From 434a42b9d096abb436cac1d9788157c16b8ccab0 Mon Sep 17 00:00:00 2001
From: Patrick Simianer <p@simianer.de>
Date: Wed, 20 May 2015 22:01:16 +0200
Subject: dtrain: hope/fear sgd

---
 training/dtrain/dtrain.cc                   |  6 ++++-
 training/dtrain/dtrain.h                    |  1 +
 training/dtrain/example/standard/dtrain.ini |  5 ++--
 training/dtrain/update.h                    | 38 +++++++++++++++++++++++++++++
 4 files changed, 47 insertions(+), 3 deletions(-)

diff --git a/training/dtrain/dtrain.cc b/training/dtrain/dtrain.cc
index 97df530b..378c988f 100644
--- a/training/dtrain/dtrain.cc
+++ b/training/dtrain/dtrain.cc
@@ -18,6 +18,7 @@ main(int argc, char** argv)
   const weight_t eta          = conf["learning_rate"].as<weight_t>();
   const weight_t margin       = conf["margin"].as<weight_t>();
   const bool average          = conf["average"].as<bool>();
+  const bool structured       = conf["struct"].as<bool>();
   const weight_t l1_reg       = conf["l1_reg"].as<weight_t>();
   const bool keep             = conf["keep"].as<bool>();
   const string output_fn      = conf["output"].as<string>();
@@ -160,7 +161,10 @@ main(int argc, char** argv)
 
     // get pairs and update
     SparseVector<weight_t> updates;
-    num_up += CollectUpdates(samples, updates, margin);
+    if (structured)
+      num_up += CollectUpdatesStruct(samples, updates);
+    else
+      num_up += CollectUpdates(samples, updates, margin);
     SparseVector<weight_t> lambdas_copy;
     if (l1_reg)
       lambdas_copy = lambdas;
diff --git a/training/dtrain/dtrain.h b/training/dtrain/dtrain.h
index 2636fa89..dc3ac30f 100644
--- a/training/dtrain/dtrain.h
+++ b/training/dtrain/dtrain.h
@@ -59,6 +59,7 @@ dtrain_init(int argc, char** argv, po::variables_map* conf)
     ("N",                 po::value<size_t>()->default_value(4),              "N for BLEU approximation")
     ("input_weights,w",   po::value<string>(),                                      "input weights file")
     ("average,a",         po::value<bool>()->default_value(false),              "output average weights")
+    ("struct,S",          po::value<bool>()->default_value(false),                      "structured SGD")
     ("keep,K",            po::value<bool>()->default_value(false),  "output a weight file per iteration")
     ("output,o",          po::value<string>()->default_value("-"), "output weights file, '-' for STDOUT")
     ("print_weights,P",   po::value<string>()->default_value("EgivenFCoherent SampleCountF CountEF MaxLexFgivenE MaxLexEgivenF IsSingletonF IsSingletonFE Glue WordPenalty PassThrough LanguageModel LanguageModel_OOV"),
diff --git a/training/dtrain/example/standard/dtrain.ini b/training/dtrain/example/standard/dtrain.ini
index c52bef4a..dfb9b844 100644
--- a/training/dtrain/example/standard/dtrain.ini
+++ b/training/dtrain/example/standard/dtrain.ini
@@ -4,7 +4,8 @@ decoder_conf=./cdec.ini   # config for cdec
 iterations=3              # run over input 3 times
 k=100                     # use 100best lists
 N=4                       # optimize (approx.) BLEU4
-learning_rate=0.1         # learning rate
-margin=1.0                # margin for margin perceptron
+learning_rate=0.0001         # learning rate
+margin=0                # margin for margin perceptron
 print_weights=Glue WordPenalty LanguageModel LanguageModel_OOV PhraseModel_0 PhraseModel_1 PhraseModel_2 PhraseModel_3 PhraseModel_4 PhraseModel_5 PhraseModel_6 PassThrough
 score=nakov
+struct=true
diff --git a/training/dtrain/update.h b/training/dtrain/update.h
index d7224cca..6f42e5bd 100644
--- a/training/dtrain/update.h
+++ b/training/dtrain/update.h
@@ -10,6 +10,18 @@ _cmp(ScoredHyp a, ScoredHyp b)
   return a.gold > b.gold;
 }
 
+bool
+_cmpHope(ScoredHyp a, ScoredHyp b)
+{
+  return (a.model+a.gold) > (b.model+b.gold);
+}
+
+bool
+_cmpFear(ScoredHyp a, ScoredHyp b)
+{
+  return (a.model-a.gold) > (b.model-b.gold);
+}
+
 inline bool
 _good(ScoredHyp& a, ScoredHyp& b, weight_t margin)
 {
@@ -20,6 +32,15 @@ _good(ScoredHyp& a, ScoredHyp& b, weight_t margin)
   return false;
 }
 
+inline bool
+_goodS(ScoredHyp& a, ScoredHyp& b)
+{
+  if (a.gold==b.gold)
+    return true;
+
+  return false;
+}
+
 /*
  * multipartite ranking
  *  sort (descending) by bleu
@@ -56,6 +77,23 @@ CollectUpdates(vector<ScoredHyp>* s,
   return num_up;
 }
 
+inline size_t
+CollectUpdatesStruct(vector<ScoredHyp>* s,
+                     SparseVector<weight_t>& updates,
+                     weight_t unused=-1)
+{
+  // hope
+  sort(s->begin(), s->end(), _cmpHope);
+  ScoredHyp hope = (*s)[0];
+  // fear
+  sort(s->begin(), s->end(), _cmpFear);
+  ScoredHyp fear = (*s)[0];
+  if (!_goodS(hope, fear))
+    updates += hope.f - fear.f;
+
+  return updates.size();
+}
+
 } // namespace
 
 #endif
-- 
cgit v1.2.3