1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
|
/*
* Data.cpp
* mert - Minimum Error Rate Training
*
* Created by Nicola Bertoldi on 13/05/08.
*
*/
#include <algorithm>
#include "util/check.hh"
#include <cmath>
#include <fstream>
#include "Data.h"
#include "FileStream.h"
#include "Scorer.h"
#include "ScorerFactory.h"
#include "Util.h"
Data::Data()
: theScorer(NULL),
number_of_scores(0),
_sparse_flag(false),
scoredata(),
featdata() {}
Data::Data(Scorer& ptr)
: theScorer(&ptr),
score_type(theScorer->getName()),
number_of_scores(0),
_sparse_flag(false),
scoredata(new ScoreData(*theScorer)),
featdata(new FeatureData)
{
TRACE_ERR("Data::score_type " << score_type << std::endl);
TRACE_ERR("Data::Scorer type from Scorer: " << theScorer->getName() << endl);
}
//ADDED BY TS
void Data::remove_duplicates() {
size_t nSentences = featdata->size();
assert(scoredata->size() == nSentences);
for (size_t s=0; s < nSentences; s++) {
FeatureArray& feat_array = featdata->get(s);
ScoreArray& score_array = scoredata->get(s);
assert(feat_array.size() == score_array.size());
//serves as a hash-map:
std::map<double, std::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()) {
//std::cerr << "hit" << std::endl;
std::vector<size_t>& cur_list = lookup[sum];
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;
// }
// }
}
if (nRemoved > 0) {
feat_array.resize(end_pos+1);
score_array.resize(end_pos+1);
}
}
}
//END_ADDED
void Data::loadnbest(const std::string &file)
{
TRACE_ERR("loading nbest from " << file << std::endl);
ScoreStats scoreentry;
inputfilestream inp(file); // matches a stream with a file. Opens the file
if (!inp.good())
throw runtime_error("Unable to open: " + file);
std::string subsubstring, stringBuf;
std::string sentence_index, sentence, feature_str;
std::string::size_type loc;
while (getline(inp,stringBuf,'\n')) {
if (stringBuf.empty()) continue;
// adding statistics for error measures
scoreentry.clear();
getNextPound(stringBuf, sentence_index, "|||"); // first field
getNextPound(stringBuf, sentence, "|||"); // second field
getNextPound(stringBuf, feature_str, "|||"); // third field
theScorer->prepareStats(sentence_index, sentence, scoreentry);
scoredata->add(scoreentry, sentence_index);
// examine first line for name of features
if (!existsFeatureNames()) {
InitFeatureMap(feature_str);
}
AddFeatures(feature_str, sentence_index);
}
inp.close();
}
void Data::InitFeatureMap(const string& str) {
string buf = str;
string substr;
string features = "";
string tmp_name = "";
size_t tmp_index = 0;
string::size_type loc;
char tmp[64]; // for snprintf();
while (!buf.empty()) {
getNextPound(buf, substr);
// string ending with ":" are skipped, because they are the names of the features
if ((loc = substr.find_last_of(":")) != substr.length()-1) {
snprintf(tmp, sizeof(tmp), "%s_%lu ", tmp_name.c_str(), tmp_index);
features.append(tmp);
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);
}
}
featdata->setFeatureMap(features);
}
void Data::AddFeatures(const string& str,
const string& sentence_index) {
string::size_type loc;
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 ((loc = substr.find_last_of(":")) != substr.length()-1) {
feature_entry.add(ConvertStringToFeatureStatsType(substr));
} else if (substr.find("_") != string::npos) {
// sparse feature name? store as well
std::string name = substr;
getNextPound(buf, substr);
feature_entry.addSparse(name, atof(substr.c_str()));
_sparse_flag = true;
}
}
featdata->add(feature_entry, sentence_index);
}
// TODO
void Data::mergeSparseFeatures() {
std::cerr << "ERROR: sparse features can only be trained with pairwise ranked optimizer (PRO), not traditional MERT\n";
exit(1);
}
void Data::createShards(size_t shard_count, float shard_size, const string& scorerconfig,
std::vector<Data>& shards)
{
CHECK(shard_count);
CHECK(shard_size >= 0);
CHECK(shard_size <= 1);
size_t data_size = scoredata->size();
CHECK(data_size == featdata->size());
shard_size *= data_size;
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 * static_cast<float>(data_size) / shard_count);
const size_t shard_end = floor(0.5 + (shard_id + 1) * static_cast<float>(data_size) / shard_count);
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(rand() % data_size);
}
}
Scorer* scorer = ScorerFactory::getScorer(score_type, scorerconfig);
shards.push_back(Data(*scorer));
shards.back().score_type = score_type;
shards.back().number_of_scores = number_of_scores;
shards.back()._sparse_flag = _sparse_flag;
for (size_t i = 0; i < shard_contents.size(); ++i) {
shards.back().featdata->add(featdata->get(shard_contents[i]));
shards.back().scoredata->add(scoredata->get(shard_contents[i]));
}
//cerr << endl;
}
}
|