/* For the historical case:
One experiment replicate per line, containing all responses in one line
can't allow different sigma for each response and the
config vars are repeated per experiment. */
inline void
read_historical_data(const std::string& expDataFileName,
const std::string& context_message,
size_t numExperiments,
IntVector& numReplicates,
size_t numExpConfigVars,
size_t numFunctions,
size_t numExpStdDeviationsRead,
bool expDataFileAnnotated,
bool calc_sigma_from_data,
RealMatrix& xObsData,
RealMatrixArray& yObsData,
RealMatrixArray& yStdData)
{
//using boost::multi_array;
//using boost::extents;
size_t i,j,k, total_num_rows=0, max_replicates=1;
//for now, numExperiments are the same for all functions
//numReplicates can vary per experiment
for (i=0; i<numExperiments; i++){
total_num_rows+=numReplicates[i];
max_replicates = (numReplicates[i]>max_replicates) ? numReplicates[i] : max_replicates;
}
Cout << "total number of rows " << total_num_rows << '\n';
Cout << "max_replicates " << max_replicates << '\n';
Cout << "numFunctions " << numFunctions << '\n';
//yObsFull(extents[numFunctions][numExperiments][max_replicates]);
// Read from a matrix with numExperiments rows and a number of cols
// columns: numExpConfigVars X, numFunctions Y, [numFunctions Sigma]
RealMatrix experimental_data;
size_t num_cols = numExpConfigVars + numFunctions + numExpStdDeviationsRead;
TabularIO::read_data_tabular(expDataFileName, context_message,
experimental_data, total_num_rows , num_cols,
expDataFileAnnotated);
// Get views of the data in 3 matrices for convenience
size_t start_row, start_col;
if (numExpConfigVars > 0) {
start_row = 0;
start_col = 0;
RealMatrix x_obs_data(Teuchos::View, experimental_data,
total_num_rows, numExpConfigVars,
start_row, start_col);
xObsData.reshape(total_num_rows, numExpConfigVars);
for (i=0; i<total_num_rows; i++)
for (j=0; j<numExpConfigVars; j++)
xObsData(i,j) = x_obs_data(i,j);
}
start_row = 0;
start_col = numExpConfigVars;
RealMatrix y_obs_data(Teuchos::View, experimental_data,
total_num_rows, numFunctions,
start_row, start_col);
yObsData.resize(numFunctions);
size_t numrows_thusfar = 0;
for (j=0; j<numFunctions; j++)
yObsData[j].reshape(numExperiments,max_replicates);
for (j=0; j<numFunctions; j++){
size_t numrows_thusfar = 0;
for (i=0; i<numExperiments; i++){
for (k=0; k<numReplicates(i); k++){
yObsData[j](i,k) = y_obs_data(numrows_thusfar+k,j);
Cout << yObsData[j];
}
numrows_thusfar +=numReplicates(i);
}
}
// BMA TODO: The number of experimental functions may not match the
// user functions, so can't assume numFunctions
yStdData.resize(numFunctions);
for (j=0; j<numFunctions; j++)
yStdData[j].reshape(numExperiments,max_replicates);
if (numExpStdDeviationsRead > 0) {
start_row = 0;
start_col = numExpConfigVars + numFunctions;
RealMatrix y_std_data(Teuchos::View, experimental_data,
total_num_rows, numExpStdDeviationsRead,
start_row, start_col);
size_t numrows_thusfar = 0;
// We allow 1 or numFunctions sigmas
for (j=0; j<numFunctions; j++){
size_t numrows_thusfar = 0;
for (i=0; i<numExperiments; i++){
for (k=0; k<numReplicates(i); k++){
if (numExpStdDeviationsRead == 1)
yStdData[j](i,k) = y_std_data(numrows_thusfar+k,0);
else
yStdData[j](i,k) = y_std_data(numrows_thusfar+k,j);
Cout << yStdData[j];
}
numrows_thusfar +=numReplicates(i);
}
}
// user values? Commenting out as we don't currently support
// input file-specified errors.
// if (expStdDeviations.length()==1) {
// for (int i=0; i<numExperiments; i++)
// for (int j=0; j<numFunctions; j++)
// yStdData(i,j) = expStdDeviations(0);
// }
// else if (expStdDeviations.length()==numFunctions) {
// for (int i=0; i<numExperiments; i++)
// for (int j=0; j<numFunctions; j++)
// yStdData(i,j) = expStdDeviations(j);
// }
}
else if (calc_sigma_from_data) {
// calculate sigma terms
Real mean_est, var_est;
for (j=0; j<numFunctions; j++){
for (i=0; i<numExperiments; i++) {
mean_est = 0;
for (k=0; k<numReplicates(i); k++)
mean_est += yObsData[j](i,k);
mean_est = mean_est / ((Real)numReplicates(i));
var_est = 0;
for (k=0; k<numReplicates(i); k++)
var_est += (yObsData[j](i,k)-mean_est)*(yObsData[j](i,k)-mean_est);
// If only one data, point, use 1.0 in the likelihood (no weight)
for (k=0; k<numReplicates(i); k++)
yStdData[j](i,k) = (numReplicates(i) > 1) ?
std::sqrt(var_est/(Real)(numReplicates(i)-1)) : 1.0;
}
}
}
else {
// Default: use 1.0 in the likelihood (no weight)
for (j=0; j<numFunctions; j++)
for (i=0; i<numExperiments; i++)
for (k=0; k<numReplicates(i); k++)
yStdData[j](i,k) = 1.0;
}
}
