vector< vector<u32> > kmeans(const vector< vector<f64> >& points, u32 k,
vector< vector<f64> >& centers)
{
if (points.size() == 0)
throw eInvalidArgument("There must be points to cluster!");
if (points[0].size() == 0)
throw eInvalidArgument("There must be at least one dimension!");
for (size_t i = 1; i < points.size(); i++)
if (points[i].size() != points[0].size())
throw eInvalidArgument("All data points must have the same dimensionality.");
if (k == 0)
throw eInvalidArgument("Clustering into zero clusters makes no sense.");
if ((u32)centers.size() != k)
throw eInvalidArgument("You must supply k initial centers to this version of kmeans().");
for (u32 i = 0; i < k; i++)
if (centers[i].size() != points[0].size())
throw eInvalidArgument("All initial centers must have the same dimensionality as the data points.");
vector< vector<u32> > clusters(k);
while (true)
{
vector< vector<u32> > newClusters(k);
for (size_t i = 0; i < points.size(); i++)
{
f64 dist = s_distanceSquared(points[i], centers[0]);
u32 closest = 0;
for (u32 c = 1; c < k; c++)
{
f64 distHere = s_distanceSquared(points[i], centers[c]);
if (distHere < dist)
{
closest = c;
dist = distHere;
}
}
newClusters[closest].push_back((u32)i);
}
for (u32 i = 0; i < k; i++)
if (newClusters[i].size() > 0)
centers[i] = s_calcCenter(points, newClusters[i]);
// Else, what should I do? Leave it? Randomize a new center?
if (clusters == newClusters)
break;
clusters = newClusters;
}
return clusters;
}
DEBaseAlgorithm::DEBaseAlgorithm(DEBaseParameters& para)
:_parameters(para),
_bestClusters(para.N,para.atomTypes,para.atomNumber)
{
int i = 0;
for(i = 0; i < para._numberOfClusters; i++)
{
DEBaseClusters newClusters(para.N,para.atomTypes,para.atomNumber);
_X.push_back(newClusters);
_V.push_back(newClusters);
_U.push_back(newClusters);
}
}
// Zhu et al. "A Rank-Order Distance based Clustering Algorithm for Face Tagging", CVPR 2011
br::Clusters br::ClusterGallery(const QStringList &simmats, float aggressiveness, const QString &csv)
{
qDebug("Clustering %d simmat(s)", simmats.size());
// Read in gallery parts, keeping top neighbors of each template
Neighborhood neighborhood = getNeighborhood(simmats);
const int cutoff = neighborhood.first().size();
const float threshold = 3*cutoff/4 * aggressiveness/5;
// Initialize clusters
Clusters clusters(neighborhood.size());
for (int i=0; i<neighborhood.size(); i++)
clusters[i].append(i);
bool done = false;
while (!done) {
// nextClusterIds[i] = j means that cluster i is set to merge into cluster j
QVector<int> nextClusterIDs(neighborhood.size());
for (int i=0; i<neighborhood.size(); i++) nextClusterIDs[i] = i;
// For each cluster
for (int clusterID=0; clusterID<neighborhood.size(); clusterID++) {
const Neighbors &neighbors = neighborhood[clusterID];
int nextClusterID = nextClusterIDs[clusterID];
// Check its neighbors
foreach (const Neighbor &neighbor, neighbors) {
int neighborID = neighbor.first;
int nextNeighborID = nextClusterIDs[neighborID];
// Don't bother if they have already merged
if (nextNeighborID == nextClusterID) continue;
// Flag for merge if similar enough
if (normalizedROD(neighborhood, clusterID, neighborID) < threshold) {
if (nextClusterID < nextNeighborID) nextClusterIDs[neighborID] = nextClusterID;
else nextClusterIDs[clusterID] = nextNeighborID;
}
}
}
// Transitive merge
for (int i=0; i<neighborhood.size(); i++) {
int nextClusterID = i;
while (nextClusterID != nextClusterIDs[nextClusterID]) {
assert(nextClusterIDs[nextClusterID] < nextClusterID);
nextClusterID = nextClusterIDs[nextClusterID];
}
nextClusterIDs[i] = nextClusterID;
}
// Construct new clusters
QHash<int, int> clusterIDLUT;
QList<int> allClusterIDs = QSet<int>::fromList(nextClusterIDs.toList()).values();
for (int i=0; i<neighborhood.size(); i++)
clusterIDLUT[i] = allClusterIDs.indexOf(nextClusterIDs[i]);
Clusters newClusters(allClusterIDs.size());
Neighborhood newNeighborhood(allClusterIDs.size());
for (int i=0; i<neighborhood.size(); i++) {
int newID = clusterIDLUT[i];
newClusters[newID].append(clusters[i]);
newNeighborhood[newID].append(neighborhood[i]);
}
// Update indices and trim
for (int i=0; i<newNeighborhood.size(); i++) {
Neighbors &neighbors = newNeighborhood[i];
int size = qMin(neighbors.size(),cutoff);
std::partial_sort(neighbors.begin(), neighbors.begin()+size, neighbors.end(), compareNeighbors);
for (int j=0; j<size; j++)
neighbors[j].first = clusterIDLUT[j];
neighbors = neighbors.mid(0, cutoff);
}
// Update results
done = true; //(newClusters.size() >= clusters.size());
clusters = newClusters;
neighborhood = newNeighborhood;
}
