C++ Col::rows方法代码示例

inline void
CoverTree<MetricType, StatisticType, MatType, RootPointPolicy>::CreateChildren(
    arma::Col<size_t>& indices,
    arma::vec& distances,
    size_t nearSetSize,
    size_t& farSetSize,
    size_t& usedSetSize)
{
  // Determine the next scale level.  This should be the first level where there
  // are any points in the far set.  So, if we know the maximum distance in the
  // distances array, this will be the largest i such that
  //   maxDistance > pow(base, i)
  // and using this for the scale factor should guarantee we are not creating an
  // implicit node.  If the maximum distance is 0, every point in the near set
  // will be created as a leaf, and a child to this node.  We also do not need
  // to change the furthestChildDistance or furthestDescendantDistance.
  const ElemType maxDistance = max(distances.rows(0,
      nearSetSize + farSetSize - 1));
  if (maxDistance == 0)
  {
    // Make the self child at the lowest possible level.
    // This should not modify farSetSize or usedSetSize.
    size_t tempSize = 0;
    children.push_back(new CoverTree(*dataset, base, point, INT_MIN, this, 0,
        indices, distances, 0, tempSize, usedSetSize, *metric));
    distanceComps += children.back()->DistanceComps();

    // Every point in the near set should be a leaf.
    for (size_t i = 0; i < nearSetSize; ++i)
    {
      // farSetSize and usedSetSize will not be modified.
      children.push_back(new CoverTree(*dataset, base, indices[i],
          INT_MIN, this, distances[i], indices, distances, 0, tempSize,
          usedSetSize, *metric));
      distanceComps += children.back()->DistanceComps();
      usedSetSize++;
    }

    // The number of descendants is just the number of children, because each of
    // them are leaves and contain one point.
    numDescendants = children.size();

    // Re-sort the dataset.  We have
    // [ used | far | other used ]
    // and we want
    // [ far | all used ].
    SortPointSet(indices, distances, 0, usedSetSize, farSetSize);

    return;
  }

  const int nextScale = std::min(scale,
      (int) ceil(log(maxDistance) / log(base))) - 1;
  const ElemType bound = pow(base, nextScale);

  // First, make the self child.  We must split the given near set into the near
  // set and far set for the self child.
  size_t childNearSetSize =
      SplitNearFar(indices, distances, bound, nearSetSize);

  // Build the self child (recursively).
  size_t childFarSetSize = nearSetSize - childNearSetSize;
  size_t childUsedSetSize = 0;
  children.push_back(new CoverTree(*dataset, base, point, nextScale, this, 0,
      indices, distances, childNearSetSize, childFarSetSize, childUsedSetSize,
      *metric));
  // Don't double-count the self-child (so, subtract one).
  numDescendants += children[0]->NumDescendants();

  // The self-child can't modify the furthestChildDistance away from 0, but it
  // can modify the furthestDescendantDistance.
  furthestDescendantDistance = children[0]->FurthestDescendantDistance();

  // Remove any implicit nodes we may have created.
  RemoveNewImplicitNodes();

  distanceComps += children[0]->DistanceComps();

  // Now the arrays, in memory, look like this:
  // [ childFar | childUsed | far | used ]
  // but we need to move the used points past our far set:
  // [ childFar | far | childUsed + used ]
  // and keeping in mind that childFar = our near set,
  // [ near | far | childUsed + used ]
  // is what we are trying to make.
  SortPointSet(indices, distances, childFarSetSize, childUsedSetSize,
      farSetSize);

  // Update size of near set and used set.
  nearSetSize -= childUsedSetSize;
  usedSetSize += childUsedSetSize;

  // Now for each point in the near set, we need to make children.  To save
  // computation later, we'll create an array holding the points in the near
  // set, and then after each run we'll check which of those (if any) were used
  // and we will remove them.  ...if that's faster.  I think it is.
  while (nearSetSize > 0)
  {
    size_t newPointIndex = nearSetSize - 1;

//.........这里部分代码省略.........