+     PhraseModel_3 = +0.104
+     PhraseModel_4 = -0.14308
+     PhraseModel_5 = +0.0247
+     PhraseModel_6 = -0.012
+       PassThrough = -0.2161
         ---
-       1best avg score: 0.17508 (+0.17508)
- 1best avg model score: -1.2392 (-1.2392)
-           avg # pairs: 1329.8
-        avg # rank err: 649.1
-     avg # margin viol: 677.5
-    non0 feature count: 874
-           avg list sz: 88.6
-           avg f count: 85.643
-(time 0.25 min, 0.75 s/S)
+       1best avg score: 0.16872 (+0.16872)
+ 1best avg model score: -1.8276 (-1.8276)
+           avg # pairs: 1121.1
+        avg # rank err: 555.6
+     avg # margin viol: 0
+    non0 feature count: 277
+           avg list sz: 77.2
+           avg f count: 90.96
+(time 0.1 min, 0.6 s/S)
 
 Iteration #2 of 3.
- .. 20
+ . 10
 WEIGHTS
-              Glue = -0.0792
-       WordPenalty = -0.056198
-     LanguageModel = +0.31038
- LanguageModel_OOV = -0.4011
-     PhraseModel_0 = +0.072188
-     PhraseModel_1 = +0.11473
-     PhraseModel_2 = +0.049774
-     PhraseModel_3 = -0.18448
-     PhraseModel_4 = -0.12092
-     PhraseModel_5 = +0.1599
-     PhraseModel_6 = -0.0606
-       PassThrough = -0.3848
+              Glue = -0.3526
+       WordPenalty = +0.067576
+     LanguageModel = +1.155
+ LanguageModel_OOV = -0.2728
+     PhraseModel_0 = -0.025529
+     PhraseModel_1 = +0.095869
+     PhraseModel_2 = +0.094567
+     PhraseModel_3 = +0.12482
+     PhraseModel_4 = -0.36533
+     PhraseModel_5 = +0.1068
+     PhraseModel_6 = -0.1517
+       PassThrough = -0.286
         ---
-       1best avg score: 0.24015 (+0.065075)
- 1best avg model score: -10.131 (-8.8914)
-           avg # pairs: 1324.7
-        avg # rank err: 558.65
-     avg # margin viol: 752.85
-    non0 feature count: 1236
-           avg list sz: 84.9
-           avg f count: 88.306
-(time 0.22 min, 0.65 s/S)
+       1best avg score: 0.18394 (+0.015221)
+ 1best avg model score: 3.205 (+5.0326)
+           avg # pairs: 1168.3
+        avg # rank err: 594.8
+     avg # margin viol: 0
+    non0 feature count: 543
+           avg list sz: 77.5
+           avg f count: 85.916
+(time 0.083 min, 0.5 s/S)
 
 Iteration #3 of 3.
- .. 20
+ . 10
 WEIGHTS
-              Glue = -0.051
-       WordPenalty = -0.077956
-     LanguageModel = +0.33699
- LanguageModel_OOV = -0.4726
-     PhraseModel_0 = +0.040228
-     PhraseModel_1 = +0.18
-     PhraseModel_2 = +0.15618
-     PhraseModel_3 = -0.098908
-     PhraseModel_4 = -0.036555
-     PhraseModel_5 = +0.1619
-     PhraseModel_6 = +0.0078
-       PassThrough = -0.4563
+              Glue = -0.392
+       WordPenalty = +0.071963
+     LanguageModel = +0.81266
+ LanguageModel_OOV = -0.4177
+     PhraseModel_0 = -0.2649
+     PhraseModel_1 = -0.17931
+     PhraseModel_2 = +0.038261
+     PhraseModel_3 = +0.20261
+     PhraseModel_4 = -0.42621
+     PhraseModel_5 = +0.3198
+     PhraseModel_6 = -0.1437
+       PassThrough = -0.4309
         ---
-       1best avg score: 0.25527 (+0.015113)
- 1best avg model score: -13.906 (-3.7756)
-           avg # pairs: 1356.3
-        avg # rank err: 562.1
-     avg # margin viol: 757.35
-    non0 feature count: 1482
-           avg list sz: 86.65
-           avg f count: 87.475
-(time 0.23 min, 0.7 s/S)
+       1best avg score: 0.2962 (+0.11225)
+ 1best avg model score: -36.274 (-39.479)
+           avg # pairs: 1109.6
+        avg # rank err: 515.9
+     avg # margin viol: 0
+    non0 feature count: 741
+           avg list sz: 77
+           avg f count: 88.982
+(time 0.083 min, 0.5 s/S)
 
 Writing weights file to '-' ...
 done
 
 ---
-Best iteration: 3 [SCORE 'stupid_bleu'=0.25527].
-This took 0.7 min.
+Best iteration: 3 [SCORE 'stupid_bleu'=0.2962].
+This took 0.26667 min.