bool VLFeat::Initialize_kmeans() {
bin_data dat(kmeansname);
kmeans = vl_kmeans_new(VL_TYPE_FLOAT, VlDistanceL2);
vl_kmeans_set_centers(kmeans, dat.vector_address(0),
dat.vdim(), dat.nobjects());
if (MethodVerbose())
cout << "VLFeat::Initialize_kmeans() read in kmeans <" << kmeansname
<< "> dim=" << descriptor_dim() << " nclust=" << nclusters()
<< endl;
if (kdtree_ntrees) {
kdtree = vl_kdforest_new(VL_TYPE_FLOAT, descriptor_dim(), kdtree_ntrees,
VlDistanceL2);
vl_kdforest_build(kdtree, nclusters(), means());
vl_kdforest_set_max_num_comparisons(kdtree, kdtree_maxcomps);
}
do_vlad = true;
return true;
}
void CodeBook::SetUp()
{
if (kmeans_model_ != NULL)
vl_kmeans_delete(kmeans_model_);
kmeans_model_ = vl_kmeans_new(VL_TYPE_FLOAT, dist_type_);
// use the ANN for fast computation
vl_kmeans_set_max_num_iterations(kmeans_model_, max_iter_);
vl_kmeans_set_algorithm(kmeans_model_, VlKMeansANN);
vl_kmeans_set_num_trees(kmeans_model_, num_kdtrees_);
vl_kmeans_set_max_num_comparisons(kmeans_model_, max_comp_);
has_setup_ = true;
}
void kmeans()
{
const vl_size data_dim = 2;
const vl_size num_data = 1000;
float data[data_dim * num_data] = { 0.0f, };
for (vl_size i = 0; i < data_dim * num_data; ++i)
data[i] = (float)std::rand() / RAND_MAX;
//
std::cout << "start processing ..." << std::endl;
const vl_size num_clusters = 3;
const VlVectorComparisonType distance = VlDistanceL2;
const vl_size num_max_iterations = 100;
VlKMeans *kmeans = vl_kmeans_new(VL_TYPE_DOUBLE, distance);
vl_kmeans_set_max_num_iterations(kmeans, num_max_iterations);
if (true)
{
// initialization.
#if 1
vl_kmeans_init_centers_with_rand_data(kmeans, (void const *)data, data_dim, num_data, num_clusters);
#elif 0
vl_kmeans_init_centers_plus_plus(kmeans, (void const *)data, data_dim, num_data, num_clusters);
#else
{
float init_centers[data_dim * num_clusters] = { 0.0f, };
for (vl_size i = 0; i < data_dim * num_clusters; ++i)
init_centers[i] = (float)std::rand() / RAND_MAX;
vl_kmeans_set_centers(kmeans, (void const *)init_centers, data_dim, num_clusters);
}
#endif
// clustering.
const double energy = vl_kmeans_refine_centers(kmeans, data, num_data);
}
else
{
const VlKMeansAlgorithm algorithm = VlKMeansLloyd; // VlKMeansLloyd, VlKMeansElkan, VlKMeansANN.
const VlKMeansInitialization initialization = VlKMeansRandomSelection; // VlKMeansRandomSelection, VlKMeansPlusPlus.
const vl_size num_repetitions = 100;
vl_kmeans_set_algorithm(kmeans, VlKMeansLloyd);
vl_kmeans_set_initialization(kmeans, VlKMeansRandomSelection);
vl_kmeans_set_num_repetitions(kmeans, num_repetitions);
// clustering.
const double energy = vl_kmeans_cluster(kmeans, (void const *)data, data_dim, num_data, num_clusters);
}
//
const vl_size num_iterations = vl_kmeans_get_num_repetitions(kmeans);
//const vl_type data_type = vl_kmeans_get_data_type(kmeans);
//
{
const float *centers = (float *)vl_kmeans_get_centers(kmeans);
for (int i = 0; i < num_clusters; ++i)
{
std::cout << '(';
for (int j = 0; j < data_dim; ++j)
{
if (j) std::cout << ',';
std::cout << centers[i * data_dim + j];
}
std::cout << ')' << std::endl;
}
}
//
{
vl_uint32 assignments[num_data] = { 0, };
double distances[num_data] = { 0, };
vl_kmeans_quantize(kmeans, assignments, (void *)distances, (void const *)data, num_data);
for (int i = 0; i < num_data; ++i)
{
std::cout << '(';
for (int j = 0; j < data_dim; ++j)
{
if (j) std::cout << ',';
std::cout << data[i * data_dim + j]; // TODO [check] >> is it correct?
}
std::cout << ") => " << assignments[i] << std::endl;
}
}
std::cout << "end processing ..." << std::endl;
//
if (kmeans)
{
vl_kmeans_delete(kmeans);
kmeans = NULL;
}
}
