unsigned kd_tree_ln_search (kd_tree_t *tree, const float *pt, unsigned *cnt)
{
unsigned k, c = 0;
const float *cm = tree->means;
float d, min = l2sqr(cm, pt, tree->dim);
cm += tree->dim;
for (k = 1; k < tree->K; k++, cm += tree->dim) {
d = l2sqr(cm, pt, tree->dim);
if (d < min) {
c = k;
min = d;
}
}
*cnt = tree->K;
return c;
}
static inline void kd_search_leaf_or_node (union kd_leaf_or_node lon,
struct kd_search_stat *stat,
float d2b) {
if (kd_is_leaf(lon, stat->K)) {
float l = l2sqr(stat->pt, stat->means + stat->dim * lon.leaf, stat->dim);
stat->cnt++;
if (l < stat->nn_dist) {
stat->nn = lon.leaf;
stat->nn_dist = l;
}
}
else {
kd_search_node(lon.node, stat, d2b);
}
}
// ${LSHKIT_HOME}/tools/fitdata.cpp
void fitdata_example()
{
const std::string data_file("./data/search_algorithm/lshkit/audio.data");
const unsigned N = 0; // number of points to use.
const unsigned P = 50000; // number of pairs to sample.
unsigned Q = 1000; // number of queries to sample.
unsigned K = 100; // search for K nearest neighbors.
const unsigned F = 10; // divide the sample to F folds.
// load matrix.
lshkit::Matrix<float> data(data_file);
std::vector<unsigned> idx(data.getSize());
for (unsigned i = 0; i < idx.size(); ++i) idx[i] = i;
random_shuffle(idx.begin(), idx.end());
if (N > 0 && N < data.getSize()) idx.resize(N);
lshkit::metric::l2sqr<float> l2sqr(data.getDim());
lshkit::DefaultRng rng;
boost::variate_generator<lshkit::DefaultRng &, lshkit::UniformUnsigned> gen(rng, lshkit::UniformUnsigned(0, idx.size()-1));
double gM = 0.0;
double gG = 0.0;
{
// sample P pairs of points
for (unsigned k = 0; k < P; ++k)
{
double dist, logdist;
for (;;)
{
unsigned i = gen();
unsigned j = gen();
if (i == j) continue;
dist = l2sqr(data[idx[i]], data[idx[j]]);
logdist = std::log(dist);
if (local::is_good_value(logdist)) break;
}
gM += dist;
gG += logdist;
}
gM /= P;
gG /= P;
gG = std::exp(gG);
}
if (Q > idx.size()) Q = idx.size();
if (K > idx.size() - Q) K = idx.size() - Q;
// sample query.
std::vector<unsigned> qry(Q);
lshkit::SampleQueries(&qry, idx.size(), rng);
// do the queries.
std::vector<lshkit::Topk<unsigned> > topks(Q);
for (unsigned i = 0; i < Q; ++i) topks[i].reset(K);
/* ... */
gsl_matrix *X = gsl_matrix_alloc(F * K, 3);
gsl_vector *yM = gsl_vector_alloc(F * K);
gsl_vector *yG = gsl_vector_alloc(F * K);
gsl_vector *pM = gsl_vector_alloc(3);
gsl_vector *pG = gsl_vector_alloc(3);
gsl_matrix *cov = gsl_matrix_alloc(3,3);
std::vector<double> M(K);
std::vector<double> G(K);
boost::progress_display progress(F, std::cerr);
unsigned m = 0;
for (unsigned l = 0; l < F; l++)
{
// Scan
for (unsigned i = l; i< idx.size(); i += F)
{
for (unsigned j = 0; j < Q; j++)
{
int id = qry[j];
if (i != id)
{
float d = l2sqr(data[idx[id]], data[idx[i]]);
if (local::is_good_value(std::log(double(d)))) topks[j] << lshkit::Topk<unsigned>::Element(i, d);
}
}
}
std::fill(M.begin(), M.end(), 0.0);
std::fill(G.begin(), G.end(), 0.0);
for (unsigned i = 0; i < Q; i++)
{
for (unsigned k = 0; k < K; k++)
{
M[k] += topks[i][k].dist;
G[k] += std::log(topks[i][k].dist);
}
}
for (unsigned k = 0; k < K; k++)
{
M[k] = std::log(M[k]/Q);
G[k] /= Q;
gsl_matrix_set(X, m, 0, 1.0);
gsl_matrix_set(X, m, 1, std::log(double(data.getSize() * (l + 1)) / double(F)));
gsl_matrix_set(X, m, 2, std::log(double(k + 1)));
gsl_vector_set(yM, m, M[k]);
gsl_vector_set(yG, m, G[k]);
++m;
}
++progress;
}
gsl_multifit_linear_workspace *work = gsl_multifit_linear_alloc(F * K, 3);
double chisq;
gsl_multifit_linear(X, yM, pM, cov, &chisq, work);
gsl_multifit_linear(X, yG, pG, cov, &chisq, work);
std::cout << gM << '\t' << gG << std::endl;
std::cout << gsl_vector_get(pM, 0) << '\t'
<< gsl_vector_get(pM, 1) << '\t'
<< gsl_vector_get(pM, 2) << std::endl;
std::cout << gsl_vector_get(pG, 0) << '\t'
<< gsl_vector_get(pG, 1) << '\t'
<< gsl_vector_get(pG, 2) << std::endl;
gsl_matrix_free(X);
gsl_matrix_free(cov);
gsl_vector_free(pM);
gsl_vector_free(pG);
gsl_vector_free(yM);
gsl_vector_free(yG);
}
