void CGibbsSamplingInfEngine::
Sampling( int statTime, int endTime )
{
intVector ndsForSampling;
GetNdsForSampling( &ndsForSampling );
boolVector sampleIsNeed;
GetSamplingNdsFlags( &sampleIsNeed );
int numNdsForSampling = ndsForSampling.size();
pEvidencesVector currentEvidences;
GetCurrentEvidences( ¤tEvidences );
CEvidence * pCurrentEvidence;
int t;
int i;
for( t = statTime; t < endTime; t++ )
{
int series;
for( series = 0; series < GetNumStreams(); series++ )
{
pCurrentEvidence = currentEvidences[series];
for( i = 0; i < numNdsForSampling; i++ )
{
if( sampleIsNeed[i] )
{
pCurrentEvidence->ToggleNodeStateBySerialNumber(1, &i);
bool canBeSample = ConvertingFamilyToPot( ndsForSampling[i], pCurrentEvidence );
if(canBeSample)
{
GetPotToSampling(ndsForSampling[i])->GenerateSample( pCurrentEvidence, m_bMaximize );
}
else
{
pCurrentEvidence->ToggleNodeStateBySerialNumber(1, &i);
}
}
}
}
if( t > GetBurnIn())
{
pFactorVector queryFactors;
GetQueryFactors( &queryFactors );
int i;
for( i = 0; i < queryFactors.size(); i++ )
{
queryFactors[i]->UpdateStatisticsML( &(GetCurrentEvidences()->front()), GetNumStreams() );
}
}
}
}
void
CPersistEvidence::TraverseSubobject(CPNLBase *pObj, CContext *pContext)
{
CPersistNodeValues::TraverseSubobject(pObj, pContext);
CEvidence *pEvidence = static_cast<CEvidence*>(pObj);
intVector *pV = new intVector;
pV->resize(pEvidence->GetNumberObsNodes());
memcpy((void*)&pV->front(), pEvidence->GetAllObsNodes(),
pEvidence->GetNumberObsNodes()*sizeof(pV->front()));
pContext->Put(const_cast<CModelDomain*>(pEvidence->GetModelDomain()), "ModelDomain");
pContext->Put(new CCoverDel<intVector>(pV), "ObservedNodes", true);
}
int testEvidence()
{
int ret = TRS_OK;
int nnodes = 0;
int nObsNodes = 0;
int i,j;
while(nnodes <= 0)
{
trsiRead( &nnodes, "10", "Number of nodes in Model" );
}
while((nObsNodes <= 0)||(nObsNodes>nnodes))
{
trsiRead( &nObsNodes, "2", "Number of Observed nodes from all nodes in model");
}
int seed1 = pnlTestRandSeed();
/*create string to display the value*/
char value[42];
sprintf(value, "%i", seed1);
trsiRead(&seed1, value, "Seed for srand to define NodeTypes etc.");
trsWrite(TW_CON|TW_RUN|TW_DEBUG|TW_LST, "seed for rand = %d\n", seed1);
CNodeType *modelNodeType = new CNodeType[2];
modelNodeType[0] = CNodeType( 1, 4 );
modelNodeType[1] = CNodeType( 0, 3 );
int *NodeAssociat=new int [nnodes+1];
for(i=0; i<(nnodes+1)/2; i++)
{
NodeAssociat[2*i]=0;
NodeAssociat[2*i+1]=1;
}
//create random graph - number of nodes for every node is rand too
int lowBorder = nnodes - 1;
int upperBorder = int((nnodes * (nnodes - 1)) / 2);
int numEdges = rand()%(upperBorder - lowBorder)+lowBorder;
CGraph* theGraph = tCreateRandomDAG( nnodes, numEdges, 1 );
CBNet *grModel = CBNet::Create(nnodes, 2, modelNodeType, NodeAssociat,theGraph);
int *obsNodes = (int*)trsGuardcAlloc(nObsNodes, sizeof(int));
srand ((unsigned int)seed1);
intVector residuaryNodes;
for (i=0; i<nnodes; i++)
{
residuaryNodes.push_back(i);
}
int num = 0;
valueVector Values;
Values.reserve(3*nnodes);
Value val;
for (i = 0; i<nObsNodes; i++)
{
num = rand()%(nnodes-i);
obsNodes[i] = residuaryNodes[num];
residuaryNodes.erase(residuaryNodes.begin()+num);
CNodeType nt = modelNodeType[NodeAssociat[obsNodes[i]]];
if(nt.IsDiscrete())
{
val.SetInt(1);
Values.push_back(val);
}
else
{
val.SetFlt(1.0f);
Values.push_back(val);
val.SetFlt(2.0f);
Values.push_back(val);
val.SetFlt(3.0f);
Values.push_back(val);
}
}
residuaryNodes.clear();
CEvidence *pMyEvid = CEvidence::Create(grModel,
nObsNodes, obsNodes, Values) ;
int nObsNodesFromEv = pMyEvid->GetNumberObsNodes();
const int *pObsNodesNow = pMyEvid->GetObsNodesFlags();
// const int *myOffset = pMyEvid->GetOffset();
const int *myNumAllObsNodes = pMyEvid->GetAllObsNodes();
valueVector ev;
pMyEvid->GetRawData(&ev);
const Value* vall = pMyEvid->GetValue(obsNodes[0]);
if( NodeAssociat[obsNodes[0]] == 0 )
{
if( (vall)[0].GetInt() != 1 )
{
ret = TRS_FAIL;
}
}
else
{
for( j=0; j<3; j++)
{
if( (vall)[j].GetFlt() != (j+1)*1.0f )
{
ret = TRS_FAIL;
break;
}
}
}
if(nObsNodesFromEv == nObsNodes)
{
intVector numbersOfReallyObsNodes;
int numReallyObsNodes=0;
for ( i=0; i<nObsNodesFromEv; i++)
{
if (pObsNodesNow[i])
{
numbersOfReallyObsNodes.push_back(myNumAllObsNodes[i]);
numReallyObsNodes++;
}
}
#if 0
const CNodeType ** AllNodeTypesFromModel= new const CNodeType*[nnodes];
for (i=0; i<nnodes; i++)
{
AllNodeTypesFromModel[i] = grModel->GetNodeType(i);
}
for (i=0; i<nObsNodesFromEv; i++)
{
//Test the values which are keep in Evidence
CNodeType nt = *AllNodeTypesFromModel[myNumAllObsNodes[i]];
int IsDiscreteNode = nt.IsDiscrete();
if(IsDiscreteNode)
{
int valFromEv = (ev[myOffset[i]].GetInt());
if(!(Values[i].GetInt() == valFromEv))
{
ret=TRS_FAIL;
break;
}
}
else
{
;
for (j=0; j<3; j++)
{
if(!((ev[myOffset[i]+j]).GetFlt() == Values[i+j].GetFlt()))
{
ret=TRS_FAIL;
break;
}
}
}
}
delete []AllNodeTypesFromModel;
#endif
}
else
{
ret = TRS_FAIL;
}
//Toggle some Node
int someNumber = (int)(rand()*nObsNodesFromEv/RAND_MAX);
int *someOfNodes = new int[someNumber];
intVector residuaryNums = intVector(myNumAllObsNodes,
myNumAllObsNodes+nObsNodesFromEv);
num=0;
for(i=0; i<someNumber;i++)
{
num = (int)(rand()%(nObsNodes-i));
someOfNodes[i] = residuaryNums[num];
residuaryNums.erase(residuaryNums.begin()+num);
}
residuaryNums.clear();
pMyEvid->ToggleNodeState(someNumber, someOfNodes);
const int *pObsNodesAfterToggle = pMyEvid->GetObsNodesFlags();
for (i=0; i<nObsNodesFromEv; i++)
{
//Test the ToggleNode method...
if(pObsNodesAfterToggle[i])
{
for(j=0; j<someNumber;j++)
{
if(myNumAllObsNodes[i]==someOfNodes[j])
{
ret=TRS_FAIL;
break;
}
}
}
}
delete grModel;
delete pMyEvid;
delete []modelNodeType;
delete []NodeAssociat;
delete []someOfNodes;
int obsNodes_memory_flag = trsGuardCheck( obsNodes );
if( obsNodes_memory_flag)
{
return trsResult( TRS_FAIL, "Dirty memory");
}
trsGuardFree( obsNodes );
return trsResult( ret, ret == TRS_OK ? "No errors" : "Bad test on Values");
}
void Infer_Process(const CBNet* pBnet)
{
//create simple evidence for node 0 from BNet
CEvidence* pEvidForWS = CreateEvidenceForBNet(pBnet);
//create Naive inference for BNet
CNaiveInfEngine* pNaiveInf = CNaiveInfEngine::Create( pBnet );
//enter evidence created before
pNaiveInf->EnterEvidence( pEvidForWS );
//set the query node
int numQueryNds = 1;//*<-
int queryNds[] = { 3 };//*<-
//get a marginal for query set of nodes
pNaiveInf->MarginalNodes( queryNds, numQueryNds );
const CPotential* pMarg = pNaiveInf->GetQueryJPD();
//display the evidence node and such velue of BNet
intVector obsNds;
pConstValueVector obsVls;
pEvidForWS->GetObsNodesWithValues(&obsNds, &obsVls);
int i;
for( i = 0; i < obsNds.size(); i++ )
{
std::cout<<" observed value for node "<<obsNds[i];
std::cout<<" is "<<obsVls[i]->GetInt()<<std::endl;
}
//display the query node and such velue of BNet
int nnodes;
const int* domain;
pMarg->GetDomain( &nnodes, &domain );
std::cout<<" inference results: \n";
std::cout<<" probability distribution for nodes [ ";
for( i = 0; i < nnodes; i++ )
{
std::cout<<domain[i]<<" ";
}
std::cout<<"]"<<std::endl;
CMatrix<float>* pMat = pMarg->GetMatrix(matTable);
// graphical model hase been created using dense matrix
// so, the marginal is also dense
EMatrixClass type = pMat->GetMatrixClass();
if( ! ( type == mcDense || type == mcNumericDense || type == mc2DNumericDense ) )
{
assert(0);
}
int nEl;
const float* data;
static_cast<CNumericDenseMatrix<float>*>(pMat)->GetRawData(&nEl, &data);
for( i = 0; i < nEl; i++ )
{
std::cout<<" "<<data[i];
}
std::cout<<std::endl;
delete pEvidForWS;
delete pNaiveInf;
}
void CLWSamplingInfEngine::
MarginalNodes( const int *queryIn, int querySz, int notExpandJPD)
{
PNL_CHECK_IS_NULL_POINTER(queryIn);
if( m_bNormalized == false ) NormalizeWeight();
int i, j, k;
int offset;
int nsamples = m_particleCount;
const CBNet *pBNet = static_cast<const CBNet *>( m_pGraphicalModel );
int type = 0;
int totalnodesizes = 0;
intVector nodesize(querySz);
intVector mulnodesize(querySz, 1);
for( i = querySz; --i >=0; )
{
const CNodeType* pNodeType = pBNet->GetNodeType(queryIn[i]);
nodesize[i] = pNodeType->GetNodeSize();
if(i == querySz-1)
mulnodesize[i] = 1;
else
mulnodesize[i] *= nodesize[i+1];
if(pNodeType->IsDiscrete())
{
type++;
if(totalnodesizes == 0)
totalnodesizes = nodesize[i];
else
totalnodesizes = totalnodesizes * nodesize[i];
}
else
{
totalnodesizes = totalnodesizes + nodesize[i];
}
}
if(type == querySz)
{
//all query nodes are discrete
float *tab = new float[totalnodesizes];
for(i = 0; i < totalnodesizes; i++) tab[i] = 0;
for( i = 0; i < nsamples; i++)
{
CEvidence* pEvidence = m_currentEvVec[i];
int index = 0;
for(j = 0; j < querySz; j++)
{
Value* pValue = pEvidence->GetValue(queryIn[j]);
index += pValue->GetInt() * mulnodesize[j];
}
tab[index] += m_particleWeight[i];
}
m_pQueryJPD = CTabularPotential::Create(queryIn, querySz, pBNet->GetModelDomain () );
m_pQueryJPD->AllocMatrix( tab, matTable );
delete []tab;
}
else if(type == 0)
{
//all query nodes are gaussian
float* val = new float[totalnodesizes];
float* mean = new float[totalnodesizes];
float* cov = new float[totalnodesizes * totalnodesizes];
for( i = 0; i < totalnodesizes; i++) mean[i] = 0;
for( i = 0; i < totalnodesizes * totalnodesizes; i++) cov[i] = 0;
// mean
for( i = 0; i < nsamples; i++)
{
CEvidence* pEvidence = m_currentEvVec[i];
for(j = 0, offset = 0; j < querySz; j++)
{
Value* pValue = pEvidence->GetValue(queryIn[j]);
for(k = 0; k < nodesize[j]; k++)
{
mean[offset] += m_particleWeight[i] * pValue[k].GetFlt();
offset++;
}
}
}
// covariance
for( i = 0; i < nsamples; i++)
{
CEvidence* pEvidence = m_currentEvVec[i];
for(j = 0, offset = 0; j < querySz; j++)
{
Value* pValue = pEvidence->GetValue(queryIn[j]);
for(k = 0; k < nodesize[j]; k++)
{
val[offset] = pValue[k].GetFlt();
offset++;
}
}
for(j = 0; j < totalnodesizes; j++)
{
for(k = j; k < totalnodesizes; k++)
{
cov[k*totalnodesizes+j] += ( m_particleWeight[i] * ( val[j]- mean[j]) * (val[k] - mean[k]) );
cov[j*totalnodesizes+k] = cov[k*totalnodesizes+j];
}
}
}
m_pQueryJPD = CGaussianPotential::Create( queryIn, querySz, pBNet->GetModelDomain () );
m_pQueryJPD->AllocMatrix( mean, matMean );
m_pQueryJPD->AllocMatrix( cov, matCovariance);
delete []val;
delete []mean;
delete []cov;
}
//Get MPE
delete m_pEvidenceMPE;
m_pEvidenceMPE = NULL;
m_pEvidenceMPE = m_pQueryJPD->GetMPE();
}
//-----------------------------------------------------------------------------
CPotential *CSoftMaxCPD::ConvertToTabularPotential(const CEvidence* pEvidence) const
{
//searching discrite nodes in domain
intVector discriteNodesPosInDom;
int domSize = m_Domain.size();
int numSoftMaxNode;
int discrDomSize = 0;
int *parentIndexes;
int SoftMaxSize;
const pConstNodeTypeVector* ntVec = GetDistribFun()->GetNodeTypesVector();
int i;
for (i = 0; i < domSize; i++)
{
if ((*ntVec)[i]->IsDiscrete())
{
discriteNodesPosInDom.push_back(m_Domain[i]);
SoftMaxSize = (*ntVec)[i]->GetNodeSize();
numSoftMaxNode = i;
}
};
discrDomSize = discriteNodesPosInDom.size();
//fill parents indexes vector
parentIndexes = new int[discrDomSize-1];
for( i = 0; i < discrDomSize-1; i++ )
{
parentIndexes[i] = discriteNodesPosInDom[i];
}
// creating new evidece that contain all observed nodes in this domain
intVector pObsNodes;
pConstValueVector pObsValues;
pConstNodeTypeVector pNodeTypes;
pEvidence->GetObsNodesWithValues(&pObsNodes,&pObsValues,&pNodeTypes);
int *obsNodes;
int obsNodesSize;
obsNodesSize=pObsNodes.size();
obsNodes = new int[obsNodesSize];
for(i = 0;i < obsNodesSize; i++)
{
obsNodes[i] = pObsNodes[i];
}
CEvidence *pCopyEvidence;
valueVector cpyValVect(0);
for(i = 0; i < obsNodesSize; i++)
{
cpyValVect.push_back(*(pObsValues[i]));
}
pCopyEvidence = CEvidence::Create(pEvidence->GetModelDomain(), obsNodesSize,
obsNodes,(const valueVector&)cpyValVect);
for(i = 0; i < pObsNodes.size(); i++)
{
if((std::find(m_Domain.begin(),m_Domain.end(), pObsNodes[i])) == m_Domain.end())
{
pCopyEvidence->MakeNodeHidden(pObsNodes[i]);
}
};
//creating tabular potential
CTabularPotential *resFactor = CTabularPotential::Create(
GetModelDomain(),discriteNodesPosInDom);
if( m_DistributionType == dtSoftMax)
{
resFactor->AttachMatrix(((CSoftMaxDistribFun*)m_CorrespDistribFun)->
GetProbMatrix(pCopyEvidence),matTable);
}
else
{
if(m_DistributionType == dtCondSoftMax)
{
resFactor->AttachMatrix(((CCondSoftMaxDistribFun*)m_CorrespDistribFun)->
GetProbMatrix(pCopyEvidence),matTable);
}
else
{
PNL_THROW( CInconsistentType,
"distribution must be SoftMax or conditional SoftMax" )
}
}
delete [] parentIndexes;
delete [] obsNodes;
delete pCopyEvidence;
return resFactor;
}