mosesdecoder/mert/Data.cpp
Jeroen Vermeulen 38d790cac0 Add cross-platform randomizer module.
The code uses two mechanisms for generating random numbers: srand()/rand(),
which is not thread-safe, and srandom()/random(), which is POSIX-specific.

Here I add a util/random.cc module that centralizes these calls, and unifies
some common usage patterns.  If the implementation is not good enough, we can
now change it in a single place.

To keep things simple, this uses the portable srand()/rand() but protects them
with a lock to avoid concurrency problems.

The hard part was to keep the regression tests passing: they rely on fixed
sequences of random numbers, so a small code change could break them very
thoroughly.  Util::rand(), for wide types like size_t, calls std::rand() not
once but twice.  This behaviour was generalized into utils::wide_rand() and
friends.
2015-04-23 23:46:04 +07:00

308 lines
8.4 KiB
C++

/*
* Data.cpp
* mert - Minimum Error Rate Training
*
* Created by Nicola Bertoldi on 13/05/08.
*
*/
#include <algorithm>
#include <cmath>
#include <fstream>
#include "Data.h"
#include "Scorer.h"
#include "ScorerFactory.h"
#include "Util.h"
#include "util/exception.hh"
#include "util/file_piece.hh"
#include "util/random.hh"
#include "util/tokenize_piece.hh"
#include "util/string_piece.hh"
#include "FeatureDataIterator.h"
using namespace std;
namespace MosesTuning
{
Data::Data(Scorer* scorer, const string& sparse_weights_file)
: m_scorer(scorer),
m_score_type(m_scorer->getName()),
m_num_scores(0),
m_score_data(new ScoreData(m_scorer)),
m_feature_data(new FeatureData)
{
TRACE_ERR("Data::m_score_type " << m_score_type << endl);
TRACE_ERR("Data::Scorer type from Scorer: " << m_scorer->getName() << endl);
if (sparse_weights_file.size()) {
m_sparse_weights.load(sparse_weights_file);
ostringstream msg;
msg << "Data::sparse_weights {";
m_sparse_weights.write(msg,"=");
msg << "}";
TRACE_ERR(msg.str() << std::endl);
}
}
//ADDED BY TS
// TODO: This is too long; consider creating additional functions to
// reduce the lines of this function.
void Data::removeDuplicates()
{
size_t nSentences = m_feature_data->size();
assert(m_score_data->size() == nSentences);
for (size_t s = 0; s < nSentences; s++) {
FeatureArray& feat_array = m_feature_data->get(s);
ScoreArray& score_array = m_score_data->get(s);
assert(feat_array.size() == score_array.size());
//serves as a hash-map:
map<double, vector<size_t> > lookup;
size_t end_pos = feat_array.size() - 1;
size_t nRemoved = 0;
for (size_t k = 0; k <= end_pos; k++) {
const FeatureStats& cur_feats = feat_array.get(k);
double sum = 0.0;
for (size_t l = 0; l < cur_feats.size(); l++)
sum += cur_feats.get(l);
if (lookup.find(sum) != lookup.end()) {
//cerr << "hit" << endl;
vector<size_t>& cur_list = lookup[sum];
// TODO: Make sure this is correct because we have already used 'l'.
// If this does not impact on the removing duplicates, it is better
// to change
size_t l = 0;
for (l = 0; l < cur_list.size(); l++) {
size_t j = cur_list[l];
if (cur_feats == feat_array.get(j)
&& score_array.get(k) == score_array.get(j)) {
if (k < end_pos) {
feat_array.swap(k,end_pos);
score_array.swap(k,end_pos);
k--;
}
end_pos--;
nRemoved++;
break;
}
}
if (l == lookup[sum].size())
cur_list.push_back(k);
} else {
lookup[sum].push_back(k);
}
// for (size_t j=0; j < k; j++) {
// if (feat_array.get(k) == feat_array.get(j)
// && score_array.get(k) == score_array.get(j)) {
// if (k < end_pos) {
// feat_array.swap(k,end_pos);
// score_array.swap(k,end_pos);
// k--;
// }
// end_pos--;
// nRemoved++;
// break;
// }
// }
} // end for k
if (nRemoved > 0) {
feat_array.resize(end_pos+1);
score_array.resize(end_pos+1);
}
}
}
//END_ADDED
void Data::load(const std::string &featfile, const std::string &scorefile)
{
m_feature_data->load(featfile, m_sparse_weights);
m_score_data->load(scorefile);
}
void Data::loadNBest(const string &file, bool oneBest)
{
TRACE_ERR("loading nbest from " << file << endl);
util::FilePiece in(file.c_str());
ScoreStats scoreentry;
string sentence, feature_str, alignment;
int sentence_index;
while (true) {
try {
StringPiece line = in.ReadLine();
if (line.empty()) continue;
// adding statistics for error measures
scoreentry.clear();
util::TokenIter<util::MultiCharacter> it(line, util::MultiCharacter("|||"));
sentence_index = ParseInt(*it);
if (oneBest && m_score_data->exists(sentence_index)) continue;
++it;
sentence = it->as_string();
++it;
feature_str = it->as_string();
++it;
if (it) {
++it; // skip model score.
if (it) {
alignment = it->as_string(); //fifth field (if present) is either phrase or word alignment
++it;
if (it) {
alignment = it->as_string(); //sixth field (if present) is word alignment
}
}
}
//TODO check alignment exists if scorers need it
if (m_scorer->useAlignment()) {
sentence += "|||";
sentence += alignment;
}
m_scorer->prepareStats(sentence_index, sentence, scoreentry);
m_score_data->add(scoreentry, sentence_index);
// examine first line for name of features
if (!existsFeatureNames()) {
InitFeatureMap(feature_str);
}
AddFeatures(feature_str, sentence_index);
} catch (util::EndOfFileException &e) {
PrintUserTime("Loaded N-best lists");
break;
}
}
}
void Data::save(const std::string &featfile, const std::string &scorefile, bool bin)
{
if (bin)
cerr << "Binary write mode is selected" << endl;
else
cerr << "Binary write mode is NOT selected" << endl;
m_feature_data->save(featfile, bin);
m_score_data->save(scorefile, bin);
}
void Data::InitFeatureMap(const string& str)
{
string buf = str;
string substr;
string features = "";
string tmp_name = "";
size_t tmp_index = 0;
while (!buf.empty()) {
getNextPound(buf, substr);
// string ending with "=" are skipped, because they are the names of the features
if (!EndsWith(substr, "=")) {
stringstream ss;
ss << tmp_name << "_" << tmp_index << " ";
features.append(ss.str());
tmp_index++;
} else if (substr.find("_") != string::npos) {
// ignore sparse feature name and its value
getNextPound(buf, substr);
} else { // update current feature name
tmp_index = 0;
tmp_name = substr.substr(0, substr.size() - 1);
}
}
m_feature_data->setFeatureMap(features);
}
void Data::AddFeatures(const string& str,
int sentence_index)
{
string buf = str;
string substr;
FeatureStats feature_entry;
feature_entry.reset();
while (!buf.empty()) {
getNextPound(buf, substr);
// no ':' -> feature value that needs to be stored
if (!EndsWith(substr, "=")) {
feature_entry.add(ConvertStringToFeatureStatsType(substr));
} else if (substr.find("_") != string::npos) {
// sparse feature name? store as well
string name = substr;
getNextPound(buf, substr);
feature_entry.addSparse(name, atof(substr.c_str()));
}
}
m_feature_data->add(feature_entry, sentence_index);
}
void Data::createShards(size_t shard_count, float shard_size, const string& scorerconfig,
vector<Data>& shards)
{
UTIL_THROW_IF(shard_count == 0, util::Exception, "Must have at least 1 shard");
UTIL_THROW_IF(shard_size < 0 || shard_size > 1,
util::Exception,
"Shard size must be between 0 and 1, inclusive. Currently " << shard_size);
size_t data_size = m_score_data->size();
UTIL_THROW_IF(data_size != m_feature_data->size(),
util::Exception,
"Error");
shard_size *= data_size;
const float coeff = static_cast<float>(data_size) / shard_count;
for (size_t shard_id = 0; shard_id < shard_count; ++shard_id) {
vector<size_t> shard_contents;
if (shard_size == 0) {
//split into roughly equal size shards
const size_t shard_start = floor(0.5 + shard_id * coeff);
const size_t shard_end = floor(0.5 + (shard_id + 1) * coeff);
for (size_t i = shard_start; i < shard_end; ++i) {
shard_contents.push_back(i);
}
} else {
//create shards by randomly sampling
for (size_t i = 0; i < floor(shard_size+0.5); ++i) {
shard_contents.push_back(util::rand_excl(data_size));
}
}
Scorer* scorer = ScorerFactory::getScorer(m_score_type, scorerconfig);
shards.push_back(Data(scorer));
shards.back().m_score_type = m_score_type;
shards.back().m_num_scores = m_num_scores;
for (size_t i = 0; i < shard_contents.size(); ++i) {
shards.back().m_feature_data->add(m_feature_data->get(shard_contents[i]));
shards.back().m_score_data->add(m_score_data->get(shard_contents[i]));
}
//cerr << endl;
}
}
}