/** The gather function computes XtX and Xy */
gather_type gather(icontext_type& context, const vertex_type& vertex,
edge_type& edge) const {
if(edge.data().role == edge_data::TRAIN) {
const vertex_type other_vertex = get_other_vertex(edge, vertex);
return gather_type(other_vertex.data().factor, edge.data().obs);
} else return gather_type();
} // end of gather function
factor_type gather(icontext_type& context, const vertex_type& vertex,
edge_type& edge) const {
vertex_type other_vertex = get_other_vertex(edge, vertex);
// VIOLATING THE ABSTRACTION!
vertex_data& vdata = graph_type::vertex_type(vertex).data();
// VIOLATING THE ABSTRACTION!
vertex_data& other_vdata = other_vertex.data();
factor_type& doc_topic_count =
is_doc(vertex) ? vdata.factor : other_vdata.factor;
factor_type& word_topic_count =
is_word(vertex) ? vdata.factor : other_vdata.factor;
ASSERT_EQ(doc_topic_count.size(), NTOPICS);
ASSERT_EQ(word_topic_count.size(), NTOPICS);
// run the actual gibbs sampling
factor_type& belief = edge.data().belief;
const uint32_t count = edge.data().count;
// Resample the topics
double sum = 0, old_sum = 0;
for(size_t t = 0; t < NTOPICS; ++t) {
old_sum += belief[t];
doc_topic_count[t] -= belief[t];
word_topic_count[t] -= belief[t];
GLOBAL_TOPIC_COUNT[t] -= belief[t];
const double n_dt =
std::max(count_type(doc_topic_count[t]), count_type(0));
ASSERT_GE(n_dt, 0);
const double n_wt =
std::max(count_type(word_topic_count[t]), count_type(0));
ASSERT_GE(n_wt, 0);
const double n_t =
std::max(count_type(GLOBAL_TOPIC_COUNT[t]), count_type(0));
ASSERT_GE(n_t, 0);
belief[t] = (ALPHA + n_dt) * (BETA + n_wt) / (BETA * NWORDS + n_t);
sum += belief[t];
} // End of loop over each token
ASSERT_GT(sum, 0);
if(old_sum == 0) {
size_t asg = graphlab::random::multinomial(belief);
for(size_t i = 0; i < NTOPICS; ++i) belief[i] = 0;
belief[asg] = count;
return belief;
}
for(size_t t = 0; t < NTOPICS; ++t) {
belief[t] = count * (belief[t]/sum);
doc_topic_count[t] += belief[t];
word_topic_count[t] += belief[t];
GLOBAL_TOPIC_COUNT[t] += belief[t];
}
return belief;
} // end of gather
gather_type gather(icontext_type& context, const vertex_type& vertex, edge_type& edge) const
{
cout << "gather(), edge=" << edge.source().id() << "->" << edge.target().id() << ", called from vid=" << vertex.id() << endl;
gather_type gathered;
// add id of other vertex of edge and add id->beta to map if message source
if (edge.target().id()==vertex.id())
{
// incoming edge, outgoing message, only if target is non-observed
if (!edge.source().data().is_observed)
{
gathered.message_targets.insert(edge.source().id());
cout << "added " << edge.source().id() << " as message target" << endl;
}
}
else
{
// outgoing edge, incoming message with beta
gathered.message_source_betas[edge.target().id()]=edge.data().beta;
cout << "added " << edge.target().id() << " as message source" << endl;
}
cout << "gathered=" << gathered << endl;
return gathered;
}
/**
* \brief The scatter function just signal adjacent pages
*/
void scatter(icontext_type& context, const vertex_type& vertex,
edge_type& edge) const {
const vertex_type other = get_other_vertex(edge, vertex);
distance_type newd = vertex.data().dist + edge.data().dist;
if (other.data().dist > newd) {
const min_distance_type msg(newd);
context.signal(other, msg);
}
} // end of scatter
void scatter(icontext_type& context, const vertex_type& vertex,
edge_type& edge) const
{
const vertex_type other = edge.target();
distance_type newd = vertex.data().dist + edge.data().dist;
const min_distance_type msg(newd);
context.signal(other, msg);
}
/* Gather the weighted rank of the adjacent page */
double gather(icontext_type& context, const vertex_type& vertex,
edge_type& edge) const {
if (edge.data().role == edge_data::PREDICT)
return 0;
bool brows = vertex.id() < (uint)info.get_start_node(false);
if (info.is_square())
brows = !mi.A_transpose;
if (mi.A_offset && mi.x_offset >= 0){
double val = edge.data().obs * (brows ? edge.target().data().pvec[mi.x_offset] :
edge.source().data().pvec[mi.x_offset]);
//printf("gather edge on vertex %d val %lg obs %lg\n", vertex.id(), val, edge.data().obs);
return val;
}
//printf("edge on vertex %d val %lg\n", vertex.id(), 0.0);
return 0;
}
void scatter(icontext_type& context, const vertex_type& vertex, edge_type& edge) const {
EData ed = edge.data();
float weight = ed.weight;
if ( !edge.source().data().actived && ((double)rand() / RAND_MAX) < weight){
// if ( !edge.source().data().actived && 0.5 <= weight){
context.signal(edge.source());
}
}
void scatter(icontext_type& context, const vertex_type& vertex, edge_type& edge) const {
float weight = edge.data();
if ( ((double)rand() / RAND_MAX) < weight){
context.signalVid(edge.target().id());
}
}
void scatter(icontext_type& context, const vertex_type& vertex, edge_type& edge) const
{
if (vertex.data() + edge.data() < edge.target().data())
context.signal(edge.target());
}
gather_type gather(icontext_type& context, const vertex_type& vertex, edge_type& edge) const
{
float newval = edge.data() + edge.source().data();
return gather_type(newval);
}
void scatter(icontext_type& context, const vertex_type& vertex, edge_type& edge) const
{
cout << "scatter(), edge=" << edge.source().id() << "->" << edge.target().id() << ", called from vid=" << vertex.id() << endl;
cout << "computing message from vid=" << vertex.id() << " to vid=" << edge.source().id() << endl;
vertex_id_type message_target=edge.source().id();
// find out whether full rank or incomplete Cholesky mode
// distinguish case this node being observed or not
VectorXd new_beta;
if (edge.target().data().is_observed)
{
cout << "observed target" << endl;
// extract system solutions and observation kernel vector, base on full rank or incomplete Cholesky
if (edge.data().full_rank)
{
cout << "full rank case" << endl;
MatrixXd L_s=edge.data().solution_matrices["L_s"];
cout << "L_s:" << L_s << endl;
MatrixXd L_t=edge.data().solution_matrices["L_t"];
cout << "L_t:" << L_t << endl;
VectorXd k=vertex.data().kernel_dict_obs.at(message_target);
cout << "k:" << k << endl;
// L_{s}^{-T}(L_{s}^{-1}(L_{t}^{-T}(L_{t}^{-1}k_{t}^{s}), from right to left, 4 solver calls
new_beta=k;
new_beta=L_t.triangularView<Lower>().solve(new_beta);
new_beta=L_t.transpose().triangularView<Upper>().solve(new_beta);
new_beta=L_s.triangularView<Lower>().solve(new_beta);
new_beta=L_s.transpose().triangularView<Upper>().solve(new_beta);
}
else
{
cout << "incomplete Cholesky case" << endl;
MatrixXd Q_s=edge.data().solution_matrices["Q_s"];
cout << "Q_s:" << Q_s << endl;
MatrixXd R_s=edge.data().solution_matrices["R_s"];
cout << "R_s:" << R_s << endl;
MatrixXd P_s=edge.data().solution_matrices["P_s"];
cout << "P_s:" << P_s << endl;
MatrixXd Q_t=edge.data().solution_matrices["Q_t"];
cout << "Q_t:" << Q_t << endl;
MatrixXd R_t=edge.data().solution_matrices["R_t"];
cout << "R_t:" << R_t << endl;
MatrixXd P_t=edge.data().solution_matrices["P_t"];
cout << "P_t:" << P_t << endl;
MatrixXd W=edge.data().solution_matrices["W"];
cout << "W:" << W << endl;
VectorXd k=vertex.data().kernel_dict_obs.at(message_target);
cout << "k:" << k << endl;
// R_{s}^{-1}(Q_{s}^{T}((P_{s}(W_{s}W_{t}^{T}))(R_{t}^{-1}(Q_{t}^{T}(P_{t}k_{\mathcal{I}_{t}}^{(s)})))
new_beta=k;
new_beta=P_t.transpose()*new_beta;
new_beta=Q_t.transpose()*new_beta;
new_beta=R_t.triangularView<Upper>().solve(new_beta);
new_beta=W*new_beta;
new_beta=P_s.transpose()*new_beta;
new_beta=Q_s.transpose()*new_beta;
new_beta=R_s.triangularView<Upper>().solve(new_beta);
}
}
else
{
cout << "non-observed target" << endl;
cout << "multiplied_incoming_messages: " << vertex.data().multiplied_incoming_messages << endl;
// extract system solutions, depending on full rank or incomplete Cholesky
if (edge.data().full_rank)
{
cout << "full rank case" << endl;
MatrixXd L_s=edge.data().solution_matrices["L_s"];
cout << "L_s:" << L_s << endl;
VectorXd k;
if (!vertex.data().multiplied_incoming_messages.size())
{
cout << "no incoming messages, using constant unit norm vector" << endl;
k=VectorXd::Constant(L_s.cols(), 1.0/sqrt(L_s.cols()));
}
else
{
k=vertex.data().multiplied_incoming_messages.at(message_target);
}
cout << "k:" << k << endl;
// (K_{s}+\lambda I){}^{-1}k_{ut}^{(s)}=L_{s}^{-T}(L_{s}^{-1}k_{ut}^{(s)}) from right to left, 2 solver calls
new_beta=k;
new_beta=L_s.triangularView<Lower>().solve(new_beta);
new_beta=L_s.transpose().triangularView<Upper>().solve(new_beta);
}
else
{
cout << "incomplete Cholesky case" << endl;
MatrixXd Q_s=edge.data().solution_matrices["Q_s"];
cout << "Q_s:" << Q_s << endl;
MatrixXd R_s=edge.data().solution_matrices["R_s"];
cout << "R_s:" << R_s << endl;
MatrixXd P_s=edge.data().solution_matrices["P_s"];
cout << "P_s:" << P_s << endl;
MatrixXd W=edge.data().solution_matrices["W"];
cout << "W:" << W << endl;
VectorXd k;
if (!vertex.data().multiplied_incoming_messages.size())
{
cout << "no incoming messages, using constant unit norm vector" << endl;
k=VectorXd::Constant(W.cols(), 1.0/sqrt(W.cols()));
}
else
{
k=vertex.data().multiplied_incoming_messages.at(message_target);
}
cout << "k:" << k << endl;
// R_{s}^{-1}(Q_{s}^{T}(P_{s}^{T}k_{t}^{(s)}))
new_beta=k;
new_beta=W*new_beta;
new_beta=P_s.transpose()*new_beta;
new_beta=Q_s.transpose()*new_beta;
new_beta=R_s.triangularView<Upper>().solve(new_beta);
}
}
// normalise
new_beta=new_beta/new_beta.norm();
// check whether has changed or not yet existed
double difference;
if (!edge.data().beta.rows())
difference=numeric_limits<double>::infinity();
else
difference=(new_beta-edge.data().beta).norm();
cout << "beta norm difference is " << difference << endl;
if (difference>BETA_EPSILON)
{
// store new message and signal depending node if beta has changed or has not yet existed
edge.data().beta=new_beta;
context.signal(edge.source());
cout << "beta has changed, new_beta=" << new_beta << "\nhas norm=" << new_beta.norm() << ", signalling vid=" << edge.source().id() << endl;
}
else
{
cout << "converged!\n";
}
cout << "beta: " << edge.source().id() << "->" << edge.target().id() << ": " << edge.data().beta << endl;
}