Working fasttext style embeddings learning

neural_network/include/NeuralNetwork.hpp

@@ -159,6 +159,13 @@ class NeuralNetwork
   ///
   /// @return The model of this NeuralNetwork.
   dynet::ParameterCollection & getModel();
+
+  /// \brief How much input neurons a certain feature will take.
+  ///
+  /// \param fv The FeatureValue to measure the size of.
+  ///
+  /// \return The number of input neurons taken by fv.
+  static unsigned int featureSize(const FeatureModel::FeatureValue & fv);
 };
 
 #endif

neural_network/src/NeuralNetwork.cpp

@@ -47,6 +47,9 @@ dynet::Expression NeuralNetwork::featValue2Expression(dynet::ComputationGraph &
       expressions.emplace_back(dynet::lookup(cg, lu, index));
   }
 
+  if (fv.func == FeatureModel::Function::Mean)
+    return dynet::average(expressions);
+
   return dynet::concatenate(expressions);
 }
 
@@ -172,3 +175,16 @@ dynet::Expression NeuralNetwork::activate(dynet::Expression h, Activation f)
   return h;
 }
 
+unsigned int NeuralNetwork::featureSize(const FeatureModel::FeatureValue & fv)
+{
+  unsigned int res = 0;
+
+  if (fv.func == FeatureModel::Function::Concat)
+    for (auto dict : fv.dicts)
+      res += dict->getDimension();
+  else if (fv.func == FeatureModel::Function::Mean)
+    res = fv.dicts[0]->getDimension();
+
+  return res;
+}
+

transition_machine/src/Classifier.cpp

@@ -145,8 +145,7 @@ void Classifier::initClassifier(Config & config)
   int nbOutputs = as->actions.size();
 
   for (auto feat : fd.values)
-    for (auto dict : feat.dicts)
-      nbInputs += dict->getDimension();
+    nbInputs += NeuralNetwork::featureSize(feat);
 
   nn->init(nbInputs, topology, nbOutputs);
 }

transition_machine/src/FeatureBank.cpp

@@ -519,7 +519,6 @@ FeatureModel::FeatureValue FeatureBank::aggregateStack(Config & c, int from, con
   return result;
 }
 
-//TODO : ne pas utiliser une feature value pour word mais un string, pour que ça marche avec les mots inconnus
 FeatureModel::FeatureValue FeatureBank::fasttext(Config & c, const FeatureModel::FeatureValue & word)
 {
   FeatureModel::FeatureValue result(FeatureModel::Function::Mean);
@@ -531,7 +530,9 @@ FeatureModel::FeatureValue FeatureBank::fasttext(Config & c, const FeatureModel:
     return {lettersDict, word.names[0], Dict::nullValueStr, policy};
 
   unsigned int wordLength = getNbSymbols(word.values[0]);
-  unsigned int gramLength = 2;
+  unsigned int gramLength = 4;
+
+  bool slidingMode = false;
 
   if (wordLength < gramLength)
   {
@@ -542,6 +543,22 @@ FeatureModel::FeatureValue FeatureBank::fasttext(Config & c, const FeatureModel:
     result.policies.emplace_back(value.policies[0]);
   }
   else
+  {
+    if (!slidingMode)
+    {
+      int nbGrams = wordLength / gramLength + (wordLength % gramLength ? 1 : 0);
+      for (int i = 0; i < nbGrams; i++)
+      {
+        int from = i * gramLength;
+        int to = i == nbGrams-1 ? wordLength-1 : (i+1)*gramLength-1;
+        auto value = getLetters(c, word, from, to);
+        result.dicts.emplace_back(value.dicts[0]);
+        result.names.emplace_back(value.names[0]);
+        result.values.emplace_back(value.values[0]);
+        result.policies.emplace_back(value.policies[0]);
+      }
+    }
+    else
     {
       for (unsigned int i = 0; i+gramLength-1 < wordLength; i++)
       {
@@ -552,6 +569,7 @@ FeatureModel::FeatureValue FeatureBank::fasttext(Config & c, const FeatureModel:
         result.policies.emplace_back(value.policies[0]);
       }
     }
+  }
 
   return result;
 }