-
Notifications
You must be signed in to change notification settings - Fork 208
/
learner.cpp
320 lines (271 loc) · 9.59 KB
/
learner.cpp
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
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
// MeCab -- Yet Another Part-of-Speech and Morphological Analyzer
//
//
// Copyright(C) 2001-2006 Taku Kudo <taku@chasen.org>
// Copyright(C) 2004-2006 Nippon Telegraph and Telephone Corporation
#include <fstream>
#include <string>
#include <vector>
#include "common.h"
#include "feature_index.h"
#include "freelist.h"
#include "lbfgs.h"
#include "learner_tagger.h"
#include "param.h"
#include "string_buffer.h"
#include "thread.h"
#include "utils.h"
namespace MeCab {
namespace {
#define DCONF(file) create_filename(dicdir, std::string(file)).c_str()
#ifdef MECAB_USE_THREAD
class learner_thread: public thread {
public:
unsigned short start_i;
unsigned short thread_num;
size_t size;
size_t micro_p;
size_t micro_r;
size_t micro_c;
size_t err;
double f;
EncoderLearnerTagger **x;
std::vector<double> expected;
void run() {
micro_p = micro_r = micro_c = err = 0;
f = 0.0;
std::fill(expected.begin(), expected.end(), 0.0);
for (size_t i = start_i; i < size; i += thread_num) {
f += x[i]->gradient(&expected[0]);
err += x[i]->eval(µ_c, µ_p, µ_r);
}
}
};
#endif
class CRFLearner {
public:
static int run(Param *param) {
const std::string dicdir = param->get<std::string>("dicdir");
CHECK_DIE(param->load(DCONF(DICRC)))
<< "no such file or directory: " << DCONF(DICRC);
const std::vector<std::string> &files = param->rest_args();
if (files.size() != 2) {
std::cout << "Usage: " <<
param->program_name() << " corpus model" << std::endl;
return -1;
}
const std::string ifile = files[0];
const std::string model = files[1];
const std::string old_model = param->get<std::string>("old-model");
EncoderFeatureIndex feature_index;
std::vector<double> expected;
std::vector<double> observed;
std::vector<double> alpha;
std::vector<double> old_alpha;
std::vector<EncoderLearnerTagger *> x;
Tokenizer<LearnerNode, LearnerPath> tokenizer;
Allocator<LearnerNode, LearnerPath> allocator;
CHECK_DIE(tokenizer.open(*param)) << "cannot open tokenizer";
CHECK_DIE(feature_index.open(*param)) << "cannot open feature index";
if (!old_model.empty()) {
std::cout << "Using previous model: " << old_model << std::endl;
std::cout << "--cost --freq and --eta options are overwritten."
<< std::endl;
CHECK_DIE(tokenizer.dictionary_info());
const char *dic_charset = tokenizer.dictionary_info()->charset;
feature_index.reopen(old_model.c_str(),
dic_charset, &old_alpha, param);
}
const double C = param->get<double>("cost");
const double eta = param->get<double>("eta");
const size_t eval_size = param->get<size_t>("eval-size");
const size_t unk_eval_size = param->get<size_t>("unk-eval-size");
const size_t thread_num = param->get<size_t>("thread");
const size_t freq = param->get<size_t>("freq");
CHECK_DIE(C > 0) << "cost parameter is out of range: " << C;
CHECK_DIE(eta > 0) "eta is out of range: " << eta;
CHECK_DIE(eval_size > 0) << "eval-size is out of range: " << eval_size;
CHECK_DIE(unk_eval_size > 0) <<
"unk-eval-size is out of range: " << unk_eval_size;
CHECK_DIE(freq > 0) <<
"freq is out of range: " << unk_eval_size;
CHECK_DIE(thread_num > 0 && thread_num <= 512)
<< "# thread is invalid: " << thread_num;
std::cout.setf(std::ios::fixed, std::ios::floatfield);
std::cout.precision(5);
std::cout << "reading corpus ..." << std::flush;
std::ifstream ifs(WPATH(ifile.c_str()));
CHECK_DIE(ifs) << "no such file or directory: " << ifile;
while (ifs) {
EncoderLearnerTagger *tagger = new EncoderLearnerTagger();
CHECK_DIE(tagger->open(&tokenizer,
&allocator,
&feature_index,
eval_size,
unk_eval_size));
CHECK_DIE(tagger->read(&ifs, &observed));
if (!tagger->empty()) {
x.push_back(tagger);
} else {
delete tagger;
}
if (x.size() % 100 == 0) {
std::cout << x.size() << "... " << std::flush;
}
}
feature_index.shrink(freq, &observed);
feature_index.clearcache();
const size_t psize = feature_index.size();
observed.resize(psize);
expected.resize(psize);
alpha.resize(psize);
old_alpha.resize(psize);
alpha = old_alpha;
feature_index.set_alpha(&alpha[0]);
std::cout << std::endl;
std::cout << "Number of sentences: " << x.size() << std::endl;
std::cout << "Number of features: " << psize << std::endl;
std::cout << "eta: " << eta << std::endl;
std::cout << "freq: " << freq << std::endl;
std::cout << "eval-size: " << eval_size << std::endl;
std::cout << "unk-eval-size: " << unk_eval_size << std::endl;
#ifdef MECAB_USE_THREAD
std::cout << "threads: " << thread_num << std::endl;
#endif
std::cout << "charset: " <<
tokenizer.dictionary_info()->charset << std::endl;
std::cout << "C(sigma^2): " << C << std::endl
<< std::endl;
#ifdef MECAB_USE_THREAD
std::vector<learner_thread> thread;
if (thread_num > 1) {
thread.resize(thread_num);
for (size_t i = 0; i < thread_num; ++i) {
thread[i].start_i = i;
thread[i].size = x.size();
thread[i].thread_num = thread_num;
thread[i].x = &x[0];
thread[i].expected.resize(expected.size());
}
}
#endif
int converge = 0;
double prev_obj = 0.0;
LBFGS lbfgs;
for (size_t itr = 0; ; ++itr) {
std::fill(expected.begin(), expected.end(), 0.0);
double obj = 0.0;
size_t err = 0;
size_t micro_p = 0;
size_t micro_r = 0;
size_t micro_c = 0;
#ifdef MECAB_USE_THREAD
if (thread_num > 1) {
for (size_t i = 0; i < thread_num; ++i) {
thread[i].start();
}
for (size_t i = 0; i < thread_num; ++i) {
thread[i].join();
}
for (size_t i = 0; i < thread_num; ++i) {
obj += thread[i].f;
err += thread[i].err;
micro_r += thread[i].micro_r;
micro_p += thread[i].micro_p;
micro_c += thread[i].micro_c;
for (size_t k = 0; k < psize; ++k) {
expected[k] += thread[i].expected[k];
}
}
} else
#endif
{
for (size_t i = 0; i < x.size(); ++i) {
obj += x[i]->gradient(&expected[0]);
err += x[i]->eval(µ_c, µ_p, µ_r);
}
}
const double p = 1.0 * micro_c / micro_p;
const double r = 1.0 * micro_c / micro_r;
const double micro_f = 2 * p * r / (p + r);
for (size_t i = 0; i < psize; ++i) {
const double penalty = (alpha[i] - old_alpha[i]);
obj += (penalty * penalty / (2.0 * C));
expected[i] = expected[i] - observed[i] + penalty / C;
}
const double diff = (itr == 0 ? 1.0 :
std::fabs(1.0 * (prev_obj - obj)) / prev_obj);
std::cout << "iter=" << itr
<< " err=" << 1.0 * err/x.size()
<< " F=" << micro_f
<< " target=" << obj
<< " diff=" << diff << std::endl;
prev_obj = obj;
if (diff < eta) {
converge++;
} else {
converge = 0;
}
if (converge == 3) {
break; // 3 is ad-hoc
}
const int ret = lbfgs.optimize(psize,
&alpha[0], obj,
&expected[0], false, C);
CHECK_DIE(ret >= 0) << "unexpected error in LBFGS routin";
if (ret == 0) {
break;
}
}
std::cout << "\nDone! writing model file ... " << std::endl;
std::ostringstream oss;
oss << "eta: " << eta << std::endl;
oss << "freq: " << freq << std::endl;
oss << "C: " << C << std::endl;
oss.setf(std::ios::fixed, std::ios::floatfield);
oss.precision(16);
oss << "eval-size: " << eval_size << std::endl;
oss << "unk-eval-size: " << unk_eval_size << std::endl;
oss << "charset: " << tokenizer.dictionary_info()->charset << std::endl;
const std::string header = oss.str();
CHECK_DIE(feature_index.save(model.c_str(), header.c_str()))
<< "permission denied: " << model;
return 0;
}
};
class Learner {
public:
static bool run(int argc, char **argv) {
static const MeCab::Option long_options[] = {
{ "dicdir", 'd', ".", "DIR",
"set DIR as dicdir(default \".\" )" },
{ "old-model", 'M', 0, "FILE",
"set FILE as old CRF model file" },
{ "cost", 'c', "1.0", "FLOAT",
"set FLOAT for cost C for constraints violatoin" },
{ "freq", 'f', "1", "INT",
"set the frequency cut-off (default 1)" },
{ "eta", 'e', "0.00005", "DIR",
"set FLOAT for tolerance of termination criterion" },
{ "thread", 'p', "1", "INT", "number of threads(default 1)" },
{ "version", 'v', 0, 0, "show the version and exit" },
{ "help", 'h', 0, 0, "show this help and exit." },
{ 0, 0, 0, 0 }
};
Param param;
if (!param.open(argc, argv, long_options)) {
std::cout << param.what() << "\n\n" << COPYRIGHT
<< "\ntry '--help' for more information." << std::endl;
return -1;
}
if (!param.help_version()) {
return 0;
}
return CRFLearner::run(¶m);
}
};
}
}
int mecab_cost_train(int argc, char **argv) {
return MeCab::Learner::run(argc, argv);
}