diff --git a/trainer/include/Trainer.hpp b/trainer/include/Trainer.hpp
index 2b4a666496dc3c417ebc73f4b0c7465969bc67f2..7568a73278ddf1cc9bb9d4eef93d45a283ffaa40 100644
--- a/trainer/include/Trainer.hpp
+++ b/trainer/include/Trainer.hpp
@@ -16,6 +16,24 @@ class Trainer
{
private :
+ struct EndOfIteration : public std::exception
+ {
+ const char * what() const throw()
+ {
+ return "Iteration must end because an oracle could not find a zero-cost action.";
+ }
+ };
+
+ struct EndOfTraining : public std::exception
+ {
+ const char * what() const throw()
+ {
+ return "Training must end because every epoch has happened.";
+ }
+ };
+
+ private :
+
/// @brief The TransitionMachine that will be trained.
TransitionMachine & tm;
/// @brief The BD initialized with training examples.
@@ -32,8 +50,20 @@ class Trainer
/// Can be nullptr if dev is not used in this training.
Config * devConfig;
+ /// @brief Lots of informations about the current training.
TrainInfos TI;
+ /// @brief Number of training steps done so far.
+ int nbSteps;
+ /// @brief Number of Actions taken so far.
+ int nbActions;
+ /// @brief Number of Actions needed to compute speed.
+ int nbActionsCutoff;
+ /// @brief Current training speed in actions per second.
+ float currentSpeed;
+ /// @brief The date the last time the speed has been computed.
+ std::chrono::time_point<std::chrono::high_resolution_clock> pastTime;
+
public :
/// @brief The FeatureDescritpion of a Config.
@@ -43,9 +73,16 @@ class Trainer
/// @brief Compute and print scores for each Classifier on this epoch, and save the Classifier if they achieved their all time best score.
void printScoresAndSave(FILE * output);
-
/// @brief Get the scores of the classifiers on the dev dataset.
void computeScoreOnDev();
+ /// @brief Read the input file again and shuffle it.
+ void resetAndShuffle();
+ /// @brief Run the current classifier and take the next transition, no training.
+ void doStepNoTrain();
+ /// @brief Run the current classifier and take the next transition, training the classifier.
+ void doStepTrain();
+ /// @brief Compute and print dev scores, increase epoch counter.
+ void prepareNextEpoch();
public :
diff --git a/trainer/src/Trainer.cpp b/trainer/src/Trainer.cpp
index ed834948f5c28224f2e9d8a6daef7d22a2cc5246..fe4a4171cf2201dbd2a39ca442db580e6f28dddd 100644
--- a/trainer/src/Trainer.cpp
+++ b/trainer/src/Trainer.cpp
@@ -6,10 +6,21 @@ Trainer::Trainer(TransitionMachine & tm, BD & bd, Config & config)
{
this->devBD = nullptr;
this->devConfig = nullptr;
+
+ nbSteps = 0;
+ nbActions = 0;
+ nbActionsCutoff = 2*ProgramParameters::batchSize;
+ currentSpeed = 0.0;
+ pastTime = std::chrono::high_resolution_clock::now();
}
Trainer::Trainer(TransitionMachine & tm, BD & bd, Config & config, BD * devBD, Config * devConfig) : tm(tm), trainBD(bd), trainConfig(config), devBD(devBD), devConfig(devConfig)
{
+ nbSteps = 0;
+ nbActions = 0;
+ nbActionsCutoff = 2*ProgramParameters::batchSize;
+ currentSpeed = 0.0;
+ pastTime = std::chrono::high_resolution_clock::now();
}
void Trainer::computeScoreOnDev()
@@ -163,210 +174,290 @@ void Trainer::computeScoreOnDev()
fprintf(stderr, "End of %s\n", __func__);
}
-void Trainer::train()
+void Trainer::resetAndShuffle()
{
- Dict::createFiles(ProgramParameters::expPath, "");
+ tm.reset();
+ trainConfig.reset();
- fprintf(stderr, "%sTraining of \'%s\' :\n",
- ProgramParameters::printTime ? ("["+getTime()+"] ").c_str() : "",
- tm.name.c_str());
+ if(ProgramParameters::shuffleExamples)
+ trainConfig.shuffle(ProgramParameters::sequenceDelimiterTape, ProgramParameters::sequenceDelimiter);
+
+ TI.resetCounters();
+}
- auto resetAndShuffle = [this]()
+void Trainer::doStepNoTrain()
+{
+ int neededActionIndex = tm.getCurrentClassifier()->getOracleActionIndex(trainConfig);
+ std::string neededActionName = tm.getCurrentClassifier()->getActionName(neededActionIndex);
+ if (ProgramParameters::debug)
{
- tm.reset();
- trainConfig.reset();
+ trainConfig.printForDebug(stderr);
+ fprintf(stderr, "action=<%s>\n", neededActionName.c_str());
+ }
- if(ProgramParameters::shuffleExamples)
- trainConfig.shuffle(ProgramParameters::sequenceDelimiterTape, ProgramParameters::sequenceDelimiter);
+ Action * action = tm.getCurrentClassifier()->getAction(neededActionName);
+ TransitionMachine::Transition * transition = tm.getTransition(neededActionName);
+ action->setInfos(transition->headMvt, tm.getCurrentState());
- TI.resetCounters();
- };
+ action->apply(trainConfig);
+ tm.takeTransition(transition);
+}
- int nbSteps = 0;
- int nbActions = 0;
- int nbActionsCutoff = 2*ProgramParameters::batchSize;
- float currentSpeed = 0.0;
- auto pastTime = std::chrono::high_resolution_clock::now();
- while (TI.getEpoch() <= ProgramParameters::nbIter)
+void Trainer::doStepTrain()
+{
+ if (!TI.isTopologyPrinted(tm.getCurrentClassifier()->name))
{
- resetAndShuffle();
- while (!trainConfig.isFinal())
+ TI.setTopologyPrinted(tm.getCurrentClassifier()->name);
+ tm.getCurrentClassifier()->printTopology(stderr);
+ }
+
+ // Print current iter advancement in percentage
+ if (ProgramParameters::interactive && !ProgramParameters::featureExtraction)
+ {
+ int totalSize = ProgramParameters::iterationSize == -1 ? ProgramParameters::tapeSize : ProgramParameters::iterationSize;
+ int steps = ProgramParameters::iterationSize == -1 ? trainConfig.getHead() : nbSteps;
+ if (steps % nbActionsCutoff == 0 || totalSize-steps < nbActionsCutoff)
{
- trainConfig.setCurrentStateName(tm.getCurrentState());
- Dict::currentClassifierName = tm.getCurrentClassifier()->name;
- tm.getCurrentClassifier()->initClassifier(trainConfig);
+ fprintf(stderr, " \r");
+ fprintf(stderr, "Current Iteration : %.2f%% speed : %s actions/s\r", 100.0*steps/totalSize, int2humanStr((int)currentSpeed).c_str());
+ }
+ }
- if(!tm.getCurrentClassifier()->needsTrain())
- {
- int neededActionIndex = tm.getCurrentClassifier()->getOracleActionIndex(trainConfig);
- std::string neededActionName = tm.getCurrentClassifier()->getActionName(neededActionIndex);
- if (ProgramParameters::debug)
- {
- trainConfig.printForDebug(stderr);
- fprintf(stderr, "action=<%s>\n", neededActionName.c_str());
- }
+ std::string pAction = "";
+ std::string oAction = "";
+ bool pActionIsZeroCost = false;
- Action * action = tm.getCurrentClassifier()->getAction(neededActionName);
- TransitionMachine::Transition * transition = tm.getTransition(neededActionName);
- action->setInfos(transition->headMvt, tm.getCurrentState());
+
+ std::string actionName = "";
+ float loss = 0.0;
- action->apply(trainConfig);
- tm.takeTransition(transition);
- }
- else
- {
- if (!TI.isTopologyPrinted(tm.getCurrentClassifier()->name))
- {
- TI.setTopologyPrinted(tm.getCurrentClassifier()->name);
- tm.getCurrentClassifier()->printTopology(stderr);
- }
-
- // Print current iter advancement in percentage
- if (ProgramParameters::interactive && !ProgramParameters::featureExtraction)
+ Classifier::WeightedActions weightedActions;
+ if (tm.getCurrentClassifier()->name.rfind("Error_", 0) != 0)
+ {
+ if (!ProgramParameters::featureExtraction)
+ {
+ weightedActions = tm.getCurrentClassifier()->weightActions(trainConfig);
+
+ for (auto & it : weightedActions)
+ if (it.first)
{
- int totalSize = ProgramParameters::iterationSize == -1 ? ProgramParameters::tapeSize : ProgramParameters::iterationSize;
- int steps = ProgramParameters::iterationSize == -1 ? trainConfig.getHead() : nbSteps;
- if (steps % nbActionsCutoff == 0 || totalSize-steps < nbActionsCutoff)
+ if (pAction == "")
+ pAction = it.second.second;
+
+ if (tm.getCurrentClassifier()->getActionCost(trainConfig, it.second.second) == 0)
{
- fprintf(stderr, " \r");
- fprintf(stderr, "Current Iteration : %.2f%% speed : %s actions/s\r", 100.0*steps/totalSize, int2humanStr((int)currentSpeed).c_str());
+ oAction = it.second.second;
+ break;
}
}
+
+ if (pAction == oAction)
+ pActionIsZeroCost = true;
+ }
+ else
+ {
+ oAction = tm.getCurrentClassifier()->getZeroCostActions(trainConfig)[0];
+ }
+
+ if (oAction.empty())
+ oAction = tm.getCurrentClassifier()->getDefaultAction();
+
+ if (oAction.empty())
+ {
+ if (trainConfig.endOfTapes())
+ {
+ while (!trainConfig.stackEmpty())
+ trainConfig.stackPop();
+ throw EndOfIteration();
+ }
+
+ fprintf(stderr, "ERROR (%s) : Unable to find any zero cost action. Aborting.\n", ERRINFO);
+ fprintf(stderr, "State : %s\n", tm.getCurrentState().c_str());
+ trainConfig.printForDebug(stderr);
+ tm.getCurrentClassifier()->explainCostOfActions(stderr, trainConfig);
+ exit(1);
+ }
- std::string pAction = "";
- std::string oAction = "";
- bool pActionIsZeroCost = false;
+ if (!ProgramParameters::featureExtraction)
+ loss = tm.getCurrentClassifier()->trainOnExample(trainConfig, tm.getCurrentClassifier()->getActionIndex(oAction));
+
+ TI.addTrainExample(tm.getCurrentClassifier()->name, loss);
+ if (pActionIsZeroCost)
+ TI.addTrainSuccess(tm.getCurrentClassifier()->name);
+
+ int k = ProgramParameters::dynamicEpoch;
+
+
+ if (ProgramParameters::featureExtraction)
+ {
+ auto features = tm.getCurrentClassifier()->getFeatureModel()->getFeatureDescription(trainConfig).featureValues();
+ fprintf(stdout, "%s\t%s\n", oAction.c_str(), features.c_str());
+ }
+
+ if (TI.getEpoch() >= k && choiceWithProbability(ProgramParameters::dynamicProbability))
+ {
+ actionName = pAction;
+ }
+ else
+ {
+ if (pActionIsZeroCost)
+ actionName = pAction;
+ else
+ actionName = oAction;
+ }
+
+ if (ProgramParameters::debug)
+ {
+ trainConfig.printForDebug(stderr);
+ fprintf(stderr, "pAction=<%s> oAction=<%s> action=<%s>\n", pAction.c_str(), oAction.c_str(), actionName.c_str());
+ }
- Classifier::WeightedActions weightedActions;
- if (!ProgramParameters::featureExtraction)
- {
- weightedActions = tm.getCurrentClassifier()->weightActions(trainConfig);
-
- for (auto & it : weightedActions)
- if (it.first)
- {
- if (pAction == "")
- pAction = it.second.second;
-
- if (tm.getCurrentClassifier()->getActionCost(trainConfig, it.second.second) == 0)
- {
- oAction = it.second.second;
- break;
- }
- }
-
- if (pAction == oAction)
- pActionIsZeroCost = true;
- }
- else
+ }
+ else
+ {
+ if (!ProgramParameters::featureExtraction)
+ {
+ weightedActions = tm.getCurrentClassifier()->weightActions(trainConfig);
+
+ for (auto & it : weightedActions)
+ if (it.first)
{
- oAction = tm.getCurrentClassifier()->getZeroCostActions(trainConfig)[0];
+ pAction = it.second.second;
+ break;
}
- if (oAction.empty())
- oAction = tm.getCurrentClassifier()->getDefaultAction();
+ auto zeroCosts = tm.getCurrentClassifier()->getZeroCostActions(trainConfig);
+ oAction = zeroCosts[rand() % zeroCosts.size()];
+ }
+ else
+ {
+ oAction = tm.getCurrentClassifier()->getZeroCostActions(trainConfig)[0];
+ }
+
+ if (oAction.empty())
+ oAction = tm.getCurrentClassifier()->getDefaultAction();
+
+ if (oAction.empty())
+ {
+ if (trainConfig.endOfTapes())
+ {
+ while (!trainConfig.stackEmpty())
+ trainConfig.stackPop();
+ throw EndOfIteration();
+ }
+ }
- if (oAction.empty())
- {
- if (trainConfig.endOfTapes())
- {
- while (!trainConfig.stackEmpty())
- trainConfig.stackPop();
- break;
- }
+ //ici
+
+ float loss = 0.0;
+ if (!ProgramParameters::featureExtraction)
+ loss = tm.getCurrentClassifier()->trainOnExample(trainConfig, tm.getCurrentClassifier()->getActionIndex(oAction));
+
+ TI.addTrainExample(tm.getCurrentClassifier()->name, loss);
+ if (pActionIsZeroCost)
+ TI.addTrainSuccess(tm.getCurrentClassifier()->name);
+
+ int k = ProgramParameters::dynamicEpoch;
+
+
+ if (ProgramParameters::featureExtraction)
+ {
+ auto features = tm.getCurrentClassifier()->getFeatureModel()->getFeatureDescription(trainConfig).featureValues();
+ fprintf(stdout, "%s\t%s\n", oAction.c_str(), features.c_str());
+ }
+
+ if (TI.getEpoch() >= k && choiceWithProbability(ProgramParameters::dynamicProbability))
+ {
+ actionName = pAction;
+ }
+ else
+ {
+ if (pActionIsZeroCost)
+ actionName = pAction;
+ else
+ actionName = oAction;
+ }
+
+ if (ProgramParameters::debug)
+ {
+ trainConfig.printForDebug(stderr);
+ fprintf(stderr, "pAction=<%s> oAction=<%s> action=<%s>\n", pAction.c_str(), oAction.c_str(), actionName.c_str());
+ }
- fprintf(stderr, "ERROR (%s) : Unable to find any zero cost action. Aborting.\n", ERRINFO);
- fprintf(stderr, "State : %s\n", tm.getCurrentState().c_str());
- trainConfig.printForDebug(stderr);
- tm.getCurrentClassifier()->explainCostOfActions(stderr, trainConfig);
- exit(1);
- }
+ actionName = "BANANE";
+ }
- float loss = 0.0;
- if (!ProgramParameters::featureExtraction)
- loss = tm.getCurrentClassifier()->trainOnExample(trainConfig, tm.getCurrentClassifier()->getActionIndex(oAction));
+ Action * action = tm.getCurrentClassifier()->getAction(actionName);
+ TransitionMachine::Transition * transition = tm.getTransition(actionName);
+ action->setInfos(transition->headMvt, tm.getCurrentState());
- TI.addTrainExample(tm.getCurrentClassifier()->name, loss);
- if (pActionIsZeroCost)
- TI.addTrainSuccess(tm.getCurrentClassifier()->name);
+ action->apply(trainConfig);
+ tm.takeTransition(transition);
- int k = ProgramParameters::dynamicEpoch;
+ nbActions++;
- std::string actionName = "";
+ if (nbActions >= nbActionsCutoff)
+ {
+ auto actualTime = std::chrono::high_resolution_clock::now();
- if (ProgramParameters::featureExtraction)
- {
- auto features = tm.getCurrentClassifier()->getFeatureModel()->getFeatureDescription(trainConfig).featureValues();
- fprintf(stdout, "%s\t%s\n", oAction.c_str(), features.c_str());
- }
+ double seconds = std::chrono::duration<double, std::milli>(actualTime-pastTime).count() / 1000.0;
+ currentSpeed = nbActions / seconds;
- if (TI.getEpoch() >= k && choiceWithProbability(ProgramParameters::dynamicProbability))
- {
- actionName = pAction;
- }
- else
- {
- if (pActionIsZeroCost)
- actionName = pAction;
- else
- actionName = oAction;
- }
+ pastTime = actualTime;
- if (ProgramParameters::debug)
- {
- trainConfig.printForDebug(stderr);
- fprintf(stderr, "pAction=<%s> oAction=<%s> action=<%s>\n", pAction.c_str(), oAction.c_str(), actionName.c_str());
- }
-
- Action * action = tm.getCurrentClassifier()->getAction(actionName);
- TransitionMachine::Transition * transition = tm.getTransition(actionName);
- action->setInfos(transition->headMvt, tm.getCurrentState());
+ nbActions = 0;
+ }
- action->apply(trainConfig);
- tm.takeTransition(transition);
+ float entropy = Classifier::computeEntropy(weightedActions);
+ trainConfig.addToEntropyHistory(entropy);
+}
- nbActions++;
+void Trainer::prepareNextEpoch()
+{
+ printScoresAndSave(stderr);
+ nbSteps = 0;
+ TI.nextEpoch();
- if (nbActions >= nbActionsCutoff)
- {
- auto actualTime = std::chrono::high_resolution_clock::now();
+ if (TI.getEpoch() > ProgramParameters::nbIter)
+ throw EndOfTraining();
+}
- double seconds = std::chrono::duration<double, std::milli>(actualTime-pastTime).count() / 1000.0;
- currentSpeed = nbActions / seconds;
+void Trainer::train()
+{
+ Dict::createFiles(ProgramParameters::expPath, "");
- pastTime = actualTime;
+ fprintf(stderr, "%sTraining of \'%s\' :\n",
+ ProgramParameters::printTime ? ("["+getTime()+"] ").c_str() : "",
+ tm.name.c_str());
- nbActions = 0;
- }
+ while (TI.getEpoch() <= ProgramParameters::nbIter)
+ {
+ resetAndShuffle();
+ while (!trainConfig.isFinal())
+ {
+ trainConfig.setCurrentStateName(tm.getCurrentState());
+ Dict::currentClassifierName = tm.getCurrentClassifier()->name;
+ tm.getCurrentClassifier()->initClassifier(trainConfig);
- float entropy = Classifier::computeEntropy(weightedActions);
- trainConfig.addToEntropyHistory(entropy);
- }
+ if(!tm.getCurrentClassifier()->needsTrain())
+ doStepNoTrain();
+ else
+ try {doStepTrain();}
+ catch (EndOfIteration &) {break;}
nbSteps++;
- if (ProgramParameters::iterationSize != -1 && nbSteps >= ProgramParameters::iterationSize)
- {
- printScoresAndSave(stderr);
- nbSteps = 0;
- TI.nextEpoch();
- if (TI.getEpoch() > ProgramParameters::nbIter)
- break;
- }
+ if (ProgramParameters::iterationSize != -1 && nbSteps >= ProgramParameters::iterationSize)
+ try {prepareNextEpoch();}
+ catch (EndOfTraining &) {break;}
}
if (ProgramParameters::debug)
fprintf(stderr, "Config is final\n");
if (ProgramParameters::iterationSize == -1)
- {
- printScoresAndSave(stderr);
- nbSteps = 0;
- TI.nextEpoch();
-
- if (TI.getEpoch() > ProgramParameters::nbIter)
- break;
- }
+ try {prepareNextEpoch();}
+ catch (EndOfTraining &) {break;}
if (ProgramParameters::debug)
fprintf(stderr, "End of epoch\n");