/** 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
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);
}
std::string save_edge(const edge_type& edge) const {
std::stringstream strm;
const double prediction =
edge.source().data().factor.dot(edge.target().data().factor);
strm << edge.source().id() << '\t'
<< edge.target().id() << '\t'
<< prediction << '\n';
return strm.str();
}
std::string var(const edge_type& v) const {
assert(v.size() == r);
std::stringstream s("V");
typename edge_type::const_iterator i = v.begin();
s << *(i++) + 1;
for (; i != v.end(); ++i)
s << "S" << *i + 1;
return s.str();
}
// Scatter to scatter_edges edges with the new message value.
void scatter(icontext_type& context, const vertex_type& vertex, edge_type& edge) const {
bool isEdgeSource = (vertex.id() == edge.source().id());
bool hasSameData = isEdgeSource ? (vertex.data() == edge.target().data()) : (vertex.data() == edge.source().data()) ;
if (!hasSameData) {
min_combiner combiner;
combiner.value = message_value;
context.signal(isEdgeSource ? edge.target() : edge.source(), combiner);
}
}
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;
}
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());
}
}
gather_type gather(icontext_type& context, const vertex_type& vertex, edge_type& edge) const {
if (context.iteration() == 0) {
if (vertex.id() == edge.source().id()) {
return gather_type(edge.target().id());
} else {
return gather_type(edge.source().id());
}
} else {
return gather_type();
}
}
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
void scatter(icontext_type& context, const vertex_type& vertex, edge_type& edge) const {
if (context.iteration() == 0) {
pair<size_t, size_t> p = count_triangles(edge.source(), edge.target());
if (p.first > 0) {
context.signal(edge.source(), p.first);
}
if (p.second > 0) {
context.signal(edge.target(), p.second);
}
} else {
//
}
}
double gather(icontext_type& context, const vertex_type& vertex,
edge_type& edge) const
{
edge_data e = edge.parse<edge_data>();
int topic = rand()%env_inst.NTOPICS; //sample topic from Mean(1/K)
e.assignment=topic;
edge.update<edge_data>(e);
vertex_data vs=edge.source().parse<vertex_data>();
vertex_data vt=edge.target().parse<vertex_data>();
vs.n[topic]++;
vt.n[topic]++;
env_inst.nwsum[topic]++;
env_inst.ndsum[k2id(vs.id)]++;
return 0;
}
/**
* \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
/* 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 {
const vertex_type other = edge.target();
pagerank_type value = vertex.data().pagerank / vertex.num_out_edges();
assert(other.id() != vertex.id());
const sum_pagerank_type msg(value);
context.signal(other, msg);
}
/**
* \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 = edge.target();
/*if (USE_DELTA_CACHE) {
gather_data_type delta;
delta.dist = lastchange;
context.post_delta(other, delta);
}*/
context.signal(other);
}
double gather(icontext_type& context, const vertex_type& vertex,
edge_type& edge) const
{
// remove topic from the count variables
edge_data e=edge.parse<edge_data>();
int topic = e.assignment;
vertex_data vd=edge.source().parse<vertex_data>();
vertex_data vw=edge.target().parse<vertex_data>();
vw.n[topic]--;
vd.n[topic]--;
env_inst.nwsum[topic]--;
env_inst.ndsum[k2id(vd.id)]--;
// do multinomial sampling via cumulative method:
double *p=new double[env_inst.NTOPICS];
for (int k = 0; k < env_inst.NTOPICS; k++) {
p[k] = (vw.n[k] + env_inst.BETA) / (env_inst.nwsum[k] + env_inst.NWORDS * env_inst.BETA)
* (vd.n[k] + env_inst.ALPHA) / (env_inst.ndsum[k2id(vd.id)] + env_inst.NTOPICS * env_inst.ALPHA);
}
// cumulate multinomial parameters
for (int k = 1; k < env_inst.NTOPICS; k++) {
p[k] += p[k - 1];
}
// scaled sample because of unnormalised p[]
double u = (double)rand()/RAND_MAX * p[env_inst.NTOPICS - 1];
for (topic = 0; topic < env_inst.NTOPICS; topic++) {
if (u <= p[topic])
break;
}
delete []p;
// add newly estimated topic to count variables
vw.n[topic]++;
vd.n[topic]++;
env_inst.nwsum[topic]++;
env_inst.ndsum[k2id(vd.id)]++;
e.assignment=topic;
edge.update<edge_data>(e);
}
void scatter(icontext_type& context, const vertex_type& vertex,
edge_type& edge) const {
context.signal(edge.target());
}
void scatter(icontext_type& context, const vertex_type& vertex, edge_type& edge) const {
const vertex_type& other = edge.source().id() == vertex.id() ? edge.target() : edge.source();
context.signal(other);
}
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);
}
sum_pagerank_type gather(icontext_type& context, const vertex_type& vertex,
edge_type& edge) const
{
return sum_pagerank_type(edge.source().data().pagerank / edge.source().num_out_edges());
}
min_color_type gather(icontext_type& context, const vertex_type& vertex,
edge_type& edge) const
{
return min_color_type(edge.source().data().color);
}
/**
* \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 = edge.target();
context.signal(other);
}
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;
}
void scatter(icontext_type& context, const vertex_type& vertex, edge_type& edge) const {
if( edge.target().data().in_core ) {
context.signal(edge.target());
}
}
int gather(icontext_type& context, const vertex_type& vertex, edge_type& edge) const {
return ( edge.source().data().in_core ? 1 : 0 );
}
gather_type gather(icontext_type& context, const vertex_type& vertex, edge_type& edge) const {
const vertex_type& other = edge.source().id() == vertex.id() ? edge.target() : edge.source();
return gather_type(other.data());
}
double gather(icontext_type& context, const vertex_type& vertex,
edge_type& edge) const {
return ((1.0 - RESET_PROB) / edge.source().num_out_edges())
* edge.source().data();
}
