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
|
#include "dtrain.h"
#include "sample.h"
#include "score.h"
#include "update.h"
using namespace dtrain;
int
main(int argc, char** argv)
{
// get configuration
po::variables_map conf;
if (!dtrain_init(argc, argv, &conf))
exit(1); // something is wrong
const size_t k = conf["k"].as<size_t>();
const size_t N = conf["N"].as<size_t>();
const size_t T = conf["iterations"].as<size_t>();
const weight_t eta = conf["learning_rate"].as<weight_t>();
const weight_t margin = conf["margin"].as<weight_t>();
const bool average = conf["average"].as<bool>();
const weight_t l1_reg = conf["l1_reg"].as<weight_t>();
const bool keep = conf["keep"].as<bool>();
const string output_fn = conf["output"].as<string>();
vector<string> print_weights;
boost::split(print_weights, conf["print_weights"].as<string>(),
boost::is_any_of(" "));
// setup decoder
register_feature_functions();
SetSilent(true);
ReadFile f(conf["decoder_conf"].as<string>());
Decoder decoder(f.stream());
ScoredKbest* observer = new ScoredKbest(k, new PerSentenceBleuScorer(N));
// weights
vector<weight_t>& decoder_weights = decoder.CurrentWeightVector();
SparseVector<weight_t> lambdas, w_average;
if (conf.count("input_weights")) {
Weights::InitFromFile(conf["input_weights"].as<string>(), &decoder_weights);
Weights::InitSparseVector(decoder_weights, &lambdas);
}
// input
string input_fn = conf["bitext"].as<string>();
ReadFile input(input_fn);
vector<string> buf; // decoder only accepts strings as input
vector<vector<Ngrams> > buf_ngs; // compute ngrams and lengths of references
vector<vector<size_t> > buf_ls; // just once
size_t input_sz = 0;
cerr << _p4;
// output configuration
cerr << "dtrain" << endl << "Parameters:" << endl;
cerr << setw(25) << "k " << k << endl;
cerr << setw(25) << "N " << N << endl;
cerr << setw(25) << "T " << T << endl;
cerr << setw(25) << "learning rate " << eta << endl;
cerr << setw(25) << "margin " << margin << endl;
cerr << setw(25) << "l1 reg " << l1_reg << endl;
cerr << setw(25) << "decoder conf " << "'"
<< conf["decoder_conf"].as<string>() << "'" << endl;
cerr << setw(25) << "input " << "'" << input_fn << "'" << endl;
cerr << setw(25) << "output " << "'" << output_fn << "'" << endl;
if (conf.count("input_weights")) {
cerr << setw(25) << "weights in " << "'"
<< conf["input_weights"].as<string>() << "'" << endl;
}
cerr << "(1 dot per processed input)" << endl;
// meta
weight_t best=0., gold_prev=0.;
size_t best_iteration = 0;
time_t total_time = 0.;
for (size_t t = 0; t < T; t++) // T iterations
{
time_t start, end;
time(&start);
weight_t gold_sum=0., model_sum=0.;
size_t i=0, num_up=0, feature_count=0, list_sz=0;
cerr << "Iteration #" << t+1 << " of " << T << "." << endl;
while(true)
{
bool next = true;
// getting input
if (t == 0) {
string in;
if(!getline(*input, in)) {
next = false;
} else {
vector<string> parts;
boost::algorithm::split_regex(parts, in, boost::regex(" \\|\\|\\| "));
buf.push_back(parts[0]);
parts.erase(parts.begin());
buf_ngs.push_back({});
buf_ls.push_back({});
for (auto s: parts) {
vector<WordID> r;
vector<string> toks;
boost::split(toks, s, boost::is_any_of(" "));
for (auto tok: toks)
r.push_back(TD::Convert(tok));
buf_ngs.back().emplace_back(MakeNgrams(r, N));
buf_ls.back().push_back(r.size());
}
}
} else {
next = i<input_sz;
}
// produce some pretty output
if (next) {
if (i%20==0)
cerr << " ";
cerr << ".";
if ((i+1)%20==0)
cerr << " " << i+1 << endl;
} else {
cerr << " " << i << endl;
}
cerr.flush();
// stop iterating
if (!next) break;
// decode
if (t > 0 || i > 0)
lambdas.init_vector(&decoder_weights);
observer->SetReference(buf_ngs[i], buf_ls[i]);
decoder.Decode(buf[i], observer);
vector<ScoredHyp>* samples = observer->GetSamples();
// stats for 1best
gold_sum += samples->front().gold;
model_sum += samples->front().model;
feature_count += observer->GetFeatureCount();
list_sz += observer->GetSize();
// get pairs and update
SparseVector<weight_t> updates;
num_up += CollectUpdates(samples, updates, margin);
SparseVector<weight_t> lambdas_copy;
if (l1_reg)
lambdas_copy = lambdas;
lambdas.plus_eq_v_times_s(updates, eta);
// l1 regularization
// NB: regularization is done after each sentence,
// not after every single pair!
if (l1_reg) {
SparseVector<weight_t>::iterator it = lambdas.begin();
for (; it != lambdas.end(); ++it) {
weight_t v = it->second;
if (!v)
continue;
if (!lambdas_copy.get(it->first) // new or..
|| lambdas_copy.get(it->first)!=v) // updated feature
{
if (v > 0) {
it->second = max(0., v - l1_reg);
} else {
it->second = min(0., v + l1_reg);
}
}
}
}
i++;
} // input loop
if (t == 0)
input_sz = i; // remember size of input (# lines)
// update average
if (average)
w_average += lambdas;
// stats
weight_t gold_avg = gold_sum/(weight_t)input_sz;
cerr << _p << "WEIGHTS" << endl;
for (auto name: print_weights)
cerr << setw(18) << name << " = " << lambdas.get(FD::Convert(name)) << endl;
cerr << " ---" << endl;
cerr << _np << " 1best avg score: " << gold_avg*100;
cerr << _p << " (" << (gold_avg-gold_prev)*100 << ")" << endl;
cerr << " 1best avg model score: "
<< model_sum/(weight_t)input_sz << endl;
cerr << " avg # updates: ";
cerr << _np << num_up/(float)input_sz << endl;
cerr << " non-0 feature count: " << lambdas.num_nonzero() << endl;
cerr << " avg f count: " << feature_count/(float)list_sz << endl;
cerr << " avg list sz: " << list_sz/(float)input_sz << endl;
if (gold_avg > best) {
best = gold_avg;
best_iteration = t;
}
gold_prev = gold_avg;
time (&end);
time_t time_diff = difftime(end, start);
total_time += time_diff;
cerr << "(time " << time_diff/60. << " min, ";
cerr << time_diff/input_sz << " s/S)" << endl;
if (t+1 != T) cerr << endl;
if (keep) { // keep intermediate weights
lambdas.init_vector(&decoder_weights);
string w_fn = "weights." + boost::lexical_cast<string>(t) + ".gz";
Weights::WriteToFile(w_fn, decoder_weights, true);
}
} // outer loop
// final weights
if (average) {
w_average /= T;
w_average.init_vector(decoder_weights);
} else if (!keep) {
lambdas.init_vector(decoder_weights);
}
if (average || !keep)
Weights::WriteToFile(output_fn, decoder_weights, true);
cerr << endl << "---" << endl << "Best iteration: ";
cerr << best_iteration+1 << " [GOLD = " << best*100 << "]." << endl;
cerr << "This took " << total_time/60. << " min." << endl;
return 0;
}
|
