blob: 3314f74a64a30deac09d70ca282cee4494fa6b4c (
plain)
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
|
package phrase;
import java.io.PrintStream;
import java.util.Arrays;
import java.util.List;
import optimization.gradientBasedMethods.ProjectedGradientDescent;
import optimization.gradientBasedMethods.ProjectedObjective;
import optimization.gradientBasedMethods.stats.OptimizerStats;
import optimization.linesearch.ArmijoLineSearchMinimizationAlongProjectionArc;
import optimization.linesearch.InterpolationPickFirstStep;
import optimization.linesearch.LineSearchMethod;
import optimization.linesearch.WolfRuleLineSearch;
import optimization.projections.SimplexProjection;
import optimization.stopCriteria.CompositeStopingCriteria;
import optimization.stopCriteria.ProjectedGradientL2Norm;
import optimization.stopCriteria.StopingCriteria;
import optimization.stopCriteria.ValueDifference;
import optimization.util.MathUtils;
public class PhraseObjective extends ProjectedObjective
{
static final double GRAD_DIFF = 0.00002;
static double INIT_STEP_SIZE = 300;
static double VAL_DIFF = 1e-4; // FIXME needs to be tuned - and this might be too weak
static int ITERATIONS = 100;
//private double c1=0.0001; // wolf stuff
//private double c2=0.9;
private static double lambda[][];
private PhraseCluster c;
/**@brief
* for debugging purposes
*/
//public static PrintStream ps;
/**@brief current phrase being optimzed*/
public int phrase;
/**@brief un-regularized posterior
* unnormalized
* p[edge][tag]
* P(tag|edge) \propto P(tag|phrase)P(context|tag)
*/
private double[][]p;
/**@brief regularized posterior
* q[edge][tag] propto p[edge][tag]*exp(-lambda)
*/
private double q[][];
private List<Corpus.Edge> data;
/**@brief log likelihood of the associated phrase
*
*/
private double loglikelihood;
private SimplexProjection projection;
double[] newPoint ;
private int n_param;
/**@brief likelihood under p
*
*/
public double llh;
public PhraseObjective(PhraseCluster cluster, int phraseIdx){
phrase=phraseIdx;
c=cluster;
data=c.c.getEdgesForPhrase(phrase);
n_param=data.size()*c.K;
if (lambda==null){
lambda=new double[c.c.getNumPhrases()][];
}
if (lambda[phrase]==null){
lambda[phrase]=new double[n_param];
}
parameters=lambda[phrase];
newPoint = new double[n_param];
gradient = new double[n_param];
initP();
projection=new SimplexProjection(c.scalePT);
q=new double [data.size()][c.K];
setParameters(parameters);
}
private void initP(){
p=new double[data.size()][];
for(int edge=0;edge<data.size();edge++){
p[edge]=c.posterior(data.get(edge));
llh += data.get(edge).getCount() * Math.log(arr.F.l1norm(p[edge])); // Was bug here - count inside log!
arr.F.l1normalize(p[edge]);
}
}
@Override
public void setParameters(double[] params) {
super.setParameters(params);
updateFunction();
}
private void updateFunction(){
updateCalls++;
loglikelihood=0;
for(int tag=0;tag<c.K;tag++){
for(int edge=0;edge<data.size();edge++){
q[edge][tag]=p[edge][tag]*
Math.exp(-parameters[tag*data.size()+edge]/data.get(edge).getCount());
}
}
for(int edge=0;edge<data.size();edge++){
loglikelihood+=data.get(edge).getCount() * Math.log(arr.F.l1norm(q[edge]));
arr.F.l1normalize(q[edge]);
}
for(int tag=0;tag<c.K;tag++){
for(int edge=0;edge<data.size();edge++){
gradient[tag*data.size()+edge]=-q[edge][tag];
}
}
}
@Override
public double[] projectPoint(double[] point)
{
double toProject[]=new double[data.size()];
for(int tag=0;tag<c.K;tag++){
for(int edge=0;edge<data.size();edge++){
toProject[edge]=point[tag*data.size()+edge];
}
projection.project(toProject);
for(int edge=0;edge<data.size();edge++){
newPoint[tag*data.size()+edge]=toProject[edge];
}
}
return newPoint;
}
@Override
public double[] getGradient() {
gradientCalls++;
return gradient;
}
@Override
public double getValue() {
functionCalls++;
return loglikelihood;
}
@Override
public String toString() {
return "No need for pointless toString";
}
public double [][]posterior(){
return q;
}
public void optimizeWithProjectedGradientDescent(){
LineSearchMethod ls =
new ArmijoLineSearchMinimizationAlongProjectionArc
(new InterpolationPickFirstStep(INIT_STEP_SIZE));
//LineSearchMethod ls = new WolfRuleLineSearch(
// (new InterpolationPickFirstStep(INIT_STEP_SIZE)), c1, c2);
OptimizerStats stats = new OptimizerStats();
ProjectedGradientDescent optimizer = new ProjectedGradientDescent(ls);
StopingCriteria stopGrad = new ProjectedGradientL2Norm(GRAD_DIFF);
StopingCriteria stopValue = new ValueDifference(VAL_DIFF*(-llh));
CompositeStopingCriteria compositeStop = new CompositeStopingCriteria();
compositeStop.add(stopGrad);
compositeStop.add(stopValue);
optimizer.setMaxIterations(ITERATIONS);
updateFunction();
boolean succed = optimizer.optimize(this,stats,compositeStop);
// System.out.println("Ended optimzation Projected Gradient Descent\n" + stats.prettyPrint(1));
if(succed){
//System.out.println("Ended optimization in " + optimizer.getCurrentIteration());
}else{
System.out.println("Failed to optimize");
}
lambda[phrase]=parameters;
// ps.println(Arrays.toString(parameters));
// for(int edge=0;edge<data.getSize();edge++){
// ps.println(Arrays.toString(q[edge]));
// }
}
public double KL_divergence()
{
return -loglikelihood + MathUtils.dotProduct(parameters, gradient);
}
public double loglikelihood()
{
return llh;
}
public double l1lmax()
{
double sum=0;
for(int tag=0;tag<c.K;tag++){
double max=0;
for(int edge=0;edge<data.size();edge++){
if(q[edge][tag]>max)
max=q[edge][tag];
}
sum+=max;
}
return sum;
}
public double primal()
{
return loglikelihood() - KL_divergence() - c.scalePT * l1lmax();
}
}
|
