Java Code Examples for gnu.trove.map.hash.TIntIntHashMap#adjustOrPutValue()

@Override public TIntObjectHashMap<TIntIntHashMap> getEntityKeywordIntersectionCount(Entities entities) throws EntityLinkingDataAccessException {
  TIntObjectHashMap<TIntIntHashMap> isec = new TIntObjectHashMap<TIntIntHashMap>();

  for (String[] eKps : allEntityKeyphrases) {
    int entity = DataAccess.getInternalIdForKBEntity(getTestKBEntity(eKps[0]));
    TIntIntHashMap counts = new TIntIntHashMap();
    isec.put(entity, counts);

    if (eKps.length > 1) {
      String[] tokens = null;
      for (int i = 1; i < eKps.length; ++i) {
        if (i % 2 == 1) {
          tokens = eKps[i].split(" ");
        } else {
          int count = Integer.parseInt(eKps[i]);
          for (String token : tokens) {
            counts.adjustOrPutValue(DataAccess.getIdForWord(token), count, count);
          }
        }
      }
    }
  }
  return isec;
}
 

public void startCollectingStatsInMemory(UnitType unitType) throws SQLException, EntityLinkingDataAccessException {
  TIntObjectHashMap<TIntHashSet> entityUnits = UnitUtil.loadEntityUnits(unitType.getUnitSize());

  // Count the occurrences of units with respect to entities
  TIntIntHashMap unitCounts = new TIntIntHashMap();
  for (int entity : entityUnits.keys()) {
    TIntHashSet units = entityUnits.get(entity);
    for (int unit : units.toArray()) {
      unitCounts.adjustOrPutValue(unit, 1, 1);
    }
  }

  logger.info("Storing data ...");
  storeUnitsIntoDB(unitCounts, unitType);
}
 

private TIntObjectHashMap<TIntIntHashMap> getEntityUnitIntersectionCountInternal(Entities entities, UnitType unitType)
    throws EntityLinkingDataAccessException {
  TIntObjectHashMap<TIntIntHashMap> isec = new TIntObjectHashMap<>();

  for (String[] eKps : allEntityKeyphrases) {
    int entity = DataAccess.getInternalIdForKBEntity(getTestKBEntity(eKps[0]));
    TIntIntHashMap counts = new TIntIntHashMap();
    isec.put(entity, counts);

    if (eKps.length > 1) {
      String[] keyphrase = null;

      for (int i = 1; i < eKps.length; ++i) {
        if (i % 2 == 1) {
          keyphrase = eKps[i].split(" ");
        } else {
          int count = Integer.parseInt(eKps[i]);

          Set<String[]> ngrams = StringUtils.getNgrams(keyphrase, unitType.getUnitSize());
          for (String[] ngram : ngrams) {
            String ngramString = String.join(" ", ngram);
            counts.adjustOrPutValue(DataAccess.getIdForWord(ngramString), count, count);
          }
        }
      }
    }
  }
  return isec;
}
 

public InputTextWrapper(Context context, UnitType unitType, boolean removeStopwords) throws EntityLinkingDataAccessException {
  logger.debug("Wrapping input text.");
  mentionToIgnore = null;
  this.unitType = unitType;
  int unitLength = unitType.getUnitSize();
  if (context.getTokenCount() < unitLength) return;
  List<String> unitStrings = new ArrayList<>(context.getTokenCount());
  Queue<String> curTokens = new ArrayDeque<>(unitLength);
  String[] curTokensArray = new String[unitLength];
  for (String token : context.getTokens()) {
    curTokens.add(token);
    if (curTokens.size() == unitLength || (!curTokens.isEmpty() && curTokens.size() - 1 == unitLength)) {
      unitStrings.add(UnitBuilder.buildUnit(curTokens.toArray(curTokensArray)));
      curTokens.remove();
    }
  }

  logger.debug("Get ids for words.");
  TObjectIntHashMap<String> wordIds = DataAccess.getIdsForWords(unitStrings);
  units = new int[unitStrings.size()];
  unitCounts = new TIntIntHashMap((int) (wordIds.size() / Constants.DEFAULT_LOAD_FACTOR), Constants.DEFAULT_LOAD_FACTOR);
  numOfUnits = 0;
  for (int i = 0; i < unitStrings.size(); i++) {
    int unitId = wordIds.get(unitStrings.get(i));
    if (unitId == 0) continue;

    logger.debug("Get contract term for unit id {}.", unitId);
    int contractedUnitId = DataAccess.contractTerm(unitId);
    if (contractedUnitId != 0) unitId = contractedUnitId;
    if (removeStopwords && StopWord.isStopwordOrSymbol(unitId, Language.getLanguageForString("en")))  continue;
    units[i] = unitId;
    unitCounts.adjustOrPutValue(unitId, 1, 1);
    numOfUnits++;
  }
}
 

/**
 * 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;
	}
}