summaryrefslogtreecommitdiff
path: root/gi/posterior-regularisation/prjava/src/phrase/PhraseContextObjective.java
blob: 5947c4be9229fc69999e7a79ca6bd984bf29a35e (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
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
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
package phrase;

import java.util.ArrayList;
import java.util.Arrays;
import java.util.HashMap;
import java.util.List;
import java.util.Map;
import java.util.concurrent.ExecutionException;
import java.util.concurrent.ExecutorService;
import java.util.concurrent.Future;

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.projections.SimplexProjection;
import optimization.stopCriteria.CompositeStopingCriteria;
import optimization.stopCriteria.ProjectedGradientL2Norm;
import optimization.stopCriteria.StopingCriteria;
import optimization.stopCriteria.ValueDifference;
import optimization.util.MathUtils;
import phrase.Corpus.Edge;

public class PhraseContextObjective extends ProjectedObjective
{
	private static final double GRAD_DIFF = 0.00002;
	private static double INIT_STEP_SIZE = 300;
	private static double VAL_DIFF = 1e-4; // FIXME needs to be tuned
	private static int ITERATIONS = 100;
	
	private PhraseCluster c;
	
	// un-regularized unnormalized posterior, p[edge][tag]
	// P(tag|edge) \propto P(tag|phrase)P(context|tag)
	private double p[][];

	// regularized unnormalized posterior 
	// q[edge][tag] propto p[edge][tag]*exp(-lambda)
	private double q[][];
	private List<Corpus.Edge> data;
	
	// log likelihood under q
	private double loglikelihood;
	private SimplexProjection projectionPhrase;
	private SimplexProjection projectionContext;
	
	double[] newPoint;
	private int n_param;
	
	// likelihood under p
	public double llh;
	
	private Map<Corpus.Edge, Integer> edgeIndex;
	
	private long projectionTime;
	private long objectiveTime;
	private long actualProjectionTime;
	private ExecutorService pool;
	
	double scalePT;
	double scaleCT;
	
	public PhraseContextObjective(PhraseCluster cluster, double[] startingParameters, ExecutorService pool,
			double scalePT, double scaleCT)
	{
		c=cluster;
		data=c.c.getEdges();
		n_param=data.size()*c.K*2;
		this.pool=pool;
		this.scalePT = scalePT;
		this.scaleCT = scaleCT;
		
		parameters = startingParameters;
		if (parameters == null)
			parameters = new double[n_param];
		
		System.out.println("Num parameters " + n_param);
		newPoint = new double[n_param];
		gradient = new double[n_param];
		initP();
		projectionPhrase = new SimplexProjection(scalePT);
		projectionContext = new SimplexProjection(scaleCT);
		q=new double [data.size()][c.K];
		
		edgeIndex = new HashMap<Edge, Integer>();
		for (int e=0; e<data.size(); e++)
			edgeIndex.put(data.get(e), e);

		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]));
			arr.F.l1normalize(p[edge]);
		}
	}
	
	@Override
	public void setParameters(double[] params) {
		//System.out.println("setParameters " + Arrays.toString(parameters));
		// TODO: test if params have changed and skip update otherwise
		super.setParameters(params);
		updateFunction();
	}
	
	private void updateFunction()
	{
		updateCalls++;
		loglikelihood=0;
		System.out.print(".");
		System.out.flush();

		long begin = System.currentTimeMillis();
		for (int e=0; e<data.size(); e++) 
		{
			Edge edge = data.get(e);
			int offset = edgeIndex.get(edge)*c.K*2;
			for(int tag=0; tag<c.K; tag++)
			{
				int ip = offset + tag*2;
				int ic = ip + 1;
				q[e][tag] = p[e][tag]*
					Math.exp((-parameters[ip]-parameters[ic]) / 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 e=0; e<data.size(); e++) 
		{
			Edge edge = data.get(e);
			int offset = edgeIndex.get(edge)*c.K*2;
			for(int tag=0; tag<c.K; tag++)
			{
				int ip = offset + tag*2;
				int ic = ip + 1;		
				gradient[ip]=-q[e][tag];
				gradient[ic]=-q[e][tag];
			}
		}
		//System.out.println("objective " + loglikelihood + " ||gradient||_2: " + arr.F.l2norm(gradient));		
		objectiveTime += System.currentTimeMillis() - begin;
	}
	
	@Override
	public double[] projectPoint(double[] point) 
	{
		long begin = System.currentTimeMillis();
		List<Future<?>> tasks = new ArrayList<Future<?>>();
		
		System.out.print(",");
		System.out.flush();

		//System.out.println("\t\tprojectPoint: " + Arrays.toString(point));
		Arrays.fill(newPoint, 0, newPoint.length, 0);
		
		// first project using the phrase-tag constraints,
		// for all p,t: sum_c lambda_ptc < scaleP 
		if (pool == null)
		{
			for (int p = 0; p < c.c.getNumPhrases(); ++p)
			{
				List<Edge> edges = c.c.getEdgesForPhrase(p);
				double[] toProject = new double[edges.size()];
				for(int tag=0;tag<c.K;tag++)
				{
					for(int e=0; e<edges.size(); e++)
						toProject[e] = point[index(edges.get(e), tag, true)];
					long lbegin = System.currentTimeMillis();
					projectionPhrase.project(toProject);
					actualProjectionTime += System.currentTimeMillis() - lbegin;
					for(int e=0; e<edges.size(); e++)
						newPoint[index(edges.get(e), tag, true)] = toProject[e];
				}
			}
		}
		else // do above in parallel using thread pool
		{	
			for (int p = 0; p < c.c.getNumPhrases(); ++p)
			{
				final int phrase = p;
				final double[] inPoint = point;
				Runnable task = new Runnable()
				{
					public void run()
					{
						List<Edge> edges = c.c.getEdgesForPhrase(phrase);
						double toProject[] = new double[edges.size()];
						for(int tag=0;tag<c.K;tag++)
						{
							for(int e=0; e<edges.size(); e++)
								toProject[e] = inPoint[index(edges.get(e), tag, true)];
							projectionPhrase.project(toProject);
							for(int e=0; e<edges.size(); e++)
								newPoint[index(edges.get(e), tag, true)] = toProject[e];
						}
					}		
				};
				tasks.add(pool.submit(task));
			}
		}
		//System.out.println("after PT " + Arrays.toString(newPoint));
	
		// now project using the context-tag constraints,
		// for all c,t: sum_p omega_pct < scaleC
		if (pool == null)
		{
			for (int ctx = 0; ctx < c.c.getNumContexts(); ++ctx)
			{
				List<Edge> edges = c.c.getEdgesForContext(ctx);
				double toProject[] = new double[edges.size()];
				for(int tag=0;tag<c.K;tag++)
				{
					for(int e=0; e<edges.size(); e++)
						toProject[e] = point[index(edges.get(e), tag, false)];
					long lbegin = System.currentTimeMillis();
					projectionContext.project(toProject);
					actualProjectionTime += System.currentTimeMillis() - lbegin;
					for(int e=0; e<edges.size(); e++)
						newPoint[index(edges.get(e), tag, false)] = toProject[e];
				}
			}
		}
		else
		{
			// do above in parallel using thread pool
			for (int ctx = 0; ctx < c.c.getNumContexts(); ++ctx)
			{
				final int context = ctx;
				final double[] inPoint = point;
				Runnable task = new Runnable()
				{
					public void run()
					{
						List<Edge> edges = c.c.getEdgesForContext(context);
						double toProject[] = new double[edges.size()];
						for(int tag=0;tag<c.K;tag++)
						{
							for(int e=0; e<edges.size(); e++)
								toProject[e] = inPoint[index(edges.get(e), tag, false)];
							projectionContext.project(toProject);
							for(int e=0; e<edges.size(); e++)
								newPoint[index(edges.get(e), tag, false)] = toProject[e];
						}
					}
				};
				tasks.add(pool.submit(task));
			}
		}
		
		if (pool != null)
		{
			// wait for all the jobs to complete
			Exception failure = null;
			for (Future<?> task: tasks)
			{
				try {
					task.get();
				} catch (InterruptedException e) {
					System.err.println("ERROR: Projection thread interrupted");
					e.printStackTrace();
					failure = e;
				} catch (ExecutionException e) {
					System.err.println("ERROR: Projection thread died");
					e.printStackTrace();
					failure = e;
				}
			}
			// rethrow the exception
			if (failure != null)
			{
				pool.shutdownNow();
				throw new RuntimeException(failure);
			}
		}
		
		double[] tmp = newPoint;
		newPoint = point;
		projectionTime += System.currentTimeMillis() - begin;
		
		//System.out.println("\t\treturning " + Arrays.toString(tmp));
		return tmp;
	}
	
	private int index(Edge edge, int tag, boolean phrase)
	{
		// NB if indexing changes must also change code in updateFunction and constructor
		if (phrase)
			return edgeIndex.get(edge)*c.K*2 + tag*2;
		else
			return edgeIndex.get(edge)*c.K*2 + tag*2 + 1;
	}

	@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(int edgeIndex){
		return q[edgeIndex];
	}
	
	public boolean optimizeWithProjectedGradientDescent()
	{
		projectionTime = 0;
		actualProjectionTime = 0;
		objectiveTime = 0;
		long start = System.currentTimeMillis();

		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 success = optimizer.optimize(this,stats,compositeStop);
//		System.out.println("Ended optimzation Projected Gradient Descent\n" + stats.prettyPrint(1));

		System.out.println();
		
		if (success)
			System.out.print("\toptimization took " + optimizer.getCurrentIteration() + " iterations");
	 	else
			System.out.print("\toptimization failed to converge");
		long total = System.currentTimeMillis() - start;
		System.out.println(" and " + total + " ms: projection " + projectionTime + 
				" actual " + actualProjectionTime + " objective " + objectiveTime);

		return success;
	}
	
	double loglikelihood()
	{
		return llh;
	}
	
	double KL_divergence()
	{
		return -loglikelihood + MathUtils.dotProduct(parameters, gradient);
	}
	
	double phrase_l1lmax()
	{
		// \sum_{tag,phrase} max_{context} P(tag|context,phrase)
		double sum=0;
		for (int p = 0; p < c.c.getNumPhrases(); ++p)
		{
			List<Edge> edges = c.c.getEdgesForPhrase(p);
			for(int tag=0;tag<c.K;tag++)
			{
				double max=0;
				for (Edge edge: edges)
					max = Math.max(max, q[edgeIndex.get(edge)][tag]);
				sum+=max;
			}	
		}
		return sum;
	}
	
	double context_l1lmax()
	{
		// \sum_{tag,context} max_{phrase} P(tag|context,phrase)
		double sum=0;
		for (int ctx = 0; ctx < c.c.getNumContexts(); ++ctx)
		{
			List<Edge> edges = c.c.getEdgesForContext(ctx);
			for(int tag=0; tag<c.K; tag++)
			{
				double max=0;
				for (Edge edge: edges)
					max = Math.max(max, q[edgeIndex.get(edge)][tag]);
				sum+=max;
			}	
		}
		return sum;
	}
	
	// L - KL(q||p) - scalePT * l1lmax_phrase - scaleCT * l1lmax_context
	public double primal()
	{
		return loglikelihood() - KL_divergence() - scalePT * phrase_l1lmax() - scaleCT * context_l1lmax();
	}
}