/**
* Return an instance of the bmrminnersolver based on user's argument in the
* configuration file.
*
*/
static CBMRMInnerSolver* GetBMRMInnerSolver(CModel &model, double lambda)
{
CBMRMInnerSolver* innerSolver = 0;
Configuration &config = Configuration::GetInstance();
std::string innerSolverType = "";
if(config.IsSet("BMRM.innerSolverType"))
innerSolverType = config.GetString("BMRM.innerSolverType");
else
throw CBMRMException("No BMRM inner solver specified", "CBMRMInnerSolverFactory::GetBMRMInnerSolver()");
// select the innersolver specified by user (in configuration file)
if(innerSolverType == "L2N2_DaiFletcherPGM")
{
innerSolver = new CL2N2_DaiFletcherPGM(lambda);
}
else if(innerSolverType == "L2N2_prLOQO")
{
innerSolver = new CL2N2_prLOQO(lambda);
}
else if(innerSolverType == "L2N2_LineSearch")
{
innerSolver = new CL2N2_LineSearch(lambda);
}
#ifdef HAVE_L1N1_INNER_SOLVER
else if(innerSolverType == "L1N1_Clp")
{
int wLength = model.GetW().Length();
innerSolver = new CL1N1_Clp(lambda, wLength);
}
#endif
else if(innerSolverType == "L2N2_qld")
{
innerSolver = new CL2N2_qld(lambda);
}
else
{
throw CBMRMException("unknown innerSolverType <" + innerSolverType + ">",
"CBMRMInnerSolverFactory::GetBMRMInnerSolver()");
}
return innerSolver;
}
/** Return an instance of loss function based on user's argument in configuration file
*
* @param model [read] Pointer to model object
* @param data [read] Pointer to data object
* @return loss object
*/
static CLoss* GetLoss(CModel* &model, CData* &data)
{
CLoss* loss = 0;
Configuration &config = Configuration::GetInstance();
// select the loss function specified by user (in configuration file)
if(config.IsSet("Loss.lossFunctionType"))
{
std::string lossFunctionType = config.GetString("Loss.lossFunctionType");
if(lossFunctionType == "LINESEARCH_HINGE")
{
CVecData *vecdata = 0;
if(! (vecdata = dynamic_cast<CVecData*>(data)))
{
throw CBMRMException("unable to cast data into CVecData",
"CLineSearchLossFactory::GetLoss()");
}
loss = new CLinesearch_HingeLoss(model, vecdata);
}
else if(lossFunctionType == "LINESEARCH_MULTI_LABEL_CLASSIFICATION")
{
CMultilabelVecData *mlvecdata = 0;
if(! (mlvecdata = dynamic_cast<CMultilabelVecData*>(data)))
{
throw CBMRMException("unable to cast data into CVecData",
"CLineSearchLossFactory::GetLoss()");
}
loss = new CLinesearch_MultilabelLoss(model, mlvecdata);
}
else
{
throw CBMRMException("ERROR: unrecognised loss function ("+lossFunctionType+")\n",
"CLineSearchLossFactory::GetLoss()");
}
}
else
{
throw CBMRMException("ERROR: no loss function specified!\n", "CLineSearchLossFactory::GetLoss()");
}
return loss;
}
CConsVecData(
unsigned int start = 0,
unsigned int nparts = 1) {
if (nparts != 1) {
string msg =
"CConsVecData does not support distribution of data (yet)";
throw CBMRMException(msg, "CConsVecData::CConsVecData");
}
LoadConstraintData(NumOfLabel());
}
/** Return a data object based on user's argument in configuration file.
* For serial computation, start and nparts are dummies.
* For distributed/parallel computation,
* load a portion of whole dataset: divide dataset into "nparts" parts and
* load only the "start"-th part.
*
* @param start [read] The part of dataset (divided into nparts) this machine should load
* @param nparts [read] The number of parts the original dataset will be divided into
* @return data object
*/
static CData* GetData(unsigned int start=0, unsigned int nparts=1)
{
CData *ds = 0;
Configuration &config = Configuration::GetInstance();
// default to this format
std::string dataFormat = "VECTOR_LABEL_VECTOR_FEATURE";
// default when a constraints file is given
if(config.IsSet("Data.constraintsFile"))
dataFormat = "CONSTRAINTS_LABEL_VECTOR_FEATURE";
// unless the user specifies otherwise in the config
if(config.IsSet("Data.format"))
dataFormat = config.GetString("Data.format");
if(dataFormat == "VECTOR_LABEL_VECTOR_FEATURE")
{
ds = new CVecData(start,nparts);
}
else if(dataFormat == "VARIABLE_LENGTH_VECTOR_LABEL_VECTOR_FEATURE")
{
ds = new CMultilabelVecData(start,nparts);
}
else if(dataFormat == "CONSTRAINTS_LABEL_VECTOR_FEATURE")
{
ds = new CConsVecData(start,nparts);
}
// else if(dataFormat == "YOUR_DATA_FORMAT")
//{
// ds = new CYourDataFormat();
//}
else
{
throw CBMRMException("ERROR: unrecognised data format ("+dataFormat+")\n", "CDataFactory::GetData()");
}
return ds;
}
int main(int argc, char** argv)
{
// sanity check
if(argc < 2)
{
std::cout << "Usage: ramp-bmrm-predict config.file" << std::endl;
std::cout << "Check the configfiles directory for examples" << std::endl;
std::cout << "ERROR: No configuration file given!" << std::endl;
exit(EXIT_FAILURE);
}
// the very first thing to do!
Configuration &config = Configuration::GetInstance();
config.ReadFromFile(argv[1]);
CData* data = 0;
CLoss* loss_vex = 0;
CLoss* loss_cav = 0;
CModel* model = 0;
try {
// serial computation with centralised data
config.SetString("Computation.mode", "SERIAL_CENTRALISED_DS");
std::string modelFilename = config.GetString("Model.modelFile");
std::string programMode = config.GetString("Program.mode");
data = CDataFactory::GetData();
model = CModelFactory::GetModel();
model->Initialize(modelFilename, data->dim());
CRampLossFactory::GetRampLoss(model,data, loss_vex, loss_cav);
if(programMode == "PREDICTION")
loss_vex->Predict(model);
else if(programMode == "EVALUATION")
loss_vex->Evaluate(model);
else
throw CBMRMException("unknown program mode <" + programMode +">","main()");
// compute ramp loss function value
Scalar lossVal_vex = 0.0;
Scalar lossVal_cav = 0.0;
loss_vex->ComputeLoss(lossVal_vex);
loss_cav->ComputeLoss(lossVal_cav);
std::cout << "a) Convex loss function value: " << lossVal_vex << std::endl;
std::cout << "b) Concave loss function linearization value: " << lossVal_cav << std::endl;
std::cout << "c) Ramp loss function value (a-b) : " << lossVal_vex - lossVal_cav << std::endl;
// cleaning up
delete model;
delete loss_vex;
delete loss_cav;
delete data;
}
catch(CBMRMException e) {
cout << e.Report() << endl;
}
return EXIT_SUCCESS;
}
void ComputeLoss(Scalar& loss)
{
throw CBMRMException("ERROR: not implemented!\n", "CGraphMatchLoss::ComputeLoss()");
}
/** Instantiate one convex loss function and another linearization of the concave loss
* based on user's argument in configuration file
*
* @param model [read] Pointer to model object
* @param data [read] Pointer to data object
* @param loss_vex [write] Convex loss function
* @param loss_cav [write] Linearization of concave loss function corresponding to loss_vex
*/
static void GetRampLoss(CModel* &model, CData* &data, CLoss* &loss_vex, CLoss* &loss_cav)
{
Configuration &config = Configuration::GetInstance();
// select the loss function specified by user (in configuration file)
if(config.IsSet("Loss.lossFunctionType"))
{
std::string lossFunctionType = config.GetString("Loss.lossFunctionType");
if(lossFunctionType == "WTA_MULTICLASS")
{
CVecData *vecdata = 0;
if(not (vecdata = dynamic_cast<CVecData*>(data)))
{
throw CBMRMException("unable to cast data into CVecData",
"CLossFactory::GetLoss()");
}
loss_vex = new CWTAMulticlassLoss(model, vecdata);
// in the ramp losses, users must know which loss_cav to use for their specific loss_vex!
loss_cav = new CWTAMulticlassLoss(model, vecdata, false); // with additive label loss switched off
}
else if(lossFunctionType == "ROC_SCORE")
{
CVecData *vecdata = 0;
if(not (vecdata = dynamic_cast<CVecData*>(data)))
{
throw CBMRMException("unable to cast data into CVecData",
"CLossFactory::GetLoss()");
}
loss_vex = new CROCScoreLoss(model, vecdata);
// in the ramp losses, users must know which loss_cav to use for their specific loss_vex!
loss_cav = new CROCScoreLoss(model, vecdata, false); // with additive label loss switched off
}
else if(lossFunctionType == "NDCG_RANK")
{
CVecData *vecdata = 0;
if(not (vecdata = dynamic_cast<CVecData*>(data)))
{
throw CBMRMException("unable to cast data into CVecData",
"CLossFactory::GetLoss()");
}
loss_vex = new CNDCGRankLoss(model, vecdata);
// in the ramp losses, users must know which loss_cav to use for their specific loss_vex!
loss_cav = new CRampNDCGRankLoss(model, vecdata);
}
//else if(lossFunctionType == "YOUR_LOSS_FUNCTION")
//{
// loss = new CYourLoss(w, data);
//}
else
{
throw CBMRMException("ERROR: unrecognised loss function ("+lossFunctionType+")\n",
"CLossFactory::GetLoss()");
}
}
else
{
throw CBMRMException("ERROR: no loss function specified!\n", "CLossFactory::GetLoss()");
}
}
/** Return an instance of loss function based on user's argument in configuration file
*
* @param data [read] Pointer to data object
* @return loss object
*/
static CLoss* GetLoss(CModel* &model, CData* &data)
{
CLoss* loss = 0;
Configuration &config = Configuration::GetInstance();
// select the loss function specified by user (in configuration file)
if(config.IsSet("Loss.lossFunctionType"))
{
std::string lossFunctionType = config.GetString("Loss.lossFunctionType");
if(lossFunctionType == "HINGE")
{
CVecData *vecdata = 0;
if(! (vecdata = dynamic_cast<CVecData*>(data)))
{
throw CBMRMException("unable to cast data into CVecData",
"CLossFactory::GetLoss()");
}
loss = new CHingeLoss(model, vecdata);
}
else if(lossFunctionType == "SQUARED_HINGE")
{
CVecData *vecdata = 0;
if(! (vecdata = dynamic_cast<CVecData*>(data)))
{
throw CBMRMException("unable to cast data into CVecData",
"CLossFactory::GetLoss()");
}
loss = new CSquaredHingeLoss(model, vecdata);
}
else if(lossFunctionType == "HUBER_HINGE")
{
CVecData *vecdata = 0;
if(! (vecdata = dynamic_cast<CVecData*>(data)))
{
throw CBMRMException("unable to cast data into CVecData",
"CLossFactory::GetLoss()");
}
loss = new CHuberHingeLoss(model, vecdata);
}
else if(lossFunctionType == "LOGISTIC")
{
CVecData *vecdata = 0;
if(! (vecdata = dynamic_cast<CVecData*>(data)))
{
throw CBMRMException("unable to cast data into CVecData",
"CLossFactory::GetLoss()");
}
loss = new CLogisticLoss(model, vecdata);
}
else if(lossFunctionType == "EXPONENTIAL")
{
CVecData *vecdata = 0;
if(! (vecdata = dynamic_cast<CVecData*>(data)))
{
throw CBMRMException("unable to cast data into CVecData",
"CLossFactory::GetLoss()");
}
loss = new CExponentialLoss(model, vecdata);
}
#ifndef PARALLEL_BMRM
else if(lossFunctionType == "ROC_SCORE")
{
CVecData *vecdata = 0;
if(! (vecdata = dynamic_cast<CVecData*>(data)))
{
throw CBMRMException("unable to cast data into CVecData",
"CLossFactory::GetLoss()");
}
loss = new CROCScoreLoss(model, vecdata);
}
else if(lossFunctionType == "F_BETA")
{
CVecData *vecdata = 0;
if(! (vecdata = dynamic_cast<CVecData*>(data)))
{
throw CBMRMException("unable to cast data into CVecData",
"CLossFactory::GetLoss()");
}
loss = new CFBetaLoss(model, vecdata);
}
#endif
else if(lossFunctionType == "EPSILON_INSENSITIVE")
{
CVecData *vecdata = 0;
if(! (vecdata = dynamic_cast<CVecData*>(data)))
{
throw CBMRMException("unable to cast data into CVecData",
"CLossFactory::GetLoss()");
}
loss = new CEpsilonInsensitiveLoss(model, vecdata);
}
else if(lossFunctionType == "LEAST_SQUARES")
{
CVecData *vecdata = 0;
if(! (vecdata = dynamic_cast<CVecData*>(data)))
{
throw CBMRMException("unable to cast data into CVecData",
"CLossFactory::GetLoss()");
}
loss = new CLeastSquaresLoss(model, vecdata);
}
else if(lossFunctionType == "LEAST_ABSOLUTE_DEVIATION")
{
CVecData *vecdata = 0;
if(! (vecdata = dynamic_cast<CVecData*>(data)))
{
throw CBMRMException("unable to cast data into CVecData",
"CLossFactory::GetLoss()");
}
loss = new CLeastAbsDevLoss(model, vecdata);
}
else if(lossFunctionType == "QUANTILE_REGRESSION")
{
CVecData *vecdata = 0;
if(! (vecdata = dynamic_cast<CVecData*>(data)))
{
throw CBMRMException("unable to cast data into CVecData",
"CLossFactory::GetLoss()");
}
loss = new CQuantileLoss(model, vecdata);
}
else if(lossFunctionType == "POISSON_REGRESSION")
{
CVecData *vecdata = 0;
if(! (vecdata = dynamic_cast<CVecData*>(data)))
{
throw CBMRMException("unable to cast data into CVecData",
"CLossFactory::GetLoss()");
}
loss = new CPoissonLoss(model, vecdata);
}
else if(lossFunctionType == "HUBER_ROBUST_REGRESSION")
{
CVecData *vecdata = 0;
if(! (vecdata = dynamic_cast<CVecData*>(data)))
{
throw CBMRMException("unable to cast data into CVecData",
"CLossFactory::GetLoss()");
}
loss = new CHuberRobustLoss(model, vecdata);
}
else if(lossFunctionType == "NOVELTY_DETECTION")
{
CVecData *vecdata = 0;
if(! (vecdata = dynamic_cast<CVecData*>(data)))
{
throw CBMRMException("unable to cast data into CVecData",
"CLossFactory::GetLoss()");
}
loss = new CNoveltyLoss(model, vecdata);
}
else if(lossFunctionType == "WTA_MULTICLASS")
{
CVecData *vecdata = 0;
if(! (vecdata = dynamic_cast<CVecData*>(data)))
{
throw CBMRMException("unable to cast data into CVecData",
"CLossFactory::GetLoss()");
}
loss = new CWTAMulticlassLoss(model, vecdata);
}
else if(lossFunctionType == "MULTI_LABEL_CLASSIFICATION")
{
CMultilabelVecData *mlvecdata = 0;
if(! (mlvecdata = dynamic_cast<CMultilabelVecData*>(data)))
{
throw CBMRMException("unable to cast data into CMultilabelVecData",
"CLossFactory::GetLoss()");
}
loss = new CMultilabelLoss(model, mlvecdata);
}
else if(lossFunctionType == "NDCG_RANK")
{
CVecData *vecdata = 0;
if(! (vecdata = dynamic_cast<CVecData*>(data)))
{
throw CBMRMException("unable to cast data into CVecData",
"CLossFactory::GetLoss()");
}
loss = new CNDCGRankLoss(model, vecdata);
}
else if(lossFunctionType == "SOFT_MARGIN")
{
CConsVecData *consvecdata = 0;
if(! (consvecdata = dynamic_cast<CConsVecData*>(data)))
{
throw CBMRMException("unable to cast data into CConsVecData",
"CLossFactory::GetLoss()");
}
loss = new SoftMarginLoss(model, consvecdata);
}
else
{
throw CBMRMException("ERROR: unrecognised loss function ("+lossFunctionType+")\n",
"CLossFactory::GetLoss()");
}
}
else
{
throw CBMRMException("ERROR: no loss function specified!\n", "CLossFactory::GetLoss()");
}
return loss;
}
/** Read examples into memory
*/
void CSeqFeature::LoadFeatures()
{
unsigned int tmpFidx = 0;
Scalar tmpFval = 0;
unsigned int featureCnt = 0;
unsigned int seqNum = 0;
unsigned int phiNum = 0;
unsigned int posNum1 = 0, posNum2 = 0;
std::string line = "";
std::string token = "";
std::ifstream featureFp;
featureFp.open(featureFile.c_str());
if(!featureFp.good())
{
string msg = "Cannot open feature file <" + featureFile + ">!";
throw CBMRMException(msg, "CSeqFeature::ScanFeatureFile()");
}
// read header information
int headerInfoCnt = 3; // min duration, max duration, feature dimension
do {
getline(featureFp, line);
trim(line);
if(IsBlankLine(line)) continue; // blank line
if(line[0] == '#') continue; // comment line
if(sscanf(line.c_str(),"maxDuration:%d",&maxDuration)==1) headerInfoCnt--;
if(sscanf(line.c_str(),"minDuration:%d",&minDuration)==1) headerInfoCnt--;
if(sscanf(line.c_str(),"globalFeatureDim:%d",&featureDimension)==1) headerInfoCnt--;
} while(!featureFp.eof() && (headerInfoCnt != 0));
assert(maxDuration >= minDuration);
assert(featureDimension < (1<<30)); // featureDimension is normally less then 1 billion
if(featureFp.eof())
throw CBMRMException("Feature file does not contain valid examples","CSeqFeature::LoadFeatures()");
// read sequences
nnz = 0;
while(!featureFp.eof())
{
// read sequence number
do {
getline(featureFp, line);
trim(line);
if(IsBlankLine(line)) continue; // blank line
if(line[0] == '#') continue; // comment line
if(sscanf(line.c_str(),"sequence:%d",&seqNum)==1) break;
} while(!featureFp.eof());
if(featureFp.eof())
throw CBMRMException("Feature file does not contain valid phi:*","CSeqFeature::LoadFeatures()");
// read phi:1 tag
phiNum = 0;
do {
getline(featureFp, line);
trim(line);
if(IsBlankLine(line)) continue; // blank line
if(line[0] == '#') continue; // comment line
if(sscanf(line.c_str(),"phi:%d",&phiNum)==1) break;
} while(!featureFp.eof());
if(featureFp.eof() || (phiNum != 1))
throw CBMRMException("Feature file does not contain valid phi:1","CSeqFeature::LoadFeatures()");
// read phi:1 sparse vectors
do {
getline(featureFp, line);
trim(line);
if(IsBlankLine(line)) continue; // blank line
if(line[0] == '#') continue; // comment line
if(sscanf(line.c_str(),"phi:%d",&phiNum) == 1)
break;
istringstream iss(line);
iss >> token;
if((sscanf(token.c_str(),"pos:%d",&posNum1) != 1))
throw CBMRMException("Feature file does not contain valid pos tag in phi:1","CSeqFeature::LoadFeatures()");
TheMatrix svec(1,featureDimension,SML::SPARSE);
featureCnt = 0;
while(!iss.eof())
{
iss >> token;
if(sscanf(token.c_str(),svec_feature_index_and_value_format.c_str(),&tmpFidx, &tmpFval) != 2)
{
ostringstream msg;
msg << "Invalid #" << featureCnt + 1 << " sparse vector element in phi:"<< phiNum << " seq:" << seqNum << " pos:" << posNum1;
throw CBMRMException(msg.str(),"CSeqFeature::LoadFeatures()");
}
svec.Set(0,tmpFidx,tmpFval);
nnz++;
}
if(featureCnt == 0)
throw CBMRMException("Feature file does not contain valid phi:2 sparse vector","CSeqFeature::LoadFeatures()");
phi_1.push_back(svec);
} while(!featureFp.eof());
if(phi_1.size() < 1)
throw CBMRMException("Feature file does not contain valid phi:1","CSeqFeature::LoadFeatures()");
numOfSeq = phi_1.size();
if(featureFp.eof() || (phiNum != 2))
throw CBMRMException("Feature file does not contain valid phi:2","CSeqFeature::LoadFeatures()");
// read phi:2 sparse vectors
unsigned int prevPosNum1 = 0, prevPosNum2 = 0;
vector<TheMatrix> tmp_phi_2_svecs;
featureCnt = 0;
do {
getline(featureFp, line);
trim(line);
if(IsBlankLine(line)) continue; // blank line
if(line[0] == '#') continue; // comment line
if((sscanf(line.c_str(),"phi:%d",&phiNum) == 1))
break;
istringstream iss(line);
iss >> token;
if((sscanf(token.c_str(),"pos:%d,%d",&posNum1,&posNum2) != 2))
throw CBMRMException("Feature file does not containt valid pos tag in phi:2","CSeqFeature::LoadFeatures()");
if(prevPosNum2 >= posNum2)
{
ostringstream msg;
msg << "previous posNum2 must be > current posNum2 in phi:2 (phi:2 pos:" << posNum1 << "," << posNum2;
throw CBMRMException(msg.str(),"CSeqFeature::LoadFeatures()");
}
if(prevPosNum1 >= posNum1)
{
ostringstream msg;
msg << "previous posNum1 must be > current posNum1 in phi:2 (phi:2 pos:" << posNum1 << "," << posNum2;
throw CBMRMException(msg.str(),"CSeqFeature::LoadFeatures()");
}
if(posNum1 != prevPosNum1)
{
phi_2.push_back(tmp_phi_2_svecs);
tmp_phi_2_svecs.clear();
}
TheMatrix svec(1,featureDimension,SML::SPARSE);
featureCnt = 0;
while(!iss.eof())
{
iss >> token;
if(sscanf(token.c_str(),svec_feature_index_and_value_format.c_str(),&tmpFidx, &tmpFval) != 2)
{
ostringstream msg;
msg << "Invalid #" << featureCnt + 1 << " sparse vector element in phi:"<< phiNum << " seq:" << seqNum << " pos:" << posNum1;
throw CBMRMException(msg.str(),"CSeqFeature::LoadFeatures()");
}
svec.Set(0,tmpFidx,tmpFval);
nnz++;
}
if(featureCnt == 0)
throw CBMRMException("Feature file does not containt valid phi:2 sparse vector","CSeqFeature::LoadFeatures()");
tmp_phi_2_svecs.push_back(svec);
} while(!featureFp.eof());
if(phi_2.size() < 1)
throw CBMRMException("Feature file does not contain phi:2","CSeqFeature::LoadFeatures()");
}
// data matrix density
density = ((double)nnz/featureDimension)/numOfSeq;
featureFp.close();
}
