-
Notifications
You must be signed in to change notification settings - Fork 0
/
emission.h
224 lines (183 loc) · 6.65 KB
/
emission.h
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
/*
* emission.h
* Supervised training of the emission WFST
*
* Created on: Oct 23, 2019
* Author: Maria Ryskina
*/
#ifndef EMISSION_H_
#define EMISSION_H_
#include <list>
#include <math.h>
#include <fst/fstlib.h>
#include "base_fst.h"
#include "fst_utils.h"
#include "data_utils.h"
using namespace fst;
// Trainer for the emission model in the expectation semiring
class EmissionLogExpSemiring {
public:
EmissionFst<ExpVecArc> fst;
int origAlphSize;
int latinAlphSize;
int max_delay;
// Initializing emission parameters with uniform + random noise (from seed)
EmissionLogExpSemiring(int md, size_t oa, size_t la, int seed) :
fst(md, oa, la, VecWeight::One()) {
origAlphSize = oa;
latinAlphSize = la;
max_delay = md;
if (md == 0) {
EpsilonTotalFilter<ExpVecArc, NUM_EPS_TOTAL> epsFilterInput(origAlphSize, fst.orig_epsilon);
EpsilonTotalFilter<ExpVecArc, NUM_EPS_TOTAL> epsFilterOutput(latinAlphSize, fst.latin_epsilon);
Compose(epsFilterInput, fst, &fst);
Compose(fst, epsFilterOutput, &fst);
}
VecWeight initLp = addNoise(fst.arcIndexer.size(), fst.getLogProbs(), seed);
fst = EmissionFst<ExpVecArc>(md, oa, la, initLp);
};
// Initializing emission parameters with fixed values (e.g. from prior)
EmissionLogExpSemiring(int md, size_t oa, size_t la, VecWeight lp = VecWeight::One()) :
fst(md, oa, la, lp) {
origAlphSize = oa;
latinAlphSize = la;
max_delay = md;
if (md == 0) {
EpsilonTotalFilter<ExpVecArc, NUM_EPS_TOTAL> epsFilterInput(origAlphSize, fst.orig_epsilon);
EpsilonTotalFilter<ExpVecArc, NUM_EPS_TOTAL> epsFilterOutput(latinAlphSize, fst.latin_epsilon);
Compose(epsFilterInput, fst, &fst);
Compose(fst, epsFilterOutput, &fst);
}
};
VecWeight train(std::vector<std::vector<int>> origIndicesVector, std::vector<std::vector<int>> latinIndicesVector,
bool verbose = false) {
std::clock_t start;
double elapsed;
start = std::clock();
int max_iter = 5;
VecWeight emProbs;
float mll = 0;
float prevMll = INFINITY;
int iter = 1;
float convergenceThreshold = 1.05;
int step = div(origIndicesVector.size(), 10).quot;
while (true) {
std::cout << "ITERATION " << iter << std::endl;
float numTokens = 0;
int skipCount = 0;
int diffSkipCount = 0;
Adder<VecWeight> final;
for (int i = 0; i < origIndicesVector.size(); i++) {
std::vector<int> origIndices = origIndicesVector[i];
std::vector<int> latinIndices = latinIndicesVector[i];
VectorFst<ExpVecArc> input = constructInput<ExpVecArc>(origIndices, fst.orig_epsilon);
VectorFst<ExpVecArc> output = constructOutput<ExpVecArc>(latinIndices, fst.latin_epsilon);
VectorFst<ExpVecArc> lattice;
Compose<ExpVecArc>(input, (VectorFst<ExpVecArc>)fst, &lattice);
Compose<ExpVecArc>(lattice, output, &lattice);
if (lattice.NumStates() == 0) {
int diff = latinIndices.size() - origIndices.size();
if (diff > max_delay) diffSkipCount++;
skipCount++;
continue;
}
numTokens += latinIndices.size();
VecWeight out;
LogWeight ll;
// Collecting expected counts
std::vector<ExpVecWeight> dist;
ShortestDistance(lattice, &dist, true);
out = dist[0].Value2();
ll = dist[0].Value1();
mll += ll.Value();
final.Add(Divide(out, ll));
if (((i+1) % step == 0 || i == origIndicesVector.size() - 1) && verbose) {
elapsed = (std::clock() - start) / (double) CLOCKS_PER_SEC;
std::cout << "String pairs processed: " << (i + 1) << "; of them skipped: " <<
skipCount << "; time elapsed: " << elapsed << std::endl;
}
}
std::cout<<"Log-likelihood of training data: "<< mll << std::endl;
if (skipCount > 0) {
std::cout << "Skipped due to composition failure: " << skipCount <<
" out of " << origIndicesVector.size();
std::cout << "; of them " << diffSkipCount << " due to delay over "
<< max_delay << std::endl;
}
if (verbose) {
elapsed = (std::clock() - start) / (double) CLOCKS_PER_SEC;
std::cout << "Time elapsed: "<< elapsed << "; tokens per second: " <<
numTokens / elapsed << std::endl;
}
// Normalizing expected counts collected over entire corpus
emProbs = fst.normalize(final.Sum());
fst = EmissionFst<ExpVecArc>(max_delay, origAlphSize, latinAlphSize, emProbs);
if (prevMll - mll <= log(convergenceThreshold) || iter == max_iter) break;
prevMll = mll;
mll = 0;
iter++;
}
return emProbs;
}
protected:
VecWeight addNoise(int numArcs, VecWeight lp, int seed) {
srand(seed);
double delta = 1e-2;
for (int i = 0; i < numArcs; i++) {
double f = (double)rand() / RAND_MAX;
if (lp.Value(i) == LogWeight::One()) {
lp.SetValue(i, - f * delta); // using Taylor approximation
} else {
LogWeight noise = -log(f * delta);
lp.SetValue(i, Plus(lp.Value(i), noise));
}
}
return fst.normalize(lp);
}
};
class EmissionTropicalSemiring {
public:
EmissionFst<StdArc> fst;
int origAlphSize;
int latinAlphSize;
int max_delay;
EmissionTropicalSemiring(int md, size_t oa, size_t la, VecWeight lp) :
fst(md, oa, la, lp) {
origAlphSize = oa;
latinAlphSize = la;
max_delay = md;
if (md == 0) {
EpsilonTotalFilter<StdArc, NUM_EPS_TOTAL> epsFilterInput(origAlphSize, fst.orig_epsilon);
EpsilonTotalFilter<StdArc, NUM_EPS_TOTAL> epsFilterOutput(latinAlphSize, fst.latin_epsilon);
Compose(epsFilterInput, fst, &fst);
Compose(fst, epsFilterOutput, &fst);
}
}
// Collecting a set of all allowed emission labels
std::vector<bool> getOIndices() {
std::vector<bool> res(latinAlphSize + 1, false);
for (ArcIterator<VectorFst<StdArc>> aiter(fst, fst.Start()); !aiter.Done(); aiter.Next()) {
const StdArc &arc = aiter.Value();
res[arc.olabel] = true;
}
return res;
}
};
// Training the emission WFST on supervised data
EmissionTropicalSemiring trainEmission(IndexedStrings data, int max_delay, int origAlphSize, int latinAlphSize,
int seed, std::string output_dir, bool no_save = false) {
EmissionLogExpSemiring logExpEm(max_delay, origAlphSize, latinAlphSize, seed);
VecWeight emProbs = logExpEm.train(data.origIndices, data.latinIndices, true);
std::cout << "Emission model (expectation semiring): " << logExpEm.fst.NumStates() << " states, "
<< logExpEm.fst.getNumArcs() << " arcs\n";
EmissionTropicalSemiring tropicalEm(max_delay, origAlphSize, latinAlphSize, emProbs);
std::cout << "Emission model (tropical semiring): " << tropicalEm.fst.NumStates() << " states, "
<< tropicalEm.fst.getNumArcs() << " arcs\n";
if (!no_save) {
std::string emission_outfile = output_dir + "/emission.fst";
std::cout << "Saving the trained emission model to: " << emission_outfile << std::endl;
tropicalEm.fst.Write(emission_outfile);
}
return tropicalEm;
}
#endif /* EMISSION_H_ */