diff --git a/MLP/include/MLP.hpp b/MLP/include/MLP.hpp
index 5c419724bd908267a28cb7a684757ea368cc96f6..484d88100e481a5e53ab8d098c5578d7bd2c7946 100644
--- a/MLP/include/MLP.hpp
+++ b/MLP/include/MLP.hpp
@@ -12,7 +12,6 @@
#include <dynet/timing.h>
#include <dynet/expr.h>
#include "FeatureModel.hpp"
-#include "TrainingExamples.hpp"
/// @brief Multi Layer Perceptron.
///
diff --git a/trainer/include/Trainer.hpp b/trainer/include/Trainer.hpp
index 0f66aeb3b8b6025d01a428505ddabf9ddb453fd0..6efd1574028d7b19dbec7f93c9d221580d3f0c85 100644
--- a/trainer/include/Trainer.hpp
+++ b/trainer/include/Trainer.hpp
@@ -9,7 +9,6 @@
#include "TransitionMachine.hpp"
#include "BD.hpp"
#include "Config.hpp"
-#include "TrainingExamples.hpp"
/// @brief An object capable of training a TransitionMachine given a BD initialized with training examples.
class Trainer
@@ -68,17 +67,6 @@ class Trainer
private :
- /// @brief Train the TransitionMachine one example at a time.
- ///
- /// For each epoch all the Classifier of the TransitionMachine are fed all the
- /// training examples, at the end of the epoch Classifier are evaluated on
- /// the devBD if available, and each Classifier will be saved only if its score
- /// on the current epoch is its all time best.\n
- /// When a Classifier is saved that way, all the Dict involved are also saved.
- /// @param nbIter The number of epochs.
- /// @param mustShuffle Whether or not to shuffle examples between each epoch.
- void trainUnbatched(int nbIter, bool mustShuffle);
-
/// @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);
@@ -107,13 +95,17 @@ class Trainer
/// @param debugMode If true, infos will be printed on stderr.
/// @param removeDuplicates If true, duplicates examples will be removed from the training set.
Trainer(TransitionMachine & tm, BD & bd, Config & config, BD * devBD, Config * devConfig, bool debugMode, bool removeDuplicates);
- /// @brief Train the TransitionMachine.
+
+ /// @brief Train the TransitionMachine one example at a time.
///
- /// @param nbIter The number of training epochs.
- /// @param batchSize The size of each batch.
- /// @param mustShuffle Will the examples be shuffled after every epoch ?
- /// @param batched True if we feed the training algorithm with batches of examples
- void train(int nbIter, int batchSize, bool mustShuffle, bool batched);
+ /// For each epoch all the Classifier of the TransitionMachine are fed all the
+ /// training examples, at the end of the epoch Classifier are evaluated on
+ /// the devBD if available, and each Classifier will be saved only if its score
+ /// on the current epoch is its all time best.\n
+ /// When a Classifier is saved that way, all the Dict involved are also saved.
+ /// @param nbIter The number of epochs.
+ /// @param mustShuffle Whether or not to shuffle examples between each epoch.
+ void train(int nbIter, bool mustShuffle);
};
#endif
diff --git a/trainer/src/Trainer.cpp b/trainer/src/Trainer.cpp
index 1bc497b31f8fa3763c78f74958a5f9dd15b7115c..c541924e6826d064fdeff3ffcbaeb3ba5651f5da 100644
--- a/trainer/src/Trainer.cpp
+++ b/trainer/src/Trainer.cpp
@@ -26,6 +26,9 @@ std::map<std::string, float> Trainer::getScoreOnDev()
std::map< std::string, std::pair<int, int> > counts;
+ if (debugMode)
+ fprintf(stderr, "Computing score on dev set\n");
+
while (!devConfig->isFinal())
{
TransitionMachine::State * currentState = tm.getCurrentState();
@@ -73,6 +76,12 @@ std::map<std::string, float> Trainer::getScoreOnDev()
std::string actionName = pAction;
Action * action = classifier->getAction(actionName);
+ if (debugMode)
+ {
+ devConfig->printForDebug(stderr);
+ fprintf(stderr, "pAction=<%s> action=<%s>\n", pAction.c_str(), actionName.c_str());
+ }
+
action->apply(*devConfig);
TransitionMachine::Transition * transition = tm.getTransition(actionName);
tm.takeTransition(transition);
@@ -87,7 +96,7 @@ std::map<std::string, float> Trainer::getScoreOnDev()
return scores;
}
-void Trainer::trainUnbatched(int nbIter, bool mustShuffle)
+void Trainer::train(int nbIter, bool mustShuffle)
{
this->nbIter = nbIter;
@@ -98,6 +107,7 @@ void Trainer::trainUnbatched(int nbIter, bool mustShuffle)
for (curIter = 0; curIter < nbIter; curIter++)
{
tm.reset();
+
trainConfig.reset();
if(mustShuffle)
@@ -144,6 +154,7 @@ void Trainer::trainUnbatched(int nbIter, bool mustShuffle)
if (zeroCostActions.empty())
{
fprintf(stderr, "ERROR (%s) : Unable to find any zero cost action. Aborting.\n", ERRINFO);
+ fprintf(stderr, "State : %s\n", currentState->name.c_str());
trainConfig.printForDebug(stderr);
exit(1);
}
@@ -278,8 +289,3 @@ void Trainer::printScoresAndSave(FILE * output)
printColumns(output, {names, acc, train, dev, savedStr});
}
-void Trainer::train(int nbIter, int batchSize, bool mustShuffle, bool batched)
-{
- trainUnbatched(nbIter, mustShuffle);
-}
-
diff --git a/trainer/src/macaon_train.cpp b/trainer/src/macaon_train.cpp
index 9a11afaf141e68f33107128cb2737da4a3e82e85..9c8e31d513c7d6bad15ce165a8301a875f14a2b5 100644
--- a/trainer/src/macaon_train.cpp
+++ b/trainer/src/macaon_train.cpp
@@ -43,14 +43,10 @@ po::options_description getOptionsDescription()
"Language you are working with")
("nbiter,n", po::value<int>()->default_value(5),
"Number of training epochs (iterations)")
- ("batchsize,b", po::value<int>()->default_value(256),
- "Size of each training batch (in number of examples)")
("seed,s", po::value<int>()->default_value(100),
"The random seed that will initialize RNG")
("duplicates", po::value<bool>()->default_value(true),
"Remove identical training examples")
- ("batched", po::value<bool>()->default_value(true),
- "Uses batch of training examples")
("shuffle", po::value<bool>()->default_value(true),
"Shuffle examples after each iteration");
@@ -115,10 +111,8 @@ int main(int argc, char * argv[])
std::string expName = vm["expName"].as<std::string>();
std::string lang = vm["lang"].as<std::string>();
int nbIter = vm["nbiter"].as<int>();
- int batchSize = vm["batchsize"].as<int>();
int randomSeed = vm["seed"].as<int>();
bool mustShuffle = vm["shuffle"].as<bool>();
- bool batched = vm["batched"].as<bool>();
bool removeDuplicates = vm["duplicates"].as<bool>();
bool debugMode = vm.count("debug") == 0 ? false : true;
@@ -159,7 +153,7 @@ int main(int argc, char * argv[])
}
trainer->expPath = expPath;
- trainer->train(nbIter, batchSize, mustShuffle, batched);
+ trainer->train(nbIter, mustShuffle);
return 0;
}
diff --git a/transition_machine/include/TrainingExamples.hpp b/transition_machine/include/TrainingExamples.hpp
deleted file mode 100644
index 4fd52a177419eca6c80a0a951748ea305e8dfa5e..0000000000000000000000000000000000000000
--- a/transition_machine/include/TrainingExamples.hpp
+++ /dev/null
@@ -1,33 +0,0 @@
-/// @file TrainingExamples.hpp
-/// @author Franck Dary
-/// @version 1.0
-/// @date 2018-08-09
-
-#ifndef TRAININGEXAMPLES__H
-#define TRAININGEXAMPLES__H
-
-#include <vector>
-#include "FeatureModel.hpp"
-
-class TrainingExamples
-{
- public :
-
- std::vector<unsigned int> order;
- std::vector<FeatureModel::FeatureDescription> examples;
- std::vector<int> classes;
- int nextIndex;
-
- public :
-
- TrainingExamples();
- void add(const FeatureModel::FeatureDescription & example, int gold);
- unsigned int size();
- TrainingExamples getBatch(unsigned int batchSize);
- void reset();
- void shuffle();
- void remove(int index);
- void removeDuplicates();
-};
-
-#endif
diff --git a/transition_machine/src/ActionBank.cpp b/transition_machine/src/ActionBank.cpp
index 7877255e0fa20fc98bf80ffd89e318247256eab1..f1c28ee8285321524deec997ecc8d063443ea875 100644
--- a/transition_machine/src/ActionBank.cpp
+++ b/transition_machine/src/ActionBank.cpp
@@ -102,8 +102,8 @@ std::vector<Action::BasicAction> ActionBank::str2sequence(const std::string & na
{c.stack.push_back(c.head);};
auto undo = [](Config & c, Action::BasicAction &)
{c.stack.pop_back();};
- auto appliable = [](Config &, Action::BasicAction &)
- {return true;};
+ auto appliable = [](Config & c, Action::BasicAction &)
+ {return c.head < (int)c.tapes[0].ref.size()-1;};
Action::BasicAction basicAction =
{Action::BasicAction::Type::Push, "", apply, undo, appliable};
diff --git a/transition_machine/src/Config.cpp b/transition_machine/src/Config.cpp
index d90f4f6bd68e70a29d1f734e6e0ea0fe3b4597cf..badbbbada01dbc681968b8d02f8996de0cb4c82d 100644
--- a/transition_machine/src/Config.cpp
+++ b/transition_machine/src/Config.cpp
@@ -56,6 +56,9 @@ void Config::readInput(const std::string & filename)
tape.ref.emplace_back();
tape.hyp.resize(tape.ref.size());
+
+ tape.ref.emplace_back("0");
+ tape.hyp.emplace_back("");
}
}
@@ -134,23 +137,23 @@ void Config::printAsOutput(FILE * output)
if(bd.mustPrintLine(j))
lastToPrint = j;
- for (unsigned int i = 0; i < tapes[0].hyp.size(); i++)
- {
+ for (unsigned int i = 0; i < tapes[0].hyp.size() - 1; i++)
for (unsigned int j = 0; j < tapes.size(); j++)
+ {
if(bd.mustPrintLine(j))
fprintf(output, "%s%s", tapes[j][i].empty() ? "0" : tapes[j][i].c_str(), j == lastToPrint ? "\n" : "\t");
- }
+ }
}
void Config::moveHead(int mvt)
{
- if (mvt + head < (int)getTapeByInputCol(0).hyp.size())
+ if (head + mvt < (int)tapes[0].ref.size())
head += mvt;
}
bool Config::isFinal()
{
- return head >= (int)getTapeByInputCol(0).hyp.size() -1;
+ return head >= (int)getTapeByInputCol(0).hyp.size()-1 && stack.empty();
}
void Config::reset()
@@ -210,6 +213,8 @@ void Config::shuffle(const std::string & delimiterTape, const std::string & deli
previousIndex = i+1;
}
+ std::pair<unsigned int, unsigned int> suffix = {delimiters.back().second+1, tape.ref.size()-1};
+
std::random_shuffle(delimiters.begin(), delimiters.end());
std::vector<Tape> newTapes = tapes;
@@ -224,6 +229,12 @@ void Config::shuffle(const std::string & delimiterTape, const std::string & deli
std::copy(tapes[tape].ref.begin()+delimiter.first, tapes[tape].ref.begin()+delimiter.second+1, std::back_inserter(newTapes[tape].ref));
std::copy(tapes[tape].hyp.begin()+delimiter.first, tapes[tape].hyp.begin()+delimiter.second+1, std::back_inserter(newTapes[tape].hyp));
}
+
+ if (suffix.first <= suffix.second)
+ {
+ std::copy(tapes[tape].ref.begin()+suffix.first, tapes[tape].ref.begin()+suffix.second+1, std::back_inserter(newTapes[tape].ref));
+ std::copy(tapes[tape].hyp.begin()+suffix.first, tapes[tape].hyp.begin()+suffix.second+1, std::back_inserter(newTapes[tape].hyp));
+ }
}
tapes = newTapes;
diff --git a/transition_machine/src/Oracle.cpp b/transition_machine/src/Oracle.cpp
index 3e4bb79269bbec035aa5f1d4fb8305312676a9fb..bd262315a32f8dc3ea1402727926935e95877bee 100644
--- a/transition_machine/src/Oracle.cpp
+++ b/transition_machine/src/Oracle.cpp
@@ -86,7 +86,7 @@ void Oracle::createDatabase()
},
[](Config & c, Oracle *, const std::string & action)
{
- return action == "WRITE 0 POS " + c.getTape("POS").ref[c.head];
+ return action == "WRITE 0 POS " + c.getTape("POS").ref[c.head] || c.head >= (int)c.tapes[0].ref.size()-1;
})));
str2oracle.emplace("morpho", std::unique_ptr<Oracle>(new Oracle(
@@ -102,7 +102,7 @@ void Oracle::createDatabase()
},
[](Config & c, Oracle *, const std::string & action)
{
- return action == "WRITE 0 MORPHO " + c.getTape("MORPHO").ref[c.head];
+ return action == "WRITE 0 MORPHO " + c.getTape("MORPHO").ref[c.head] || c.head >= (int)c.tapes[0].ref.size()-1;
})));
str2oracle.emplace("signature", std::unique_ptr<Oracle>(new Oracle(
@@ -211,7 +211,7 @@ void Oracle::createDatabase()
const std::string & lemma = c.getTape("LEMMA").ref[c.head];
std::string rule = getRule(form, lemma);
- return action == std::string("RULE LEMMA ON FORM ") + rule;
+ return action == std::string("RULE LEMMA ON FORM ") + rule || c.head >= (int)c.tapes[0].ref.size()-1;
})));
str2oracle.emplace("parser", std::unique_ptr<Oracle>(new Oracle(
@@ -244,6 +244,8 @@ void Oracle::createDatabase()
sentenceStart++;
while(sentenceEnd < (int)eos.ref.size() && eos.ref[sentenceEnd] != "1")
sentenceEnd++;
+ if (sentenceEnd == (int)eos.ref.size())
+ sentenceEnd--;
auto parts = split(action);
diff --git a/transition_machine/src/TrainingExamples.cpp b/transition_machine/src/TrainingExamples.cpp
deleted file mode 100644
index a887acafb0e1222e975fdeb0a4717003790cb621..0000000000000000000000000000000000000000
--- a/transition_machine/src/TrainingExamples.cpp
+++ /dev/null
@@ -1,72 +0,0 @@
-#include "TrainingExamples.hpp"
-#include <algorithm>
-
-TrainingExamples::TrainingExamples()
-{
- nextIndex = 0;
-}
-
-void TrainingExamples::add(const FeatureModel::FeatureDescription & example, int gold)
-{
- examples.emplace_back(example);
- classes.emplace_back(gold);
- order.emplace_back(order.size());
-}
-
-unsigned int TrainingExamples::size()
-{
- return order.size();
-}
-
-TrainingExamples TrainingExamples::getBatch(unsigned int batchSize)
-{
- TrainingExamples batch;
-
- for(unsigned int i = 0; i < batchSize && (unsigned)nextIndex < order.size()-1; i++)
- {
- batch.add(examples[order[nextIndex]], classes[order[nextIndex]]);
- nextIndex++;
- }
-
- return batch;
-}
-
-void TrainingExamples::reset()
-{
- nextIndex = 0;
-}
-
-void TrainingExamples::shuffle()
-{
- std::random_shuffle(order.begin(), order.end());
-}
-
-void TrainingExamples::removeDuplicates()
-{
- std::map<std::string, int> lastIndex;
- std::map<int, bool> toRemove;
-
- for (unsigned int i = 0; i < examples.size(); i++)
- {
- std::string example = examples[i].toString();
-
- if (lastIndex.count(example))
- toRemove[i] = true;
- else
- lastIndex[example] = i;
- }
-
- for (auto & it : toRemove)
- remove(it.first);
-}
-
-void TrainingExamples::remove(int index)
-{
- for (unsigned int i = 0; i < order.size(); i++)
- if ((int)order[i] == index)
- {
- order[i] = order.back();
- order.pop_back();
- }
-}
-