neulab · msperber · Jan 11, 2019 · Jan 9, 2019 · Jan 9, 2019 · Jan 9, 2019
diff --git a/docs/experiment_config_files.rst b/docs/experiment_config_files.rst
@@ -252,56 +252,48 @@ Scoring N-best lists
 .. literalinclude:: examples/15_score.yaml
     :language: yaml
 
-Transformer
-~~~~~~~~~~~
-
-(this is currently broken)
-
-.. literalinclude:: examples/16_transformer.yaml
-    :language: yaml
-
 Ensembling
 ~~~~~~~~~~
 
-.. literalinclude:: examples/17_ensembling.yaml
+.. literalinclude:: examples/16_ensembling.yaml
     :language: yaml
 
 Minimum risk training
 ~~~~~~~~~~~~~~~~~~~~~
 
-.. literalinclude:: examples/18_minrisk.yaml
+.. literalinclude:: examples/17_minrisk.yaml
     :language: yaml
 
 Biased Lexicon
 ~~~~~~~~~~~~~~
 
 (this is currently broken)
 
-.. literalinclude:: examples/19_lexiconbias.yaml
+.. literalinclude:: examples/18_lexiconbias.yaml
     :language: yaml
 
 Subword Sampling
 ~~~~~~~~~~~~~~~~
 
-.. literalinclude:: examples/20_subword_sample.yaml
+.. literalinclude:: examples/19_subword_sample.yaml
     :language: yaml
 
 Self Attention
 ~~~~~~~~~~~~~~
 
-.. literalinclude:: examples/21_self_attention.yaml
+.. literalinclude:: examples/20_self_attention.yaml
     :language: yaml
 
 Char Segment
 ~~~~~~~~~~~~
 
-.. literalinclude:: examples/22_char_segment.yaml
+.. literalinclude:: examples/21_char_segment.yaml
     :language: yaml
 
 Switchout
 ~~~~~~~~~
 
-.. literalinclude:: examples/23_switchout.yaml
+.. literalinclude:: examples/22_switchout.yaml
     :language: yaml
 
 Autobatching

diff --git a/examples/17_ensembling.yaml → examples/16_ensembling.yaml b/examples/17_ensembling.yaml → examples/16_ensembling.yaml
diff --git a/examples/16_transformer.yaml b/examples/16_transformer.yaml
diff --git a/examples/18_minrisk.yaml → examples/17_minrisk.yaml b/examples/18_minrisk.yaml → examples/17_minrisk.yaml
@@ -59,7 +59,7 @@ exp2-finetune-minrisk: !LoadSerialized
   - path: train.loss_calculator
     val: !MinRiskLoss
       alpha: 0.005
-  - path: model.search_strategy
+  - path: model.inference.search_strategy
     val: !SamplingSearch
       sample_size: 10
       max_len: 50

diff --git a/examples/19_lexiconbias.yaml → examples/18_lexiconbias.yaml b/examples/19_lexiconbias.yaml → examples/18_lexiconbias.yaml
diff --git a/examples/20_subword_sample.yaml → examples/19_subword_sample.yaml b/examples/20_subword_sample.yaml → examples/19_subword_sample.yaml
diff --git a/examples/21_self_attention.yaml → examples/20_self_attention.yaml b/examples/21_self_attention.yaml → examples/20_self_attention.yaml
diff --git a/examples/22_char_segment.yaml → examples/21_char_segment.yaml b/examples/22_char_segment.yaml → examples/21_char_segment.yaml
diff --git a/examples/23_switchout.yaml → examples/22_switchout.yaml b/examples/23_switchout.yaml → examples/22_switchout.yaml
diff --git a/examples/24_autobatch.yaml → examples/23_autobatch.yaml b/examples/24_autobatch.yaml → examples/23_autobatch.yaml
diff --git a/xnmt/input_readers.py b/xnmt/input_readers.py
@@ -216,7 +216,7 @@ def read_sent(self, line: str, idx: numbers.Integral) -> sent.SimpleSentence:
       words = self.subword_model.SampleEncodeAsPieces(line.strip(), self.l, self.alpha)
     else:
       words = self.subword_model.EncodeAsPieces(line.strip())
-    words = [w.decode('utf-8') for w in words]
+    #words = [w.decode('utf-8') for w in words]
     return sent.SimpleSentence(idx=idx,
                                words=[self.vocab.convert(word) for word in words] + [self.vocab.convert(vocabs.Vocab.ES_STR)],
                                vocab=self.vocab,

diff --git a/xnmt/modelparts/scorers.py b/xnmt/modelparts/scorers.py
@@ -312,4 +312,5 @@ def calc_log_probs(self, x: dy.Expression) -> dy.Expression:
       return dy.log_softmax(self.calc_scores(x))
 
   def can_loss_be_derived_from_scores(self):
+    # TODO: this line is broken
     return self.lexicon_type == 'bias' and super().is_modifying_softmax_layer()
diff --git a/xnmt/models/translators.py b/xnmt/models/translators.py
@@ -476,16 +476,7 @@ def set_trg_vocab(self, trg_vocab: Optional[vocabs.Vocab] = None) -> None:
     self._proxy.set_trg_vocab(trg_vocab=trg_vocab)
 
   def calc_nll(self, src: Union[batchers.Batch, sent.Sentence], trg: Union[batchers.Batch, sent.Sentence]) -> dy.Expression:
-    sub_losses = collections.defaultdict(list)
-    for model in self.models:
-      for loss_name, loss in model.calc_nll(src, trg).expr_factors.items():
-        sub_losses[loss_name].append(loss)
-    model_loss = losses.FactoredLossExpr()
-    for loss_name, losslist in sub_losses.items():
-      # TODO: dy.average(losslist)  _or_  dy.esum(losslist) / len(self.models) ?
-      #       -- might not be the same if not all models return all losses
-      model_loss.add_loss(loss_name, dy.average(losslist))
-    return model_loss
+    return dy.average([model.calc_nll(src, trg) for model in self.models])
 
   def generate(self,
                src: batchers.Batch,