Eigen::VectorXd RBM::getParameters()
{
return currentParameters();
}
void Net::save(std::ostream& stream)
{
stream << architecture.str() << "parameters " << currentParameters();
}
bool QMCSHLinearOptimize::run()
{
start();
//size of matrix
numParams = optTarget->NumParams();
N = numParams + 1;
// initialize our parameters
vector<RealType> currentOvlp(N,0);
vector<RealType> currentHOvlp(N,0);
vector<RealType> currentParameters(numParams,0);
RealType E_avg(0);
vector<RealType> bestParameters(currentParameters);
optdir.resize(numParams,0);
optparm.resize(numParams,0);
for (int i=0; i<numParams; i++) optparm[i] = currentParameters[i] = optTarget->Params(i);
bool acceptedOneMove(false);
int tooManyTries(20);
int failedTries(0);
RealType lastCost(0);
RealType startCost(0);
startCost = lastCost = optTarget->Cost(false);
app_log()<<"Starting cost: "<<startCost<<endl;
Matrix<RealType> OM;
OM.resize(N,N);
dmcEngine->fillVectors(currentOvlp,currentHOvlp,E_avg,OM);
for (int i=0; i<numParams; i++) optdir[i]=currentOvlp[i];
std::vector<RealType> dP(N,1);
for (int i=0; i<numParams; i++) dP[i+1]=optdir[i];
Lambda = getNonLinearRescale(dP,OM);
app_log()<<"rescaling factor :"<<Lambda<<endl;
RealType bigOptVec(std::abs(optdir[0]));
for (int i=1; i<numParams; i++) bigOptVec =std::max(std::abs(optdir[i]),bigOptVec);
// app_log()<<"currentOvlp"<<endl;
// for (int i=0; i<numParams; i++)
// app_log()<<optdir[i]<<" ";
// app_log()<<endl;
//
// app_log()<<"optparam"<<endl;
// for (int i=0; i<numParams; i++)
// app_log()<<optparm[i]<<" ";
// app_log()<<endl;
if (MinMethod=="rescale")
{
if (bigOptVec*std::abs(Lambda)>bigChange)
app_log()<<" Failed Step. Largest LM parameter change:"<<bigOptVec*std::abs(Lambda)<<endl;
else
{
for (int i=0; i<numParams; i++) bestParameters[i] = optTarget->Params(i) = currentParameters[i] + Lambda*optdir[i];
acceptedOneMove = true;
}
}
else if (MinMethod=="overlap")
{
orthoScale(optdir,OM);
if (bigOptVec>bigChange)
app_log()<<" Failed Step. Largest LM parameter change:"<<bigOptVec<<endl;
else
{
for (int i=0; i<numParams; i++) bestParameters[i] = optTarget->Params(i) = currentParameters[i] + optdir[i];
acceptedOneMove = true;
}
}
else if (MinMethod=="average")
{
for (int i=0; i<numParams; i++) bestParameters[i] = optTarget->Params(i) = currentParameters[i] + currentHOvlp[i];
acceptedOneMove = true;
}
else
{
TOL = param_tol/bigOptVec;
AbsFuncTol=true;
largeQuarticStep=bigChange/bigOptVec;
quadstep = stepsize*Lambda; //bigOptVec;
// initial guess for line min bracketing
LambdaMax = quadstep;
myTimers[3]->start();
if (MinMethod=="quartic")
{
int npts(7);
lineoptimization3(npts,startCost);
}
else lineoptimization2();
myTimers[3]->stop();
RealType biggestParameterChange = bigOptVec*std::abs(Lambda);
if (biggestParameterChange>bigChange)
{
app_log()<<" Failed Step. Largest LM parameter change:"<<biggestParameterChange<<endl;
for (int i=0; i<numParams; i++) optTarget->Params(i) = bestParameters[i] = currentParameters[i];
}
else
{
for (int i=0; i<numParams; i++) optTarget->Params(i) = optparm[i] + Lambda * optdir[i];
lastCost = optTarget->Cost(false);
app_log()<<" Costs: "<<startCost<<" "<<lastCost<<endl;
app_log()<<" Optimal rescaling factor :"<<Lambda<<endl;
if (lastCost<startCost)
acceptedOneMove = true;
else
{
for (int i=0; i<numParams; i++) optTarget->Params(i) = currentParameters[i];
app_log()<<" Failed Step. Cost increase "<<endl;
}
}
}
// if (acceptedOneMove)
// for (int i=0; i<numParams; i++) optTarget->Params(i) = bestParameters[i];
// else
// for (int i=0; i<numParams; i++) optTarget->Params(i) = currentParameters[i];
// if (W.getActiveWalkers()>NumOfVMCWalkers)
// {
// W.destroyWalkers(W.getActiveWalkers()-NumOfVMCWalkers);
// app_log() << " QMCLinearOptimize::generateSamples removed walkers." << endl;
// app_log() << " Number of Walkers per node " << W.getActiveWalkers() << endl;
// }
finish();
return (optTarget->getReportCounter() > 0);
}
