void CFactorGraph::loss_augmentation(SGVector<int32_t> states_gt, SGVector<float64_t> loss)
{
if (loss.size() == 0)
{
loss.resize_vector(states_gt.size());
SGVector<float64_t>::fill_vector(loss.vector, loss.vlen, 1.0 / states_gt.size());
}
int32_t num_vars = states_gt.size();
ASSERT(num_vars == loss.size());
SGVector<int32_t> var_flags(num_vars);
var_flags.zero();
// augment loss to incorrect states in the first factor containing the variable
// since += L_i for each variable if it takes wrong state ever
// TODO: augment unary factors
for (int32_t fi = 0; fi < m_factors->get_num_elements(); ++fi)
{
CFactor* fac = dynamic_cast<CFactor*>(m_factors->get_element(fi));
SGVector<int32_t> vars = fac->get_variables();
for (int32_t vi = 0; vi < vars.size(); vi++)
{
int32_t vv = vars[vi];
ASSERT(vv < num_vars);
if (var_flags[vv])
continue;
SGVector<float64_t> energies = fac->get_energies();
for (int32_t ei = 0; ei < energies.size(); ei++)
{
CTableFactorType* ftype = fac->get_factor_type();
int32_t vstate = ftype->state_from_index(ei, vi);
SG_UNREF(ftype);
if (states_gt[vv] == vstate)
continue;
// -delta(y_n, y_i_n)
fac->set_energy(ei, energies[ei] - loss[vv]);
}
var_flags[vv] = 1;
}
SG_UNREF(fac);
}
// make sure all variables have been checked
int32_t min_var = SGVector<int32_t>::min(var_flags.vector, var_flags.vlen);
ASSERT(min_var == 1);
}
SGVector< float64_t > CFactorGraphModel::get_joint_feature_vector(int32_t feat_idx, CStructuredData* y)
{
// factor graph instance
CFactorGraphFeatures* mf = CFactorGraphFeatures::obtain_from_generic(m_features);
CFactorGraph* fg = mf->get_sample(feat_idx);
// ground truth states
CFactorGraphObservation* fg_states = CFactorGraphObservation::obtain_from_generic(y);
SGVector<int32_t> states = fg_states->get_data();
// initialize psi
SGVector<float64_t> psi(get_dim());
psi.zero();
// construct unnormalized psi
CDynamicObjectArray* facs = fg->get_factors();
for (int32_t fi = 0; fi < facs->get_num_elements(); ++fi)
{
CFactor* fac = dynamic_cast<CFactor*>(facs->get_element(fi));
CTableFactorType* ftype = fac->get_factor_type();
int32_t id = ftype->get_type_id();
SGVector<int32_t> w_map = get_params_mapping(id);
ASSERT(w_map.size() == ftype->get_w_dim());
SGVector<float64_t> dat = fac->get_data();
int32_t dat_size = dat.size();
ASSERT(w_map.size() == dat_size * ftype->get_num_assignments());
int32_t ei = ftype->index_from_universe_assignment(states, fac->get_variables());
for (int32_t di = 0; di < dat_size; di++)
psi[w_map[ei*dat_size + di]] += dat[di];
SG_UNREF(ftype);
SG_UNREF(fac);
}
// negation (-E(x,y) = <w,phi(x,y)>)
psi.scale(-1.0);
SG_UNREF(facs);
SG_UNREF(fg);
return psi;
}
void CFactorGraph::connect_components()
{
if (m_dset->get_connected())
return;
// need to be reset once factor graph is updated
m_dset->make_sets();
bool flag = false;
for (int32_t fi = 0; fi < m_factors->get_num_elements(); ++fi)
{
CFactor* fac = dynamic_cast<CFactor*>(m_factors->get_element(fi));
SGVector<int32_t> vars = fac->get_variables();
int32_t r0 = m_dset->find_set(vars[0]);
for (int32_t vi = 1; vi < vars.size(); vi++)
{
// for two nodes in a factor, should be an edge between them
// but this time link() isn't performed, if they are linked already
// means there is another path connected them, so cycle detected
int32_t ri = m_dset->find_set(vars[vi]);
if (r0 == ri)
{
flag = true;
continue;
}
r0 = m_dset->link_set(r0, ri);
}
SG_UNREF(fac);
}
m_has_cycle = flag;
m_dset->set_connected(true);
}