TreeEP::TreeEP( const FactorGraph &fg, const PropertySet &opts ) : JTree(fg, opts("updates",string("HUGIN")), false), _maxdiff(0.0), _iters(0), props(), _Q() {
setProperties( opts );
if( opts.hasKey("tree") ) {
construct( fg, opts.getAs<RootedTree>("tree") );
} else {
if( props.type == Properties::TypeType::ORG || props.type == Properties::TypeType::ALT ) {
// ORG: construct weighted graph with as weights a crude estimate of the
// mutual information between the nodes
// ALT: construct weighted graph with as weights an upper bound on the
// effective interaction strength between pairs of nodes
WeightedGraph<Real> wg;
// in order to get a connected weighted graph, we start
// by connecting every variable to the zero'th variable with weight 0
for( size_t i = 1; i < fg.nrVars(); i++ )
wg[UEdge(i,0)] = 0.0;
for( size_t i = 0; i < fg.nrVars(); i++ ) {
SmallSet<size_t> delta_i = fg.bipGraph().delta1( i, false );
const Var& v_i = fg.var(i);
bforeach( size_t j, delta_i )
if( i < j ) {
const Var& v_j = fg.var(j);
VarSet v_ij( v_i, v_j );
SmallSet<size_t> nb_ij = fg.bipGraph().nb1Set( i ) | fg.bipGraph().nb1Set( j );
Factor piet;
bforeach( size_t I, nb_ij ) {
const VarSet& Ivars = fg.factor(I).vars();
if( props.type == Properties::TypeType::ORG ) {
if( (Ivars == v_i) || (Ivars == v_j) )
piet *= fg.factor(I);
else if( Ivars >> v_ij )
piet *= fg.factor(I).marginal( v_ij );
} else {
if( Ivars >> v_ij )
piet *= fg.factor(I);
}
}
if( props.type == Properties::TypeType::ORG ) {
if( piet.vars() >> v_ij ) {
piet = piet.marginal( v_ij );
Factor pietf = piet.marginal(v_i) * piet.marginal(v_j);
wg[UEdge(i,j)] = dist( piet, pietf, DISTKL );
} else {
// this should never happen...
DAI_ASSERT( 0 == 1 );
wg[UEdge(i,j)] = 0;
}
} else
wg[UEdge(i,j)] = piet.strength(v_i, v_j);
}
void RegionGraph::constructCVM( const FactorGraph &fg, const std::vector<VarSet> &cl, size_t verbose ) {
if( verbose )
cerr << "constructCVM called (" << fg.nrVars() << " vars, " << fg.nrFactors() << " facs, " << cl.size() << " clusters)" << endl;
// Retain only maximal clusters
if( verbose )
cerr << " Constructing ClusterGraph" << endl;
ClusterGraph cg( cl );
if( verbose )
cerr << " Erasing non-maximal clusters" << endl;
cg.eraseNonMaximal();
// Create inner regions - first pass
if( verbose )
cerr << " Creating inner regions (first pass)" << endl;
set<VarSet> betas;
for( size_t alpha = 0; alpha < cg.nrClusters(); alpha++ )
for( size_t alpha2 = alpha; (++alpha2) != cg.nrClusters(); ) {
VarSet intersection = cg.cluster(alpha) & cg.cluster(alpha2);
if( intersection.size() > 0 )
betas.insert( intersection );
}
// Create inner regions - subsequent passes
if( verbose )
cerr << " Creating inner regions (next passes)" << endl;
set<VarSet> new_betas;
do {
new_betas.clear();
for( set<VarSet>::const_iterator gamma = betas.begin(); gamma != betas.end(); gamma++ )
for( set<VarSet>::const_iterator gamma2 = gamma; (++gamma2) != betas.end(); ) {
VarSet intersection = (*gamma) & (*gamma2);
if( (intersection.size() > 0) && (betas.count(intersection) == 0) )
new_betas.insert( intersection );
}
betas.insert(new_betas.begin(), new_betas.end());
} while( new_betas.size() );
// Create inner regions - final phase
if( verbose )
cerr << " Creating inner regions (final phase)" << endl;
vector<Region> irs;
irs.reserve( betas.size() );
for( set<VarSet>::const_iterator beta = betas.begin(); beta != betas.end(); beta++ )
irs.push_back( Region(*beta,0.0) );
// Create edges
if( verbose )
cerr << " Creating edges" << endl;
vector<pair<size_t,size_t> > edges;
for( size_t beta = 0; beta < irs.size(); beta++ )
for( size_t alpha = 0; alpha < cg.nrClusters(); alpha++ )
if( cg.cluster(alpha) >> irs[beta] )
edges.push_back( pair<size_t,size_t>(alpha,beta) );
// Construct region graph
if( verbose )
cerr << " Constructing region graph" << endl;
construct( fg, cg.clusters(), irs, edges );
// Calculate counting numbers
if( verbose )
cerr << " Calculating counting numbers" << endl;
calcCVMCountingNumbers();
if( verbose )
cerr << "Done." << endl;
}
void TestHSMM::test_marginal_cut(const char* filename, size_t ID){
HSMMparam param(filename);
FactorGraph *graph;
JTree *jt;
// Set some constants
size_t maxiter = 10000;
Real tol = 1e-9;
size_t verb = 0;
//window size
size_t W = 20;
size_t start = 0;
// Store the constants in a PropertySet object
PropertySet opts;
opts.set("maxiter",maxiter); // Maximum number of iterations
opts.set("tol",tol); // Tolerance for convergence
opts.set("verbose",verb); // Verbosity (amount of output generated)
Factor O_last;
vector< vector<Real> > all_marginal;
vector<Real> sequence_marginal;
all_marginal.reserve(test_data.size());
cout << "Now we do testing...\n";
for(size_t i=0; i<test_data.size(); i++) {
//allocate memory
sequence_marginal.reserve(test_data[i].size());
//initialize HSMM of ever increasing size up to test_data[i].size()
graph = new FactorGraph();
graph->createHSMMFactorGraph(param.init, param.dist, test_data[i].size());
jt = new JTree(*graph, opts("updates",string("HUGIN"))("heuristic",string("MINWEIGHT")) );
jt->init();
jt->run();
O_last = jt->calcMarginal(graph->var(4));
sequence_marginal.push_back( log(O_last.p().get(test_data[i][0].second)) );
delete jt;
for(size_t k=1; k < test_data[i].size(); k++){
jt = new JTree(*graph, opts("updates",string("HUGIN"))("heuristic",string("MINWEIGHT")) );
//clamp a window of observable variables to their values, except last variable which is not clamped
start = k-W;
if(start < 0) start = 0;
for(size_t j = start; j <= k-1; j++){
jt->clamp(test_data[i][j].first, test_data[i][j].second);
}
jt->init();
jt->run();
//compute p(o_last=c | o_1...o_{last-1})
//this will give us a distribution: {o_last=1, o_last=2, ... o_last=M}
O_last = jt->calcMarginal(graph->var(3*k+4));
//since we have a specific observation at last time step: o_last=c, get its probability:
sequence_marginal.push_back( log(O_last.p().get(test_data[i][k].second)) );
delete jt;
}
cout << "Tested point " << i << " out of " << test_data.size() <<"\n";
all_marginal.push_back(sequence_marginal);
sequence_marginal.clear();
delete graph;
}
cout << "Testing done.\n";
ofstream os;
stringstream result;
result << string("data/HSMMmarginal_test_") << ID << string(".txt");
os.open(result.str().c_str(), ios::trunc);
for(size_t i=0; i<all_marginal.size(); i++){
for(size_t j=0; j<all_marginal[i].size(); j++){
os << all_marginal[i][j]<<" ";
}
os << "\n";
}
}
void TestHSMM::test_loglik(const char* filename, size_t ID, string type, int dummy, int ID2){
//"type" specifies the suffix of the output file "test" or "true"
//read in HSMM parameters from textfile
HSMMparam param(filename, 0);
vector<Real> likelihood_test;
FactorGraph *graph;
JTree *jt;
VarSet X;
// Set some constants
size_t maxiter = 10000;
Real tol = 1e-9;
size_t verb = 0;
// Store the constants in a PropertySet object
PropertySet opts;
opts.set("maxiter",maxiter); // Maximum number of iterations
opts.set("tol",tol); // Tolerance for convergence
opts.set("verbose",verb); // Verbosity (amount of output generated)
cout << "Now we do testing...\n";
for(size_t i=0; i<test_data.size(); i++) {
//initialize HSMM of size equal the number of observations
graph = new FactorGraph();
graph->createHSMMFactorGraph(param.init, param.dist, test_data[i].size(), 0);
jt = new JTree(*graph, opts("updates",string("HUGIN"))("heuristic",string("MINWEIGHT")) );
//clamp the observation variables to their observed values
for(size_t j = 0; j < test_data[i].size(); j++ ){
//cout << "clamping var" << test_data[i][j].first << " to value " << test_data[i][j].second << "\n";
jt->clamp(test_data[i][j].first, test_data[i][j].second);
}
jt->init();
jt->run();
//compute normalized loglikelihood
//likelihood_test.push_back(jt->logZ()/test_data[i].size());
likelihood_test.push_back(jt->logZ());
delete jt;
delete graph;
cout << "Tested point " << i << " out of " << test_data.size() <<"\n";
//cout << "jt->logZ() = " << likelihood_test.at(0) << "\n";
//exit(1);
}
cout << "done.\n";
ofstream os;
stringstream result;
//ID2 is used to wirte test results with fixed number of training iterations
if(ID2 >= 0){
result << string("data/HSMMlikelihood_") << type << string("_") << ID << string("-") << ID2 << string(".txt");
}
else{
result << string("data/HSMMlikelihood_") << type << string("_") << ID << string(".txt");
}
os.open(result.str().c_str(), ios::trunc);
os.unsetf ( std::ios::floatfield );
os.precision(18);
for(size_t i=0; i<likelihood_test.size(); i++){
os << likelihood_test.at(i)<<"\n";
}
}
/// Main function
int main( int argc, char *argv[] ) {
// Variables to store command line options
// Filename of factor graph
string filename;
// Filename for aliases
string aliases;
// Approximate Inference methods to use
vector<string> methods;
// Which marginals to output
MarginalsOutputType marginals;
// Output number of iterations?
bool report_iters = true;
// Output calculation time?
bool report_time = true;
// Define required command line options
po::options_description opts_required("Required options");
opts_required.add_options()
("filename", po::value< string >(&filename), "Filename of factor graph")
("methods", po::value< vector<string> >(&methods)->multitoken(), "DAI methods to perform")
;
// Define allowed command line options
po::options_description opts_optional("Allowed options");
opts_optional.add_options()
("help", "Produce help message")
("aliases", po::value< string >(&aliases), "Filename for aliases")
("marginals", po::value< MarginalsOutputType >(&marginals), "Output marginals? (NONE/VAR/FAC/VARFAC/ALL, default=NONE)")
("report-time", po::value< bool >(&report_time), "Output calculation time (default==1)?")
("report-iters", po::value< bool >(&report_iters), "Output iterations needed (default==1)?")
;
// Define all command line options
po::options_description cmdline_options;
cmdline_options.add(opts_required).add(opts_optional);
// Parse command line
po::variables_map vm;
po::store(po::parse_command_line(argc, argv, cmdline_options), vm);
po::notify(vm);
// Display help message if necessary
if( vm.count("help") || !(vm.count("filename") && vm.count("methods")) ) {
cout << "This program is part of libDAI - http://www.libdai.org/" << endl << endl;
cout << "Usage: ./testdai --filename <filename.fg> --methods <method1> [<method2> <method3> ...]" << endl << endl;
cout << "Reads factor graph <filename.fg> and performs the approximate inference algorithms" << endl;
cout << "<method*>, reporting for each method:" << endl;
cout << " o the calculation time needed, in seconds (if report-time == 1);" << endl;
cout << " o the number of iterations needed (if report-iters == 1);" << endl;
cout << " o the maximum (over all variables) total variation error in the variable marginals;" << endl;
cout << " o the average (over all variables) total variation error in the variable marginals;" << endl;
cout << " o the maximum (over all factors) total variation error in the factor marginals;" << endl;
cout << " o the average (over all factors) total variation error in the factor marginals;" << endl;
cout << " o the error (difference) of the logarithm of the partition sums;" << endl << endl;
cout << "All errors are calculated by comparing the results of the current method with" << endl;
cout << "the results of the first method (the base method). If marginals==VAR, additional" << endl;
cout << "output consists of the variable marginals, if marginals==FAC, the factor marginals" << endl;
cout << "if marginals==VARFAC, both variable and factor marginals, and if marginals==ALL, all" << endl;
cout << "marginals calculated by the method are reported." << endl << endl;
cout << "<method*> should be a list of one or more methods, seperated by spaces, in the format:" << endl << endl;
cout << " name[key1=val1,key2=val2,key3=val3,...,keyn=valn]" << endl << endl;
cout << "where name should be the name of an algorithm in libDAI (or an alias, if an alias" << endl;
cout << "filename is provided), followed by a list of properties (surrounded by rectangular" << endl;
cout << "brackets), where each property consists of a key=value pair and the properties are" << endl;
cout << "seperated by commas. If an alias file is specified, alias substitution is performed." << endl;
cout << "This is done by looking up the name in the alias file and substituting the alias" << endl;
cout << "by its corresponding method as defined in the alias file. Properties are parsed from" << endl;
cout << "left to right, so if a property occurs repeatedly, the right-most value is used." << endl << endl;
cout << opts_required << opts_optional << endl;
#ifdef DAI_DEBUG
cout << "Note: this is a debugging build of libDAI." << endl << endl;
#endif
cout << "Example: ./testdai --filename testfast.fg --aliases aliases.conf --methods JTREE_HUGIN BP_SEQFIX BP_PARALL[maxiter=5]" << endl;
return 1;
}
try {
// Read aliases
map<string,string> Aliases;
if( !aliases.empty() )
Aliases = readAliasesFile( aliases );
// Read factor graph
FactorGraph fg;
fg.ReadFromFile( filename.c_str() );
// Declare variables used for storing variable factor marginals and log partition sum of base method
vector<Factor> varMarginals0;
vector<Factor> facMarginals0;
Real logZ0 = 0.0;
// Output header
cout.setf( ios_base::scientific );
cout.precision( 3 );
cout << "# " << filename << endl;
cout.width( 39 );
cout << left << "# METHOD" << "\t";
if( report_time )
cout << right << "SECONDS " << "\t";
if( report_iters )
cout << "ITERS" << "\t";
cout << "MAX VAR ERR" << "\t";
cout << "AVG VAR ERR" << "\t";
cout << "MAX FAC ERR" << "\t";
cout << "AVG FAC ERR" << "\t";
cout << "LOGZ ERROR" << "\t";
cout << "MAXDIFF" << "\t";
cout << endl;
// For each method...
for( size_t m = 0; m < methods.size(); m++ ) {
// Parse method
pair<string, PropertySet> meth = parseNameProperties( methods[m], Aliases );
// Construct object for running the method
TestDAI testdai(fg, meth.first, meth.second );
// Run the method
testdai.doDAI();
// For the base method, store its variable marginals and logarithm of the partition sum
if( m == 0 ) {
varMarginals0 = testdai.varMarginals;
facMarginals0 = testdai.facMarginals;
logZ0 = testdai.logZ;
}
// Calculate errors relative to base method
testdai.calcErrors( varMarginals0, facMarginals0 );
// Output method name
cout.width( 39 );
cout << left << methods[m] << "\t";
// Output calculation time, if requested
if( report_time )
cout << right << testdai.time << "\t";
// Output number of iterations, if requested
if( report_iters ) {
if( testdai.has_iters ) {
cout << testdai.iters << "\t";
} else {
cout << "N/A \t";
}
}
// If this is not the base method
if( m > 0 ) {
cout.setf( ios_base::scientific );
cout.precision( 3 );
// Output maximum error in variable marginals
Real mev = clipReal( testdai.maxVarErr(), 1e-9 );
cout << mev << "\t";
// Output average error in variable marginals
Real aev = clipReal( testdai.avgVarErr(), 1e-9 );
cout << aev << "\t";
// Output maximum error in factor marginals
Real mef = clipReal( testdai.maxFacErr(), 1e-9 );
if( mef == INFINITY )
cout << "N/A \t";
else
cout << mef << "\t";
// Output average error in factor marginals
Real aef = clipReal( testdai.avgFacErr(), 1e-9 );
if( aef == INFINITY )
cout << "N/A \t";
else
cout << aef << "\t";
// Output error in log partition sum
if( testdai.has_logZ ) {
cout.setf( ios::showpos );
Real le = clipReal( testdai.logZ - logZ0, 1e-9 );
cout << le << "\t";
cout.unsetf( ios::showpos );
} else
cout << "N/A \t";
// Output maximum difference in last iteration
if( testdai.has_maxdiff ) {
Real md = clipReal( testdai.maxdiff, 1e-9 );
if( dai::isnan( mev ) )
md = mev;
if( dai::isnan( aev ) )
md = aev;
if( md == INFINITY )
md = 1.0;
cout << md << "\t";
} else
cout << "N/A \t";
}
cout << endl;
// Output marginals, if requested
if( marginals == MarginalsOutputType::VAR || marginals == MarginalsOutputType::VARFAC )
for( size_t i = 0; i < testdai.varMarginals.size(); i++ )
cout << "# " << testdai.varMarginals[i] << endl;
if( marginals == MarginalsOutputType::FAC || marginals == MarginalsOutputType::VARFAC )
for( size_t I = 0; I < testdai.facMarginals.size(); I++ )
cout << "# " << testdai.facMarginals[I] << endl;
if( marginals == MarginalsOutputType::ALL )
for( size_t I = 0; I < testdai.allMarginals.size(); I++ )
cout << "# " << testdai.allMarginals[I] << endl;
}
return 0;
} catch( string &s ) {
// Abort with error message
cerr << "Exception: " << s << endl;
return 2;
}
}
void MaximumCompositeLikelihood::SetupTrainingData(
const std::vector<labeled_instance_type>& training_data,
const std::vector<InferenceMethod*> inference_methods) {
assert(comp_training_data.size() == 0);
assert(comp_inference_methods.size() == 0);
assert(inference_methods.size() == training_data.size());
// Number of times each component will be covered
unsigned int cover_count = 1;
assert(decomp >= -1);
if (decomp == DecomposePseudolikelihood) {
cover_count = 1;
} else if (decomp > 0) {
cover_count = decomp;
}
// Produce composite factor graphs
boost::timer decomp_timer;
int training_data_size = static_cast<int>(training_data.size());
fg_cc_var_label.resize(cover_count * training_data_size);
fg_cc_count.resize(cover_count * training_data_size);
fg_orig_index.resize(cover_count * training_data_size);
std::fill(fg_cc_count.begin(), fg_cc_count.end(), 0);
unsigned int cn = 0;
for (int n = 0; n < training_data_size; ++n) {
FactorGraph* fg = training_data[n].first;
size_t var_count = fg->Cardinalities().size();
// Get observation
const FactorGraphObservation* obs = training_data[n].second;
// Obtain one or more decomposition(s)
for (unsigned int cover_iter = 0; cover_iter < cover_count;
++cover_iter) {
VAcyclicDecomposition vac(fg);
std::vector<bool> factor_is_removed;
if (decomp == DecomposePseudolikelihood) {
factor_is_removed.resize(fg->Factors().size());
std::fill(factor_is_removed.begin(),
factor_is_removed.end(), true);
} else {
std::vector<double> factor_weight(fg->Factors().size(), 0.0);
if (decomp == DecomposeUniform) {
// Use constant weights
std::fill(factor_weight.begin(), factor_weight.end(), 1.0);
} else {
// Use uniform random weights
boost::uniform_real<double> uniform_dist(0.0, 1.0);
boost::variate_generator<boost::mt19937&,
boost::uniform_real<double> >
rgen(RandomSource::GlobalRandomSampler(), uniform_dist);
for (unsigned int fi = 0; fi < factor_weight.size(); ++fi)
factor_weight[fi] = rgen();
}
vac.ComputeDecompositionSP(factor_weight, factor_is_removed);
}
// Shatter factor graph into trees
fg_cc_count[cn] += FactorGraphStructurizer::ConnectedComponents(
fg, factor_is_removed, fg_cc_var_label[cn]);
#if 0
std::cout << "MCL, instance " << n << " decomposed into " << cc_count
<< " components" << std::endl;
#endif
// Add each component as separate factor graph
for (unsigned int ci = 0; ci < fg_cc_count[cn]; ++ci) {
std::vector<unsigned int> cond_var_set;
cond_var_set.reserve(var_count);
// Add all variables not in this component to the conditioning set
for (size_t vi = 0; vi < var_count; ++vi) {
if (fg_cc_var_label[cn][vi] != ci)
cond_var_set.push_back(static_cast<unsigned int>(vi));
}
AddTrainingComponentCond(fg, obs, inference_methods[n],
cond_var_set);
}
fg_orig_index[cn] = n;
cn += 1;
}
}
std::cout << "MCL, decomposed " << training_data.size() << " instances "
<< "into " << comp_training_data.size() << " instances "
<< (decomp == DecomposeUniform ? "(uniform)" : "(randomized)")
<< " in " << decomp_timer.elapsed() << "s." << std::endl;
// Initialize MLE training data from created components
SetupMLETrainingData();
}
VarIds
readQueryAndEvidence (
FactorGraph& fg,
int argc,
const char* argv[],
int start)
{
VarIds queryIds;
for (int i = start; i < argc; i++) {
const string& arg = argv[i];
if (arg.find ('=') == std::string::npos) {
if (Util::isInteger (arg) == false) {
cerr << "error: `" << arg << "' " ;
cerr << "is not a variable id" ;
cerr << endl;
exit (0);
}
VarId vid = Util::stringToUnsigned (arg);
VarNode* queryVar = fg.getVarNode (vid);
if (queryVar == false) {
cerr << "error: unknow variable with id " ;
cerr << "`" << vid << "'" << endl;
exit (0);
}
queryIds.push_back (vid);
} else {
size_t pos = arg.find ('=');
string leftArg = arg.substr (0, pos);
string rightArg = arg.substr (pos + 1);
if (leftArg.empty()) {
cerr << "error: missing left argument" << endl;
cerr << USAGE << endl;
exit (0);
}
if (Util::isInteger (leftArg) == false) {
cerr << "error: `" << leftArg << "' " ;
cerr << "is not a variable id" << endl ;
exit (0);
continue;
}
VarId vid = Util::stringToUnsigned (leftArg);
VarNode* observedVar = fg.getVarNode (vid);
if (observedVar == false) {
cerr << "error: unknow variable with id " ;
cerr << "`" << vid << "'" << endl;
exit (0);
}
if (rightArg.empty()) {
cerr << "error: missing right argument" << endl;
cerr << USAGE << endl;
exit (0);
}
if (Util::isInteger (rightArg) == false) {
cerr << "error: `" << rightArg << "' " ;
cerr << "is not a state index" << endl ;
exit (0);
}
unsigned stateIdx = Util::stringToUnsigned (rightArg);
if (observedVar->isValidState (stateIdx) == false) {
cerr << "error: `" << stateIdx << "' " ;
cerr << "is not a valid state index for variable with id " ;
cerr << "`" << vid << "'" << endl;
exit (0);
}
observedVar->setEvidence (stateIdx);
}
}
return queryIds;
}
int main( int argc, char *argv[] ) {
if( argc != 3 ) {
cout << "Usage: " << argv[0] << " <in.fg> <tw>" << endl << endl;
cout << "Reports some characteristics of the .fg network." << endl;
cout << "Also calculates treewidth (which may take some time) unless <tw> == 0." << endl;
return 1;
} else {
// Read factorgraph
FactorGraph fg;
char *infile = argv[1];
int calc_tw = atoi(argv[2]);
fg.ReadFromFile( infile );
cout << "Number of variables: " << fg.nrVars() << endl;
cout << "Number of factors: " << fg.nrFactors() << endl;
cout << "Connected: " << fg.isConnected() << endl;
cout << "Tree: " << fg.isTree() << endl;
cout << "Has short loops: " << hasShortLoops(fg.factors()) << endl;
cout << "Has negatives: " << hasNegatives(fg.factors()) << endl;
cout << "Binary variables? " << fg.isBinary() << endl;
cout << "Pairwise interactions? " << fg.isPairwise() << endl;
if( calc_tw ) {
std::pair<size_t,size_t> tw = treewidth(fg);
cout << "Treewidth: " << tw.first << endl;
cout << "Largest cluster for JTree has " << tw.second << " states " << endl;
}
double stsp = 1.0;
for( size_t i = 0; i < fg.nrVars(); i++ )
stsp *= fg.var(i).states();
cout << "Total state space: " << stsp << endl;
double cavsum_lcbp = 0.0;
double cavsum_lcbp2 = 0.0;
size_t max_Delta_size = 0;
map<size_t,size_t> cavsizes;
for( size_t i = 0; i < fg.nrVars(); i++ ) {
VarSet di = fg.delta(i);
if( cavsizes.count(di.size()) )
cavsizes[di.size()]++;
else
cavsizes[di.size()] = 1;
size_t Ds = fg.Delta(i).nrStates();
if( Ds > max_Delta_size )
max_Delta_size = Ds;
cavsum_lcbp += di.nrStates();
for( VarSet::const_iterator j = di.begin(); j != di.end(); j++ )
cavsum_lcbp2 += j->states();
}
cout << "Maximum pancake has " << max_Delta_size << " states" << endl;
cout << "LCBP with full cavities needs " << cavsum_lcbp << " BP runs" << endl;
cout << "LCBP with only pairinteractions needs " << cavsum_lcbp2 << " BP runs" << endl;
cout << "Cavity sizes: ";
for( map<size_t,size_t>::const_iterator it = cavsizes.begin(); it != cavsizes.end(); it++ )
cout << it->first << "(" << it->second << ") ";
cout << endl;
cout << "Type: " << (fg.isPairwise() ? "pairwise" : "higher order") << " interactions, " << (fg.isBinary() ? "binary" : "nonbinary") << " variables" << endl;
if( fg.isPairwise() ) {
bool girth_reached = false;
size_t loopdepth;
for( loopdepth = 2; loopdepth <= fg.nrVars() && !girth_reached; loopdepth++ ) {
size_t nr_loops = countLoops( fg, loopdepth );
cout << "Loops up to " << loopdepth << " variables: " << nr_loops << endl;
if( nr_loops > 0 )
girth_reached = true;
}
if( girth_reached )
cout << "Girth: " << loopdepth-1 << endl;
else
cout << "Girth: infinity" << endl;
}
return 0;
}
}
int main( int argc, char *argv[] ) {
try {
size_t N, K, k, d, j, n1, n2, n3;
size_t prime;
size_t seed;
Real beta, sigma_w, sigma_th, noise, mean_w, mean_th;
string type;
size_t states = 2;
// Declare the supported options.
po::options_description desc("Allowed options");
desc.add_options()
("help", "produce help message")
("type", po::value<string>(&type), "factor graph type:\n\t'full', 'grid', 'grid_torus', 'dreg', 'loop', 'tree', 'hoi', 'ldpc_random', 'ldpc_group', 'ldpc_small', 'potts3d'")
("seed", po::value<size_t>(&seed), "random number seed (tries to read from /dev/urandom if not specified)")
("N", po::value<size_t>(&N), "number of variables (not for type=='ldpc_small')")
("n1", po::value<size_t>(&n1), "width of 3D grid (only for type=='potts3d')")
("n2", po::value<size_t>(&n2), "height of 3D grid (only for type=='potts3d')")
("n3", po::value<size_t>(&n3), "length of 3D grid (only for type=='potts3d')")
("K", po::value<size_t>(&K), "number of factors\n\t(only for type=='hoi' and 'type=='ldpc_{random,group}')")
("k", po::value<size_t>(&k), "number of variables per factor\n\t(only for type=='hoi' and type=='ldpc_{random,group}')")
("d", po::value<size_t>(&d), "variable connectivity\n\t(only for type=='dreg')")
("j", po::value<size_t>(&j), "number of parity checks per bit\n\t(only for type=='ldpc_{random,group}')")
("prime", po::value<size_t>(&prime), "prime number for construction of LDPC code\n\t(only for type=='ldpc_group')")
("beta", po::value<Real>(&beta), "stddev of log-factor entries\n\t(only for type=='hoi', 'potts3d', 'grid' if states>2)")
("mean_w", po::value<Real>(&mean_w), "mean of pairwise interactions w_{ij}\n\t(not for type=='hoi', 'ldpc_*', 'potts3d')")
("mean_th", po::value<Real>(&mean_th), "mean of singleton interactions th_i\n\t(not for type=='hoi', 'ldpc_*', 'potts3d')")
("sigma_w", po::value<Real>(&sigma_w), "stddev of pairwise interactions w_{ij}\n\t(not for type=='hoi', 'ldpc_*', 'potts3d')")
("sigma_th", po::value<Real>(&sigma_th), "stddev of singleton interactions th_i\n\t(not for type=='hoi', 'ldpc_*', 'potts3d'")
("noise", po::value<Real>(&noise), "bitflip probability for binary symmetric channel (only for type=='ldpc')")
("states", po::value<size_t>(&states), "number of states of each variable (should be 2 for all but type=='grid', 'grid_torus', 'loop', 'potts3d')")
;
po::variables_map vm;
po::store(po::parse_command_line(argc, argv, desc), vm);
po::notify(vm);
if( vm.count("help") || !vm.count("type") ) {
if( vm.count("type") ) {
if( type == FULL_TYPE ) {
cout << "Creates fully connected pairwise graphical model of <N> binary variables;" << endl;
} else if( type == GRID_TYPE ) {
cout << "Creates (non-periodic) 2D Ising grid of (approx.) <N> variables (which need not be binary);" << endl;
} else if( type == GRID_TORUS_TYPE ) {
cout << "Creates periodic 2D Ising grid of (approx.) <N> variables (which need not be binary);" << endl;
} else if( type == DREG_TYPE ) {
cout << "Creates random d-regular graph of <N> binary variables with uniform degree <d>" << endl;
cout << "(where <d><N> should be even);" << endl;
} else if( type == LOOP_TYPE ) {
cout << "Creates a pairwise graphical model consisting of a single loop of" << endl;
cout << "<N> variables (which need not be binary);" << endl;
} else if( type == TREE_TYPE ) {
cout << "Creates a pairwise, connected graphical model without cycles (i.e., a tree)" << endl;
cout << "of <N> binary variables;" << endl;
} else if( type == HOI_TYPE ) {
cout << "Creates a random factor graph of <N> binary variables and" << endl;
cout << "<K> factors, each factor being an interaction of <k> variables." << endl;
cout << "The entries of the factors are exponentials of i.i.d. Gaussian" << endl;
cout << "variables with mean 0 and standard deviation <beta>." << endl;
} else if( type == LDPC_RANDOM_TYPE ) {
cout << "Simulates LDPC decoding problem, using a LDPC code of <N> bits and <K> parity" << endl;
cout << "checks, with <k> bits per check and <j> checks per bit, transmitted on a binary" << endl;
cout << "symmetric channel with probability <noise> of flipping a bit. The transmitted" << endl;
cout << "codeword has all bits set to zero. The LDPC code is randomly generated." << endl;
} else if( type == LDPC_GROUP_TYPE ) {
cout << "Simulates LDPC decoding problem, using a LDPC code of <N> bits and <K> parity" << endl;
cout << "checks, with <k> bits per check and <j> checks per bit, transmitted on a binary" << endl;
cout << "symmetric channel with probability <noise> of flipping a bit. The transmitted" << endl;
cout << "codeword has all bits set to zero. The LDPC code is constructed (using group" << endl;
cout << "theory) using a parameter <prime>; <j> and <k> should both be divisors of <prime>-1." << endl;
} else if( type == LDPC_SMALL_TYPE ) {
cout << "Simulates LDPC decoding problem, using a LDPC code of 4 bits and 4 parity" << endl;
cout << "checks, with 3 bits per check and 3 checks per bit, transmitted on a binary" << endl;
cout << "symmetric channel with probability <noise> of flipping a bit. The transmitted" << endl;
cout << "codeword has all bits set to zero. The LDPC code is fixed." << endl;
} else if( type == POTTS3D_TYPE ) {
cout << "Builds 3D Potts model of size <n1>x<n2>x<n3> with nearest-neighbour Potts" << endl;
cout << "interactions with <states> states and inverse temperature <beta>." << endl;
} else
cerr << "Unknown type (should be one of 'full', 'grid', 'grid_torus', 'dreg', 'loop', 'tree', 'hoi', 'ldpc_random', 'ldpc_group', 'ldpc_small', 'potts3d')" << endl;
if( type == FULL_TYPE || type == GRID_TYPE || type == GRID_TORUS_TYPE || type == DREG_TYPE || type == LOOP_TYPE || type == TREE_TYPE ) {
if( type == GRID_TYPE || type == GRID_TORUS_TYPE || type == LOOP_TYPE ) {
cout << "if <states> > 2: factor entries are exponents of Gaussians with mean 0 and standard deviation beta; otherwise," << endl;
}
cout << "singleton interactions are Gaussian with mean <mean_th> and standard" << endl;
cout << "deviation <sigma_th>; pairwise interactions are Gaussian with mean" << endl;
cout << "<mean_w> and standard deviation <sigma_w>." << endl;
}
}
cout << endl << desc << endl;
return 1;
}
if( !vm.count("states") )
states = 2;
if( !vm.count("seed") ) {
ifstream infile;
bool success;
infile.open( "/dev/urandom" );
success = infile.is_open();
if( success ) {
infile.read( (char *)&seed, sizeof(size_t) / sizeof(char) );
success = infile.good();
infile.close();
}
if( !success )
throw "Please specify random number seed.";
}
rnd_seed( seed );
FactorGraph fg;
cout << "# Factor graph made by " << argv[0] << endl;
cout << "# type = " << type << endl;
if( type == FULL_TYPE ) {
if( !vm.count("N") || !vm.count("mean_w") || !vm.count("mean_th") || !vm.count("sigma_w") || !vm.count("sigma_th") )
throw "Please specify all required arguments";
MakeFullFG( N, mean_w, mean_th, sigma_w, sigma_th, fg );
cout << "# N = " << N << endl;
cout << "# mean_w = " << mean_w << endl;
cout << "# mean_th = " << mean_th << endl;
cout << "# sigma_w = " << sigma_w << endl;
cout << "# sigma_th = " << sigma_th << endl;
} else if( type == GRID_TYPE || type == GRID_TORUS_TYPE ) {
#define NEED_ARG(name, desc) do { if(!vm.count(name)) throw "Please specify " desc " with --" name; } while(0);
if( states > 2 ) {
NEED_ARG("N", "number of nodes");
NEED_ARG("beta", "stddev of log-factor entries");
} else {
NEED_ARG("N", "number of nodes");
NEED_ARG("mean_w", "mean of pairwise interactions");
NEED_ARG("mean_th", "mean of singleton interactions");
NEED_ARG("sigma_w", "stddev of pairwise interactions");
NEED_ARG("sigma_th", "stddev of singleton interactions");
}
size_t n = (size_t)sqrt((long double)N);
N = n * n;
bool periodic = false;
if( type == GRID_TYPE )
periodic = false;
else
periodic = true;
if( states > 2 )
MakeGridNonbinaryFG( periodic, n, states, beta, fg );
else
MakeGridFG( periodic, n, mean_w, mean_th, sigma_w, sigma_th, fg );
cout << "# n = " << n << endl;
cout << "# N = " << N << endl;
if( states > 2 )
cout << "# beta = " << beta << endl;
else {
cout << "# mean_w = " << mean_w << endl;
cout << "# mean_th = " << mean_th << endl;
cout << "# sigma_w = " << sigma_w << endl;
cout << "# sigma_th = " << sigma_th << endl;
}
} else if( type == DREG_TYPE ) {
if( !vm.count("N") || !vm.count("mean_w") || !vm.count("mean_th") || !vm.count("sigma_w") || !vm.count("sigma_th") || !vm.count("d") )
throw "Please specify all required arguments";
MakeDRegFG( N, d, mean_w, mean_th, sigma_w, sigma_th, fg );
cout << "# N = " << N << endl;
cout << "# d = " << d << endl;
cout << "# mean_w = " << mean_w << endl;
cout << "# mean_th = " << mean_th << endl;
cout << "# sigma_w = " << sigma_w << endl;
cout << "# sigma_th = " << sigma_th << endl;
} else if( type == LOOP_TYPE ) {
if( states > 2 ) {
if( !vm.count("N") || !vm.count("beta") )
throw "Please specify all required arguments";
} else {
if( !vm.count("N") || !vm.count("mean_w") || !vm.count("mean_th") || !vm.count("sigma_w") || !vm.count("sigma_th") )
throw "Please specify all required arguments";
}
if( states > 2 )
MakeLoopNonbinaryFG( N, states, beta, fg );
else
MakeLoopFG( N, mean_w, mean_th, sigma_w, sigma_th, fg );
cout << "# N = " << N << endl;
if( states > 2 )
cout << "# beta = " << beta << endl;
else {
cout << "# mean_w = " << mean_w << endl;
cout << "# mean_th = " << mean_th << endl;
cout << "# sigma_w = " << sigma_w << endl;
cout << "# sigma_th = " << sigma_th << endl;
}
} else if( type == TREE_TYPE ) {
if( !vm.count("N") || !vm.count("mean_w") || !vm.count("mean_th") || !vm.count("sigma_w") || !vm.count("sigma_th") )
throw "Please specify all required arguments";
MakeTreeFG( N, mean_w, mean_th, sigma_w, sigma_th, fg );
cout << "# N = " << N << endl;
cout << "# mean_w = " << mean_w << endl;
cout << "# mean_th = " << mean_th << endl;
cout << "# sigma_w = " << sigma_w << endl;
cout << "# sigma_th = " << sigma_th << endl;
} else if( type == HOI_TYPE ) {
if( !vm.count("N") || !vm.count("K") || !vm.count("k") || !vm.count("beta") )
throw "Please specify all required arguments";
do {
MakeHOIFG( N, K, k, beta, fg );
} while( !fg.isConnected() );
cout << "# N = " << N << endl;
cout << "# K = " << K << endl;
cout << "# k = " << k << endl;
cout << "# beta = " << beta << endl;
} else if( type == LDPC_RANDOM_TYPE || type == LDPC_GROUP_TYPE || type == LDPC_SMALL_TYPE ) {
if( !vm.count("noise") )
throw "Please specify all required arguments";
if( type == LDPC_RANDOM_TYPE ) {
if( !vm.count("N") || !vm.count("K") || !vm.count("j") || !vm.count("k") )
throw "Please specify all required arguments";
if( N * j != K * k )
throw "Parameters should satisfy N * j == K * k";
} else if( type == LDPC_GROUP_TYPE ) {
if( !vm.count("prime") || !vm.count("j") || !vm.count("k") )
throw "Please specify all required arguments";
if( !isPrime(prime) )
throw "Parameter <prime> should be prime";
if( !((prime-1) % j == 0 ) )
throw "Parameters should satisfy (prime-1) % j == 0";
if( !((prime-1) % k == 0 ) )
throw "Parameters should satisfy (prime-1) % k == 0";
N = prime * k;
K = prime * j;
} else if( type == LDPC_SMALL_TYPE ) {
N = 4;
K = 4;
j = 3;
k = 3;
}
cout << "# N = " << N << endl;
cout << "# K = " << K << endl;
cout << "# j = " << j << endl;
cout << "# k = " << k << endl;
if( type == LDPC_GROUP_TYPE )
cout << "# prime = " << prime << endl;
cout << "# noise = " << noise << endl;
// p = 31, j = 3, k = 5
// p = 37, j = 3, k = 4
// p = 7 , j = 2, k = 3
// p = 29, j = 2, k = 4
// Construct likelihood and paritycheck factors
Real likelihood[4] = {1.0 - noise, noise, noise, 1.0 - noise};
Real *paritycheck = new Real[1 << k];
MakeParityCheck(paritycheck, k, 0.0);
// Create LDPC structure
BipartiteGraph ldpcG;
bool regular;
do {
if( type == LDPC_GROUP_TYPE )
ldpcG = CreateGroupStructuredLDPCGraph( prime, j, k );
else if( type == LDPC_RANDOM_TYPE )
ldpcG = CreateRandomBipartiteGraph( N, K, j, k );
else if( type == LDPC_SMALL_TYPE )
ldpcG = CreateSmallLDPCGraph();
regular = true;
for( size_t i = 0; i < N; i++ )
if( ldpcG.nb1(i).size() != j )
regular = false;
for( size_t I = 0; I < K; I++ )
if( ldpcG.nb2(I).size() != k )
regular = false;
} while( !regular && !ldpcG.isConnected() );
// Convert to FactorGraph
vector<Factor> factors;
for( size_t I = 0; I < K; I++ ) {
VarSet vs;
for( size_t _i = 0; _i < k; _i++ ) {
size_t i = ldpcG.nb2(I)[_i];
vs |= Var( i, 2 );
}
factors.push_back( Factor( vs, paritycheck ) );
}
delete paritycheck;
// Generate noise vector
vector<char> noisebits(N,0);
size_t bitflips = 0;
for( size_t i = 0; i < N; i++ ) {
if( rnd_uniform() < noise ) {
noisebits[i] = 1;
bitflips++;
}
}
cout << "# bitflips = " << bitflips << endl;
// Simulate transmission of all-zero codeword
vector<char> input(N,0);
vector<char> output(N,0);
for( size_t i = 0; i < N; i++ )
output[i] = (input[i] + noisebits[i]) & 1;
// Add likelihoods
for( size_t i = 0; i < N; i++ )
factors.push_back( Factor(Var(i,2), likelihood + output[i]*2) );
// Construct Factor Graph
fg = FactorGraph( factors );
} else if( type == POTTS3D_TYPE ) {
if( !vm.count("n1") || !vm.count("n2") || !vm.count("n3") || !vm.count("beta") || !vm.count("states") )
throw "Please specify all required arguments";
Make3DPotts( n1, n2, n3, states, beta, fg );
cout << "# N = " << n1*n2*n3 << endl;
cout << "# n1 = " << n1 << endl;
cout << "# n2 = " << n2 << endl;
cout << "# n3 = " << n3 << endl;
cout << "# beta = " << beta << endl;
cout << "# states = " << states << endl;
} else {
throw "Invalid type";
}
cout << "# seed = " << seed << endl;
cout << fg;
} catch( const char *e ) {
cerr << "Error: " << e << endl;
return 1;
}
return 0;
}
