void CLMNN::train(SGMatrix<float64_t> init_transform)
{
SG_DEBUG("Entering CLMNN::train().\n")
// Check training data and arguments, initializing, if necessary, init_transform
CLMNNImpl::check_training_setup(m_features, m_labels, init_transform);
// Initializations
// cast is safe, check_training_setup ensures features are dense
CDenseFeatures<float64_t>* x = static_cast<CDenseFeatures<float64_t>*>(m_features);
CMulticlassLabels* y = CLabelsFactory::to_multiclass(m_labels);
SG_DEBUG("%d input vectors with %d dimensions.\n", x->get_num_vectors(), x->get_num_features());
// Use Eigen matrix for the linear transform L. The Mahalanobis distance is L^T*L
MatrixXd L = Map<MatrixXd>(init_transform.matrix, init_transform.num_rows,
init_transform.num_cols);
// Compute target or genuine neighbours
SG_DEBUG("Finding target nearest neighbors.\n")
SGMatrix<index_t> target_nn = CLMNNImpl::find_target_nn(x, y, m_k);
// Initialize (sub-)gradient
SG_DEBUG("Summing outer products for (sub-)gradient initialization.\n")
MatrixXd gradient = (1-m_regularization)*CLMNNImpl::sum_outer_products(x, target_nn);
// Value of the objective function at every iteration
SGVector<float64_t> obj(m_maxiter);
// The step size is modified depending on how the objective changes, leave the
// step size member unchanged and use a local one
float64_t stepsize = m_stepsize;
// Last active set of impostors computed exactly, current and previous impostors sets
ImpostorsSetType exact_impostors, cur_impostors, prev_impostors;
// Iteration counter
uint32_t iter = 0;
// Criterion for termination
bool stop = false;
// Make space for the training statistics
m_statistics->resize(m_maxiter);
// Main loop
while (!stop)
{
SG_PROGRESS(iter, 0, m_maxiter)
// Find current set of impostors
SG_DEBUG("Finding impostors.\n")
cur_impostors = CLMNNImpl::find_impostors(x,y,L,target_nn,iter,m_correction);
SG_DEBUG("Found %d impostors in the current set.\n", cur_impostors.size())
// (Sub-) gradient computation
SG_DEBUG("Updating gradient.\n")
CLMNNImpl::update_gradient(x, gradient, cur_impostors, prev_impostors, m_regularization);
// Take gradient step
SG_DEBUG("Taking gradient step.\n")
CLMNNImpl::gradient_step(L, gradient, stepsize, m_diagonal);
// Compute the objective, trace of Mahalanobis distance matrix (L squared) times the gradient
// plus the number of current impostors to account for the margin
SG_DEBUG("Computing objective.\n")
obj[iter] = TRACE(L.transpose()*L,gradient) + m_regularization*cur_impostors.size();
// Correct step size
CLMNNImpl::correct_stepsize(stepsize, obj, iter);
// Check termination criterion
stop = CLMNNImpl::check_termination(stepsize, obj, iter, m_maxiter, m_stepsize_threshold, m_obj_threshold);
// Update iteration counter
iter = iter + 1;
// Update previous set of impostors
prev_impostors = cur_impostors;
// Store statistics for this iteration
m_statistics->set(iter-1, obj[iter-1], stepsize, cur_impostors.size());
SG_DEBUG("iteration=%d, objective=%.4f, #impostors=%4d, stepsize=%.4E\n",
iter, obj[iter-1], cur_impostors.size(), stepsize)
}
// Truncate statistics in case convergence was reached in less than maxiter
m_statistics->resize(iter);
// Store the transformation found in the class attribute
int32_t nfeats = x->get_num_features();
float64_t* cloned_data = SGMatrix<float64_t>::clone_matrix(L.data(), nfeats, nfeats);
m_linear_transform = SGMatrix<float64_t>(cloned_data, nfeats, nfeats);
SG_DEBUG("Leaving CLMNN::train().\n")
}
void CKMeans::clustknb(bool use_old_mus, float64_t *mus_start)
{
ASSERT(distance && distance->get_feature_type()==F_DREAL);
CDenseFeatures<float64_t>* lhs = (CDenseFeatures<float64_t>*) distance->get_lhs();
ASSERT(lhs && lhs->get_num_features()>0 && lhs->get_num_vectors()>0);
int32_t XSize=lhs->get_num_vectors();
dimensions=lhs->get_num_features();
int32_t i, changed=1;
const int32_t XDimk=dimensions*k;
int32_t iter=0;
R=SGVector<float64_t>(k);
mus=SGMatrix<float64_t>(dimensions, k);
int32_t *ClList=SG_CALLOC(int32_t, XSize);
float64_t *weights_set=SG_CALLOC(float64_t, k);
float64_t *dists=SG_CALLOC(float64_t, k*XSize);
///replace rhs feature vectors
CDenseFeatures<float64_t>* rhs_mus = new CDenseFeatures<float64_t>(0);
CFeatures* rhs_cache = distance->replace_rhs(rhs_mus);
int32_t vlen=0;
bool vfree=false;
float64_t* vec=NULL;
/* ClList=zeros(XSize,1) ; */
memset(ClList, 0, sizeof(int32_t)*XSize);
/* weights_set=zeros(k,1) ; */
memset(weights_set, 0, sizeof(float64_t)*k);
/* cluster_centers=zeros(dimensions, k) ; */
memset(mus.matrix, 0, sizeof(float64_t)*XDimk);
if (!use_old_mus)
{
for (i=0; i<XSize; i++)
{
const int32_t Cl=CMath::random(0, k-1);
int32_t j;
float64_t weight=Weights.vector[i];
weights_set[Cl]+=weight;
ClList[i]=Cl;
vec=lhs->get_feature_vector(i, vlen, vfree);
for (j=0; j<dimensions; j++)
mus.matrix[Cl*dimensions+j] += weight*vec[j];
lhs->free_feature_vector(vec, i, vfree);
}
for (i=0; i<k; i++)
{
int32_t j;
if (weights_set[i]!=0.0)
for (j=0; j<dimensions; j++)
mus.matrix[i*dimensions+j] /= weights_set[i];
}
}
else
{
ASSERT(mus_start);
/// set rhs to mus_start
rhs_mus->copy_feature_matrix(SGMatrix<float64_t>(mus_start,dimensions,k));
float64_t* p_dists=dists;
for(int32_t idx=0;idx<XSize;idx++,p_dists+=k)
distances_rhs(p_dists,0,k,idx);
p_dists=NULL;
for (i=0; i<XSize; i++)
{
float64_t mini=dists[i*k];
int32_t Cl = 0, j;
for (j=1; j<k; j++)
{
if (dists[i*k+j]<mini)
{
Cl=j;
mini=dists[i*k+j];
}
}
ClList[i]=Cl;
}
/* Compute the sum of all points belonging to a cluster
* and count the points */
for (i=0; i<XSize; i++)
{
const int32_t Cl = ClList[i];
float64_t weight=Weights.vector[i];
weights_set[Cl]+=weight;
#ifndef MUSRECALC
vec=lhs->get_feature_vector(i, vlen, vfree);
for (j=0; j<dimensions; j++)
mus.matrix[Cl*dimensions+j] += weight*vec[j];
lhs->free_feature_vector(vec, i, vfree);
#endif
}
#ifndef MUSRECALC
/* normalization to get the mean */
for (i=0; i<k; i++)
{
if (weights_set[i]!=0.0)
{
int32_t j;
for (j=0; j<dimensions; j++)
mus.matrix[i*dimensions+j] /= weights_set[i];
}
}
#endif
}
while (changed && (iter<max_iter))
{
iter++;
if (iter==max_iter-1)
SG_WARNING("kmeans clustering changed throughout %d iterations stopping...\n", max_iter-1);
if (iter%1000 == 0)
SG_INFO("Iteration[%d/%d]: Assignment of %i patterns changed.\n", iter, max_iter, changed);
changed=0;
#ifdef MUSRECALC
/* mus=zeros(dimensions, k) ; */
memset(mus.matrix, 0, sizeof(float64_t)*XDimk);
for (i=0; i<XSize; i++)
{
int32_t j;
int32_t Cl=ClList[i];
float64_t weight=Weights.vector[i];
vec=lhs->get_feature_vector(i, vlen, vfree);
for (j=0; j<dimensions; j++)
mus.matrix[Cl*dimensions+j] += weight*vec[j];
lhs->free_feature_vector(vec, i, vfree);
}
for (i=0; i<k; i++)
{
int32_t j;
if (weights_set[i]!=0.0)
for (j=0; j<dimensions; j++)
mus.matrix[i*dimensions+j] /= weights_set[i];
}
#endif
///update rhs
rhs_mus->copy_feature_matrix(mus);
for (i=0; i<XSize; i++)
{
/* ks=ceil(rand(1,XSize)*XSize) ; */
const int32_t Pat= CMath::random(0, XSize-1);
const int32_t ClList_Pat=ClList[Pat];
int32_t imini, j;
float64_t mini, weight;
weight=Weights.vector[Pat];
/* compute the distance of this point to all centers */
for(int32_t idx_k=0;idx_k<k;idx_k++)
dists[idx_k]=distance->distance(Pat,idx_k);
/* [mini,imini]=min(dists(:,i)) ; */
imini=0 ; mini=dists[0];
for (j=1; j<k; j++)
if (dists[j]<mini)
{
mini=dists[j];
imini=j;
}
if (imini!=ClList_Pat)
{
changed= changed + 1;
/* weights_set(imini) = weights_set(imini) + weight ; */
weights_set[imini]+= weight;
/* weights_set(j) = weights_set(j) - weight ; */
weights_set[ClList_Pat]-= weight;
vec=lhs->get_feature_vector(Pat, vlen, vfree);
for (j=0; j<dimensions; j++)
{
mus.matrix[imini*dimensions+j]-=(vec[j]
-mus.matrix[imini*dimensions+j])
*(weight/weights_set[imini]);
}
lhs->free_feature_vector(vec, Pat, vfree);
/* mu_new = mu_old - (x - mu_old)/(n-1) */
/* if weights_set(j)~=0 */
if (weights_set[ClList_Pat]!=0.0)
{
vec=lhs->get_feature_vector(Pat, vlen, vfree);
for (j=0; j<dimensions; j++)
{
mus.matrix[ClList_Pat*dimensions+j]-=
(vec[j]
-mus.matrix[ClList_Pat
*dimensions+j])
*(weight/weights_set[ClList_Pat]);
}
lhs->free_feature_vector(vec, Pat, vfree);
}
else
/* mus(:,j)=zeros(dimensions,1) ; */
for (j=0; j<dimensions; j++)
mus.matrix[ClList_Pat*dimensions+j]=0;
/* ClList(i)= imini ; */
ClList[Pat] = imini;
}
}
}
/* compute the ,,variances'' of the clusters */
for (i=0; i<k; i++)
{
float64_t rmin1=0;
float64_t rmin2=0;
bool first_round=true;
for (int32_t j=0; j<k; j++)
{
if (j!=i)
{
int32_t l;
float64_t dist = 0;
for (l=0; l<dimensions; l++)
{
dist+=CMath::sq(
mus.matrix[i*dimensions+l]
-mus.matrix[j*dimensions+l]);
}
if (first_round)
{
rmin1=dist;
rmin2=dist;
first_round=false;
}
else
{
if ((dist<rmin2) && (dist>=rmin1))
rmin2=dist;
if (dist<rmin1)
{
rmin2=rmin1;
rmin1=dist;
}
}
}
}
R.vector[i]=(0.7*CMath::sqrt(rmin1)+0.3*CMath::sqrt(rmin2));
}
distance->replace_rhs(rhs_cache);
delete rhs_mus;
SG_FREE(ClList);
SG_FREE(weights_set);
SG_FREE(dists);
SG_UNREF(lhs);
}