// This is specific to auto-search
std::vector<double> AutoSearchCrossValidationDriver::doCrossValidationLoop(
CyclicCoordinateDescent& ccd,
AbstractSelector& selector,
const CCDArguments& allArguments,
int nThreads,
std::vector<CyclicCoordinateDescent*>& ccdPool,
std::vector<AbstractSelector*>& selectorPool) {
const auto& arguments = allArguments.crossValidation;
double tryvalue = (arguments.startingVariance > 0) ?
arguments.startingVariance :
modelData.getNormalBasedDefaultVar();
std::ostringstream stream;
stream << "Starting var = " << tryvalue;
if (arguments.startingVariance == -1) {
stream << " (default)";
}
logger->writeLine(stream);
const double tolerance = 1E-2; // TODO Make Cyclops argument
int nDim = ccd.getHyperprior().size();
std::vector<double> currentOptimal(nDim, tryvalue);
bool globalFinished = false;
std::vector<double> savedOptimal;
while (!globalFinished) {
if (nDim > 1) {
savedOptimal = currentOptimal; // make copy
}
for (int dim = 0; dim < nDim; ++dim) {
// Local search
UniModalSearch searcher(10, 0.01, log(1.5));
int step = 0;
bool dimFinished = false;
while (!dimFinished) {
ccd.setHyperprior(dim, currentOptimal[dim]);
selector.reseed();
std::vector<double> predLogLikelihood;
// Newly re-located code
double pointEstimate = doCrossValidationStep(ccd, selector, allArguments, step,
nThreads, ccdPool, selectorPool,
predLogLikelihood);
double stdDevEstimate = computeStDev(predLogLikelihood, pointEstimate);
std::ostringstream stream;
stream << "AvgPred = " << pointEstimate << " with stdev = " << stdDevEstimate << std::endl;
searcher.tried(currentOptimal[dim], pointEstimate, stdDevEstimate);
pair<bool,double> next = searcher.step();
stream << "Completed at " << currentOptimal[dim] << std::endl;
stream << "Next point at " << next.second << " and " << next.first;
logger->writeLine(stream);
currentOptimal[dim] = next.second;
if (!next.first) {
dimFinished = true;
}
std::ostringstream stream1;
stream1 << searcher;
logger->writeLine(stream1);
step++;
if (step >= maxSteps) {
std::ostringstream stream;
stream << "Max steps reached!";
logger->writeLine(stream);
dimFinished = true;
}
}
}
if (nDim == 1) {
globalFinished = true;
} else {
double diff = 0.0;
for (int i = 0; i < nDim; ++i) {
diff += std::abs((currentOptimal[i] - savedOptimal[i]) / savedOptimal[i]);
}
std::ostringstream stream;
stream << "Absolute percent difference in cycle: " << diff << std::endl;
globalFinished = (diff < tolerance);
}
}
return currentOptimal;
}
void HierarchyAutoSearchCrossValidationDriver::drive(
CyclicCoordinateDescent& ccd,
AbstractSelector& selector,
const CCDArguments& arguments) {
// TODO Check that selector is type of CrossValidationSelector
std::vector<real> weights;
double tryvalue = modelData.getNormalBasedDefaultVar();
double tryvalueClass = tryvalue; // start with same variance at the class and element level; // for hierarchy class variance
UniModalSearch searcher(10, 0.01, log(1.5));
UniModalSearch searcherClass(10, 0.01, log(1.5)); // Need a better way to do this.
// const double eps = 0.05; //search stopper
std::ostringstream stream;
stream << "Default var = " << tryvalue;
logger->writeLine(stream);
bool finished = false;
bool drugLevelFinished = false;
bool classLevelFinished = false;
int step = 0;
while (!finished) {
// More hierarchy logic
ccd.setHyperprior(tryvalue);
ccd.setClassHyperprior(tryvalueClass);
std::vector<double> predLogLikelihood;
// Newly re-located code
double pointEstimate = doCrossValidation(ccd, selector, arguments, step, predLogLikelihood);
double stdDevEstimate = computeStDev(predLogLikelihood, pointEstimate);
std::ostringstream stream;
stream << "AvgPred = " << pointEstimate << " with stdev = " << stdDevEstimate;
logger->writeLine(stream);
// alternate adapting the class and element level, unless one is finished
if ((step % 2 == 0 && !drugLevelFinished) || classLevelFinished){
searcher.tried(tryvalue, pointEstimate, stdDevEstimate);
pair<bool,double> next = searcher.step();
tryvalue = next.second;
std::ostringstream stream;
stream << "Next point at " << next.second << " and " << next.first;
logger->writeLine(stream);
if (!next.first) {
drugLevelFinished = true;
}
} else {
searcherClass.tried(tryvalueClass, pointEstimate, stdDevEstimate);
pair<bool,double> next = searcherClass.step();
tryvalueClass = next.second;
std::ostringstream stream;
stream << "Next Class point at " << next.second << " and " << next.first;
logger->writeLine(stream);
if (!next.first) {
classLevelFinished = true;
}
}
// if everything is finished, end.
if (drugLevelFinished && classLevelFinished){
finished = true;
}
std::ostringstream stream2;
stream2 << searcher;
logger->writeLine(stream2);
step++;
if (step >= maxSteps) {
std::ostringstream stream;
stream << "Max steps reached!";
logger->writeLine(stream);
finished = true;
}
}
maxPoint = tryvalue;
maxPointClass = tryvalueClass;
// Report results
std::ostringstream stream2;
stream2 << std::endl;
stream2 << "Maximum predicted log likelihood estimated at:" << std::endl;
stream2 << "\t" << maxPoint << " (variance)" << std::endl;
stream2 << "class level = " << maxPointClass;
logger->writeLine(stream2);
if (!arguments.useNormalPrior) {
double lambda = convertVarianceToHyperparameter(maxPoint);
std::ostringstream stream;
stream << "\t" << lambda << " (lambda)";
logger->writeLine(stream);
}
std::ostringstream stream3;
logger->writeLine(stream3);
}