int main(int argv, char** argc){
// parse arguments
dd::CmdParser cmd_parser = parse_input(argv, argc);
// run gibbs sampler
if(cmd_parser.app_name == "gibbs"){
gibbs(cmd_parser);
} else if (cmd_parser.app_name == "em") {
em(cmd_parser);
}
}
void SimulatedAnnealingInference::PerformInference() {
// Initialize best solution
unsigned int var_count =
static_cast<unsigned int>(fg->Cardinalities().size());
primal_best.resize(var_count);
primal_best_energy = std::numeric_limits<double>::infinity();
// Initialize Gibbs sampler
GibbsSampler gibbs(fg);
// Calculate exponential multiplier alpha such that T(sa_steps)=Tfinal.
double alpha = std::exp(std::log(Tfinal / T0) /
static_cast<double>(sa_steps));
gibbs.SetInverseTemperature(1.0 / T0);
gibbs.Sweep(1);
// Perform annealing
for (unsigned int k = 1; k <= sa_steps; ++k) {
// Logarithmic schedule (Geman and Geman)
// temperature = anneal_C / std::log(1.0 + static_cast<double>(k));
// Exponential schedule
double temperature = T0 * std::pow(alpha, static_cast<double>(k));
gibbs.SetInverseTemperature(1.0 / temperature);
double cur_energy = fg->EvaluateEnergy(gibbs.State());
double energy_delta;
for (unsigned int vi = 0; vi < var_count; ++vi) {
unsigned int new_state = gibbs.SampleSite(vi, energy_delta);
gibbs.SetState(vi, new_state);
cur_energy += energy_delta;
if (cur_energy < primal_best_energy) {
// Re-evaluate for numerical reasons
cur_energy = fg->EvaluateEnergy(gibbs.State());
if (cur_energy < primal_best_energy) {
primal_best_energy = cur_energy;
std::copy(gibbs.State().begin(), gibbs.State().end(),
primal_best.begin());
}
}
}
}
}
int main() {
const double HBC = Constants::HBCFmMeV;
const double PI = Constants::Pi;
//std::vector<double> Ska35S2009{-172.485, 172.087, -1767.71, 0.282732,
// 12899.2, 0.413266, 0.35};
double Tmin = 1.e-2;
double Tmax = 2.e2;
EOSSkyrme eos = EOSSkyrme::FromErmalSkyrme(ErmalSkyrmeParameters::Ska35S2009);
EOSSingleNucleus eosSN(eos, false, Tmin, Tmax);
GibbsPhaseConstruct gibbs(eos, false, Tmin, Tmax);
EOSElectron eosEl;
// Read in pre-calculated EoS data
std::ifstream ifs("eos_data.arx");
if (ifs) {
boost::archive::binary_iarchive ia(ifs);
ia >> gibbs;
ifs.close();
} else {
void gibbs(dd::CmdParser & cmd_parser){
// number of NUMA nodes
int n_numa_node = numa_max_node() + 1;
// number of max threads per NUMA node
int n_thread_per_numa = (sysconf(_SC_NPROCESSORS_CONF))/(n_numa_node);
// get command line arguments
std::string fg_file = cmd_parser.fg_file->getValue();
std::string weight_file = cmd_parser.weight_file->getValue();
std::string variable_file = cmd_parser.variable_file->getValue();
std::string factor_file = cmd_parser.factor_file->getValue();
std::string edge_file = cmd_parser.edge_file->getValue();
std::string meta_file = cmd_parser.meta_file->getValue();
std::string output_folder = cmd_parser.output_folder->getValue();
int n_learning_epoch = cmd_parser.n_learning_epoch->getValue();
int n_samples_per_learning_epoch = cmd_parser.n_samples_per_learning_epoch->getValue();
int n_inference_epoch = cmd_parser.n_inference_epoch->getValue();
double stepsize = cmd_parser.stepsize->getValue();
double stepsize2 = cmd_parser.stepsize2->getValue();
// hack to support two parameters to specify step size
if (stepsize == 0.01) stepsize = stepsize2;
double decay = cmd_parser.decay->getValue();
int n_datacopy = cmd_parser.n_datacopy->getValue();
double reg_param = cmd_parser.reg_param->getValue();
double reg1_param = cmd_parser.reg1_param->getValue();
bool is_quiet = cmd_parser.quiet->getValue();
bool sample_evidence = cmd_parser.sample_evidence->getValue();
int burn_in = cmd_parser.burn_in->getValue();
bool learn_non_evidence = cmd_parser.learn_non_evidence->getValue();
Meta meta = read_meta(fg_file);
if (is_quiet) {
std::cout << "Running in quiet mode..." << std::endl;
} else {
std::cout << std::endl;
std::cout << "#################MACHINE CONFIG#################" << std::endl;
std::cout << "# # NUMA Node : " << n_numa_node << std::endl;
std::cout << "# # Thread/NUMA Node : " << n_thread_per_numa << std::endl;
std::cout << "################################################" << std::endl;
std::cout << std::endl;
std::cout << "#################GIBBS SAMPLING#################" << std::endl;
std::cout << "# fg_file : " << fg_file << std::endl;
std::cout << "# edge_file : " << edge_file << std::endl;
std::cout << "# weight_file : " << weight_file << std::endl;
std::cout << "# variable_file : " << variable_file << std::endl;
std::cout << "# factor_file : " << factor_file << std::endl;
std::cout << "# meta_file : " << meta_file << std::endl;
std::cout << "# output_folder : " << output_folder << std::endl;
std::cout << "# n_learning_epoch : " << n_learning_epoch << std::endl;
std::cout << "# n_samples/l. epoch : " << n_samples_per_learning_epoch << std::endl;
std::cout << "# n_inference_epoch : " << n_inference_epoch << std::endl;
std::cout << "# stepsize : " << stepsize << std::endl;
std::cout << "# decay : " << decay << std::endl;
std::cout << "# regularization : " << reg_param << std::endl;
std::cout << "# l1 regularization : " << reg1_param << std::endl;
std::cout << "################################################" << std::endl;
std::cout << "# IGNORE -s (n_samples/l. epoch). ALWAYS -s 1. #" << std::endl;
std::cout << "# IGNORE -t (threads). ALWAYS USE ALL THREADS. #" << std::endl;
std::cout << "################################################" << std::endl;
std::cout << "# nvar : " << meta.num_variables << std::endl;
std::cout << "# nfac : " << meta.num_factors << std::endl;
std::cout << "# nweight : " << meta.num_weights << std::endl;
std::cout << "# nedge : " << meta.num_edges << std::endl;
std::cout << "################################################" << std::endl;
}
// run on NUMA node 0
numa_run_on_node(0);
numa_set_localalloc();
// load factor graph
dd::FactorGraph fg(meta.num_variables, meta.num_factors, meta.num_weights, meta.num_edges);
fg.load(cmd_parser, is_quiet);
dd::GibbsSampling gibbs(&fg, &cmd_parser, n_datacopy, sample_evidence, burn_in, learn_non_evidence);
// number of learning epochs
// the factor graph is copied on each NUMA node, so the total epochs =
// epochs specified / number of NUMA nodes
int numa_aware_n_learning_epoch = (int)(n_learning_epoch/n_numa_node) +
(n_learning_epoch%n_numa_node==0?0:1);
// learning
/*gibbs.learn(numa_aware_n_learning_epoch, n_samples_per_learning_epoch,
stepsize, decay, reg_param, reg1_param, meta_file, is_quiet);*/
gibbs.learn(numa_aware_n_learning_epoch, n_samples_per_learning_epoch,
stepsize, decay, reg_param, reg1_param, is_quiet);
// dump weights
gibbs.dump_weights(is_quiet);
// number of inference epochs
int numa_aware_n_epoch = (int)(n_inference_epoch/n_numa_node) +
(n_inference_epoch%n_numa_node==0?0:1);
// inference
gibbs.inference(numa_aware_n_epoch, is_quiet);
gibbs.aggregate_results_and_dump(is_quiet);
// print weights from inference result
for(long t=0;t<fg.n_weight;t++) {
std::cout<<fg.infrs->weight_values[t]<<std::endl;
}
// print weights from factor graph
std::cout<<"PRINTING WEIGHTS FROM WEIGHT VARIABLE"<<std::endl;
for(long t=0;t<fg.n_weight;t++) {
std::cout<<fg.weights[t].weight<<std::endl;
}
}
void em(dd::CmdParser & cmd_parser){
// number of NUMA nodes
int n_numa_node = numa_max_node() + 1;
// number of max threads per NUMA node
int n_thread_per_numa = (sysconf(_SC_NPROCESSORS_CONF))/(n_numa_node);
// get command line arguments
std::string fg_file = cmd_parser.fg_file->getValue();
std::string weight_file = cmd_parser.weight_file->getValue();
std::string variable_file = cmd_parser.variable_file->getValue();
std::string factor_file = cmd_parser.factor_file->getValue();
std::string edge_file = cmd_parser.edge_file->getValue();
std::string meta_file = cmd_parser.meta_file->getValue();
std::string output_folder = cmd_parser.output_folder->getValue();
int n_learning_epoch = cmd_parser.n_learning_epoch->getValue();
int n_samples_per_learning_epoch = cmd_parser.n_samples_per_learning_epoch->getValue();
int n_inference_epoch = cmd_parser.n_inference_epoch->getValue();
double stepsize = cmd_parser.stepsize->getValue();
double stepsize2 = cmd_parser.stepsize2->getValue();
// hack to support two parameters to specify step size
if (stepsize == 0.01) stepsize = stepsize2;
double decay = cmd_parser.decay->getValue();
int n_datacopy = cmd_parser.n_datacopy->getValue();
double reg_param = cmd_parser.reg_param->getValue();
double reg1_param = cmd_parser.reg1_param->getValue();
bool is_quiet = cmd_parser.quiet->getValue();
bool check_convergence = cmd_parser.check_convergence->getValue();
bool sample_evidence = cmd_parser.sample_evidence->getValue();
int burn_in = cmd_parser.burn_in->getValue();
int n_iter = cmd_parser.n_iter->getValue();
int wl_conv = cmd_parser.wl_conv->getValue();
int delta = cmd_parser.delta->getValue();
bool learn_non_evidence = cmd_parser.learn_non_evidence->getValue();
Meta meta = read_meta(fg_file);
if (is_quiet) {
std::cout << "Running in quiet mode..." << std::endl;
} else {
std::cout << std::endl;
std::cout << "#################MACHINE CONFIG#################" << std::endl;
std::cout << "# # NUMA Node : " << n_numa_node << std::endl;
std::cout << "# # Thread/NUMA Node : " << n_thread_per_numa << std::endl;
std::cout << "################################################" << std::endl;
std::cout << std::endl;
std::cout << "#################GIBBS SAMPLING#################" << std::endl;
std::cout << "# fg_file : " << fg_file << std::endl;
std::cout << "# edge_file : " << edge_file << std::endl;
std::cout << "# weight_file : " << weight_file << std::endl;
std::cout << "# variable_file : " << variable_file << std::endl;
std::cout << "# factor_file : " << factor_file << std::endl;
std::cout << "# meta_file : " << meta_file << std::endl;
std::cout << "# output_folder : " << output_folder << std::endl;
std::cout << "# n_learning_epoch : " << n_learning_epoch << std::endl;
std::cout << "# n_samples/l. epoch : " << n_samples_per_learning_epoch << std::endl;
std::cout << "# n_inference_epoch : " << n_inference_epoch << std::endl;
std::cout << "# stepsize : " << stepsize << std::endl;
std::cout << "# decay : " << decay << std::endl;
std::cout << "# regularization : " << reg_param << std::endl;
std::cout << "# l1 regularization : " << reg1_param << std::endl;
std::cout << "################################################" << std::endl;
std::cout << "# IGNORE -s (n_samples/l. epoch). ALWAYS -s 1. #" << std::endl;
std::cout << "# IGNORE -t (threads). ALWAYS USE ALL THREADS. #" << std::endl;
std::cout << "################################################" << std::endl;
std::cout << "# nvar : " << meta.num_variables << std::endl;
std::cout << "# nfac : " << meta.num_factors << std::endl;
std::cout << "# nweight : " << meta.num_weights << std::endl;
std::cout << "# nedge : " << meta.num_edges << std::endl;
std::cout << "################################################" << std::endl;
}
// run on NUMA node 0
numa_run_on_node(0);
numa_set_localalloc();
// load factor graph
dd::FactorGraph fg(meta.num_variables, meta.num_factors, meta.num_weights, meta.num_edges);
fg.load(cmd_parser, is_quiet);
dd::GibbsSampling gibbs(&fg, &cmd_parser, n_datacopy, sample_evidence, burn_in, learn_non_evidence);
// Initialize EM instance
dd::ExpMax expMax(&fg, &gibbs, wl_conv, delta, check_convergence);
// number of inference epochs
int numa_aware_n_epoch;
int numa_aware_n_learning_epoch;
// EM init -- run Maximzation (semi-supervised learning)
// Maximization step
numa_aware_n_learning_epoch = (int)(n_learning_epoch/n_numa_node) +
(n_learning_epoch%n_numa_node==0?0:1);
expMax.maximization(numa_aware_n_learning_epoch, n_samples_per_learning_epoch,
stepsize, decay, reg_param, reg1_param, is_quiet);
/*expMax.maximization(numa_aware_n_learning_epoch, n_samples_per_learning_epoch,
stepsize, decay, reg_param, reg1_param, meta_file, is_quiet);*/
while (!expMax.hasConverged && n_iter > 0) {
// Expectation step
numa_aware_n_epoch = (int)(n_inference_epoch/n_numa_node) +
(n_inference_epoch%n_numa_node==0?0:1);
expMax.expectation(numa_aware_n_epoch,is_quiet);
// Maximization step
numa_aware_n_learning_epoch = (int)(n_learning_epoch/n_numa_node) +
(n_learning_epoch%n_numa_node==0?0:1);
/*expMax.maximization(numa_aware_n_learning_epoch, n_samples_per_learning_epoch,
stepsize, decay, reg_param, reg1_param, meta_file, is_quiet);*/
expMax.maximization(numa_aware_n_learning_epoch, n_samples_per_learning_epoch,
stepsize, decay, reg_param, reg1_param, is_quiet);
//Decrement iteration counter
n_iter--;
}
expMax.dump_weights(is_quiet);
expMax.aggregate_results_and_dump(is_quiet);
}
int main() {
double HBC = Constants::HBCFmMeV;
//EOSSkyrme eos;
//std::vector<double> Ska35S2009{-172.485, 172.087, -1767.71, 0.282732,
// 12899.2, 0.413266, 0.35};
EOSSkyrme eos = EOSSkyrme::FromErmalSkyrme(ErmalSkyrmeParameters::Ska35S2009);
GibbsPhaseConstruct gibbs(eos, false);
// Calculat self bound liquid densities
std::ofstream selfBoundFile;
selfBoundFile.open("self_bound.out");
selfBoundFile << "# [1] Nb " << std::endl;
selfBoundFile << "# [2] T (MeV) " << std::endl;
selfBoundFile << "# [3] Ye " << std::endl;
selfBoundFile << "# [4] P " << std::endl;
selfBoundFile << "# [5] Mun " << std::endl;
selfBoundFile << "# [6] Mup " << std::endl;
selfBoundFile << "# [7] Nn (low density) " << std::endl;
selfBoundFile << "# [8] Np (low density) " << std::endl;
selfBoundFile << "# [9] P (low density) " << std::endl;
std::ofstream outFile;
outFile.open("gibbs_boundary.out");
outFile << "#[1] Nn_1 (fm^-3)" << std::endl;
outFile << "#[2] Nn_2 (fm^-3)" << std::endl;
outFile << "#[3] Np_1 (fm^-3)" << std::endl;
outFile << "#[4] Np_2 (fm^-3)" << std::endl;
outFile << "#[5] P (MeV fm^-3)" << std::endl;
outFile << "#[6] Mun (MeV fm^-3)" << std::endl;
outFile << "#[7] Mup (MeV fm^-3)" << std::endl;
outFile << "#[8] T (MeV)" << std::endl;
double TStart = 0.1/HBC;
double TEnd = 0.01/HBC;
auto phaseBoundFull = gibbs.FindFixedTPhaseBoundary(TStart);
std::cout << phaseBoundFull.size() << std::endl;
for (double lT = log10(TStart); lT>=log10(TEnd); lT -= 1.0) {
double T = pow(10.0, lT);
std::cout << T * HBC << " ";
//auto selfBound = gibbs.SelfBoundPoints(T);
//for (auto &ptpair : selfBound) {
// auto pt = ptpair.second;
// selfBoundFile << pt.Nb() << " " << pt.T()*HBC << " " << pt.Ye() << " "
// << pt.P() << " " << pt.Mun() << " " << pt.Mup() << " ";
// auto eosPts =
// eos.FromMuAndT(EOSData::InputFromTMunMup(pt.T(), pt.Mun(), pt.Mup()));
// if (eosPts.size()>0) {
// selfBoundFile << eosPts[0].Np() << " " << eosPts[0].Nn() << " "
// << eosPts[0].P();
// }
// selfBoundFile << std::endl;
//}
//selfBoundFile << std::endl;
// Calculate phase boundaries
auto phaseBound = phaseBoundFull;
if (T < TStart)
phaseBound = gibbs.FindFixedTPhaseBoundary(T, phaseBoundFull);
std::cout << phaseBound.size() << std::endl;
if (phaseBound.size() > 0) {
// Write out the phase boundary
for (auto &a : phaseBound) {
outFile << (a.first).Nn() << " ";
outFile << (a.second).Nn() << " ";
outFile << (a.first).Np() << " ";
outFile << (a.second).Np() << " ";
outFile << (a.first).P()*HBC << " ";
outFile << (a.first).Mun()*HBC << " ";
outFile << (a.first).Mup()*HBC << " ";
outFile << (a.first).T()*HBC << std::endl;
}
outFile << std::endl;
}
}
selfBoundFile.close();
outFile.close();
return 0;
}
int main()
{
struct stat buffer;
double HBC = Constants::HBCFmMeV;
std:: array<std::string,5> Skyrmenames09 = {"Ska35s2009.xml","Ska25s2009.xml","SKT109.xml","SKT209.xml","SKT309.xml"};
std:: array<std::string,5> Skyrmenames1 = {"Ska35s201.xml","Ska25s201.xml","SKT11.xml","SKT21.xml","SKT31.xml"};
std:: string eos_dataName_ermal("eos_data_"),eos_dataName_saturation("eos_data_sat_"), eos_dataName;
for (int counter = 0; counter < 5; counter++)
{
eos_dataName=eos_dataName_ermal+Skyrmenames09[counter];
if((stat (eos_dataName.c_str(), &buffer) != 0))
{
EOSSkyrme eos = EOSSkyrme::FromErmalSkyrme(ErmalSkyrmeParameters::ErmalAllSkyrme09[counter]);
EOSSingleNucleus gibbs(eos);
std::ofstream ofs(eos_dataName.c_str());
boost::archive::text_oarchive oa(ofs);
oa << gibbs;
ofs.close();
}
eos_dataName=eos_dataName_ermal+Skyrmenames1[counter];
if((stat (eos_dataName.c_str(), &buffer) != 0))
{
EOSSkyrme eos = EOSSkyrme::FromErmalSkyrme(ErmalSkyrmeParameters::ErmalAllSkyrme1[counter]);
EOSSingleNucleus gibbs(eos);
std::ofstream ofs(eos_dataName.c_str());
boost::archive::text_oarchive oa(ofs);
oa << gibbs;
ofs.close();
}
eos_dataName=eos_dataName_saturation+Skyrmenames09[counter];
if((stat (eos_dataName.c_str(), &buffer) != 0))
{
EOSSkyrme eos = EOSSkyrme::FromSaturation(SaturationSkyrmeParameters::SaturationSkyrme09[counter]);
EOSSingleNucleus gibbs(eos);
std::ofstream ofs(eos_dataName.c_str());
boost::archive::text_oarchive oa(ofs);
oa << gibbs;
ofs.close();
}
eos_dataName=eos_dataName_saturation+Skyrmenames1[counter];
if((stat (eos_dataName.c_str(), &buffer) != 0))
{
EOSSkyrme eos = EOSSkyrme::FromSaturation(SaturationSkyrmeParameters::SaturationSkyrme1[counter]);
EOSSingleNucleus gibbs(eos);
std::ofstream ofs(eos_dataName.c_str());
boost::archive::text_oarchive oa(ofs);
oa << gibbs;
ofs.close();
}
}
//GibbsPhaseConstruct reloadGibbs(eos, false);
//
//std::ifstream ifs("eos_data.xml");
//boost::archive::text_iarchive ia(ifs);
//ia >> reloadGibbs;
//ifs.close();
return 0;
}
