Allow to specify batchsie in cla files, and changed error detection loss function to l1

neural_network/include/MLPBase.hpp

@@ -38,6 +38,7 @@ class MLPBase
   std::vector<FeatureModel::FeatureDescription> fdsContinuous;
   /// @brief gold outputs of the current minibatch.
   std::vector< std::vector<float> > goldsContinuous;
+  int batchSize;
 
   private :
 
@@ -144,6 +145,7 @@ class MLPBase
   void printTopology(FILE * output);
   /// @brief Clear the current batch.
   void endOfIteration();
+  void setBatchSize(int batchSize);
 };
 
 #endif

neural_network/include/NeuralNetwork.hpp

@@ -103,6 +103,9 @@ class NeuralNetwork
   /// @brief The dynet model containing the parameters to be trained.
   dynet::ParameterCollection model;
 
+  /// @brief The size of batches, if not specified, use the global batchsize.
+  int batchSize;
+
   protected :
 
   /// @brief Set dynet and srand() seeds.
@@ -192,6 +195,12 @@ class NeuralNetwork
   ///
   /// \return The number of input neurons taken by fv.
   static unsigned int featureSize(const FeatureModel::FeatureValue & fv);
+
+  /// @brief Set the batchSize.
+  void setBatchSize(int batchSize);
+
+  /// @brief Get the batchSize.
+  int getBatchSize();
 };
 
 #endif

neural_network/src/MLP.cpp

@@ -19,6 +19,7 @@ MLP::MLP(const std::string & filename) : mlp("MLP")
 
 void MLP::init(int nbInputs, const std::string & topology, int nbOutputs)
 {
+  setBatchSize(0);
   mlp.init(model, nbInputs, topology, nbOutputs);
 }
 
@@ -57,6 +58,7 @@ std::vector<float> MLP::predict(FeatureModel::FeatureDescription & fd)
 
 float MLP::update(FeatureModel::FeatureDescription & fd, int gold)
 {
+  mlp.setBatchSize(getBatchSize());
   try
   {
     float loss = mlp.update(fd, gold);
@@ -70,6 +72,7 @@ float MLP::update(FeatureModel::FeatureDescription & fd, int gold)
 
 float MLP::update(FeatureModel::FeatureDescription & fd, const std::vector<float> & gold)
 {
+  mlp.setBatchSize(getBatchSize());
   try
   {
     float loss = mlp.update(fd, gold);
@@ -83,6 +86,7 @@ float MLP::update(FeatureModel::FeatureDescription & fd, const std::vector<float
 
 float MLP::getLoss(FeatureModel::FeatureDescription & fd, int gold)
 {
+  mlp.setBatchSize(getBatchSize());
   try
   {
     float loss = mlp.getLoss(fd, gold);
@@ -95,6 +99,7 @@ float MLP::getLoss(FeatureModel::FeatureDescription & fd, int gold)
 
 float MLP::getLoss(FeatureModel::FeatureDescription & fd, const std::vector<float> & gold)
 {
+  mlp.setBatchSize(getBatchSize());
   try
   {
     float loss = mlp.getLoss(fd, gold);

neural_network/src/MLPBase.cpp

@@ -4,6 +4,7 @@ MLPBase::MLPBase(std::string name)
 {
   this->name = name;
   dropoutActive = true;
+  batchSize = 0;
 }
 
 void MLPBase::init(dynet::ParameterCollection & model, int nbInputs, const std::string & topology, int nbOutputs)
@@ -94,7 +95,9 @@ float MLPBase::update(FeatureModel::FeatureDescription & fd, int gold)
   fdsOneHot.emplace_back(fd);
   goldsOneHot.emplace_back(gold);
 
-  if ((int)fdsOneHot.size() < ProgramParameters::batchSize)
+  int effectiveBatchSize = batchSize ? batchSize : ProgramParameters::batchSize;
+
+  if ((int)fdsOneHot.size() < effectiveBatchSize)
     throw BatchNotFull();
 
   std::vector<dynet::Expression> inputs;
@@ -140,7 +143,9 @@ float MLPBase::update(FeatureModel::FeatureDescription & fd, const std::vector<f
   fdsContinuous.emplace_back(fd);
   goldsContinuous.emplace_back(gold);
 
-  if ((int)fdsContinuous.size() < ProgramParameters::batchSize)
+  int effectiveBatchSize = batchSize ? batchSize : ProgramParameters::batchSize;
+
+  if ((int)fdsContinuous.size() < effectiveBatchSize)
     throw BatchNotFull();
 
   std::vector<dynet::Expression> inputs;
@@ -165,7 +170,7 @@ float MLPBase::update(FeatureModel::FeatureDescription & fd, const std::vector<f
     goldExpressions.emplace_back(dynet::input(cg, dynet::Dim({(unsigned int)gold.size()}), gold));
  
   dynet::Expression batchedGold = dynet::concatenate_to_batch(goldExpressions);
-  batchedLoss = dynet::sum_batches(dynet::squared_distance(output, batchedGold));
+  batchedLoss = dynet::sum_batches(dynet::l1_distance(output, batchedGold));
 
   cg.backward(batchedLoss);
 
@@ -182,7 +187,9 @@ float MLPBase::getLoss(FeatureModel::FeatureDescription & fd, int gold)
   fdsOneHot.emplace_back(fd);
   goldsOneHot.emplace_back(gold);
 
-  if ((int)fdsOneHot.size() < ProgramParameters::batchSize)
+  int effectiveBatchSize = batchSize ? batchSize : ProgramParameters::batchSize;
+
+  if ((int)fdsOneHot.size() < effectiveBatchSize)
     throw BatchNotFull();
 
   std::vector<dynet::Expression> inputs;
@@ -226,7 +233,9 @@ float MLPBase::getLoss(FeatureModel::FeatureDescription & fd, const std::vector<
   fdsContinuous.emplace_back(fd);
   goldsContinuous.emplace_back(gold);
 
-  if ((int)fdsContinuous.size() < ProgramParameters::batchSize)
+  int effectiveBatchSize = batchSize ? batchSize : ProgramParameters::batchSize;
+
+  if ((int)fdsContinuous.size() < effectiveBatchSize)
     throw BatchNotFull();
 
   std::vector<dynet::Expression> inputs;
@@ -251,7 +260,7 @@ float MLPBase::getLoss(FeatureModel::FeatureDescription & fd, const std::vector<
     goldExpressions.emplace_back(dynet::input(cg, dynet::Dim({1,(unsigned int)gold.size()}), gold));
  
   dynet::Expression batchedGold = dynet::concatenate_to_batch(goldExpressions);
-  batchedLoss = dynet::sum_batches(dynet::squared_distance(output, batchedGold));
+  batchedLoss = dynet::sum_batches(dynet::l1_distance(output, batchedGold));
 
   checkGradients();
 
@@ -482,3 +491,8 @@ void MLPBase::endOfIteration()
   goldsContinuous.clear();
 }
 
+void MLPBase::setBatchSize(int batchSize)
+{
+  this->batchSize = batchSize;
+}
+

neural_network/src/NeuralNetwork.cpp

@@ -188,3 +188,13 @@ unsigned int NeuralNetwork::featureSize(const FeatureModel::FeatureValue & fv)
   return res;
 }
 
+void NeuralNetwork::setBatchSize(int batchSize)
+{
+  this->batchSize = batchSize;
+}
+
+int NeuralNetwork::getBatchSize()
+{
+  return batchSize;
+}
+

transition_machine/include/Classifier.hpp

@@ -54,6 +54,8 @@ class Classifier
   std::shared_ptr<NeuralNetwork> nn;
   /// @brief A string describing the topology of the underlying neural network.
   std::string topology;
+  /// @brief The size of a batch. If not specified, use the global batchsize.
+  int batchSize;
   /// @brief The oracle being used by this Classifier.
   ///
   /// For Classifier of type Prediction, the Oracle is used in train mode, to find

transition_machine/src/Classifier.cpp

@@ -73,6 +73,10 @@ Classifier::Classifier(const std::string & filename, bool trainMode)
     badFormatAndAbort(ERRINFO);
 
   topology = buffer;
+
+  int batchsizeRead = 0;
+  if(fscanf(fd, "Batchsize : %d\n", &batchsizeRead) == 1)
+    batchSize = batchsizeRead;
 }
 
 Classifier::Type Classifier::str2type(const std::string & s)
@@ -158,6 +162,8 @@ void Classifier::initClassifier(Config & config)
     nbInputs += NeuralNetwork::featureSize(feat);
 
   nn->init(nbInputs, topology, nbOutputs);
+  if (batchSize)
+    nn->setBatchSize(batchSize);
 }
 
 FeatureModel::FeatureDescription Classifier::getFeatureDescription(Config & config)