/*
* run cross validation for folds from idxStartFold to idxEndFold
* @param idxStartFold index of the first fold
* @param idxEndFold index of the last fold
*/
void Cvlars::subrun(int idxStartFold,int idxEndFold)
{
//create test and control container
STK::CArrayXX XControl( n_ - sizePartition_[idxStartFold], p_);
STK::CVectorX yControl( n_ - sizePartition_[idxStartFold] );
STK::CArrayXX XTest(sizePartition_[idxStartFold], p_);
STK::CVectorX yTest(sizePartition_[idxStartFold] );
STK::CVectorX yPred(sizePartition_[idxStartFold] );
for(int i = idxStartFold ; i <= idxEndFold ; i++)
{
//fill the container
int index = 1;
int index2 = 1;
for(int j = 1; j <= n_; j++)
{
if(partition_[j-1] != i)
{
yControl[index] = (*p_y_)[j];
XControl.row(index)=p_X_->row(j);
index++;
}
else
{
yTest[index2] = (*p_y_)[j];
XTest.row(index2)=p_X_->row(j);
index2++;
}
}
//run lars on control data set
HD::Lars lars(XControl,yControl,maxSteps_,intercept_,eps_);
lars.run();
for(int s = 1 ; s <= (int) index_.size(); s++)
{
//we compute the prediction of the y associated to XTest
lars.predict(XTest,index_[s-1], lambdaMode_, yPred);
//compute the residuals
residuals_(s,i+1) = (yPred-yTest).square().sum()/sizePartition_[i];
}
}
}
void SparseCoding<DictionaryInitializer>::OptimizeCode()
{
// When using the Cholesky version of LARS, this is correct even if
// lambda2 > 0.
arma::mat matGram = trans(dictionary) * dictionary;
for (size_t i = 0; i < data.n_cols; ++i)
{
// Report progress.
if ((i % 100) == 0)
Log::Debug << "Optimization at point " << i << "." << std::endl;
bool useCholesky = true;
regression::LARS lars(useCholesky, matGram, lambda1, lambda2);
// Create an alias of the code (using the same memory), and then LARS will
// place the result directly into that; then we will not need to have an
// extra copy.
arma::vec code = codes.unsafe_col(i);
lars.Regress(dictionary, data.unsafe_col(i), code, false);
}
}
void Cvlars::run2()
{
//search the first and last fold with the same size
std::vector<int> startIndex(1,0),endIndex(1,k_-1);
int k = 0;
for(int i = 1; i < k_; i++)
{
if(sizePartition_[i]!= sizePartition_[startIndex[k]])
{
startIndex.push_back(i);
endIndex[k] = i-1;
endIndex.push_back(k_-1);
k++;
}
}
//run for each size of fold
//create test and control container
#pragma omp parallel
{
#pragma omp for schedule(dynamic,1)
for(int i = 0; i < k_ ; i++)
{
STK::CArrayXX XControl( n_ - sizePartition_[i], p_);
STK::CVectorX yControl( n_ - sizePartition_[i] );
STK::CArrayXX XTest(sizePartition_[i], p_);
STK::CVectorX yTest(sizePartition_[i] );
STK::CVectorX yPred(sizePartition_[i] );
//fill the container
int index = 1;
int index2 = 1;
for(int j = 1; j <= n_; j++)
{
if(partition_[j-1] != i)
{
yControl[index] = (*p_y_)[j];
XControl.row(index)=p_X_->row(j);
index++;
}
else
{
yTest[index2] = (*p_y_)[j];
XTest.row(index2)=p_X_->row(j);
index2++;
}
}
//run lars on control data set
HD::Lars lars(XControl,yControl,maxSteps_,intercept_,eps_);
lars.run();
for(int s = 1 ; s <= (int) index_.size(); s++)
{
//we compute the prediction of the y associated to XTest
lars.predict(XTest,index_[s-1], lambdaMode_, yPred);
//compute the residuals
residuals_(s,i+1) = (yPred-yTest).square().sum()/sizePartition_[i];
}
}
}//end parallel
// compute mean prediction error for each index
STK::CVectorX one(k_,1);
cv_ = (residuals_ * one) / k_;
// compute mean standard deviation of cv_ for each index
for(int i = 1; i <= (int) index_.size(); i++)
residuals_.row(i) -= cv_[i];
residuals_ = residuals_.square();
cvError_ = (residuals_ * one)/(k_-1)/k_;
cvError_ = cvError_.sqrt();
}
