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
|
/*
Copyright (C) 1999,2000,2001 Franz Josef Och (RWTH Aachen - Lehrstuhl fuer Informatik VI)
This file is part of GIZA++ ( extension of GIZA ).
This program is free software; you can redistribute it and/or
modify it under the terms of the GNU General Public License
as published by the Free Software Foundation; either version 2
of the License, or (at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with this program; if not, write to the Free Software
Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307,
USA.
*/
#ifndef NO_TRAINING
#include "ForwardBackward.h"
#include "Globals.h"
#include "myassert.h"
#include "HMMTables.h"
#include "mymath.h"
double ForwardBackwardTraining(const HMMNetwork&net,Array<double>&g,Array<Array2<double> >&E){
const int I=net.size1(),J=net.size2(),N=I*J;
Array<double> alpha(N,0),beta(N,0),sum(J);
for(int i=0;i<I;i++)
beta[N-I+i]=net.getBetainit(i);
double * cur_beta=conv<double>(beta.begin())+N-I-1;
for(int j=J-2;j>=0;--j)
for(int ti=I-1;ti>=0;--ti,--cur_beta) {
const double *next_beta=conv<double>(beta.begin())+(j+1)*I;
const double *alprob=&net.outProb(j,ti,0),*next_node=&net.nodeProb(0,j+1);
for(int ni=0;ni<I;++ni,(next_node+=J)){
massert(cur_beta<next_beta&& &net.outProb(j,ti,ni)==alprob);
massert(next_node == &net.nodeProb(ni,j+1));
/* if( VERB&&(*next_beta)*(*alprob)*(*next_node) )
cout << "B= " << (int)(cur_beta-beta.begin()) << " += " << (*next_beta) << "("
<< next_beta-beta.begin() << ") alprob:" << (*alprob) << " lexprob:" << (*next_node) << endl;*/
(*cur_beta)+=(*next_beta++)*(*alprob++)*(*next_node);
}
}
for(int i=0;i<I;i++)
alpha[i]=net.getAlphainit(i)*net.nodeProb(i,0);
double* cur_alpha=conv<double>(alpha.begin())+I;
cur_beta=conv<double>(beta.begin())+I;
for(int j=1;j<J;j++){
Array2<double>&e=E[ (E.size()==1)?0:(j-1) ];
if( (E.size()!=1) || j==1 )
{
e.resize(I,I);
fill(e.begin(),e.end(),0.0);
}
for(int ti=0;ti<I;++ti,++cur_alpha,++cur_beta) {
const double * prev_alpha=conv<double>(alpha.begin())+I*(j-1);
double *cur_e= &e(ti,0);
double this_node=net.nodeProb(ti,j);
const double* alprob= &net.outProb(j-1,0,ti);
for(int pi=0;pi<I;++pi,++prev_alpha,(alprob+=I)){
massert(prev_alpha<cur_alpha&& &net.outProb(j-1,pi,ti)==alprob);
massert(&e(ti,pi)==cur_e);
const double alpha_increment= *prev_alpha*(*alprob)*this_node;
(*cur_alpha)+=alpha_increment;
(*cur_e++)+=alpha_increment*(*cur_beta);
}
}
}
g.resize(N);
transform(alpha.begin(),alpha.end(),beta.begin(),g.begin(),multiplies<double>());
double bsum=0,esum=0,esum2;
for(int i=0;i<I;i++)
bsum+=beta[i]*net.nodeProb(i,0)*net.getAlphainit(i);
for(unsigned int j=0;j<(unsigned int)E.size();j++)
{
Array2<double>&e=E[j];
const double *epe=e.end();
for(const double*ep=e.begin();ep!=epe;++ep)
esum+=*ep;
}
if( J>1 )
esum2=esum/(J-1);
else
esum2=0.0;
if(!(esum2==0.0||mfabs(esum2-bsum)/bsum<1e-3*I))
cout << "ERROR2: " << esum2 <<" " <<bsum << " " << esum << net << endl;
double * sumptr=conv<double>(sum.begin());
double* ge=conv<double>(g.end());
for(double* gp=conv<double>(g.begin());gp!=ge;gp+=I)
{
*sumptr++=normalize_if_possible(gp,gp+I);
if(bsum && !(mfabs((*(sumptr-1)-bsum)/bsum)<1e-3*I))
cout << "ERROR: " << *(sumptr-1) << " " << bsum << " " << mfabs((*(sumptr-1)-bsum)/bsum) << ' ' << I << ' ' << J << endl;
}
for(unsigned int j=0;j<(unsigned int)E.size();j++)
{
Array2<double>&e=E[j];
double* epe=e.end();
if( esum )
for(double*ep=e.begin();ep!=epe;++ep)
*ep/=esum;
else
for(double*ep=e.begin();ep!=epe;++ep)
*ep/=1.0/(max(I*I,I*I*(J-1)));
}
if( sum.size() )
return sum[0];
else
return 1.0;
}
void HMMViterbi(const HMMNetwork&net,Array<int>&vit) {
const int I=net.size1(),J=net.size2();
vit.resize(J);
Array<double>g;
Array<Array2<double> >e(1);
ForwardBackwardTraining(net,g,e);
for(int j=0;j<J;j++) {
double * begin=conv<double>(g.begin())+I*j;
vit[j]=max_element(begin,begin+I)-begin;
}
}
void HMMViterbi(const HMMNetwork&net,Array<double>&g,Array<int>&vit) {
const int I=net.size1(),J=net.size2();
vit.resize(J);
for(int j=0;j<J;j++) {
double* begin=conv<double>(g.begin())+I*j;
vit[j]=max_element(begin,begin+I)-begin;
}
}
double HMMRealViterbi(const HMMNetwork&net,Array<int>&vitar,int pegi,int pegj,bool verbose){
const int I=net.size1(),J=net.size2(),N=I*J;
Array<double> alpha(N,-1);
Array<double*> bp(N,(double*)0);
vitar.resize(J);
if( J==0 )
return 1.0;
for(int i=0;i<I;i++)
{
alpha[i]=net.getAlphainit(i)*net.nodeProb(i,0);
if( i>I/2 )
alpha[i]=0; // only first empty word can be chosen
bp[i]=0;
}
double *cur_alpha=conv<double>(alpha.begin())+I;
double **cur_bp=conv<double*>(bp.begin())+I;
for(int j=1;j<J;j++)
{
if( pegj+1==j)
for(int ti=0;ti<I;ti++)
if( (pegi!=-1&&ti!=pegi)||(pegi==-1&&ti<I/2) )
(cur_alpha-I)[ti]=0.0;
for(int ti=0;ti<I;++ti,++cur_alpha,++cur_bp) {
double* prev_alpha=conv<double>(alpha.begin())+I*(j-1);
double this_node=net.nodeProb(ti,j);
const double *alprob= &net.outProb(j-1,0,ti);
for(int pi=0;pi<I;++pi,++prev_alpha,(alprob+=I)){
massert(prev_alpha<cur_alpha&& &net.outProb(j-1,pi,ti)==alprob);
const double alpha_increment= *prev_alpha*(*alprob)*this_node;
if( alpha_increment> *cur_alpha )
{
(*cur_alpha)=alpha_increment;
(*cur_bp)=prev_alpha;
}
}
}
}
for(int i=0;i<I;i++)
alpha[N-I+i]*=net.getBetainit(i);
if( pegj==J-1)
for(int ti=0;ti<I;ti++)
if( (pegi!=-1&&ti!=pegi)||(pegi==-1&&ti<I/2) )
(alpha)[N-I+ti]=0.0;
int j=J-1;
cur_alpha=conv<double>(alpha.begin())+j*I;
vitar[J-1]=max_element(cur_alpha,cur_alpha+I)-cur_alpha;
double ret= *max_element(cur_alpha,cur_alpha+I);
while(bp[vitar[j]+j*I])
{
cur_alpha-=I;
vitar[j-1]=bp[vitar[j]+j*I]-cur_alpha;
massert(vitar[j-1]<I&&vitar[j-1]>=0);
j--;
}
massert(j==0);
if( verbose )
{
cout << "VERB:PEG: " << pegi << ' ' << pegj << endl;
for(int j=0;j<J;j++)
cout << "NP " << net.nodeProb(vitar[j],j) << ' ' << "AP " << ((j==0)?net.getAlphainit(vitar[j]):net.outProb(j-1,vitar[j-1],vitar[j])) << " j:" << j << " i:" << vitar[j] << "; ";
cout << endl;
}
return ret;
}
double MaximumTraining(const HMMNetwork&net,Array<double>&g,Array<Array2<double> >&E){
Array<int> vitar;
double ret=HMMRealViterbi(net,vitar);
const int I=net.size1(),J=net.size2();
if( E.size()==1 )
{
Array2<double>&e=E[0];
e.resize(I,I);
g.resize(I*J);
fill(g.begin(),g.end(),0.0);
fill(e.begin(),e.end(),0.0);
for(int i=0;i<J;++i)
{
g[i*I+vitar[i]]=1.0;
if( i>0 )
e(vitar[i],vitar[i-1])++;
}
}
else
{
g.resize(I*J);
fill(g.begin(),g.end(),0.0);
for(int i=0;i<J;++i)
{
g[i*I+vitar[i]]=1.0;
if( i>0 )
{
Array2<double>&e=E[i-1];
e.resize(I,I);
fill(e.begin(),e.end(),0.0);
e(vitar[i],vitar[i-1])++;
}
}
}
return ret;
}
#endif
|
