Added an option to chose the size of each training epoch

CMakeLists.txt

@@ -21,7 +21,7 @@ endif()
 
 set(CMAKE_CXX_FLAGS "-Wall -Wextra -std=c++11")
 set(CMAKE_CXX_FLAGS_DEBUG "-g3 -Ofast")
-set(CMAKE_CXX_FLAGS_RELEASE "-Ofast")
+set(CMAKE_CXX_FLAGS_RELEASE "-Ofast -march=native")
 
 include_directories(maca_common/include)
 include_directories(transition_machine/include)

maca_common/include/ProgramParameters.hpp

@@ -37,6 +37,7 @@ struct ProgramParameters
   static int dynamicEpoch;
   static float dynamicProbability;
   static bool showFeatureRepresentation;
+  static int iterationSize;
 
   private :
 

maca_common/src/ProgramParameters.cpp

@@ -32,3 +32,4 @@ bool ProgramParameters::interactive;
 int ProgramParameters::dynamicEpoch;
 float ProgramParameters::dynamicProbability;
 bool ProgramParameters::showFeatureRepresentation;
+int ProgramParameters::iterationSize;

trainer/src/Trainer.cpp

@@ -98,15 +98,23 @@ void Trainer::train()
 
   fprintf(stderr, "Training of \'%s\' :\n", tm.name.c_str());
 
-  for (curIter = 0; curIter < ProgramParameters::nbIter; curIter++)
+  auto resetAndShuffle = [this]()
   {
     tm.reset();
-
     trainConfig.reset();
 
     if(ProgramParameters::shuffleExamples)
       trainConfig.shuffle("EOS", "1");
 
+    for (auto & it : trainCounter)
+      it.second.first = it.second.second = 0;
+  };
+
+  curIter = 0;
+  int nbSteps = 0;
+  while (curIter < ProgramParameters::nbIter)
+  {
+    resetAndShuffle();
     while (!trainConfig.isFinal())
     {
       TransitionMachine::State * currentState = tm.getCurrentState();
@@ -142,8 +150,12 @@ void Trainer::train()
 
         // Print current iter advancement in percentage
         if (ProgramParameters::interactive)
-          if (trainConfig.head % 200 == 0 || trainConfig.tapes[0].ref.size()-trainConfig.head < 200)
-            fprintf(stderr, "Current Iteration : %.2f%%\r", 100.0*trainConfig.head/trainConfig.tapes[0].ref.size());
+        {
+          int totalSize = ProgramParameters::iterationSize == -1 ? trainConfig.tapes[0].hyp.size() : ProgramParameters::iterationSize;
+          int steps = ProgramParameters::iterationSize == -1 ? trainConfig.head : nbSteps;
+          if (steps % 200 == 0 || totalSize-steps < 200)
+            fprintf(stderr, "Current Iteration : %.2f%%\r", 100.0*steps/totalSize);
+        }
 
         auto weightedActions = classifier->weightActions(trainConfig);
         std::string pAction = "";
@@ -215,9 +227,28 @@ void Trainer::train()
         tm.takeTransition(transition);
         trainConfig.moveHead(transition->headMvt);
       }
+
+      nbSteps++;
+      if (ProgramParameters::iterationSize != -1 && nbSteps >= ProgramParameters::iterationSize)
+      {
+        printScoresAndSave(stderr);
+        nbSteps = 0;
+        curIter++;
+
+        if (curIter >= ProgramParameters::nbIter)
+          break;
+      }
     }
 
+    if (ProgramParameters::iterationSize == -1)
+    {
       printScoresAndSave(stderr);
+      nbSteps = 0;
+      curIter++;
+
+      if (curIter >= ProgramParameters::nbIter)
+        break;
+    }
   }
 }
 

trainer/src/macaon_train.cpp

@@ -46,6 +46,8 @@ po::options_description getOptionsDescription()
       "Language you are working with")
     ("nbiter,n", po::value<int>()->default_value(5),
       "Number of training epochs (iterations)")
+    ("iterationSize", po::value<int>()->default_value(-1),
+      "The number of examples for each iteration. -1 means the whole training set.")
     ("lr", po::value<float>()->default_value(0.001),
       "Learning rate of the optimizer")
     ("seed,s", po::value<int>()->default_value(100),
@@ -149,6 +151,7 @@ int main(int argc, char * argv[])
   ProgramParameters::dynamicEpoch = vm["epochd"].as<int>();
   ProgramParameters::dynamicProbability = vm["proba"].as<float>();
   ProgramParameters::showFeatureRepresentation = vm["showFeatureRepresentation"].as<bool>();
+  ProgramParameters::iterationSize = vm["iterationSize"].as<int>();
 
   const char * MACAON_DIR = std::getenv("MACAON_DIR");
   std::string slash = "/";