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
|
///////////////////////////////////////////////////////////////////////////////
// //
// This file is part of ModelBlocks. Copyright 2009, ModelBlocks developers. //
// //
// ModelBlocks 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 3 of the License, or //
// (at your option) any later version. //
// //
// ModelBlocks 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 ModelBlocks. If not, see <http://www.gnu.org/licenses/>. //
// //
// ModelBlocks developers designate this particular file as subject to //
// the "Moses" exception as provided by ModelBlocks developers in //
// the LICENSE file that accompanies this code. //
// //
///////////////////////////////////////////////////////////////////////////////
#ifndef _NL_MIXTURE__
#define _NL_MIXTURE__
#include "nl-randvar.h"
#include "nl-list.h"
////////////////////////////////////////////////////////////////////////////////
//
// Mixture Model
//
////////////////////////////////////////////////////////////////////////////////
template <template <class MY> class M,class Y,class C>
class Mixture2DModel : public Generic1DModel<Y,Prob> {
private:
// Static data members...
static bool bModeled;
// Data members...
CPT1DModel<C,Prob> mC;
SimpleHash<C,M<Y> > mY_giv_C;
public:
Mixture2DModel ( ) : mC() { }
Mixture2DModel ( const string& s ) : mC(s) { }
// Specification methods...
CPT1DModel<C,Prob>& setMixingModel ( ) { return mC; }
Prob& setComponentProb ( const C& c ) { return mC.setProb(c); }
M<Y>& setComponentModel ( const C& c ) { return mY_giv_C.set(c); }
// Extraction methods...
Prob getComponentProb ( const C& c ) { return mC.getProb(c); }
const M<Y>& getComponentModel ( const C& c ) { return mY_giv_C.get(c); }
Prob getProb ( const Y& y ) const { Prob pr=0.0;
C c; for ( bool b=c.setFirst(); b; b=c.setNext() )
pr += ( mC.getProb(c) * mY_giv_C.get(c).getProb(y) );
return pr; }
// Input / output methods...
bool readFields ( char* as[], int numF ) { return ( mC.readFields(as,numF) || mY_giv_C.set(C(as[1])).readFields(as+1,numF-1) ); }
void writeFields ( FILE* pf, string sPref ) const { mC.writeFields(pf,sPref); C c; for ( bool b=c.setFirst(); b; b=c.setNext() )
mY_giv_C.get(c).writeFields(pf,sPref+" "+c.getString()); }
};
////////////////////////////////////////////////////////////////////////////////
////////////////////////////////////////////////////////////////////////////////
template <template <class MY> class M,class Y,class X,class C>
class Mixture3DModel : public Generic2DModel<Y,X,Prob> {
private:
// Static data members...
static bool bModeled;
// Data members...
CPT2DModel<C,X,Prob> mC_giv_X;
SimpleHash<Joint2DRV<X,C>,M<Y> > mY_giv_X_C;
public:
Mixture3DModel ( ) : mC_giv_X() { }
Mixture3DModel ( const string& s ) : mC_giv_X(s) { }
// Specification methods...
Prob& setComponentProb ( const X& x, const C& c ) { return mC_giv_X.setProb(c,x); }
M<Y>& setComponentModel ( const X& x, const C& c ) { return mY_giv_X_C.set(Joint2DRV<X,C>(x,c)); }
// Extraction methods...
Prob getProb ( const Y& y, const X& x ) const { Prob pr=0.0;
C c; for ( bool b=c.setFirst(); b; b=c.setNext() )
pr += ( mC_giv_X.getProb(c,x) * mY_giv_X_C.get(Joint2DRV<X,C>(x,c)).getProb(y) );
return pr; }
// Input / output methods...
bool readFields ( char* as[], int numF ) { return ( mC_giv_X.readFields(as,numF) || mY_giv_X_C.set(Joint2DRV<X,C>(X(as[1]),C(as[2]))).readFields(as+2,numF-2) ); }
void writeFields ( FILE* pf, string sPref ) const { mC_giv_X.writeFields(pf,sPref);
X x; for ( bool bx=x.setFirst(); bx; bx=x.setNext() ) {
C c; for ( bool b=c.setFirst(); b; b=c.setNext() ) {
mY_giv_X_C.get(Joint2DRV<X,C>(x,c)).writeFields(pf,sPref+" "+c.getString()); } } }
};
////////////////////////////////////////////////////////////////////////////////
//
// Trainable Mixture Model
//
////////////////////////////////////////////////////////////////////////////////
template <template <class MY> class M,class Y,class C>
class TrainableMixture2DModel : public Mixture2DModel<M,Y,C> {
// private:
// LogPDFVal logpdfPrevDataAvg;
public:
// // Constructor / destructor methods...
// TrainableMixture2DModel ( ) : logpdfPrevDataAvg(log(0.0)) { }
// Specification methods...
void updateFields ( const List<Joint2DRV<Y,Prob> >&, const PDFVal, bool& );
void train ( List<Joint2DRV<Y,Prob> >&, const int, const PDFVal );
};
////////////////////////////////////////
template <template <class MY> class M,class Y,class C>
void TrainableMixture2DModel<M,Y,C>::updateFields ( const List<Joint2DRV<Y,Prob> >& lyp, const PDFVal WEIGHT_LIMIT, bool& bShouldStop ) {
LogPDFVal logpdfData = 0.0;
CPT1DModel<C,Prob> mprPseudoEmpC; // pseudo-empirical prob marginal
SimpleHash<C,List<pair<const Y*,Prob> > > alypPseudoEmpY_giv_C; // pseudo-empirical prob conditional
//// E step: create pseudo-annotated (expectation) corpus by re-distributing empirical probabilities over C, according to model parameters...
C c;
for ( bool b=c.setFirst(); b; b=c.setNext() )
setComponentModel(c).precomputeVarianceTerms();
// For each Y...
for ( const ListedObject<Joint2DRV<Y,Prob> >* pyp=lyp.getFirst(); pyp; pyp=lyp.getNext(pyp) ) {
const Y& y = pyp->getSub1(); // data value
const Prob& prEmpY = pyp->getSub2(); // empirical prob
CPT1DModel<C,PDFVal> mpdfEstCY; // model-estimated pdf
PDFVal pdfEstY = 0.0; // model-estimated pdf marginal
// For each C, compute model-estimated pdfs...
C c; for ( bool b=c.setFirst(); b; b=c.setNext() ) {
PDFVal pdfEstCY = getComponentProb(c) * getComponentModel(c).getProb(y);
// fprintf(stderr," ???%g %d %g\n",(double)pdfEstCY,(int)c,getComponentModel(c).getProb(y)); ////y.write(stderr); fprintf(stderr,"\n");
mpdfEstCY.setProb(c) = pdfEstCY;
pdfEstY += pdfEstCY; // marginalize out C (marginal estimated pdf of each y)
}
// Update total...
logpdfData += log(pdfEstY); // product of each datum (joint estimated pdf of training corpus)
//fprintf(stderr,"..%g %g\n",pdfEstY,logpdfData);
// For each C, compute psuedo-empirical probs (model-estimated * empirical)...
for ( bool b=c.setFirst(); b; b=c.setNext() ) {
Prob prEstC_giv_Y = mpdfEstCY.getProb(c) / pdfEstY; // so these are normalized distribs over C
Prob prPseudoEmpCY = prEmpY * prEstC_giv_Y;
alypPseudoEmpY_giv_C.set(c).add() = pair<const Y*,Prob> ( &pyp->getSub1(), prPseudoEmpCY );
mprPseudoEmpC.setProb(c) += prPseudoEmpCY; // marginalize out Y
}
}
// Renormalize sub-lists...
for ( bool b=c.setFirst(); b; b=c.setNext() )
for ( ListedObject<pair<const Y*,Prob> >* pyp=alypPseudoEmpY_giv_C.set(c).setFirst(); pyp; pyp=alypPseudoEmpY_giv_C.set(c).setNext(pyp) )
pyp->second /= mprPseudoEmpC.getProb(c);
//// M step: update model parameters using MLE (relative frequency estimation) on pseudo-annotated corpus...
for ( bool b=c.setFirst(); b; b=c.setNext() ) {
// If any change exceeds thresh, continue...
if ( bShouldStop && ( getComponentProb(c)-mprPseudoEmpC.getProb(c) > WEIGHT_LIMIT ||
getComponentProb(c)-mprPseudoEmpC.getProb(c) < -WEIGHT_LIMIT ) ) bShouldStop = false;
// Update probs...
setComponentProb(c) = mprPseudoEmpC.getProb(c);
setComponentModel(c).setFields ( alypPseudoEmpY_giv_C.get(c) );
//////getComponentModel(c).writeFields(stderr,c.getString());
}
//////Mixture2DModel<M,Y,C>::writeFields(stderr,"");
fprintf(stderr," log pdf data: %g\n",(double)logpdfData);
// // Quit if total prob doesn't improve enough...
// fprintf(stderr," log avg pdf: %g - %g < %g ORLY? %d\n",
// (double)logpdfDataAvg,(double)logpdfPrevDataAvg,(double)WEIGHT_LIMIT,
// ((logpdfDataAvg - logpdfPrevDataAvg) < WEIGHT_LIMIT) );
// bShouldStop = ( (logpdfDataAvg - logpdfPrevDataAvg) < WEIGHT_LIMIT );
// logpdfPrevDataAvg = logpdfDataAvg;
}
////////////////////////////////////////
template <template <class MY> class M,class Y,class C>
void TrainableMixture2DModel<M,Y,C>::train ( List<Joint2DRV<Y,Prob> >& lyp, const int EPOCH_LIMIT, const PDFVal WEIGHT_LIMIT ) {
// Normalize model...
Mixture2DModel<M,Y,C>::setMixingModel().normalize(); // OUGHT NOT TO BE NECESSARY!!!
// Normalize input...
PDFVal pdfNorm = 0.0;
for ( const ListedObject<Joint2DRV<Y,Prob> >* pyp=lyp.getFirst(); pyp; pyp=lyp.getNext(pyp) ) pdfNorm += pyp->getSub2();
for ( ListedObject<Joint2DRV<Y,Prob> >* pyp=lyp.setFirst(); pyp; pyp=lyp.setNext(pyp) ) pyp->setSub2() /= pdfNorm;
// Iterate over each epoch...
bool bShouldStop = true;
for ( int epoch=0; epoch<EPOCH_LIMIT; epoch++ ) {
fprintf(stderr,"---------- EPOCH %d/%d ----------\n",epoch,EPOCH_LIMIT);
clock_t start = clock();
updateFields ( lyp, WEIGHT_LIMIT, bShouldStop ) ;
clock_t end = clock();
fprintf ( stderr, "epoch time: %d ticks or %f seconds\n", end-start, ((double)end - start)/CLOCKS_PER_SEC ) ;
if (bShouldStop) break;
}
}
////////////////////////////////////////////////////////////////////////////////
////////////////////////////////////////////////////////////////////////////////
template <template <class MY> class M,class Y,class X,class C>
class TrainableMixture3DModel : public Generic2DModel<Y,X,C> {
private:
string sId;
SimpleHash<X,List<Joint2DRV<Y,Prob> > > alyp;
SimpleHash<X,TrainableMixture2DModel<M,Y,C> > am;
public:
// Constructor / destructor methods...
TrainableMixture3DModel ( string s ) : sId(s) { }
// Specification methods...
void train ( const int, const PDFVal );
void train ( const List<Joint3DRV<X,Y,Prob> >&, const int, const PDFVal );
// Extraction methods...
Prob getProb ( const Y& y, const X& x ) const { return am.get(x).getProb(y); }
// Input / output methods...
bool readData ( char* as[], int );
bool readFields ( char* as[], int numF ) { return ( am.set(X(as[1])).readFields(as+1,numF-1) ); }
void writeFields ( FILE*, string sPref );
};
////////////////////////////////////////
template <template <class MY> class M,class Y,class X,class C>
void TrainableMixture3DModel<M,Y,X,C>::train ( const int EPOCH_LIMIT, const PDFVal WEIGHT_LIMIT ) {
// Update each subphone from list...
int ctr = 0;
X x; for ( bool b=x.setFirst(); b; b=x.setNext() ) {
if (OUTPUT_NOISY) fprintf(stderr,"==================== SUBMODEL %s (number %d) %d data pts ====================\n",
x.getString().c_str(),ctr++,alyp.get(x).getCard());
am.set(x).train ( alyp.set(x), EPOCH_LIMIT, WEIGHT_LIMIT );
}
}
////////////////////////////////////////
template <template <class MY> class M,class Y,class X,class C>
void TrainableMixture3DModel<M,Y,X,C>::train ( const List<Joint3DRV<X,Y,Prob> >& lxyp, const int EPOCH_LIMIT, const PDFVal WEIGHT_LIMIT ) {
// Chop list into phone-specific sub-lists...
ListedObject<Joint3DRV<X,Y,Prob> >* pxyp;
for ( pxyp=lxyp.getFirst(); pxyp; pxyp=lxyp.getNext(pxyp) )
alyp.set(pxyp->getSub1()).add() = Joint2DRV<Y,Prob> ( pxyp->getSub2(), pxyp->getSub3() );
// Update each subphone from list...
train(EPOCH_LIMIT,WEIGHT_LIMIT);
}
////////////////////////////////////////
template <template <class MY> class M,class Y,class X,class C>
bool TrainableMixture3DModel<M,Y,X,C>::readData ( char* as[], int numFields ) {
if ( /*as[0]!=sId+"dat" ||*/ numFields!=3 ) return false;
alyp.set(X(as[1])).add() = Joint2DRV<Y,Prob>(Y(as[2]),Prob(1.0));
return true;
}
////////////////////////////////////////
template <template <class MY> class M,class Y,class X,class C>
void TrainableMixture3DModel<M,Y,X,C>::writeFields ( FILE* pf, string sPref ) {
X x; for ( bool b=x.setFirst(); b; b=x.setNext() ) {
am.get(x).writeFields(pf,sPref+" "+x.getString());
}
}
#endif /*_NL_MIXTURE__*/
|
