void IndexEncoding::BestNextWords(const double* p_residual,
const SMatrix<double>& matDictionary,
int idx_start,
int idx_end,
SMatrix<double> &mat_diff,
Vector<short>& best_idx) const
{
int m_nDimension = matDictionary.Cols();
SMatrix<double> mat_each_diff(idx_end - idx_start, m_nDimension);
Heap<PAIR<double> > pq;
pq.Reserve(m_nNumBestCan);
for (int i = idx_start; i < idx_end; i++)
{
const double* p_dic = matDictionary[i];
double e = 0;
for (int j = 0; j < m_nDimension; j++)
{
double d = p_residual[j] - p_dic[j];
mat_each_diff[i - idx_start][j] = d;
e += d * d;
}
if (pq.size() >= m_nNumBestCan)
{
const PAIR<double> &p = pq.Top();
if (p.distance > e)
{
pq.popMin();
pq.insert(PAIR<double>(i - idx_start, e));
}
}
else
{
pq.insert(PAIR<double>(i - idx_start, e));
}
}
mat_diff.AllocateSpace(m_nNumBestCan, m_nDimension);
best_idx.AllocateSpace(m_nNumBestCan);
for (int i = m_nNumBestCan - 1; i >= 0; i--)
{
PAIR<double> p;
pq.popMin(p);
best_idx[i] = p.index;
memcpy(mat_diff[i], mat_each_diff[p.index], sizeof(double) * m_nDimension);
}
}
void IndexEncoding::BestNextWordsSMart(
const Vector<double> &vec_x_map,
const SMatrix<double> &matInnerProduct,
const short* prepresentation,
int idx,
short next_idx[],
double next_errors[]) const
{
Heap<PAIR<double> > pq;
pq.Reserve(m_nNumBestCan);
int sub_dic_size = matInnerProduct.Rows() / m_nNumberDictionaryEachPartition;
int idx_start = idx * sub_dic_size;
int idx_end = idx_start + sub_dic_size;
for (int i = idx_start; i < idx_end; i++)
{
// compoute the relative error
double e = -vec_x_map[i];
const double* p_inner = matInnerProduct[i];
for (int j = 0; j < idx; j++)
{
e += p_inner[prepresentation[j]];
}
e += 0.5 * p_inner[i];
if (pq.size() >= m_nNumBestCan)
{
const PAIR<double> &p = pq.Top();
if (p.distance > e)
{
pq.popMin();
pq.insert(PAIR<double>(i, e));
}
}
else
{
pq.insert(PAIR<double>(i, e));
}
}
for (int i = m_nNumBestCan - 1; i >= 0; i--)
{
PAIR<double> p;
pq.popMin(p);
next_idx[i] = p.index;
next_errors[i] = p.distance;
}
}
int IndexEncoding::SolveAdditiveQuantization(const Vector<double> &vec_x_map,
const SMatrix<double> &matPrecomputed,
int num_dic_each_partition,
short* prepresentation) const
{
int num = 0;
bool is_changed;
int num_sub_centers = matPrecomputed.Rows() / num_dic_each_partition;
SMatrix<short>* p_curr;
SMatrix<short>* p_aux;
p_curr = new SMatrix<short>(m_nNumBestCan, num_dic_each_partition);
p_aux = new SMatrix<short>(m_nNumBestCan, num_dic_each_partition);
p_curr->SetValue(-1);
{ // find the best m_aq_candidate number of points.
int idx_center = 0;
Heap<GreatPair<double> > heap;
heap.Reserve(m_nNumBestCan);
for (int idx_sub_cluster = 0; idx_sub_cluster < num_dic_each_partition; idx_sub_cluster++)
{
for (int i = 0; i < num_sub_centers; i++)
{
double s = residual_distance(vec_x_map, idx_center, idx_sub_cluster,
(*p_curr)[0], num_dic_each_partition, matPrecomputed);
if (heap.size() < m_nNumBestCan)
{
heap.insert(GreatPair<double>(idx_center, s));
}
else
{
const GreatPair<double> &top = heap.Top();
if (top.distance > s)
{
heap.popMin();
heap.insert(GreatPair<double>(idx_center, s));
}
}
idx_center++;
}
}
SMART_ASSERT(heap.size() == m_nNumBestCan).Exit();
for (int i = 0; i < m_nNumBestCan; i++)
{
GreatPair<double> v;
heap.popMin(v);
//PRINT << v.index << "\t" << v.distance << "\n";
(*p_curr)[i][v.index / num_sub_centers] = v.index;
}
//PRINT << *p_curr << "\n";
}
int idx_start_center = 0;
for (int idx_sub_cluster1 = 1; idx_sub_cluster1 < num_dic_each_partition; idx_sub_cluster1++)
{
priority_queue<Triplet<int, int, double>,
vector<Triplet<int, int, double> >,
LessTripletThird<int, int, double> > heap;
for (int idx_can = 0; idx_can < m_nNumBestCan; idx_can++)
{
short* curr_presentation = (*p_curr)[idx_can];
for (int idx_sub_cluster2 = 0; idx_sub_cluster2 < num_dic_each_partition; idx_sub_cluster2++)
{
if (curr_presentation[idx_sub_cluster2] == -1)
{
int idx_start_center = idx_sub_cluster2 * num_sub_centers;
int idx_end_center = idx_start_center + num_sub_centers;
for (int idx_center = idx_start_center; idx_center < idx_end_center; idx_center++)
{
double s = residual_distance(vec_x_map, idx_center, idx_sub_cluster2, curr_presentation,
num_dic_each_partition, matPrecomputed);
if (heap.size() < m_nNumBestCan)
{
heap.push(Triplet<int, int, double>(idx_can, idx_center, s));
}
else
{
const Triplet<int, int, double> &top = heap.top();
if (top.third > s)
{
heap.pop();
heap.push(Triplet<int, int, double>(idx_can, idx_center, s));
}
}
}
}
}
}
SMART_ASSERT(m_nNumBestCan == heap.size())(heap.size()).Exit();
for (int idx_can = 0; idx_can < m_nNumBestCan; idx_can++)
{
const Triplet<int, int, double> &top = heap.top();
int idx_origin_can = top.first;
const short* p_origin = p_curr->operator[](idx_origin_can);
short* p_now = p_aux->operator[](idx_can);
//PRINT << top.first << "\t" << top.second << "\t" << top.third << "\n";
memcpy(p_now, p_origin, sizeof(short) * num_dic_each_partition);
p_now[top.second / num_sub_centers] = top.second;
heap.pop();
}
swap(p_curr, p_aux);
//PRINT << *p_curr;
}
memcpy(prepresentation, p_curr->operator[](m_nNumBestCan - 1), sizeof(short) * num_dic_each_partition);
//PRINT << *p_curr;
delete p_curr;
delete p_aux;
return false;
}