gnu.trove.iterator.TIntIntIterator Java Examples

public  void duplicate(ArrayList<String> docs) throws Exception {
	this.docs = docs;
	dsMap = new TreeSet<DocSim>();
	feature();

	similarity();
	System.out.println("去重复");

	boolean[] dup = new boolean[docs.size()];
	for(int id1=0;id1<docs.size();id1++){
		if(dup[id1]||similarityMap[id1]==null)
			continue;


		ArrayList<Integer> ids = new ArrayList<Integer>();
		ids.add(id1);
		TIntIntIterator it = similarityMap[id1].iterator();
		for ( int i = similarityMap[id1].size(); i-- > 0; ) {
			it.advance();
			int id2 = it.key();				
			double sim = ((double)(it.value()* 2)) / (featureLen[id1] + featureLen[id2]);
			if(sim > thres ){
				dup[id2]= true;
				ids.add(id2);
			}
		}
		DocSim docSim = new DocSim(ids);
		dsMap.add(docSim);
	}
}
 

/**
 * 
 * @param freqmap
 * @return
 */
public static TreeMap<Integer, Integer> countFrequency(TIntIntHashMap freqmap) {
	TreeMap<Integer, Integer> map = new TreeMap<Integer,Integer>();
	TIntIntIterator it = freqmap.iterator();
	while(it.hasNext()){
		it.advance();
		int freq = it.value();
		if(map.containsKey(freq)){
			map.put(freq, map.get(freq)+1);
		}else
			map.put(freq, 1);
	}
	return map;		
}
 

public void addAll(MaterialCounter other) {
  for (TIntIntIterator iter = other.counts.iterator(); iter.hasNext(); ) {
    iter.advance();
    counts.adjustOrPutValue(iter.key(), iter.value(), iter.value());
  }
}
 

public void addAll(MaterialCounter other) {
    for(TIntIntIterator iter = other.counts.iterator(); iter.hasNext();) {
        iter.advance();
        counts.adjustOrPutValue(iter.key(), iter.value(), iter.value());
    }
}
 

/** See {@link gnu.trove.map.hash.TIntIntHashMap#iterator()} */
public TIntIntIterator iterator() {
  return delegate.iterator();
}
 

@Override
public TIntIntIterator iterator() {     
    return intdata.iterator();
}
 

/**
 * data 每个元素为特征空间索引位置 1 ... T T列(模板个数) 1 N行(序列长度) 2 . data[r][t] N
 * 第t个模板作用在第r个位置 得到feature的起始位置
 * 
 * target[r],predict[r] label的编号
 * 
 * @param c
 * @param weights
 */
public float update(Instance inst, float[] weights, int k,
		float[] extraweight, Object predictLabel,
		Object goldenLabel, float c) {
	int[][] data = (int[][]) inst.getData();
	int[] target;
	if (goldenLabel == null)
		target = (int[]) inst.getTarget();
	else
		target = (int[]) goldenLabel;
	int[] predict = (int[]) predictLabel;
	// 当前clique中不同label的个数
	int ne = 0;
	/**
	 * 偏移索引
	 * 
	 */
	int tS = 0, pS = 0;

	float diffW = 0;

	int loss = 0;

	int L = data.length;
	// 稀疏矩阵表示(f(x,y)-f(x,\bar{y}))
	TIntIntHashMap diffF = new TIntIntHashMap(); // 最多有2*L*numTemplets个不同

	for (int o = -templets.maxOrder - 1, l = 0; l < L; o++, l++) {
		tS = tS * numLabels % templets.numStates + target[l]; // 目标值：计算当前状态组合的y空间偏移
		pS = pS * numLabels % templets.numStates + predict[l];// 预测值：计算当前状态组合的y空间偏移
		if (predict[l] != target[l])
			ne++;
		if (o >= 0 && (predict[o] != target[o]))
			ne--; // 减去移出clique的节点的label差异

		if (ne > 0) { // 当前clique有不相同label
			loss++; // L(y,ybar)
			for (int t = 0; t < numTemplets; t++) {
				if (data[l][t] == -1)
					continue;
				int tI = data[l][t] + templets.offset[t][tS]; // 特征索引：找到对应weights的维数
				int pI = data[l][t] + templets.offset[t][pS]; // 特征索引：找到对应weights的维数
				if (tI != pI) {
					diffF.adjustOrPutValue(tI, 1, 1);
					diffF.adjustOrPutValue(pI, -1, -1);
					diffW += weights[tI] - weights[pI]; // w^T(f(x,y)-f(x,ybar))
				}
			}
		}
	}

	float diff = 0;
	TIntIntIterator it = diffF.iterator();
	for (int i = diffF.size(); i-- > 0;) {
		it.advance();
		diff += it.value() * it.value();
	}
	it = null;
	float alpha;
	float delta;
	if (useLoss) {
		delta = loss;
	} else
		delta = 1;
	if (diffW < delta) {

		tS = 0;
		pS = 0;
		ne = 0;
		alpha = (delta - diffW) / diff;
		// System.out.println(alpha);
		alpha = Math.min(c, alpha);
		it = diffF.iterator();
		for (int i = diffF.size(); i-- > 0;) {
			it.advance();
			weights[it.key()] += it.value() * alpha;
		}

		return loss;
	} else {
		return 0;
	}
}