1 | /* |
---|
2 | * This program is free software; you can redistribute it and/or modify |
---|
3 | * it under the terms of the GNU General Public License as published by |
---|
4 | * the Free Software Foundation; either version 2 of the License, or |
---|
5 | * (at your option) any later version. |
---|
6 | * |
---|
7 | * This program is distributed in the hope that it will be useful, |
---|
8 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
---|
9 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
---|
10 | * GNU General Public License for more details. |
---|
11 | * |
---|
12 | * You should have received a copy of the GNU General Public License |
---|
13 | * along with this program; if not, write to the Free Software |
---|
14 | * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. |
---|
15 | */ |
---|
16 | |
---|
17 | /* |
---|
18 | * MINND.java |
---|
19 | * Copyright (C) 2005 University of Waikato, Hamilton, New Zealand |
---|
20 | * |
---|
21 | */ |
---|
22 | |
---|
23 | package weka.classifiers.mi; |
---|
24 | |
---|
25 | import weka.classifiers.Classifier; |
---|
26 | import weka.classifiers.AbstractClassifier; |
---|
27 | import weka.core.Capabilities; |
---|
28 | import weka.core.Instance; |
---|
29 | import weka.core.DenseInstance; |
---|
30 | import weka.core.Instances; |
---|
31 | import weka.core.MultiInstanceCapabilitiesHandler; |
---|
32 | import weka.core.Option; |
---|
33 | import weka.core.OptionHandler; |
---|
34 | import weka.core.RevisionUtils; |
---|
35 | import weka.core.TechnicalInformation; |
---|
36 | import weka.core.TechnicalInformationHandler; |
---|
37 | import weka.core.Utils; |
---|
38 | import weka.core.Capabilities.Capability; |
---|
39 | import weka.core.TechnicalInformation.Field; |
---|
40 | import weka.core.TechnicalInformation.Type; |
---|
41 | |
---|
42 | import java.util.Enumeration; |
---|
43 | import java.util.Vector; |
---|
44 | |
---|
45 | /** |
---|
46 | <!-- globalinfo-start --> |
---|
47 | * Multiple-Instance Nearest Neighbour with Distribution learner.<br/> |
---|
48 | * <br/> |
---|
49 | * It uses gradient descent to find the weight for each dimension of each exeamplar from the starting point of 1.0. In order to avoid overfitting, it uses mean-square function (i.e. the Euclidean distance) to search for the weights.<br/> |
---|
50 | * It then uses the weights to cleanse the training data. After that it searches for the weights again from the starting points of the weights searched before.<br/> |
---|
51 | * Finally it uses the most updated weights to cleanse the test exemplar and then finds the nearest neighbour of the test exemplar using partly-weighted Kullback distance. But the variances in the Kullback distance are the ones before cleansing.<br/> |
---|
52 | * <br/> |
---|
53 | * For more information see:<br/> |
---|
54 | * <br/> |
---|
55 | * Xin Xu (2001). A nearest distribution approach to multiple-instance learning. Hamilton, NZ. |
---|
56 | * <p/> |
---|
57 | <!-- globalinfo-end --> |
---|
58 | * |
---|
59 | <!-- technical-bibtex-start --> |
---|
60 | * BibTeX: |
---|
61 | * <pre> |
---|
62 | * @misc{Xu2001, |
---|
63 | * address = {Hamilton, NZ}, |
---|
64 | * author = {Xin Xu}, |
---|
65 | * note = {0657.591B}, |
---|
66 | * school = {University of Waikato}, |
---|
67 | * title = {A nearest distribution approach to multiple-instance learning}, |
---|
68 | * year = {2001} |
---|
69 | * } |
---|
70 | * </pre> |
---|
71 | * <p/> |
---|
72 | <!-- technical-bibtex-end --> |
---|
73 | * |
---|
74 | <!-- options-start --> |
---|
75 | * Valid options are: <p/> |
---|
76 | * |
---|
77 | * <pre> -K <number of neighbours> |
---|
78 | * Set number of nearest neighbour for prediction |
---|
79 | * (default 1)</pre> |
---|
80 | * |
---|
81 | * <pre> -S <number of neighbours> |
---|
82 | * Set number of nearest neighbour for cleansing the training data |
---|
83 | * (default 1)</pre> |
---|
84 | * |
---|
85 | * <pre> -E <number of neighbours> |
---|
86 | * Set number of nearest neighbour for cleansing the testing data |
---|
87 | * (default 1)</pre> |
---|
88 | * |
---|
89 | <!-- options-end --> |
---|
90 | * |
---|
91 | * @author Xin Xu (xx5@cs.waikato.ac.nz) |
---|
92 | * @version $Revision: 5987 $ |
---|
93 | */ |
---|
94 | public class MINND |
---|
95 | extends AbstractClassifier |
---|
96 | implements OptionHandler, MultiInstanceCapabilitiesHandler, |
---|
97 | TechnicalInformationHandler { |
---|
98 | |
---|
99 | /** for serialization */ |
---|
100 | static final long serialVersionUID = -4512599203273864994L; |
---|
101 | |
---|
102 | /** The number of nearest neighbour for prediction */ |
---|
103 | protected int m_Neighbour = 1; |
---|
104 | |
---|
105 | /** The mean for each attribute of each exemplar */ |
---|
106 | protected double[][] m_Mean = null; |
---|
107 | |
---|
108 | /** The variance for each attribute of each exemplar */ |
---|
109 | protected double[][] m_Variance = null; |
---|
110 | |
---|
111 | /** The dimension of each exemplar, i.e. (numAttributes-2) */ |
---|
112 | protected int m_Dimension = 0; |
---|
113 | |
---|
114 | /** header info of the data */ |
---|
115 | protected Instances m_Attributes;; |
---|
116 | |
---|
117 | /** The class label of each exemplar */ |
---|
118 | protected double[] m_Class = null; |
---|
119 | |
---|
120 | /** The number of class labels in the data */ |
---|
121 | protected int m_NumClasses = 0; |
---|
122 | |
---|
123 | /** The weight of each exemplar */ |
---|
124 | protected double[] m_Weights = null; |
---|
125 | |
---|
126 | /** The very small number representing zero */ |
---|
127 | static private double m_ZERO = 1.0e-45; |
---|
128 | |
---|
129 | /** The learning rate in the gradient descent */ |
---|
130 | protected double m_Rate = -1; |
---|
131 | |
---|
132 | /** The minimum values for numeric attributes. */ |
---|
133 | private double [] m_MinArray=null; |
---|
134 | |
---|
135 | /** The maximum values for numeric attributes. */ |
---|
136 | private double [] m_MaxArray=null; |
---|
137 | |
---|
138 | /** The stopping criteria of gradient descent*/ |
---|
139 | private double m_STOP = 1.0e-45; |
---|
140 | |
---|
141 | /** The weights that alter the dimnesion of each exemplar */ |
---|
142 | private double[][] m_Change=null; |
---|
143 | |
---|
144 | /** The noise data of each exemplar */ |
---|
145 | private double[][] m_NoiseM = null, m_NoiseV = null, m_ValidM = null, |
---|
146 | m_ValidV = null; |
---|
147 | |
---|
148 | /** The number of nearest neighbour instances in the selection of noises |
---|
149 | in the training data*/ |
---|
150 | private int m_Select = 1; |
---|
151 | |
---|
152 | /** The number of nearest neighbour exemplars in the selection of noises |
---|
153 | in the test data */ |
---|
154 | private int m_Choose = 1; |
---|
155 | |
---|
156 | /** The decay rate of learning rate */ |
---|
157 | private double m_Decay = 0.5; |
---|
158 | |
---|
159 | /** |
---|
160 | * Returns a string describing this filter |
---|
161 | * |
---|
162 | * @return a description of the filter suitable for |
---|
163 | * displaying in the explorer/experimenter gui |
---|
164 | */ |
---|
165 | public String globalInfo() { |
---|
166 | return |
---|
167 | "Multiple-Instance Nearest Neighbour with Distribution learner.\n\n" |
---|
168 | + "It uses gradient descent to find the weight for each dimension of " |
---|
169 | + "each exeamplar from the starting point of 1.0. In order to avoid " |
---|
170 | + "overfitting, it uses mean-square function (i.e. the Euclidean " |
---|
171 | + "distance) to search for the weights.\n " |
---|
172 | + "It then uses the weights to cleanse the training data. After that " |
---|
173 | + "it searches for the weights again from the starting points of the " |
---|
174 | + "weights searched before.\n " |
---|
175 | + "Finally it uses the most updated weights to cleanse the test exemplar " |
---|
176 | + "and then finds the nearest neighbour of the test exemplar using " |
---|
177 | + "partly-weighted Kullback distance. But the variances in the Kullback " |
---|
178 | + "distance are the ones before cleansing.\n\n" |
---|
179 | + "For more information see:\n\n" |
---|
180 | + getTechnicalInformation().toString(); |
---|
181 | } |
---|
182 | |
---|
183 | /** |
---|
184 | * Returns an instance of a TechnicalInformation object, containing |
---|
185 | * detailed information about the technical background of this class, |
---|
186 | * e.g., paper reference or book this class is based on. |
---|
187 | * |
---|
188 | * @return the technical information about this class |
---|
189 | */ |
---|
190 | public TechnicalInformation getTechnicalInformation() { |
---|
191 | TechnicalInformation result; |
---|
192 | |
---|
193 | result = new TechnicalInformation(Type.MISC); |
---|
194 | result.setValue(Field.AUTHOR, "Xin Xu"); |
---|
195 | result.setValue(Field.YEAR, "2001"); |
---|
196 | result.setValue(Field.TITLE, "A nearest distribution approach to multiple-instance learning"); |
---|
197 | result.setValue(Field.SCHOOL, "University of Waikato"); |
---|
198 | result.setValue(Field.ADDRESS, "Hamilton, NZ"); |
---|
199 | result.setValue(Field.NOTE, "0657.591B"); |
---|
200 | |
---|
201 | return result; |
---|
202 | } |
---|
203 | |
---|
204 | /** |
---|
205 | * Returns default capabilities of the classifier. |
---|
206 | * |
---|
207 | * @return the capabilities of this classifier |
---|
208 | */ |
---|
209 | public Capabilities getCapabilities() { |
---|
210 | Capabilities result = super.getCapabilities(); |
---|
211 | result.disableAll(); |
---|
212 | |
---|
213 | // attributes |
---|
214 | result.enable(Capability.NOMINAL_ATTRIBUTES); |
---|
215 | result.enable(Capability.RELATIONAL_ATTRIBUTES); |
---|
216 | result.enable(Capability.MISSING_VALUES); |
---|
217 | |
---|
218 | // class |
---|
219 | result.enable(Capability.NOMINAL_CLASS); |
---|
220 | result.enable(Capability.MISSING_CLASS_VALUES); |
---|
221 | |
---|
222 | // other |
---|
223 | result.enable(Capability.ONLY_MULTIINSTANCE); |
---|
224 | |
---|
225 | return result; |
---|
226 | } |
---|
227 | |
---|
228 | /** |
---|
229 | * Returns the capabilities of this multi-instance classifier for the |
---|
230 | * relational data. |
---|
231 | * |
---|
232 | * @return the capabilities of this object |
---|
233 | * @see Capabilities |
---|
234 | */ |
---|
235 | public Capabilities getMultiInstanceCapabilities() { |
---|
236 | Capabilities result = super.getCapabilities(); |
---|
237 | result.disableAll(); |
---|
238 | |
---|
239 | // attributes |
---|
240 | result.enable(Capability.NUMERIC_ATTRIBUTES); |
---|
241 | result.enable(Capability.DATE_ATTRIBUTES); |
---|
242 | result.enable(Capability.MISSING_VALUES); |
---|
243 | |
---|
244 | // class |
---|
245 | result.disableAllClasses(); |
---|
246 | result.enable(Capability.NO_CLASS); |
---|
247 | |
---|
248 | return result; |
---|
249 | } |
---|
250 | |
---|
251 | /** |
---|
252 | * As normal Nearest Neighbour algorithm does, it's lazy and simply |
---|
253 | * records the exemplar information (i.e. mean and variance for each |
---|
254 | * dimension of each exemplar and their classes) when building the model. |
---|
255 | * There is actually no need to store the exemplars themselves. |
---|
256 | * |
---|
257 | * @param exs the training exemplars |
---|
258 | * @throws Exception if the model cannot be built properly |
---|
259 | */ |
---|
260 | public void buildClassifier(Instances exs)throws Exception{ |
---|
261 | // can classifier handle the data? |
---|
262 | getCapabilities().testWithFail(exs); |
---|
263 | |
---|
264 | // remove instances with missing class |
---|
265 | Instances newData = new Instances(exs); |
---|
266 | newData.deleteWithMissingClass(); |
---|
267 | |
---|
268 | int numegs = newData.numInstances(); |
---|
269 | m_Dimension = newData.attribute(1).relation().numAttributes(); |
---|
270 | m_Attributes = newData.stringFreeStructure(); |
---|
271 | m_Change = new double[numegs][m_Dimension]; |
---|
272 | m_NumClasses = exs.numClasses(); |
---|
273 | m_Mean = new double[numegs][m_Dimension]; |
---|
274 | m_Variance = new double[numegs][m_Dimension]; |
---|
275 | m_Class = new double[numegs]; |
---|
276 | m_Weights = new double[numegs]; |
---|
277 | m_NoiseM = new double[numegs][m_Dimension]; |
---|
278 | m_NoiseV = new double[numegs][m_Dimension]; |
---|
279 | m_ValidM = new double[numegs][m_Dimension]; |
---|
280 | m_ValidV = new double[numegs][m_Dimension]; |
---|
281 | m_MinArray = new double[m_Dimension]; |
---|
282 | m_MaxArray = new double[m_Dimension]; |
---|
283 | for(int v=0; v < m_Dimension; v++) |
---|
284 | m_MinArray[v] = m_MaxArray[v] = Double.NaN; |
---|
285 | |
---|
286 | for(int w=0; w < numegs; w++){ |
---|
287 | updateMinMax(newData.instance(w)); |
---|
288 | } |
---|
289 | |
---|
290 | // Scale exemplars |
---|
291 | Instances data = m_Attributes; |
---|
292 | |
---|
293 | for(int x=0; x < numegs; x++){ |
---|
294 | Instance example = newData.instance(x); |
---|
295 | example = scale(example); |
---|
296 | for (int i=0; i<m_Dimension; i++) { |
---|
297 | m_Mean[x][i] = example.relationalValue(1).meanOrMode(i); |
---|
298 | m_Variance[x][i] = example.relationalValue(1).variance(i); |
---|
299 | if(Utils.eq(m_Variance[x][i],0.0)) |
---|
300 | m_Variance[x][i] = m_ZERO; |
---|
301 | m_Change[x][i] = 1.0; |
---|
302 | } |
---|
303 | /* for(int y=0; y < m_Variance[x].length; y++){ |
---|
304 | if(Utils.eq(m_Variance[x][y],0.0)) |
---|
305 | m_Variance[x][y] = m_ZERO; |
---|
306 | m_Change[x][y] = 1.0; |
---|
307 | } */ |
---|
308 | |
---|
309 | data.add(example); |
---|
310 | m_Class[x] = example.classValue(); |
---|
311 | m_Weights[x] = example.weight(); |
---|
312 | } |
---|
313 | |
---|
314 | for(int z=0; z < numegs; z++) |
---|
315 | findWeights(z, m_Mean); |
---|
316 | |
---|
317 | // Pre-process and record "true estimated" parameters for distributions |
---|
318 | for(int x=0; x < numegs; x++){ |
---|
319 | Instance example = preprocess(data, x); |
---|
320 | if (getDebug()) |
---|
321 | System.out.println("???Exemplar "+x+" has been pre-processed:"+ |
---|
322 | data.instance(x).relationalValue(1).sumOfWeights()+ |
---|
323 | "|"+example.relationalValue(1).sumOfWeights()+ |
---|
324 | "; class:"+m_Class[x]); |
---|
325 | if(Utils.gr(example.relationalValue(1).sumOfWeights(), 0)){ |
---|
326 | for (int i=0; i<m_Dimension; i++) { |
---|
327 | m_ValidM[x][i] = example.relationalValue(1).meanOrMode(i); |
---|
328 | m_ValidV[x][i] = example.relationalValue(1).variance(i); |
---|
329 | if(Utils.eq(m_ValidV[x][i],0.0)) |
---|
330 | m_ValidV[x][i] = m_ZERO; |
---|
331 | } |
---|
332 | /* for(int y=0; y < m_ValidV[x].length; y++){ |
---|
333 | if(Utils.eq(m_ValidV[x][y],0.0)) |
---|
334 | m_ValidV[x][y] = m_ZERO; |
---|
335 | }*/ |
---|
336 | } |
---|
337 | else{ |
---|
338 | m_ValidM[x] = null; |
---|
339 | m_ValidV[x] = null; |
---|
340 | } |
---|
341 | } |
---|
342 | |
---|
343 | for(int z=0; z < numegs; z++) |
---|
344 | if(m_ValidM[z] != null) |
---|
345 | findWeights(z, m_ValidM); |
---|
346 | |
---|
347 | } |
---|
348 | |
---|
349 | /** |
---|
350 | * Pre-process the given exemplar according to the other exemplars |
---|
351 | * in the given exemplars. It also updates noise data statistics. |
---|
352 | * |
---|
353 | * @param data the whole exemplars |
---|
354 | * @param pos the position of given exemplar in data |
---|
355 | * @return the processed exemplar |
---|
356 | * @throws Exception if the returned exemplar is wrong |
---|
357 | */ |
---|
358 | public Instance preprocess(Instances data, int pos) |
---|
359 | throws Exception{ |
---|
360 | Instance before = data.instance(pos); |
---|
361 | if((int)before.classValue() == 0){ |
---|
362 | m_NoiseM[pos] = null; |
---|
363 | m_NoiseV[pos] = null; |
---|
364 | return before; |
---|
365 | } |
---|
366 | |
---|
367 | Instances after_relationInsts =before.attribute(1).relation().stringFreeStructure(); |
---|
368 | Instances noises_relationInsts =before.attribute(1).relation().stringFreeStructure(); |
---|
369 | |
---|
370 | Instances newData = m_Attributes; |
---|
371 | Instance after = new DenseInstance(before.numAttributes()); |
---|
372 | Instance noises = new DenseInstance(before.numAttributes()); |
---|
373 | after.setDataset(newData); |
---|
374 | noises.setDataset(newData); |
---|
375 | |
---|
376 | for(int g=0; g < before.relationalValue(1).numInstances(); g++){ |
---|
377 | Instance datum = before.relationalValue(1).instance(g); |
---|
378 | double[] dists = new double[data.numInstances()]; |
---|
379 | |
---|
380 | for(int i=0; i < data.numInstances(); i++){ |
---|
381 | if(i != pos) |
---|
382 | dists[i] = distance(datum, m_Mean[i], m_Variance[i], i); |
---|
383 | else |
---|
384 | dists[i] = Double.POSITIVE_INFINITY; |
---|
385 | } |
---|
386 | |
---|
387 | int[] pred = new int[m_NumClasses]; |
---|
388 | for(int n=0; n < pred.length; n++) |
---|
389 | pred[n] = 0; |
---|
390 | |
---|
391 | for(int o=0; o<m_Select; o++){ |
---|
392 | int index = Utils.minIndex(dists); |
---|
393 | pred[(int)m_Class[index]]++; |
---|
394 | dists[index] = Double.POSITIVE_INFINITY; |
---|
395 | } |
---|
396 | |
---|
397 | int clas = Utils.maxIndex(pred); |
---|
398 | if((int)before.classValue() != clas) |
---|
399 | noises_relationInsts.add(datum); |
---|
400 | else |
---|
401 | after_relationInsts.add(datum); |
---|
402 | } |
---|
403 | |
---|
404 | int relationValue; |
---|
405 | relationValue = noises.attribute(1).addRelation( noises_relationInsts); |
---|
406 | noises.setValue(0,before.value(0)); |
---|
407 | noises.setValue(1, relationValue); |
---|
408 | noises.setValue(2, before.classValue()); |
---|
409 | |
---|
410 | relationValue = after.attribute(1).addRelation( after_relationInsts); |
---|
411 | after.setValue(0,before.value(0)); |
---|
412 | after.setValue(1, relationValue); |
---|
413 | after.setValue(2, before.classValue()); |
---|
414 | |
---|
415 | |
---|
416 | if(Utils.gr(noises.relationalValue(1).sumOfWeights(), 0)){ |
---|
417 | for (int i=0; i<m_Dimension; i++) { |
---|
418 | m_NoiseM[pos][i] = noises.relationalValue(1).meanOrMode(i); |
---|
419 | m_NoiseV[pos][i] = noises.relationalValue(1).variance(i); |
---|
420 | if(Utils.eq(m_NoiseV[pos][i],0.0)) |
---|
421 | m_NoiseV[pos][i] = m_ZERO; |
---|
422 | } |
---|
423 | /* for(int y=0; y < m_NoiseV[pos].length; y++){ |
---|
424 | if(Utils.eq(m_NoiseV[pos][y],0.0)) |
---|
425 | m_NoiseV[pos][y] = m_ZERO; |
---|
426 | } */ |
---|
427 | } |
---|
428 | else{ |
---|
429 | m_NoiseM[pos] = null; |
---|
430 | m_NoiseV[pos] = null; |
---|
431 | } |
---|
432 | |
---|
433 | return after; |
---|
434 | } |
---|
435 | |
---|
436 | /** |
---|
437 | * Calculates the distance between two instances |
---|
438 | * |
---|
439 | * @param first the first instance |
---|
440 | * @param second the second instance |
---|
441 | * @return the distance between the two given instances |
---|
442 | */ |
---|
443 | private double distance(Instance first, double[] mean, double[] var, int pos) { |
---|
444 | |
---|
445 | double diff, distance = 0; |
---|
446 | |
---|
447 | for(int i = 0; i < m_Dimension; i++) { |
---|
448 | // If attribute is numeric |
---|
449 | if(first.attribute(i).isNumeric()){ |
---|
450 | if (!first.isMissing(i)){ |
---|
451 | diff = first.value(i) - mean[i]; |
---|
452 | if(Utils.gr(var[i], m_ZERO)) |
---|
453 | distance += m_Change[pos][i] * var[i] * diff * diff; |
---|
454 | else |
---|
455 | distance += m_Change[pos][i] * diff * diff; |
---|
456 | } |
---|
457 | else{ |
---|
458 | if(Utils.gr(var[i], m_ZERO)) |
---|
459 | distance += m_Change[pos][i] * var[i]; |
---|
460 | else |
---|
461 | distance += m_Change[pos][i] * 1.0; |
---|
462 | } |
---|
463 | } |
---|
464 | |
---|
465 | } |
---|
466 | |
---|
467 | return distance; |
---|
468 | } |
---|
469 | |
---|
470 | /** |
---|
471 | * Updates the minimum and maximum values for all the attributes |
---|
472 | * based on a new exemplar. |
---|
473 | * |
---|
474 | * @param ex the new exemplar |
---|
475 | */ |
---|
476 | private void updateMinMax(Instance ex) { |
---|
477 | Instances insts = ex.relationalValue(1); |
---|
478 | for (int j = 0;j < m_Dimension; j++) { |
---|
479 | if (insts.attribute(j).isNumeric()){ |
---|
480 | for(int k=0; k < insts.numInstances(); k++){ |
---|
481 | Instance ins = insts.instance(k); |
---|
482 | if(!ins.isMissing(j)){ |
---|
483 | if (Double.isNaN(m_MinArray[j])) { |
---|
484 | m_MinArray[j] = ins.value(j); |
---|
485 | m_MaxArray[j] = ins.value(j); |
---|
486 | } else { |
---|
487 | if (ins.value(j) < m_MinArray[j]) |
---|
488 | m_MinArray[j] = ins.value(j); |
---|
489 | else if (ins.value(j) > m_MaxArray[j]) |
---|
490 | m_MaxArray[j] = ins.value(j); |
---|
491 | } |
---|
492 | } |
---|
493 | } |
---|
494 | } |
---|
495 | } |
---|
496 | } |
---|
497 | |
---|
498 | /** |
---|
499 | * Scale the given exemplar so that the returned exemplar |
---|
500 | * has the value of 0 to 1 for each dimension |
---|
501 | * |
---|
502 | * @param before the given exemplar |
---|
503 | * @return the resultant exemplar after scaling |
---|
504 | * @throws Exception if given exampler cannot be scaled properly |
---|
505 | */ |
---|
506 | private Instance scale(Instance before) throws Exception{ |
---|
507 | |
---|
508 | Instances afterInsts = before.relationalValue(1).stringFreeStructure(); |
---|
509 | Instance after = new DenseInstance(before.numAttributes()); |
---|
510 | after.setDataset(m_Attributes); |
---|
511 | |
---|
512 | for(int i=0; i < before.relationalValue(1).numInstances(); i++){ |
---|
513 | Instance datum = before.relationalValue(1).instance(i); |
---|
514 | Instance inst = (Instance)datum.copy(); |
---|
515 | |
---|
516 | for(int j=0; j < m_Dimension; j++){ |
---|
517 | if(before.relationalValue(1).attribute(j).isNumeric()) |
---|
518 | inst.setValue(j, (datum.value(j) - m_MinArray[j])/(m_MaxArray[j] - m_MinArray[j])); |
---|
519 | } |
---|
520 | afterInsts.add(inst); |
---|
521 | } |
---|
522 | |
---|
523 | int attValue = after.attribute(1).addRelation(afterInsts); |
---|
524 | after.setValue(0, before.value( 0)); |
---|
525 | after.setValue(1, attValue); |
---|
526 | after.setValue(2, before.value( 2)); |
---|
527 | |
---|
528 | return after; |
---|
529 | } |
---|
530 | |
---|
531 | /** |
---|
532 | * Use gradient descent to distort the MU parameter for |
---|
533 | * the exemplar. The exemplar can be in the specified row in the |
---|
534 | * given matrix, which has numExemplar rows and numDimension columns; |
---|
535 | * or not in the matrix. |
---|
536 | * |
---|
537 | * @param row the given row index |
---|
538 | * @param mean |
---|
539 | */ |
---|
540 | public void findWeights(int row, double[][] mean){ |
---|
541 | |
---|
542 | double[] neww = new double[m_Dimension]; |
---|
543 | double[] oldw = new double[m_Dimension]; |
---|
544 | System.arraycopy(m_Change[row], 0, neww, 0, m_Dimension); |
---|
545 | //for(int z=0; z<m_Dimension; z++) |
---|
546 | //System.out.println("mu("+row+"): "+origin[z]+" | "+newmu[z]); |
---|
547 | double newresult = target(neww, mean, row, m_Class); |
---|
548 | double result = Double.POSITIVE_INFINITY; |
---|
549 | double rate= 0.05; |
---|
550 | if(m_Rate != -1) |
---|
551 | rate = m_Rate; |
---|
552 | //System.out.println("???Start searching ..."); |
---|
553 | search: |
---|
554 | while(Utils.gr((result-newresult), m_STOP)){ // Full step |
---|
555 | oldw = neww; |
---|
556 | neww= new double[m_Dimension]; |
---|
557 | |
---|
558 | double[] delta = delta(oldw, mean, row, m_Class); |
---|
559 | |
---|
560 | for(int i=0; i < m_Dimension; i++) |
---|
561 | if(Utils.gr(m_Variance[row][i], 0.0)) |
---|
562 | neww[i] = oldw[i] + rate * delta[i]; |
---|
563 | |
---|
564 | result = newresult; |
---|
565 | newresult = target(neww, mean, row, m_Class); |
---|
566 | |
---|
567 | //System.out.println("???old: "+result+"|new: "+newresult); |
---|
568 | while(Utils.gr(newresult, result)){ // Search back |
---|
569 | //System.out.println("search back"); |
---|
570 | if(m_Rate == -1){ |
---|
571 | rate *= m_Decay; // Decay |
---|
572 | for(int i=0; i < m_Dimension; i++) |
---|
573 | if(Utils.gr(m_Variance[row][i], 0.0)) |
---|
574 | neww[i] = oldw[i] + rate * delta[i]; |
---|
575 | newresult = target(neww, mean, row, m_Class); |
---|
576 | } |
---|
577 | else{ |
---|
578 | for(int i=0; i < m_Dimension; i++) |
---|
579 | neww[i] = oldw[i]; |
---|
580 | break search; |
---|
581 | } |
---|
582 | } |
---|
583 | } |
---|
584 | //System.out.println("???Stop"); |
---|
585 | m_Change[row] = neww; |
---|
586 | } |
---|
587 | |
---|
588 | /** |
---|
589 | * Delta of x in one step of gradient descent: |
---|
590 | * delta(Wij) = 1/2 * sum[k=1..N, k!=i](sqrt(P)*(Yi-Yk)/D - 1) * (MUij - |
---|
591 | * MUkj)^2 where D = sqrt(sum[j=1..P]Kkj(MUij - MUkj)^2) |
---|
592 | * N is number of exemplars and P is number of dimensions |
---|
593 | * |
---|
594 | * @param x the weights of the exemplar in question |
---|
595 | * @param rowpos row index of x in X |
---|
596 | * @param Y the observed class label |
---|
597 | * @return the delta for all dimensions |
---|
598 | */ |
---|
599 | private double[] delta(double[] x, double[][] X, int rowpos, double[] Y){ |
---|
600 | double y = Y[rowpos]; |
---|
601 | |
---|
602 | double[] delta=new double[m_Dimension]; |
---|
603 | for(int h=0; h < m_Dimension; h++) |
---|
604 | delta[h] = 0.0; |
---|
605 | |
---|
606 | for(int i=0; i < X.length; i++){ |
---|
607 | if((i != rowpos) && (X[i] != null)){ |
---|
608 | double var = (y==Y[i]) ? 0.0 : Math.sqrt((double)m_Dimension - 1); |
---|
609 | double distance=0; |
---|
610 | for(int j=0; j < m_Dimension; j++) |
---|
611 | if(Utils.gr(m_Variance[rowpos][j], 0.0)) |
---|
612 | distance += x[j]*(X[rowpos][j]-X[i][j]) * (X[rowpos][j]-X[i][j]); |
---|
613 | distance = Math.sqrt(distance); |
---|
614 | if(distance != 0) |
---|
615 | for(int k=0; k < m_Dimension; k++) |
---|
616 | if(m_Variance[rowpos][k] > 0.0) |
---|
617 | delta[k] += (var/distance - 1.0) * 0.5 * |
---|
618 | (X[rowpos][k]-X[i][k]) * |
---|
619 | (X[rowpos][k]-X[i][k]); |
---|
620 | } |
---|
621 | } |
---|
622 | //System.out.println("???delta: "+delta); |
---|
623 | return delta; |
---|
624 | } |
---|
625 | |
---|
626 | /** |
---|
627 | * Compute the target function to minimize in gradient descent |
---|
628 | * The formula is:<br/> |
---|
629 | * 1/2*sum[i=1..p](f(X, Xi)-var(Y, Yi))^2 <p/> |
---|
630 | * where p is the number of exemplars and Y is the class label. |
---|
631 | * In the case of X=MU, f() is the Euclidean distance between two |
---|
632 | * exemplars together with the related weights and var() is |
---|
633 | * sqrt(numDimension)*(Y-Yi) where Y-Yi is either 0 (when Y==Yi) |
---|
634 | * or 1 (Y!=Yi) |
---|
635 | * |
---|
636 | * @param x the weights of the exemplar in question |
---|
637 | * @param rowpos row index of x in X |
---|
638 | * @param Y the observed class label |
---|
639 | * @return the result of the target function |
---|
640 | */ |
---|
641 | public double target(double[] x, double[][] X, int rowpos, double[] Y){ |
---|
642 | double y = Y[rowpos], result=0; |
---|
643 | |
---|
644 | for(int i=0; i < X.length; i++){ |
---|
645 | if((i != rowpos) && (X[i] != null)){ |
---|
646 | double var = (y==Y[i]) ? 0.0 : Math.sqrt((double)m_Dimension - 1); |
---|
647 | double f=0; |
---|
648 | for(int j=0; j < m_Dimension; j++) |
---|
649 | if(Utils.gr(m_Variance[rowpos][j], 0.0)){ |
---|
650 | f += x[j]*(X[rowpos][j]-X[i][j]) * (X[rowpos][j]-X[i][j]); |
---|
651 | //System.out.println("i:"+i+" j: "+j+" row: "+rowpos); |
---|
652 | } |
---|
653 | f = Math.sqrt(f); |
---|
654 | //System.out.println("???distance between "+rowpos+" and "+i+": "+f+"|y:"+y+" vs "+Y[i]); |
---|
655 | if(Double.isInfinite(f)) |
---|
656 | System.exit(1); |
---|
657 | result += 0.5 * (f - var) * (f - var); |
---|
658 | } |
---|
659 | } |
---|
660 | //System.out.println("???target: "+result); |
---|
661 | return result; |
---|
662 | } |
---|
663 | |
---|
664 | /** |
---|
665 | * Use Kullback Leibler distance to find the nearest neighbours of |
---|
666 | * the given exemplar. |
---|
667 | * It also uses K-Nearest Neighbour algorithm to classify the |
---|
668 | * test exemplar |
---|
669 | * |
---|
670 | * @param ex the given test exemplar |
---|
671 | * @return the classification |
---|
672 | * @throws Exception if the exemplar could not be classified |
---|
673 | * successfully |
---|
674 | */ |
---|
675 | public double classifyInstance(Instance ex)throws Exception{ |
---|
676 | |
---|
677 | ex = scale(ex); |
---|
678 | |
---|
679 | double[] var = new double [m_Dimension]; |
---|
680 | for (int i=0; i<m_Dimension; i++) |
---|
681 | var[i]= ex.relationalValue(1).variance(i); |
---|
682 | |
---|
683 | // The Kullback distance to all exemplars |
---|
684 | double[] kullback = new double[m_Class.length]; |
---|
685 | |
---|
686 | // The first K nearest neighbours' predictions */ |
---|
687 | double[] predict = new double[m_NumClasses]; |
---|
688 | for(int h=0; h < predict.length; h++) |
---|
689 | predict[h] = 0; |
---|
690 | ex = cleanse(ex); |
---|
691 | |
---|
692 | if(ex.relationalValue(1).numInstances() == 0){ |
---|
693 | if (getDebug()) |
---|
694 | System.out.println("???Whole exemplar falls into ambiguous area!"); |
---|
695 | return 1.0; // Bias towards positive class |
---|
696 | } |
---|
697 | |
---|
698 | double[] mean = new double[m_Dimension]; |
---|
699 | for (int i=0; i<m_Dimension; i++) |
---|
700 | mean [i]=ex.relationalValue(1).meanOrMode(i); |
---|
701 | |
---|
702 | // Avoid zero sigma |
---|
703 | for(int h=0; h < var.length; h++){ |
---|
704 | if(Utils.eq(var[h],0.0)) |
---|
705 | var[h] = m_ZERO; |
---|
706 | } |
---|
707 | |
---|
708 | for(int i=0; i < m_Class.length; i++){ |
---|
709 | if(m_ValidM[i] != null) |
---|
710 | kullback[i] = kullback(mean, m_ValidM[i], var, m_Variance[i], i); |
---|
711 | else |
---|
712 | kullback[i] = Double.POSITIVE_INFINITY; |
---|
713 | } |
---|
714 | |
---|
715 | for(int j=0; j < m_Neighbour; j++){ |
---|
716 | int pos = Utils.minIndex(kullback); |
---|
717 | predict[(int)m_Class[pos]] += m_Weights[pos]; |
---|
718 | kullback[pos] = Double.POSITIVE_INFINITY; |
---|
719 | } |
---|
720 | |
---|
721 | if (getDebug()) |
---|
722 | System.out.println("???There are still some unambiguous instances in this exemplar! Predicted as: "+Utils.maxIndex(predict)); |
---|
723 | return (double)Utils.maxIndex(predict); |
---|
724 | } |
---|
725 | |
---|
726 | /** |
---|
727 | * Cleanse the given exemplar according to the valid and noise data |
---|
728 | * statistics |
---|
729 | * |
---|
730 | * @param before the given exemplar |
---|
731 | * @return the processed exemplar |
---|
732 | * @throws Exception if the returned exemplar is wrong |
---|
733 | */ |
---|
734 | public Instance cleanse(Instance before) throws Exception{ |
---|
735 | |
---|
736 | Instances insts = before.relationalValue(1).stringFreeStructure(); |
---|
737 | Instance after = new DenseInstance(before.numAttributes()); |
---|
738 | after.setDataset(m_Attributes); |
---|
739 | |
---|
740 | for(int g=0; g < before.relationalValue(1).numInstances(); g++){ |
---|
741 | Instance datum = before.relationalValue(1).instance(g); |
---|
742 | double[] minNoiDists = new double[m_Choose]; |
---|
743 | double[] minValDists = new double[m_Choose]; |
---|
744 | int noiseCount = 0, validCount = 0; |
---|
745 | double[] nDist = new double[m_Mean.length]; |
---|
746 | double[] vDist = new double[m_Mean.length]; |
---|
747 | |
---|
748 | for(int h=0; h < m_Mean.length; h++){ |
---|
749 | if(m_ValidM[h] == null) |
---|
750 | vDist[h] = Double.POSITIVE_INFINITY; |
---|
751 | else |
---|
752 | vDist[h] = distance(datum, m_ValidM[h], m_ValidV[h], h); |
---|
753 | |
---|
754 | if(m_NoiseM[h] == null) |
---|
755 | nDist[h] = Double.POSITIVE_INFINITY; |
---|
756 | else |
---|
757 | nDist[h] = distance(datum, m_NoiseM[h], m_NoiseV[h], h); |
---|
758 | } |
---|
759 | |
---|
760 | for(int k=0; k < m_Choose; k++){ |
---|
761 | int pos = Utils.minIndex(vDist); |
---|
762 | minValDists[k] = vDist[pos]; |
---|
763 | vDist[pos] = Double.POSITIVE_INFINITY; |
---|
764 | pos = Utils.minIndex(nDist); |
---|
765 | minNoiDists[k] = nDist[pos]; |
---|
766 | nDist[pos] = Double.POSITIVE_INFINITY; |
---|
767 | } |
---|
768 | |
---|
769 | int x = 0,y = 0; |
---|
770 | while((x+y) < m_Choose){ |
---|
771 | if(minValDists[x] <= minNoiDists[y]){ |
---|
772 | validCount++; |
---|
773 | x++; |
---|
774 | } |
---|
775 | else{ |
---|
776 | noiseCount++; |
---|
777 | y++; |
---|
778 | } |
---|
779 | } |
---|
780 | if(x >= y) |
---|
781 | insts.add (datum); |
---|
782 | |
---|
783 | } |
---|
784 | |
---|
785 | after.setValue(0, before.value( 0)); |
---|
786 | after.setValue(1, after.attribute(1).addRelation(insts)); |
---|
787 | after.setValue(2, before.value( 2)); |
---|
788 | |
---|
789 | return after; |
---|
790 | } |
---|
791 | |
---|
792 | /** |
---|
793 | * This function calculates the Kullback Leibler distance between |
---|
794 | * two normal distributions. This distance is always positive. |
---|
795 | * Kullback Leibler distance = integral{f(X)ln(f(X)/g(X))} |
---|
796 | * Note that X is a vector. Since we assume dimensions are independent |
---|
797 | * f(X)(g(X) the same) is actually the product of normal density |
---|
798 | * functions of each dimensions. Also note that it should be log2 |
---|
799 | * instead of (ln) in the formula, but we use (ln) simply for computational |
---|
800 | * convenience. |
---|
801 | * |
---|
802 | * The result is as follows, suppose there are P dimensions, and f(X) |
---|
803 | * is the first distribution and g(X) is the second: |
---|
804 | * Kullback = sum[1..P](ln(SIGMA2/SIGMA1)) + |
---|
805 | * sum[1..P](SIGMA1^2 / (2*(SIGMA2^2))) + |
---|
806 | * sum[1..P]((MU1-MU2)^2 / (2*(SIGMA2^2))) - |
---|
807 | * P/2 |
---|
808 | * |
---|
809 | * @param mu1 mu of the first normal distribution |
---|
810 | * @param mu2 mu of the second normal distribution |
---|
811 | * @param var1 variance(SIGMA^2) of the first normal distribution |
---|
812 | * @param var2 variance(SIGMA^2) of the second normal distribution |
---|
813 | * @return the Kullback distance of two distributions |
---|
814 | */ |
---|
815 | public double kullback(double[] mu1, double[] mu2, |
---|
816 | double[] var1, double[] var2, int pos){ |
---|
817 | int p = mu1.length; |
---|
818 | double result = 0; |
---|
819 | |
---|
820 | for(int y=0; y < p; y++){ |
---|
821 | if((Utils.gr(var1[y], 0)) && (Utils.gr(var2[y], 0))){ |
---|
822 | result += |
---|
823 | ((Math.log(Math.sqrt(var2[y]/var1[y]))) + |
---|
824 | (var1[y] / (2.0*var2[y])) + |
---|
825 | (m_Change[pos][y] * (mu1[y]-mu2[y])*(mu1[y]-mu2[y]) / (2.0*var2[y])) - |
---|
826 | 0.5); |
---|
827 | } |
---|
828 | } |
---|
829 | |
---|
830 | return result; |
---|
831 | } |
---|
832 | |
---|
833 | /** |
---|
834 | * Returns an enumeration describing the available options |
---|
835 | * |
---|
836 | * @return an enumeration of all the available options |
---|
837 | */ |
---|
838 | public Enumeration listOptions() { |
---|
839 | Vector result = new Vector(); |
---|
840 | |
---|
841 | result.addElement(new Option( |
---|
842 | "\tSet number of nearest neighbour for prediction\n" |
---|
843 | + "\t(default 1)", |
---|
844 | "K", 1, "-K <number of neighbours>")); |
---|
845 | |
---|
846 | result.addElement(new Option( |
---|
847 | "\tSet number of nearest neighbour for cleansing the training data\n" |
---|
848 | + "\t(default 1)", |
---|
849 | "S", 1, "-S <number of neighbours>")); |
---|
850 | |
---|
851 | result.addElement(new Option( |
---|
852 | "\tSet number of nearest neighbour for cleansing the testing data\n" |
---|
853 | + "\t(default 1)", |
---|
854 | "E", 1, "-E <number of neighbours>")); |
---|
855 | |
---|
856 | return result.elements(); |
---|
857 | } |
---|
858 | |
---|
859 | /** |
---|
860 | * Parses a given list of options. <p/> |
---|
861 | * |
---|
862 | <!-- options-start --> |
---|
863 | * Valid options are: <p/> |
---|
864 | * |
---|
865 | * <pre> -K <number of neighbours> |
---|
866 | * Set number of nearest neighbour for prediction |
---|
867 | * (default 1)</pre> |
---|
868 | * |
---|
869 | * <pre> -S <number of neighbours> |
---|
870 | * Set number of nearest neighbour for cleansing the training data |
---|
871 | * (default 1)</pre> |
---|
872 | * |
---|
873 | * <pre> -E <number of neighbours> |
---|
874 | * Set number of nearest neighbour for cleansing the testing data |
---|
875 | * (default 1)</pre> |
---|
876 | * |
---|
877 | <!-- options-end --> |
---|
878 | * |
---|
879 | * @param options the list of options as an array of strings |
---|
880 | * @throws Exception if an option is not supported |
---|
881 | */ |
---|
882 | public void setOptions(String[] options) throws Exception{ |
---|
883 | |
---|
884 | setDebug(Utils.getFlag('D', options)); |
---|
885 | |
---|
886 | String numNeighbourString = Utils.getOption('K', options); |
---|
887 | if (numNeighbourString.length() != 0) |
---|
888 | setNumNeighbours(Integer.parseInt(numNeighbourString)); |
---|
889 | else |
---|
890 | setNumNeighbours(1); |
---|
891 | |
---|
892 | numNeighbourString = Utils.getOption('S', options); |
---|
893 | if (numNeighbourString.length() != 0) |
---|
894 | setNumTrainingNoises(Integer.parseInt(numNeighbourString)); |
---|
895 | else |
---|
896 | setNumTrainingNoises(1); |
---|
897 | |
---|
898 | numNeighbourString = Utils.getOption('E', options); |
---|
899 | if (numNeighbourString.length() != 0) |
---|
900 | setNumTestingNoises(Integer.parseInt(numNeighbourString)); |
---|
901 | else |
---|
902 | setNumTestingNoises(1); |
---|
903 | } |
---|
904 | |
---|
905 | /** |
---|
906 | * Gets the current settings of the Classifier. |
---|
907 | * |
---|
908 | * @return an array of strings suitable for passing to setOptions |
---|
909 | */ |
---|
910 | public String[] getOptions() { |
---|
911 | Vector result; |
---|
912 | |
---|
913 | result = new Vector(); |
---|
914 | |
---|
915 | if (getDebug()) |
---|
916 | result.add("-D"); |
---|
917 | |
---|
918 | result.add("-K"); |
---|
919 | result.add("" + getNumNeighbours()); |
---|
920 | |
---|
921 | result.add("-S"); |
---|
922 | result.add("" + getNumTrainingNoises()); |
---|
923 | |
---|
924 | result.add("-E"); |
---|
925 | result.add("" + getNumTestingNoises()); |
---|
926 | |
---|
927 | return (String[]) result.toArray(new String[result.size()]); |
---|
928 | } |
---|
929 | |
---|
930 | /** |
---|
931 | * Returns the tip text for this property |
---|
932 | * |
---|
933 | * @return tip text for this property suitable for |
---|
934 | * displaying in the explorer/experimenter gui |
---|
935 | */ |
---|
936 | public String numNeighboursTipText() { |
---|
937 | return "The number of nearest neighbours to the estimate the class prediction of test bags."; |
---|
938 | } |
---|
939 | |
---|
940 | /** |
---|
941 | * Sets the number of nearest neighbours to estimate |
---|
942 | * the class prediction of tests bags |
---|
943 | * @param numNeighbour the number of citers |
---|
944 | */ |
---|
945 | public void setNumNeighbours(int numNeighbour){ |
---|
946 | m_Neighbour = numNeighbour; |
---|
947 | } |
---|
948 | |
---|
949 | /** |
---|
950 | * Returns the number of nearest neighbours to estimate |
---|
951 | * the class prediction of tests bags |
---|
952 | * @return the number of neighbours |
---|
953 | */ |
---|
954 | public int getNumNeighbours(){ |
---|
955 | return m_Neighbour; |
---|
956 | } |
---|
957 | |
---|
958 | /** |
---|
959 | * Returns the tip text for this property |
---|
960 | * |
---|
961 | * @return tip text for this property suitable for |
---|
962 | * displaying in the explorer/experimenter gui |
---|
963 | */ |
---|
964 | public String numTrainingNoisesTipText() { |
---|
965 | return "The number of nearest neighbour instances in the selection of noises in the training data."; |
---|
966 | } |
---|
967 | |
---|
968 | /** |
---|
969 | * Sets the number of nearest neighbour instances in the |
---|
970 | * selection of noises in the training data |
---|
971 | * |
---|
972 | * @param numTraining the number of noises in training data |
---|
973 | */ |
---|
974 | public void setNumTrainingNoises (int numTraining){ |
---|
975 | m_Select = numTraining; |
---|
976 | } |
---|
977 | |
---|
978 | /** |
---|
979 | * Returns the number of nearest neighbour instances in the |
---|
980 | * selection of noises in the training data |
---|
981 | * |
---|
982 | * @return the number of noises in training data |
---|
983 | */ |
---|
984 | public int getNumTrainingNoises(){ |
---|
985 | return m_Select; |
---|
986 | } |
---|
987 | |
---|
988 | /** |
---|
989 | * Returns the tip text for this property |
---|
990 | * |
---|
991 | * @return tip text for this property suitable for |
---|
992 | * displaying in the explorer/experimenter gui |
---|
993 | */ |
---|
994 | public String numTestingNoisesTipText() { |
---|
995 | return "The number of nearest neighbour instances in the selection of noises in the test data."; |
---|
996 | } |
---|
997 | |
---|
998 | /** |
---|
999 | * Returns The number of nearest neighbour instances in the |
---|
1000 | * selection of noises in the test data |
---|
1001 | * @return the number of noises in test data |
---|
1002 | */ |
---|
1003 | public int getNumTestingNoises(){ |
---|
1004 | return m_Choose; |
---|
1005 | } |
---|
1006 | |
---|
1007 | /** |
---|
1008 | * Sets The number of nearest neighbour exemplars in the |
---|
1009 | * selection of noises in the test data |
---|
1010 | * @param numTesting the number of noises in test data |
---|
1011 | */ |
---|
1012 | public void setNumTestingNoises (int numTesting){ |
---|
1013 | m_Choose = numTesting; |
---|
1014 | } |
---|
1015 | |
---|
1016 | /** |
---|
1017 | * Returns the revision string. |
---|
1018 | * |
---|
1019 | * @return the revision |
---|
1020 | */ |
---|
1021 | public String getRevision() { |
---|
1022 | return RevisionUtils.extract("$Revision: 5987 $"); |
---|
1023 | } |
---|
1024 | |
---|
1025 | /** |
---|
1026 | * Main method for testing. |
---|
1027 | * |
---|
1028 | * @param args the options for the classifier |
---|
1029 | */ |
---|
1030 | public static void main(String[] args) { |
---|
1031 | runClassifier(new MINND(), args); |
---|
1032 | } |
---|
1033 | } |
---|