Skip to content
Snippets Groups Projects
Commit 0f838ba7 authored by Franck Dary's avatar Franck Dary
Browse files

Deleted error correction as a different executable

parent 4d4c6ba3
No related branches found
No related tags found
No related merge requests found
FILE(GLOB SOURCES src/*.cpp)
add_executable(macaon_error_correction src/macaon_error_correction.cpp)
target_link_libraries(macaon_error_correction errors)
target_link_libraries(macaon_error_correction transition_machine)
target_link_libraries(macaon_error_correction ${Boost_PROGRAM_OPTIONS_LIBRARY})
install(TARGETS macaon_error_correction DESTINATION bin)
add_executable(macaon_train_error_detector src/macaon_train_error_detector.cpp)
target_link_libraries(macaon_train_error_detector transition_machine)
target_link_libraries(macaon_train_error_detector ${Boost_PROGRAM_OPTIONS_LIBRARY})
install(TARGETS macaon_train_error_detector DESTINATION bin)
add_executable(macaon_decode_error_detector src/macaon_decode_error_detector.cpp)
target_link_libraries(macaon_decode_error_detector transition_machine)
target_link_libraries(macaon_decode_error_detector ${Boost_PROGRAM_OPTIONS_LIBRARY})
install(TARGETS macaon_decode_error_detector DESTINATION bin)
#compiling library
add_library(errors STATIC ${SOURCES})
/// @file macaon_decode_error_detector.cpp
/// @author Franck Dary
/// @version 1.0
/// @date 2018-12-03
#include <cstdio>
#include <cstdlib>
#include <boost/program_options.hpp>
#include "BD.hpp"
#include "Config.hpp"
#include "TransitionMachine.hpp"
namespace po = boost::program_options;
/// @brief Get the list of mandatory and optional program arguments.
///
/// @return The lists.
po::options_description getOptionsDescription()
{
po::options_description desc("Command-Line Arguments ");
po::options_description req("Required");
req.add_options()
("expName", po::value<std::string>()->required(),
"Name of this experiment")
("tm", po::value<std::string>()->required(),
"File describing the Tape Machine to use")
("bd", po::value<std::string>()->required(),
"BD file that describes the multi-tapes buffer")
("mcd", po::value<std::string>()->required(),
"MCD file that describes the input")
("input,I", po::value<std::string>()->required(),
"Input file formated according to the mcd");
po::options_description opt("Optional");
opt.add_options()
("help,h", "Produce this help message")
("debug,d", "Print infos on stderr")
("printEntropy", "Print entropy for each sequence")
("sequenceDelimiterTape", po::value<std::string>()->default_value("EOS"),
"The name of the buffer's tape that contains the delimiter token for a sequence")
("sequenceDelimiter", po::value<std::string>()->default_value("1"),
"The value of the token that act as a delimiter for sequences")
("showFeatureRepresentation", po::value<int>()->default_value(0),
"For each state of the Config, show its feature representation")
("lang", po::value<std::string>()->default_value("fr"),
"Language you are working with");
desc.add(req).add(opt);
return desc;
}
/// @brief Store the program arguments inside a variables_map
///
/// @param od The description of all the possible options.
/// @param argc The number of arguments given to this program.
/// @param argv The values of arguments given to this program.
///
/// @return The variables map
po::variables_map checkOptions(po::options_description & od, int argc, char ** argv)
{
po::variables_map vm;
try {po::store(po::parse_command_line(argc, argv, od), vm);}
catch(std::exception& e)
{
std::cerr << "Error: " << e.what() << "\n";
od.print(std::cerr);
exit(1);
}
if (vm.count("help"))
{
std::cout << od << "\n";
exit(0);
}
try {po::notify(vm);}
catch(std::exception& e)
{
std::cerr << "Error: " << e.what() << "\n";
od.print(std::cerr);
exit(1);
}
return vm;
}
/// @brief Uses a pre-trained TransitionMachine to predict and add information to a structured input file.
///
/// @param argc The number of arguments given to this program.
/// @param argv[] Array of arguments given to this program.
///
/// @return 0 if there was no crash.
int main(int argc, char * argv[])
{
auto od = getOptionsDescription();
po::variables_map vm = checkOptions(od, argc, argv);
ProgramParameters::expName = vm["expName"].as<std::string>();
ProgramParameters::tmName = vm["tm"].as<std::string>();
ProgramParameters::bdName = vm["bd"].as<std::string>();
ProgramParameters::input = vm["input"].as<std::string>();
ProgramParameters::mcdName = vm["mcd"].as<std::string>();
ProgramParameters::debug = vm.count("debug") == 0 ? false : true;
ProgramParameters::printEntropy = vm.count("printEntropy") == 0 ? false : true;
ProgramParameters::lang = vm["lang"].as<std::string>();
ProgramParameters::sequenceDelimiterTape = vm["sequenceDelimiterTape"].as<std::string>();
ProgramParameters::sequenceDelimiter = vm["sequenceDelimiter"].as<std::string>();
ProgramParameters::showFeatureRepresentation = vm["showFeatureRepresentation"].as<int>();
const char * MACAON_DIR = std::getenv("MACAON_DIR");
std::string slash = "/";
ProgramParameters::expPath = MACAON_DIR + slash + ProgramParameters::lang + slash + "bin/" + ProgramParameters::expName + slash;
ProgramParameters::tmFilename = ProgramParameters::expPath + ProgramParameters::tmName;
ProgramParameters::bdFilename = ProgramParameters::expPath + ProgramParameters::bdName;
ProgramParameters::mcdFilename = ProgramParameters::mcdName;
TransitionMachine tm(false);
BD bd(ProgramParameters::bdFilename, ProgramParameters::mcdFilename);
Config config(bd);
File input(ProgramParameters::input, "r");
FILE * inputPtr = input.getDescriptor();
int isError, errorIndex;
while (fscanf(inputPtr, "%d\t%d\n", &isError, &errorIndex) == 2)
{
config.loadFromFile(input);
TransitionMachine::State * currentState = tm.getCurrentState();
Classifier * classifier = currentState->classifier;
config.setCurrentStateName(&currentState->name);
Dict::currentClassifierName = classifier->name;
classifier->initClassifier(config);
auto weightedActions = classifier->weightActions(config);
std::string pAction = "";
for (auto & it : weightedActions)
if (it.first)
if (pAction == "")
{
pAction = it.second.second;
break;
}
Action * action = classifier->getAction(pAction);
action->apply(config);
}
return 0;
}
/// @file macaon_error_correction.cpp
/// @author Franck Dary
/// @version 1.0
/// @date 2018-11-27
#include <cstdio>
#include <cstdlib>
#include <boost/program_options.hpp>
#include "BD.hpp"
#include "Config.hpp"
#include "TransitionMachine.hpp"
#include "util.hpp"
#include "Error.hpp"
#include "ActionBank.hpp"
namespace po = boost::program_options;
/// @brief Get the list of mandatory and optional program arguments.
///
/// @return The lists.
po::options_description getOptionsDescription()
{
po::options_description desc("Command-Line Arguments ");
po::options_description req("Required");
req.add_options()
("expName", po::value<std::string>()->required(),
"Name of this experiment")
("classifier", po::value<std::string>()->required(),
"Name of the monitored classifier")
("tm", po::value<std::string>()->required(),
"File describing the Tape Machine to use")
("bd", po::value<std::string>()->required(),
"BD file that describes the multi-tapes buffer")
("mcd", po::value<std::string>()->required(),
"MCD file that describes the input")
("input,I", po::value<std::string>()->required(),
"Input file formated according to the mcd");
po::options_description opt("Optional");
opt.add_options()
("help,h", "Produce this help message")
("debug,d", "Print infos on stderr")
("sequenceDelimiterTape", po::value<std::string>()->default_value("EOS"),
"The name of the buffer's tape that contains the delimiter token for a sequence")
("sequenceDelimiter", po::value<std::string>()->default_value("1"),
"The value of the token that act as a delimiter for sequences")
("lang", po::value<std::string>()->default_value("fr"),
"Language you are working with");
desc.add(req).add(opt);
return desc;
}
/// @brief Store the program arguments inside a variables_map
///
/// @param od The description of all the possible options.
/// @param argc The number of arguments given to this program.
/// @param argv The values of arguments given to this program.
///
/// @return The variables map
po::variables_map checkOptions(po::options_description & od, int argc, char ** argv)
{
po::variables_map vm;
try {po::store(po::parse_command_line(argc, argv, od), vm);}
catch(std::exception& e)
{
std::cerr << "Error: " << e.what() << "\n";
od.print(std::cerr);
exit(1);
}
if (vm.count("help"))
{
std::cout << od << "\n";
exit(0);
}
try {po::notify(vm);}
catch(std::exception& e)
{
std::cerr << "Error: " << e.what() << "\n";
od.print(std::cerr);
exit(1);
}
return vm;
}
/// @brief Uses a pre-trained TransitionMachine to output a pair of Config - labels. That can be used as a corpus for error detection.
///
/// @param argc The number of arguments given to this program.
/// @param argv[] Array of arguments given to this program.
///
/// @return 0 if there was no crash.
int main(int argc, char * argv[])
{
auto od = getOptionsDescription();
po::variables_map vm = checkOptions(od, argc, argv);
ProgramParameters::expName = vm["expName"].as<std::string>();
ProgramParameters::tmName = vm["tm"].as<std::string>();
ProgramParameters::bdName = vm["bd"].as<std::string>();
ProgramParameters::input = vm["input"].as<std::string>();
ProgramParameters::mcdName = vm["mcd"].as<std::string>();
ProgramParameters::debug = vm.count("debug") == 0 ? false : true;
ProgramParameters::lang = vm["lang"].as<std::string>();
ProgramParameters::sequenceDelimiterTape = vm["sequenceDelimiterTape"].as<std::string>();
ProgramParameters::sequenceDelimiter = vm["sequenceDelimiter"].as<std::string>();
ProgramParameters::classifierName = vm["classifier"].as<std::string>();
const char * MACAON_DIR = std::getenv("MACAON_DIR");
std::string slash = "/";
ProgramParameters::expPath = MACAON_DIR + slash + ProgramParameters::lang + slash + "bin/" + ProgramParameters::expName + slash;
ProgramParameters::tmFilename = ProgramParameters::expPath + ProgramParameters::tmName;
ProgramParameters::bdFilename = ProgramParameters::expPath + ProgramParameters::bdName;
ProgramParameters::mcdFilename = ProgramParameters::mcdName;
TransitionMachine tm(false);
BD bd(ProgramParameters::bdFilename, ProgramParameters::mcdFilename);
Config config(bd);
config.readInput(ProgramParameters::input);
float entropyAccumulator = 0.0;
int nbActionsInSequence = 0;
bool justFlipped = false;
bool configIsError = false;
int actionIndex = 0;
int errorIndex = 0;
Errors errors;
errors.newSequence();
while (!config.isFinal())
{
TransitionMachine::State * currentState = tm.getCurrentState();
Classifier * classifier = currentState->classifier;
config.setCurrentStateName(&currentState->name);
Dict::currentClassifierName = classifier->name;
if (ProgramParameters::debug)
{
config.printForDebug(stderr);
fprintf(stderr, "State : \'%s\'\n", currentState->name.c_str());
}
auto weightedActions = classifier->weightActions(config);
if (ProgramParameters::debug)
{
Classifier::printWeightedActions(stderr, weightedActions);
fprintf(stderr, "\n");
}
std::string & predictedAction = weightedActions[0].second.second;
Action * action = classifier->getAction(predictedAction);
for(unsigned int i = 0; i < weightedActions.size(); i++)
{
predictedAction = weightedActions[i].second.second;
action = classifier->getAction(predictedAction);
if(weightedActions[i].first)
break;
}
if(!action->appliable(config))
{
// First case the analysis is finished but without an empty stack
if (config.head == (int)config.tapes[0].ref.size()-1)
{
while (!config.stackEmpty())
config.stackPop();
continue;
}
else
{
fprintf(stderr, "ERROR (%s) : action \'%s\' is not appliable. Aborting\n", ERRINFO, predictedAction.c_str());
exit(1);
}
}
if (classifier->name == ProgramParameters::classifierName)
{
//fprintf(stderr, "%d\t%d\n", configIsError ? 1 : 0, errorIndex - actionIndex);
//config.printAsExample(stderr);
actionIndex++;
auto zeroCostActions = classifier->getZeroCostActions(config);
bool pActionIsZeroCost = false;
for (auto & s : zeroCostActions)
if (s == action->name)
{
pActionIsZeroCost = true;
break;
}
int windowSize = 5;
if (!pActionIsZeroCost)
{
if (!configIsError || (actionIndex - errorIndex > windowSize))
{
configIsError = true;
errorIndex = actionIndex-1;
}
}
else if (configIsError && (actionIndex - errorIndex > windowSize))
{
configIsError = false;
errorIndex = 0;
}
if (configIsError)
{
errors.add({action->name, zeroCostActions[0], weightedActions, classifier->getActionCost(config, action->name), ActionBank::getLinkLength(config, action->name), ActionBank::getLinkLength(config, zeroCostActions[0])});
}
}
action->apply(config);
TransitionMachine::Transition * transition = tm.getTransition(predictedAction);
tm.takeTransition(transition);
config.moveHead(transition->headMvt);
if (true)
{
nbActionsInSequence++;
float entropy = Classifier::computeEntropy(weightedActions);
config.addToEntropyHistory(entropy);
entropyAccumulator += entropy;
if (config.head >= 1 && config.getTape(ProgramParameters::sequenceDelimiterTape)[config.head-1] != ProgramParameters::sequenceDelimiter)
justFlipped = false;
if ((config.head >= 1 && config.getTape(ProgramParameters::sequenceDelimiterTape)[config.head-1] == ProgramParameters::sequenceDelimiter && !justFlipped))
{
justFlipped = true;
entropyAccumulator = 0.0;
errors.newSequence();
configIsError = false;
errorIndex = 0;
}
}
}
errors.printStats();
return 0;
}
/// @file macaon_train.cpp
/// @author Franck Dary
/// @version 1.0
/// @date 2018-08-07
#include <cstdio>
#include <cstdlib>
#include <boost/program_options.hpp>
#include "BD.hpp"
#include "Config.hpp"
#include "TransitionMachine.hpp"
#include "Trainer.hpp"
#include "ProgramParameters.hpp"
namespace po = boost::program_options;
/// @brief Get the list of mandatory and optional program arguments.
///
/// @return The lists.
po::options_description getOptionsDescription()
{
po::options_description desc("Command-Line Arguments ");
po::options_description req("Required");
req.add_options()
("expName", po::value<std::string>()->required(),
"Name of this experiment")
("templateName", po::value<std::string>()->required(),
"Name of the template folder")
("tm", po::value<std::string>()->required(),
"File describing the Tape Machine we will train")
("bd", po::value<std::string>()->required(),
"BD file that describes the multi-tapes buffer")
("mcd", po::value<std::string>()->required(),
"MCD file that describes the input")
("train,T", po::value<std::string>()->required(),
"Training corpus formated according to the MCD")
("dev", po::value<std::string>()->default_value(""),
"Development corpus formated according to the MCD");
po::options_description opt("Optional");
opt.add_options()
("help,h", "Produce this help message")
("debug,d", "Print infos on stderr")
("printEntropy", "Print mean entropy and standard deviation accross sequences")
("optimizer", po::value<std::string>()->default_value("amsgrad"),
"The learning algorithm to use : amsgrad | adam | sgd")
("loss", po::value<std::string>()->default_value("neglogsoftmax"),
"The loss function to use : neglogsoftmax | weighted")
("lang", po::value<std::string>()->default_value("fr"),
"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),
"The random seed that will initialize RNG")
("batchSize", po::value<int>()->default_value(50),
"The size of each minibatch (in number of taining examples)")
("nbTrain", po::value<int>()->default_value(0),
"The number of models that will be trained, with only the random seed changing")
("duplicates", po::value<bool>()->default_value(true),
"Remove identical training examples")
("showFeatureRepresentation", po::value<int>()->default_value(0),
"For each state of the Config, show its feature representation")
("interactive", po::value<bool>()->default_value(true),
"Is the shell interactive ? Display advancement informations")
("randomEmbeddings", po::value<bool>()->default_value(false),
"When activated, the embeddings will be randomly initialized")
("randomParameters", po::value<bool>()->default_value(true),
"When activated, the parameters will be randomly initialized")
("sequenceDelimiterTape", po::value<std::string>()->default_value("EOS"),
"The name of the buffer's tape that contains the delimiter token for a sequence")
("sequenceDelimiter", po::value<std::string>()->default_value("1"),
"The value of the token that act as a delimiter for sequences")
("printTime", "Print time on stderr")
("shuffle", po::value<bool>()->default_value(true),
"Shuffle examples after each iteration");
po::options_description oracle("Oracle related options");
oracle.add_options()
("epochd", po::value<int>()->default_value(3),
"Number of the first epoch where the oracle will be dynamic")
("proba", po::value<float>()->default_value(0.9),
"The probability that the dynamic oracle will chose the predicted action");
po::options_description ams("Amsgrad family optimizers");
ams.add_options()
("b1", po::value<float>()->default_value(0.9),
"beta1 parameter for the Amsgtad or Adam optimizer")
("b2", po::value<float>()->default_value(0.999),
"beta2 parameter for the Amsgtad or Adam optimizer")
("bias", po::value<float>()->default_value(1e-8),
"bias parameter for the Amsgtad or Adam or Adagrad optimizer");
desc.add(req).add(opt).add(oracle).add(ams);
return desc;
}
/// @brief Store the program arguments inside a variables_map
///
/// @param od The description of all the possible options.
/// @param argc The number of arguments given to this program.
/// @param argv The values of arguments given to this program.
///
/// @return The variables map
po::variables_map checkOptions(po::options_description & od, int argc, char ** argv)
{
po::variables_map vm;
try {po::store(po::parse_command_line(argc, argv, od), vm);}
catch(std::exception& e)
{
std::cerr << "Error: " << e.what() << "\n";
od.print(std::cerr);
exit(1);
}
if (vm.count("help"))
{
std::cout << od << "\n";
exit(0);
}
try {po::notify(vm);}
catch(std::exception& e)
{
std::cerr << "Error: " << e.what() << "\n";
od.print(std::cerr);
exit(1);
}
return vm;
}
/// @brief Set all the usefull paths relative to expPath
void updatePaths()
{
const char * MACAON_DIR = std::getenv("MACAON_DIR");
std::string slash = "/";
ProgramParameters::langPath = MACAON_DIR + slash + ProgramParameters::lang + slash;
ProgramParameters::expPath = ProgramParameters::langPath + "bin/" + ProgramParameters::expName + slash;
ProgramParameters::templatePath = ProgramParameters::langPath + ProgramParameters::templateName + slash;
ProgramParameters::tmFilename = ProgramParameters::expPath + ProgramParameters::tmName;
ProgramParameters::bdFilename = ProgramParameters::expPath + ProgramParameters::bdName;
ProgramParameters::mcdFilename = ProgramParameters::expPath + ProgramParameters::mcdName;
ProgramParameters::trainFilename = ProgramParameters::expPath + ProgramParameters::trainName;
ProgramParameters::devFilename = ProgramParameters::expPath + ProgramParameters::devName;
ProgramParameters::newTemplatePath = ProgramParameters::langPath + "bin/" + ProgramParameters::baseExpName + slash;
}
/// @brief Create the folder containing the current experiment from the template frolder
void createExpPath()
{
std::string decode = "\
#! /bin/bash\n\
\n\
if [ \"$#\" -lt 2 ]; then\n\
echo \"Usage : $0 input mcd\"\n\
exit\n\
fi\n\
\n\
INPUT=$1\n\
MCD=$2\n\
\n\
shift\n\
shift\n\
ARGS=\"\"\n\
for arg in \"$@\"\n\
do\n\
ARGS=\"$ARGS $arg\"\n\
done\n\
\n\
macaon_decode --lang " + ProgramParameters::lang + " --tm machine.tm --bd test.bd -I $INPUT --mcd $MCD --expName " + ProgramParameters::expName + "$ARGS";
if (system(("rm -r " + ProgramParameters::expPath + " 2> /dev/null").c_str())){}
if (system(("mkdir " + ProgramParameters::expPath).c_str())){}
if (system(("cp -r " + ProgramParameters::newTemplatePath + "* " + ProgramParameters::expPath + ".").c_str())){}
if (system(("echo \'" + decode + "\' > " + ProgramParameters::expPath + "decode.sh").c_str())){}
if (system(("chmod +x " + ProgramParameters::expPath + "decode.sh").c_str())){}
if (system(("ln -f -s " + ProgramParameters::expPath + "decode.sh " + ProgramParameters::langPath + "bin/maca_tm_" + ProgramParameters::expName).c_str())){}
}
std::map<std::string, std::pair<float, std::pair<float, float> > > getScoreOnDev(TransitionMachine & tm, std::vector<int> & devIsErrors, std::vector<int> &, File & dev, Config & devConfig)
{
dev.rewind();
FILE * devPtr = dev.getDescriptor();
tm.reset();
std::map< std::string, std::pair<int, int> > counts;
if (ProgramParameters::debug)
fprintf(stderr, "Computing score on dev set\n");
std::vector<int> predictions;
std::string classifierName;
int isError, errorIndex;
for (unsigned int i = 0; i < devIsErrors.size(); i++)
{
if (fscanf(devPtr, "%d\t%d\n", &isError, &errorIndex) != 2)
{
fprintf(stderr, "ERROR (%s) : corpus bad format. Aborting.\n", ERRINFO);
exit(1);
}
devConfig.loadFromFile(dev);
TransitionMachine::State * currentState = tm.getCurrentState();
Classifier * classifier = currentState->classifier;
devConfig.setCurrentStateName(&currentState->name);
Dict::currentClassifierName = classifier->name;
classifier->initClassifier(devConfig);
auto weightedActions = classifier->weightActions(devConfig);
std::string pAction = "";
for (auto & it : weightedActions)
if (it.first)
{
pAction = it.second.second;
break;
}
predictions.emplace_back(pAction == "ERROR" ? 1 : 0);
classifierName = classifier->name;
}
int pred1Hyp0 = 0;
int pred0Hyp1 = 0;
int pred0Hyp0 = 0;
int pred1Hyp1 = 0;
for (unsigned int i = 0; i < devIsErrors.size(); i++)
{
if (devIsErrors[i] == 0)
{
counts[classifierName].first++;
if (predictions[i] == 0)
{
counts[classifierName].second++;
pred0Hyp0++;
}
else
pred1Hyp0++;
}
else if (i > 0 && devIsErrors[i] == 1 && devIsErrors[i-1] == 0)
{
counts[classifierName].first++;
unsigned int j;
bool found = false;
for (j = i; devIsErrors[j] == 1 && j < devIsErrors.size(); j++)
{
if (predictions[j] == 1)
{
counts[classifierName].second++;
pred1Hyp1++;
found = true;
break;
}
}
i = j;
if (!found)
pred0Hyp1++;
}
}
int nbErrorsIntroduced = pred1Hyp0;
int nbErrorsCorrected = pred1Hyp1;
fprintf(stderr, "\nClass 0 nbExemples : %d\n", pred0Hyp0+pred1Hyp0);
fprintf(stderr, "Class 0 precision : %.2f%%\n", 100.0*pred0Hyp0 / (pred0Hyp0+pred0Hyp1));
fprintf(stderr, "Class 0 recall : %.2f%%\n\n", 100.0*pred0Hyp0 / (pred0Hyp0+pred1Hyp0));
fprintf(stderr, "Class 1 nbExemples : %d\n", pred0Hyp1+pred1Hyp1);
fprintf(stderr, "Class 1 precision : %.2f%%\n", 100.0*pred1Hyp1 / (pred1Hyp1+pred1Hyp0));
fprintf(stderr, "Class 1 recall : %.2f%%\n\n", 100.0*pred1Hyp1 / (pred1Hyp1+pred0Hyp1));
fprintf(stderr, "Nb errors introduced : %d\n", nbErrorsIntroduced);
fprintf(stderr, "Nb errors corrected : %d\n", nbErrorsCorrected);
fprintf(stderr, "Difference : %d\n", nbErrorsCorrected-nbErrorsIntroduced);
std::map<std::string, std::pair<float,std::pair<float,float> > > scores;
for (auto & it : counts)
scores[it.first].first = 100.0*pred1Hyp1 / (pred1Hyp1+pred1Hyp0);
return scores;
}
void printScoresAndSave(FILE * output, std::map< std::string, std::pair<int, int> > & trainCounter, std::map< std::string, float > & scores, TransitionMachine & tm, int curIter, std::map< std::string, float > & bestScores, std::vector<int> & devIsErrors, std::vector<int> & devErrorIndexes, File & devFile, Config & config, float totalLoss)
{
for (auto & it : trainCounter)
scores[it.first] = 100.0 * it.second.second / it.second.first;
std::vector<std::string> names;
std::vector<std::string> acc;
std::vector<std::string> train;
std::vector<std::string> dev;
std::vector<std::string> savedStr;
std::map<std::string, bool> saved;
auto devScores = getScoreOnDev(tm, devIsErrors, devErrorIndexes, devFile, config);
for (auto & it : devScores)
{
if (bestScores.count(it.first) == 0 || bestScores[it.first] < it.second.first)
{
bestScores[it.first] = it.second.first;
saved[it.first] = true;
}
else
saved[it.first] = false;
}
auto classifiers = tm.getClassifiers();
for (auto * cla : classifiers)
{
if (!saved.count(cla->name))
continue;
if (saved[cla->name])
{
cla->save(ProgramParameters::expPath + cla->name + ".model");
Dict::saveDicts(ProgramParameters::expPath, cla->name);
}
}
for (auto & it : saved)
{
names.emplace_back(it.first);
acc.emplace_back("accuracy");
train.emplace_back(": train(" + float2str(scores[it.first], "%.2f") + "%)");
dev.emplace_back("dev(" +float2str(devScores[it.first].first, "%.2f") + "%)");
savedStr.emplace_back(saved[it.first] ? "SAVED" : "");
if (ProgramParameters::printEntropy)
savedStr.back() += " Entropy[" + float2str(devScores[it.first].second.first, "%.2f") + "\u00B1" + float2str(devScores[it.first].second.second, "%.2f") + "]";
savedStr.back() += " Loss[" + float2str(totalLoss, "%.2f") + "]";
}
if (ProgramParameters::interactive)
fprintf(stderr, " \r");
if (ProgramParameters::printTime)
fprintf(output, "[%s] ", getTime().c_str());
fprintf(output, "Iteration %d/%d :\n", curIter+1, ProgramParameters::nbIter);
printColumns(output, {names, acc, train, dev, savedStr});
}
/// @brief Train a model according to all the ProgramParameters
void launchTraining()
{
std::map< std::string, float > scores;
std::map< std::string, float > bestScores;
TransitionMachine tm(true);
BD trainBD(ProgramParameters::bdFilename, ProgramParameters::mcdFilename);
File train(ProgramParameters::expPath + ProgramParameters::trainName, "r");
FILE * trainPtr = train.getDescriptor();
File dev(ProgramParameters::expPath + ProgramParameters::devName, "r");
FILE * devPtr = dev.getDescriptor();
Dict::createFiles(ProgramParameters::expPath, "");
fprintf(stderr, "%sTraining of \'%s\' :\n",
ProgramParameters::printTime ? ("["+getTime()+"] ").c_str() : "",
tm.name.c_str());
std::map< std::string, bool > topologyPrinted;
std::map< std::string, std::pair<int, int> > trainCounter;
int curIter = 0;
std::vector<int> isErrors;
std::vector<int> errorIndexes;
std::vector<int> devIsErrors;
std::vector<int> devErrorIndexes;
int isError;
int errorIndex;
Config config(trainBD);
fprintf(stderr, "Reading train corpus...");
while (fscanf(trainPtr, "%d\t%d\n", &isError, &errorIndex) == 2)
{
isErrors.emplace_back(isError);
errorIndexes.emplace_back(errorIndex);
config.loadFromFile(train);
}
fprintf(stderr, " done !\n");
fprintf(stderr, "Reading dev corpus...");
while (fscanf(devPtr, "%d\t%d\n", &isError, &errorIndex) == 2)
{
devIsErrors.emplace_back(isError);
devErrorIndexes.emplace_back(errorIndex);
config.loadFromFile(dev);
}
fprintf(stderr, " done !\n");
float totalLoss = 0.0;
auto resetAndShuffle = [&trainCounter,&train,&dev,&trainPtr,&totalLoss]()
{
train.rewind();
dev.rewind();
trainPtr = train.getDescriptor();
for (auto & it : trainCounter)
it.second.first = it.second.second = 0;
totalLoss = 0.0;
};
Config trainConfig(trainBD);
while (curIter < ProgramParameters::nbIter)
{
resetAndShuffle();
for (unsigned int i = 0; i < isErrors.size(); i++)
{
if (fscanf(trainPtr, "%d\t%d\n", &isError, &errorIndex) != 2)
{
fprintf(stderr, "ERROR (%s) : corpus bad format. Aborting.\n", ERRINFO);
exit(1);
}
trainConfig.loadFromFile(train);
TransitionMachine::State * currentState = tm.getCurrentState();
Classifier * classifier = currentState->classifier;
trainConfig.setCurrentStateName(&currentState->name);
Dict::currentClassifierName = classifier->name;
classifier->initClassifier(trainConfig);
if (!topologyPrinted.count(classifier->name))
{
topologyPrinted[classifier->name] = true;
classifier->printTopology(stderr);
}
// Print current iter advancement in percentage
if (ProgramParameters::interactive)
{
int totalSize = isErrors.size();
int steps = i;
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 = "";
for (auto & it : weightedActions)
if (it.first)
if (pAction == "")
{
pAction = it.second.second;
break;
}
std::string oAction = isError ? "ERROR" : "CORRECT";
totalLoss += classifier->trainOnExample(trainConfig, classifier->getActionIndex(oAction));
trainCounter[classifier->name].first++;
trainCounter[classifier->name].second += pAction == oAction ? 1 : 0;
}
printScoresAndSave(stderr, trainCounter, scores, tm, curIter, bestScores, devIsErrors, devErrorIndexes, dev, config, totalLoss);
curIter++;
}
}
void createTemplatePath()
{
if (system(("rm -r " + ProgramParameters::newTemplatePath + " 2> /dev/null").c_str())){}
if (system(("mkdir " + ProgramParameters::newTemplatePath).c_str())){}
if (system(("cp -r " + ProgramParameters::templatePath + "* " + ProgramParameters::newTemplatePath + ".").c_str())){}
}
void removeTemplatePath()
{
if (system(("rm -r " + ProgramParameters::newTemplatePath + " 2> /dev/null").c_str())){}
}
/// @brief Train a TransitionMachine to predict and add information to a structured input file, by using annotated examples.
///
/// @param argc The number of arguments given to this program.
/// @param argv[] Array of arguments given to this program.
///
/// @return 0 if there was no crash.
int main(int argc, char * argv[])
{
auto od = getOptionsDescription();
po::variables_map vm = checkOptions(od, argc, argv);
ProgramParameters::expName = vm["expName"].as<std::string>();
ProgramParameters::baseExpName = ProgramParameters::expName;
ProgramParameters::templateName = vm["templateName"].as<std::string>();
ProgramParameters::tmName = vm["tm"].as<std::string>();
ProgramParameters::bdName = vm["bd"].as<std::string>();
ProgramParameters::mcdName = vm["mcd"].as<std::string>();
ProgramParameters::debug = vm.count("debug") == 0 ? false : true;
ProgramParameters::printEntropy = vm.count("printEntropy") == 0 ? false : true;
ProgramParameters::printTime = vm.count("printTime") == 0 ? false : true;
ProgramParameters::trainName = vm["train"].as<std::string>();
ProgramParameters::devName = vm["dev"].as<std::string>();
ProgramParameters::lang = vm["lang"].as<std::string>();
ProgramParameters::nbIter = vm["nbiter"].as<int>();
ProgramParameters::seed = vm["seed"].as<int>();
ProgramParameters::batchSize = vm["batchSize"].as<int>();
ProgramParameters::nbTrain = vm["nbTrain"].as<int>();
ProgramParameters::removeDuplicates = vm["duplicates"].as<bool>();
ProgramParameters::interactive = vm["interactive"].as<bool>();
ProgramParameters::shuffleExamples = vm["shuffle"].as<bool>();
ProgramParameters::randomEmbeddings = vm["randomEmbeddings"].as<bool>();
ProgramParameters::randomParameters = vm["randomParameters"].as<bool>();
ProgramParameters::sequenceDelimiterTape = vm["sequenceDelimiterTape"].as<std::string>();
ProgramParameters::sequenceDelimiter = vm["sequenceDelimiter"].as<std::string>();
ProgramParameters::learningRate = vm["lr"].as<float>();
ProgramParameters::beta1 = vm["b1"].as<float>();
ProgramParameters::beta2 = vm["b2"].as<float>();
ProgramParameters::bias = vm["bias"].as<float>();
ProgramParameters::optimizer = vm["optimizer"].as<std::string>();
ProgramParameters::dynamicEpoch = vm["epochd"].as<int>();
ProgramParameters::loss = vm["loss"].as<std::string>();
ProgramParameters::dynamicProbability = vm["proba"].as<float>();
ProgramParameters::showFeatureRepresentation = vm["showFeatureRepresentation"].as<int>();
ProgramParameters::iterationSize = vm["iterationSize"].as<int>();
if (ProgramParameters::nbTrain)
{
updatePaths();
createTemplatePath();
for (int i = 0; i < ProgramParameters::nbTrain; i++)
{
fprintf(stderr, "Training number %d / %d :\n", i+1, ProgramParameters::nbTrain);
ProgramParameters::expName = ProgramParameters::baseExpName + "_" + std::to_string(i);
updatePaths();
createExpPath();
Dict::deleteDicts();
launchTraining();
}
removeTemplatePath();
}
else
{
updatePaths();
ProgramParameters::newTemplatePath = ProgramParameters::templatePath;
createExpPath();
Dict::deleteDicts();
launchTraining();
}
return 0;
}
0% Loading or .
You are about to add 0 people to the discussion. Proceed with caution.
Please register or to comment