IncrementalScoringFunction::Data IncrementalScoringFunction::create_data(
ParticleIndex pi, RestraintsTemp cr,
const boost::unordered_map<Restraint *, int> &all,
const Restraints &dummies) const {
IMP_LOG_TERSE("Dependent restraints for particle "
<< get_model()->get_particle_name(pi) << " are " << cr
<< std::endl);
Data ret;
for (unsigned int j = 0; j < cr.size(); ++j) {
if (all.find(cr[j]) != all.end()) {
int index = all.find(cr[j])->second;
IMP_INTERNAL_CHECK(
std::find(ret.indexes.begin(), ret.indexes.end(), index) ==
ret.indexes.end(),
"Found duplicate restraint " << Showable(cr[j]) << " in list " << cr);
ret.indexes.push_back(index);
}
}
cr += RestraintsTemp(dummies.begin(), dummies.end());
ret.sf = new IncrementalRestraintsScoringFunction(
cr, 1.0, NO_MAX, get_model()->get_particle_name(pi) + " restraints");
return ret;
}
SubsetGraph get_restraint_graph(ScoringFunctionAdaptor in,
const ParticleStatesTable *pst) {
RestraintsTemp rs =
IMP::create_decomposition(in->create_restraints());
// ScoreStatesTemp ss= get_required_score_states(rs);
SubsetGraph ret(rs.size()); // + ss.size());
IMP_LOG_TERSE("Creating restraint graph on " << rs.size() << " restraints."
<< std::endl);
boost::unordered_map<Particle *, int> map;
SubsetGraphVertexName pm = boost::get(boost::vertex_name, ret);
DependencyGraph dg = get_dependency_graph(rs[0]->get_model());
DependencyGraphVertexIndex index = IMP::get_vertex_index(dg);
/*IMP_IF_LOG(VERBOSE) {
IMP_LOG_VERBOSE( "dependency graph is \n");
IMP::internal::show_as_graphviz(dg, std::cout);
}*/
Subset ps = pst->get_subset();
for (unsigned int i = 0; i < ps.size(); ++i) {
ParticlesTemp t = get_dependent_particles(
ps[i], ParticlesTemp(ps.begin(), ps.end()), dg, index);
for (unsigned int j = 0; j < t.size(); ++j) {
IMP_USAGE_CHECK(map.find(t[j]) == map.end(),
"Currently particles which depend on more "
<< "than one particle "
<< "from the input set are not supported."
<< " Particle \"" << t[j]->get_name()
<< "\" depends on \"" << ps[i]->get_name()
<< "\" and \""
<< ps[map.find(t[j])->second]->get_name() << "\"");
map[t[j]] = i;
}
IMP_IF_LOG(VERBOSE) {
IMP_LOG_VERBOSE("Particle \"" << ps[i]->get_name() << "\" controls ");
for (unsigned int i = 0; i < t.size(); ++i) {
IMP_LOG_VERBOSE("\"" << t[i]->get_name() << "\" ");
}
IMP_LOG_VERBOSE(std::endl);
}
}
for (unsigned int i = 0; i < rs.size(); ++i) {
ParticlesTemp pl = IMP::get_input_particles(rs[i]->get_inputs());
std::sort(pl.begin(), pl.end());
pl.erase(std::unique(pl.begin(), pl.end()), pl.end());
Subset os(pl);
for (unsigned int j = 0; j < pl.size(); ++j) {
pl[j] = ps[map[pl[j]]];
}
std::sort(pl.begin(), pl.end());
pl.erase(std::unique(pl.begin(), pl.end()), pl.end());
Subset s(pl);
IMP_LOG_VERBOSE("Subset for restraint " << rs[i]->get_name() << " is " << s
<< " from " << os << std::endl);
pm[i] = s;
}
/*ScoreStatesTemp ss= get_required_score_states(rs);
for (ScoreStatesTemp::const_iterator it= ss.begin();
it != ss.end(); ++it) {
ParticlesTemp pl= (*it)->get_input_particles();
add_edges(ps, pl, map, *it, ret);
ParticlesTemp opl= (*it)->get_output_particles();
add_edges(ps, opl, map, *it, ret);
}
IMP_INTERNAL_CHECK(boost::num_vertices(ret) == ps.size(),
"Wrong number of vertices "
<< boost::num_vertices(ret)
<< " vs " << ps.size());*/
for (unsigned int i = 0; i < boost::num_vertices(ret); ++i) {
for (unsigned int j = 0; j < i; ++j) {
if (get_intersection(pm[i], pm[j]).size() > 0) {
boost::add_edge(i, j, ret);
IMP_LOG_VERBOSE("Connecting " << rs[i]->get_name() << " with "
<< rs[j]->get_name() << std::endl);
}
}
}
return ret;
}
double
get_slack_estimate(const ParticlesTemp& ps,
double upper_bound,
double step,
const RestraintsTemp &restraints,
bool derivatives,
Optimizer *opt,
ClosePairContainer *cpc) {
std::vector<Data> datas;
for (double slack=0; slack< upper_bound; slack+= step) {
IMP_LOG(VERBOSE, "Computing for " << slack << std::endl);
datas.push_back(Data());
datas.back().slack=slack;
{
boost::timer imp_timer;
int count=0;
base::SetLogState sl(opt->get_model(), SILENT);
do {
cpc->set_slack(slack);
cpc->update();
++count;
} while (imp_timer.elapsed()==0);
datas.back().ccost= imp_timer.elapsed()/count;
IMP_LOG(VERBOSE, "Close pair finding cost "
<< datas.back().ccost << std::endl);
}
{
boost::timer imp_timer;
double score=0;
int count=0;
int iters=1;
base::SetLogState sl(opt->get_model(), SILENT);
do {
for ( int j=0; j< iters; ++j) {
for (unsigned int i=0; i< restraints.size(); ++i) {
score+=restraints[i]->evaluate(derivatives);
// should restore
}
}
count += iters;
iters*=2;
} while (imp_timer.elapsed()==0);
datas.back().rcost= imp_timer.elapsed()/count;
IMP_LOG(VERBOSE, "Restraint evaluation cost "
<< datas.back().rcost << std::endl);
}
}
int ns=100;
int last_ns=1;
int opt_i=-1;
std::vector<Floats > dists(1, Floats(1,0.0));
std::vector< std::vector<algebra::Vector3D> >
pos(1, std::vector<algebra::Vector3D>(ps.size()));
for (unsigned int j=0; j< ps.size(); ++j) {
pos[0][j]=core::XYZ(ps[j]).get_coordinates();
}
do {
IMP_LOG(VERBOSE, "Stepping from " << last_ns << " to " << ns << std::endl);
dists.resize(ns, Floats(ns, 0.0));
for ( int i=0; i< last_ns; ++i) {
dists[i].resize(ns, 0.0);
}
pos.resize(ns, std::vector<algebra::Vector3D>(ps.size()));
base::SetLogState sl(opt->get_model(), SILENT);
for ( int i=last_ns; i< ns; ++i) {
opt->optimize(1);
for (unsigned int j=0; j< ps.size(); ++j) {
pos[i][j]=core::XYZ(ps[j]).get_coordinates();
}
}
for ( int i=last_ns; i< ns; ++i) {
for ( int j=0; j< i; ++j) {
double md=0;
for (unsigned int k=0; k< ps.size(); ++k) {
md= std::max(md, algebra::get_distance(pos[i][k], pos[j][k]));
}
dists[i][j]=md;
dists[j][i]=md;
}
}
// estimate lifetimes from slack
for (unsigned int i=0; i< datas.size(); ++i) {
Ints deaths;
for ( int j=0; j< ns; ++j) {
for ( int k=j+1; k < ns; ++k) {
if (dists[j][k]> datas[i].slack) {
deaths.push_back(k-j);
break;
}
}
}
std::sort(deaths.begin(), deaths.end());
// kaplan meier estimator
double ml=0;
if (deaths.empty()) {
ml= ns;
} else {
//double l=1;
IMP_INTERNAL_CHECK(deaths.size() < static_cast<unsigned int>(ns),
"Too much death");
double S=1;
for (unsigned int j=0; j< deaths.size(); ++j) {
double n= ns-j;
double t=(n-1.0)/n;
ml+=(S-t*S)*deaths[j];
S*=t;
}
}
datas[i].lifetime=ml;
IMP_LOG(VERBOSE, "Expected life of " << datas[i].slack
<< " is " << datas[i].lifetime << std::endl);
}
/**
C(s) is cost to compute
R(s) is const to eval restraints
L(s) is lifetime of slack
minimize C(s)/L(s)+R(s)
*/
// smooth
for (unsigned int i=1; i< datas.size()-1; ++i) {
datas[i].rcost=(datas[i].rcost+ datas[i-1].rcost+datas[i+1].rcost)/3.0;
datas[i].ccost=(datas[i].ccost+ datas[i-1].ccost+datas[i+1].ccost)/3.0;
datas[i].lifetime=(datas[i].lifetime
+ datas[i-1].lifetime+datas[i+1].lifetime)/3.0;
}
double min= std::numeric_limits<double>::max();
for (unsigned int i=0; i< datas.size(); ++i) {
double v= datas[i].rcost+ datas[i].ccost/datas[i].lifetime;
IMP_LOG(VERBOSE, "Cost of " << datas[i].slack << " is " << v
<< " from " << datas[i].rcost
<< " " << datas[i].ccost << " " << datas[i].lifetime
<< std::endl);
if (v < min) {
min=v;
opt_i=i;
}
}
last_ns=ns;
ns*=2;
IMP_LOG(VERBOSE, "Opt is " << datas[opt_i].slack << std::endl);
// 2 for the value, 2 for the doubling
// if it more than 1000, just decide that is enough
} while (datas[opt_i].lifetime > ns/4.0 && ns <1000);
return datas[opt_i].slack;
}
