int main(int argc, char* argv[]) {
Env = TEnv(argc, argv, TNotify::StdNotify);
Env.PrepArgs(TStr::Fmt("Rolx. build: %s, %s. Time: %s", __TIME__, __DATE__, TExeTm::GetCurTm()));
TExeTm ExeTm;
Try
const TStr InFNm = Env.GetIfArgPrefixStr("-i:", "graph.txt", "Input graph (one edge per line, tab/space separated)");
const TStr OutFNm = Env.GetIfArgPrefixStr("-o:", "roles.txt", "Output file name prefix");
const int MinRoles = Env.GetIfArgPrefixInt("-l:", 2, "Lower bound of the number of roles");
const int MaxRoles = Env.GetIfArgPrefixInt("-u:", 3, "Upper bound of the number of roles");
double Threshold = 1e-6;
if (MinRoles > MaxRoles || MinRoles < 2) {
printf("min roles and max roles should be integer and\n");
printf("2 <= min roles <= max roles\n");
exit(EXIT_SUCCESS);
}
printf("loading file...\n");
PNGraph Graph = TSnap::LoadEdgeList<PNGraph>(InFNm, 0, 1);
printf("extracting features...\n");
TIntFtrH Features = ExtractFeatures(Graph);
TIntIntH NodeIdMtxIdH = CreateNodeIdMtxIdxHash(Features);
TFltVV V = ConvertFeatureToMatrix(Features, NodeIdMtxIdH);
//printf("saving features...\n");
//FPrintMatrix(V, "v.txt");
printf("feature matrix is saved in v.txt\n");
TFlt MnError = TFlt::Mx;
TFltVV FinalG, FinalF;
int NumRoles = -1;
for (int r = MinRoles; r <= MaxRoles; ++r) {
TFltVV G, F;
printf("factorizing for %d roles...\n", r);
CalcNonNegativeFactorization(V, r, G, F, Threshold);
//FPrintMatrix(G, "g.txt");
//FPrintMatrix(F, "f.txt");
TFlt Error = CalcDescriptionLength(V, G, F);
if (Error < MnError) {
MnError = Error;
FinalG = G;
FinalF = F;
NumRoles = r;
}
}
//FPrintMatrix(FinalG, "final_g.txt");
//FPrintMatrix(FinalF, "final_f.txt");
printf("using %d roles, min error: %f\n", NumRoles, MnError());
TIntIntH Roles = FindRoles(FinalG, NodeIdMtxIdH);
FPrintRoles(Roles, OutFNm);
//PlotRoles(Graph, Roles);
Catch
printf("\nrun time: %s (%s)\n", ExeTm.GetTmStr(), TSecTm::GetCurTm().GetTmStr().CStr());
return 0;
}
void NetworkTemporal3BCPNN::TrainLayer(const vector<vector<float> >& trainingData, PopulationColumns* inputLayer, StructureMIMDSVQ* structure, int iterationsCorrs, int iterationsMDS, int iterationsVQ, int iterationsFeatures)
{
// Training phase
int nrTrainImages = trainingData.size();
structure->MDSHypercolumns()->SwitchOnOff(false);
structure->MDS()->SwitchOnOff(false);
structure->VQ()->SwitchOnOff(false);
structure->CSLLearn()->SwitchOnOff(false);
structure->GetLayer(1)->SwitchOnOff(false);
structure->SetRecording(false);
// Semi-sequential version
// 1. Training phase
// 1A. Patches creation
structure->CSLLearn()->SetMaxPatterns(nrTrainImages);
structure->CSLLearn()->SetEta(0.001);
// turn of response in 2nd layer during initial training phase for speed
structure->GetLayer(1)->SwitchOnOff(false);
int j=0;
while(j<iterationsCorrs) {
//if(j==(int)iterationsPatches*0.8)
// break;
//for(int i=0;i<trainingData.size();i++) {
ComputeCorrelation(trainingData,inputLayer,structure,j);
j++;
}
j=0;
while(j<iterationsMDS) {
//if(j==(int)iterationsPatches*0.9)
// break;
//for(int i=0;i<trainingData.size();i++) {
if(!ComputeMDS(trainingData,inputLayer,structure,j))
break;
j++;
}
structure->GetLayer(1)->SwitchOnOff(true);
j=0;
while(j<iterationsVQ) {
//for(int i=0;i<trainingData.size();i++) {
if(m_verbose && this->MPIGetNodeId() == 0)
cout << "DataPoint (patches): " << j <<endl;
ComputeVQ(trainingData,inputLayer,structure,j);
j++;
}
structure->CSLLearn()->SetMaxPatterns(nrTrainImages);
j=0;
while(j<iterationsFeatures) {
//for(int i=0;i<trainingData.size();i++) {
if(m_verbose && this->MPIGetNodeId() == 0)
cout << "DataPoint (features): " << j <<endl;
if(!ExtractFeatures(trainingData,inputLayer,structure))
break;
//}
j++;
}
this->RecordAll();
}
