Merge pull request #496 from neulab/revert-486-refactor-loss

Revert "Started refactoring loss calculation [WIP]"
neulab · Aug 2, 2018 · 9dd7f1c · 9dd7f1c
2 parents f21ddca + 53b3a96
commit 9dd7f1c
Show file tree

Hide file tree

Showing 43 changed files with 444 additions and 578 deletions.
diff --git a/test/config/assemble.yaml b/test/config/assemble.yaml
@@ -1,7 +1,5 @@
 # note: run as ``xnmt --settings=unittest <config>.yaml``
 exp1-pretrain1: !Experiment
-  exp_global: !ExpGlobal
-    default_layer_dim: 64
   model: !DefaultTranslator
     src_reader: !PlainTextReader
       vocab: !Vocab {vocab_file: examples/data/head.ja.vocab}
@@ -33,7 +31,6 @@ exp1-pretrain1: !Experiment
 exp2-pretrain2: !Experiment
   exp_global: !ExpGlobal
     model_file: 'test/tmp/{EXP}.mod'
-    default_layer_dim: 64
   model: !DefaultTranslator
     src_reader: !PlainTextReader
       vocab: !Vocab {vocab_file: examples/data/head.ja.vocab}
@@ -63,8 +60,6 @@ exp2-pretrain2: !Experiment
 
 
 exp3-assemble: !Experiment
-  exp_global: !ExpGlobal
-    default_layer_dim: 64
   model: !DefaultTranslator
     src_reader: !PlainTextReader
       vocab: !Vocab {vocab_file: examples/data/head.en.vocab}
@@ -97,8 +92,6 @@ exp3-assemble: !Experiment
 
 
 exp4-assemble: !Experiment
-  exp_global: !ExpGlobal
-    default_layer_dim: 64
   model: !DefaultTranslator
     src_reader: !PlainTextReader
       vocab: !Vocab {vocab_file: examples/data/head.en.vocab}

diff --git a/test/config/classifier.yaml b/test/config/classifier.yaml
@@ -1,7 +1,5 @@
 # This demonstrates using a sequence classifier model that takes a sequence as input and outputs a single class label.
 classifier: !Experiment
-  exp_global: !ExpGlobal
-    default_layer_dim: 64
   model: !SequenceClassifier
     src_reader: !PlainTextReader
       vocab: !Vocab {vocab_file: examples/data/head.en.vocab}

diff --git a/test/config/lm.yaml b/test/config/lm.yaml
@@ -1,10 +1,9 @@
 lm: !Experiment
-  exp_global: !ExpGlobal
-    default_layer_dim: 64
   model: !LanguageModel
     src_reader: !PlainTextReader
       vocab: !Vocab {vocab_file: examples/data/head.en.vocab}
-    src_embedder: !SimpleWordEmbedder {}
+    src_embedder: !SimpleWordEmbedder
+      emb_dim: 512
     rnn: !UniLSTMSeqTransducer
       layers: 1
     scorer: !Softmax

diff --git a/test/config/minrisk.yaml b/test/config/minrisk.yaml
@@ -7,9 +7,10 @@ minrisk: !Experiment
       vocab: !Vocab {vocab_file: examples/data/head.ja.vocab}
     trg_reader: !PlainTextReader
       vocab: !Vocab {vocab_file: examples/data/head.en.vocab}
+    search_strategy: !SamplingSearch
+      max_len: 10
+      sample_size: 10
   train: !SimpleTrainingRegimen
-    batcher: !SrcBatcher
-      batch_size: 1
     run_for_epochs: 2
     src_file: examples/data/head.ja
     trg_file: examples/data/head.en
@@ -24,9 +25,6 @@ minrisk: !Experiment
         ref_file: examples/data/head.en
     loss_calculator: !MinRiskLoss
       alpha: 0.005
-      search_strategy: !SamplingSearch
-        max_len: 10
-        sample_size: 10
   evaluate:
     - !AccuracyEvalTask
       eval_metrics: bleu

diff --git a/test/config/preproc.yaml b/test/config/preproc.yaml
@@ -1,6 +1,4 @@
 standard-preproc: !Experiment
-  exp_global: !ExpGlobal
-    default_layer_dim: 64
   preproc: !PreprocRunner
     overwrite: True
     tasks:

diff --git a/test/config/pretrained_embeddings.yaml b/test/config/pretrained_embeddings.yaml
@@ -1,7 +1,5 @@
 # An example for how to use pretrained embeddings
 pretrained_emb: !Experiment
-  exp_global: !ExpGlobal
-    default_layer_dim: 64
   model: !DefaultTranslator
     src_reader: !PlainTextReader
       vocab: !Vocab {vocab_file: examples/data/head.ja.vocab}

diff --git a/test/config/reinforce.yaml b/test/config/reinforce.yaml
@@ -5,10 +5,9 @@ reinforce: !Experiment
     src_reader: !PlainTextReader
       vocab: !Vocab {vocab_file: examples/data/head.ja.vocab}
     trg_reader: !PlainTextReader
-      vocab: !Vocab {vocab_file: examples/data/head.en.vocab} 
+      vocab: !Vocab {vocab_file: examples/data/head.en.vocab}
+    search_strategy: !SamplingSearch {}
   train: !SimpleTrainingRegimen
-    batcher: !SrcBatcher
-      batch_size: 1
     run_for_epochs: 2
     src_file: examples/data/head.ja
     trg_file: examples/data/head.en
@@ -23,7 +22,6 @@ reinforce: !Experiment
         ref_file: examples/data/head.en
     loss_calculator: !ReinforceLoss
       use_baseline: True
-      search_strategy: !SamplingSearch {}
   evaluate:
     - !AccuracyEvalTask
       eval_metrics: bleu

diff --git a/test/config/reload.yaml b/test/config/reload.yaml
@@ -1,6 +1,4 @@
 reload: !Experiment
-  exp_global: !ExpGlobal
-    default_layer_dim: 64
   model: !DefaultTranslator
     src_reader: !NpzReader
       transpose: True

diff --git a/test/config/reload_exception.yaml b/test/config/reload_exception.yaml
@@ -1,7 +1,5 @@
 # note: run as ``xnmt --settings=unittest <config>.yaml``
 reload: !Experiment
-  exp_global: !ExpGlobal
-    default_layer_dim: 64
   model: !DefaultTranslator
     src_reader: !NpzReader
       transpose: True

diff --git a/test/config/report.yaml b/test/config/report.yaml
@@ -1,6 +1,5 @@
 report: !Experiment
   exp_global: !ExpGlobal
-    default_layer_dim: 64
     compute_report: True
   model: !DefaultTranslator
     src_reader: !PlainTextReader

diff --git a/test/config/seg_debug.yaml b/test/config/seg_debug.yaml
@@ -1,6 +1,5 @@
 report: !Experiment
   exp_global: !ExpGlobal
-    default_layer_dim: 64
     compute_report: True
   model: !DefaultTranslator
     src_reader: !CharFromWordTextReader

diff --git a/test/config/seg_report.yaml b/test/config/seg_report.yaml
@@ -1,7 +1,6 @@
 report: !Experiment
   exp_global: !ExpGlobal
     compute_report: True
-    default_layer_dim: 64
   model: !DefaultTranslator
     src_reader: !CharFromWordTextReader
       vocab: !Vocab {vocab_file: examples/data/head.ja.charvocab}
@@ -14,13 +13,11 @@ report: !Experiment
       final_transducer: !BiLSTMSeqTransducer {}
       policy_learning: !PolicyGradient 
         z_normalization: True
+        sample: 2
   train: !SimpleTrainingRegimen
     run_for_epochs: 1
     src_file: examples/data/head.ja
     trg_file: examples/data/head.en
-    loss_calculator: !FeedbackLoss
-      child_loss: !MLELoss {}
-      repeat: 2
   evaluate:
     - !AccuracyEvalTask
       eval_metrics: bleu,wer

diff --git a/test/config/self_attentional_am.yaml b/test/config/self_attentional_am.yaml
@@ -1,7 +1,6 @@
 # note: run as ``xnmt --settings=unittest <config>.yaml``
 speech-self-att: !Experiment
   exp_global: !ExpGlobal
-    default_layer_dim: 64
     dropout: 0.2
   preproc: !PreprocRunner
     overwrite: False

diff --git a/test/config/seq_labeler.yaml b/test/config/seq_labeler.yaml
@@ -1,7 +1,5 @@
 seq-labeler: !Experiment
   # model architecture
-  exp_global: !ExpGlobal
-    default_layer_dim: 64
   model: !SeqLabeler
     src_reader: !PlainTextReader
       vocab: !Vocab {vocab_file: examples/data/head.en.vocab}

diff --git a/test/config/standard.yaml b/test/config/standard.yaml
@@ -7,6 +7,8 @@ standard: !Experiment
       vocab: !Vocab {vocab_file: examples/data/head.ja.vocab}
     trg_reader: !PlainTextReader
       vocab: !Vocab {vocab_file: examples/data/head.en.vocab}
+    search_strategy: !BeamSearch
+      beam_size: 5
   train: !SimpleTrainingRegimen
     run_for_epochs: 2
     src_file: examples/data/head.ja

diff --git a/test/test_beam_search.py b/test/test_beam_search.py
@@ -9,12 +9,12 @@
 from xnmt.modelparts.embedders import SimpleWordEmbedder
 from xnmt.input_readers import PlainTextReader
 from xnmt.transducers.recurrent import UniLSTMSeqTransducer, BiLSTMSeqTransducer
-from xnmt.loss_calculators import MLELoss
+from xnmt.loss_calculators import AutoRegressiveMLELoss
+from xnmt.param_collections import ParamManager
 from xnmt.modelparts.transforms import NonLinear
 from xnmt.modelparts.scorers import Softmax
 from xnmt.models.translators import DefaultTranslator
 from xnmt.search_strategies import BeamSearch, GreedySearch
-from xnmt.param_collections import ParamManager
 from xnmt.vocabs import Vocab
 
 class TestForcedDecodingOutputs(unittest.TestCase):
@@ -88,8 +88,9 @@ def setUp(self):
 
   def test_single(self):
     dy.renew_cg()
-    train_loss = self.model.calc_nll(src=self.src_data[0],
-                                     trg=self.trg_data[0]).value()
+    train_loss = self.model.calc_loss(src=self.src_data[0],
+                                      trg=self.trg_data[0],
+                                      loss_calculator=AutoRegressiveMLELoss()).value()
     dy.renew_cg()
     outputs = self.model.generate(batchers.mark_as_batch([self.src_data[0]]), [0], BeamSearch(beam_size=1),
                                   forced_trg_ids=batchers.mark_as_batch([self.trg_data[0]]))
@@ -127,8 +128,9 @@ def test_single(self):
     outputs = self.model.generate(batchers.mark_as_batch([self.src_data[0]]), [0], BeamSearch(),
                                   forced_trg_ids=batchers.mark_as_batch([self.trg_data[0]]))
     dy.renew_cg()
-    train_loss = self.model.calc_nll(src=self.src_data[0],
-                                     trg=outputs[0]).value()
+    train_loss = self.model.calc_loss(src=self.src_data[0],
+                                      trg=outputs[0],
+                                      loss_calculator=AutoRegressiveMLELoss()).value()
 
     self.assertAlmostEqual(-outputs[0].score, train_loss, places=4)
 

diff --git a/test/test_decoders.py b/test/test_decoders.py
@@ -8,7 +8,7 @@
 from xnmt.modelparts.decoders import AutoRegressiveDecoder
 from xnmt.modelparts.embedders import SimpleWordEmbedder
 from xnmt.input_readers import PlainTextReader
-from xnmt.loss_calculators import MLELoss
+from xnmt.loss_calculators import AutoRegressiveMLELoss
 from xnmt.transducers.recurrent import UniLSTMSeqTransducer, BiLSTMSeqTransducer
 from xnmt.param_collections import ParamManager
 from xnmt.modelparts.transforms import NonLinear
@@ -90,8 +90,9 @@ def setUp(self):
 
   def test_single(self):
     dy.renew_cg()
-    train_loss = self.model.calc_nll(src=self.src_data[0],
-                                     trg=self.trg_data[0]).value()
+    train_loss = self.model.calc_loss(src=self.src_data[0],
+                                      trg=self.trg_data[0],
+                                      loss_calculator=AutoRegressiveMLELoss()).value()
     dy.renew_cg()
     outputs = self.model.generate(batchers.mark_as_batch([self.src_data[0]]), [0], GreedySearch(),
                                   forced_trg_ids=batchers.mark_as_batch([self.trg_data[0]]))
@@ -132,8 +133,9 @@ def test_single(self):
     output_score = outputs[0].score
 
     dy.renew_cg()
-    train_loss = self.model.calc_nll(src=self.src_data[0],
-                                     trg=outputs[0]).value()
+    train_loss = self.model.calc_loss(src=self.src_data[0],
+                                      trg=outputs[0],
+                                      loss_calculator=AutoRegressiveMLELoss()).value()
 
     self.assertAlmostEqual(-output_score, train_loss, places=5)
 

diff --git a/test/test_evaluation.py b/test/test_evaluation.py
@@ -3,7 +3,7 @@
 
 from xnmt.eval import metrics
 from xnmt import events
-from xnmt.utils import has_cython
+from test.utils import has_cython
 from xnmt.vocabs import Vocab
 
 class TestBLEU(unittest.TestCase):
@@ -26,7 +26,7 @@ def test_bleu_1gram(self):
   def test_bleu_4gram_fast(self):
     bleu = metrics.FastBLEUEvaluator(ngram=4, smooth=1)
     exp_bleu = math.exp(math.log((3.0/5.0) * (2.0/5.0) * (1.0/4.0) * (1.0/3.0))/4.0)
-    act_bleu = bleu.evaluate_one_sent(self.ref_id, self.hyp_id)
+    act_bleu = bleu.evaluate(self.ref_id, self.hyp_id)
     self.assertEqual(act_bleu, exp_bleu)
 
 class TestGLEU(unittest.TestCase):

diff --git a/test/test_hyperparams.py b/test/test_hyperparams.py
@@ -1,6 +1,7 @@
 import unittest
 
 from xnmt import events
+
 from xnmt.hyper_params import *
 
 
@@ -48,7 +49,7 @@ def test_defined_sequence(self):
 #                                            transformer=self.tail_transformer)
 #    self.src_reader = PlainTextReader()
 #    self.trg_reader = PlainTextReader()
-#    self.loss_calculator = MLELoss()
+#    self.loss_calculator = AutoRegressiveMLELoss()
 #    self.segmenting_encoder = SegmentingSeqTransducer(
 #      embed_encoder = self.segment_embed_encoder_bilstm,
 #      segment_composer =  self.segment_composer,

diff --git a/test/test_run.py b/test/test_run.py
@@ -1,7 +1,7 @@
 import unittest
 import os, shutil
 
-from xnmt.utils import has_cython
+from test.utils import has_cython
 import xnmt.xnmt_run_experiments as run
 import xnmt.events