inline static void Initialize(const MatType& V,
const size_t r,
arma::mat& W,
arma::mat& H)
{
const size_t n = V.n_rows;
const size_t m = V.n_cols;
double avgV = 0;
size_t count = 0;
double min = DBL_MAX;
// Iterate over all elements in the matrix (for sparse matrices, this only
// iterates over nonzeros).
for (typename MatType::const_row_col_iterator it = V.begin();
it != V.end(); ++it)
{
++count;
avgV += *it;
// Track the minimum value.
if (*it < min)
min = *it;
}
avgV = sqrt(((avgV / (n * m)) - min) / r);
// Initialize to random values.
W.randu(n, r);
H.randu(r, m);
W = W + avgV;
H = H + avgV;
}
inline static void Initialize(const MatType& V,
const size_t r,
arma::mat& W,
arma::mat& H)
{
size_t n = V.n_rows;
size_t m = V.n_cols;
double V_avg = 0;
size_t count = 0;
double min = DBL_MAX;
for(typename MatType::const_row_col_iterator it = V.begin();it != V.end();it++)
{
if(*it != 0)
{
count++;
V_avg += *it;
if(*it < min) min = *it;
}
}
V_avg = sqrt(((V_avg / (n * m)) - min) / r);
// Intialize to random values.
W.randu(n, r);
H.randu(r, m);
W = W + V_avg;
H = H + V_avg;
}
inline static void Initialize(const MatType& V,
const size_t r,
arma::mat& W,
arma::mat& H)
{
// Simple implementation (left in the header file due to its simplicity).
const size_t n = V.n_rows;
const size_t m = V.n_cols;
// Initialize to random values.
W.randu(n, r);
H.randu(r, m);
}
inline static void Initialize(const MatType& V,
const size_t r,
arma::mat& W,
arma::mat& H)
{
const size_t n = V.n_rows;
const size_t m = V.n_cols;
if (columnsToAverage > m)
{
Log::Warn << "Number of random columns (columnsToAverage) is more than "
<< "the number of columns available in the V matrix; weird results "
<< "may ensue!" << std::endl;
}
W.zeros(n, r);
// Initialize W matrix with random columns.
for (size_t col = 0; col < r; col++)
{
for (size_t randCol = 0; randCol < columnsToAverage; randCol++)
{
// .col() does not work in this case, as of Armadillo 3.920.
W.unsafe_col(col) += V.col(math::RandInt(0, m));
}
}
// Now divide by p.
W /= columnsToAverage;
// Initialize H to random values.
H.randu(r, m);
}
inline static void Initialize(arma::mat& weights,
arma::vec& biases,
const size_t numFeatures,
const size_t numClasses)
{
weights.randu(numFeatures, numClasses);
biases.randu(numClasses);
}
void MVU::Unfold(const size_t newDim,
const size_t numNeighbors,
arma::mat& outputData)
{
// First we have to choose the output point. We'll take a linear projection
// of the data for now (this is probably not a good final solution).
// outputData = trans(data.rows(0, newDim - 1));
// Following Nick's idea.
outputData.randu(data.n_cols, newDim);
// The number of constraints is the number of nearest neighbors plus one.
LRSDP<arma::sp_mat> mvuSolver(numNeighbors * data.n_cols + 1, outputData);
// Set up the objective. Because we are maximizing the trace of (R R^T),
// we'll instead state it as min(-I_n * (R R^T)), meaning C() is -I_n.
mvuSolver.C().eye(data.n_cols, data.n_cols);
mvuSolver.C() *= -1;
// Now set up each of the constraints.
// The first constraint is trace(ones * R * R^T) = 0.
mvuSolver.B()[0] = 0;
mvuSolver.A()[0].ones(data.n_cols, data.n_cols);
// All of our other constraints will be sparse except the first. So set that
// vector of modes accordingly.
mvuSolver.AModes().ones();
mvuSolver.AModes()[0] = 0;
// Now all of the other constraints. We first have to run AllkNN to get the
// list of nearest neighbors.
arma::Mat<size_t> neighbors;
arma::mat distances;
AllkNN allknn(data);
allknn.Search(numNeighbors, neighbors, distances);
// Add each of the other constraints. They are sparse constraints:
// Tr(A_ij K) = d_ij;
// A_ij = zeros except for 1 at (i, i), (j, j); -1 at (i, j), (j, i).
for (size_t i = 0; i < neighbors.n_cols; ++i)
{
for (size_t j = 0; j < numNeighbors; ++j)
{
// This is the index of the constraint.
const size_t index = (i * numNeighbors) + j + 1;
arma::mat& aRef = mvuSolver.A()[index];
aRef.set_size(3, 4);
// A_ij(i, i) = 1.
aRef(0, 0) = i;
aRef(1, 0) = i;
aRef(2, 0) = 1;
// A_ij(i, j) = -1.
aRef(0, 1) = i;
aRef(1, 1) = neighbors(j, i);
aRef(2, 1) = -1;
// A_ij(j, i) = -1.
aRef(0, 2) = neighbors(j, i);
aRef(1, 2) = i;
aRef(2, 2) = -1;
// A_ij(j, j) = 1.
aRef(0, 3) = neighbors(j, i);
aRef(1, 3) = neighbors(j, i);
aRef(2, 3) = 1;
// The constraint b_ij is the distance between these two points.
mvuSolver.B()[index] = distances(j, i);
}
}
// Now on with the solving.
double objective = mvuSolver.Optimize(outputData);
Log::Info << "Final objective is " << objective << "." << std::endl;
// Revert to original data format.
outputData = trans(outputData);
}
