void OnlineRNNet::weightedParameterDerivative(RealMatrix const& pattern, const RealMatrix& coefficients, RealVector& gradient) {
SIZE_CHECK(pattern.size1()==1);//we can only process a single input at a time.
SIZE_CHECK(coefficients.size1()==1);
SIZE_CHECK(pattern.size2() == inputSize());
SIZE_CHECK(pattern.size2() == coefficients.size2());
gradient.resize(mpe_structure->parameters());
std::size_t numNeurons = mpe_structure->numberOfNeurons();
std::size_t numUnits = mpe_structure->numberOfUnits();
//first check wether this is the first call of the derivative after a change of internal structure. in this case we have to allocate A LOT
//of memory for the derivative and set it to zero.
if(m_unitGradient.size1() != mpe_structure->parameters() || m_unitGradient.size2() != numNeurons) {
m_unitGradient.resize(mpe_structure->parameters(),numNeurons);
m_unitGradient.clear();
}
//for the next steps see Kenji Doya, "Recurrent Networks: Learning Algorithms"
//calculate the derivative for all neurons f'
RealVector neuronDerivatives(numNeurons);
for(std::size_t i=0; i!=numNeurons; ++i) {
neuronDerivatives(i)=mpe_structure->neuronDerivative(m_activation(i+inputSize()+1));
}
//calculate the derivative for every weight using the derivative of the last time step
auto hiddenWeights = columns(
mpe_structure->weights(),
inputSize()+1,numUnits
);
//update the new gradient with the effect of last timestep
noalias(m_unitGradient) = prod(m_unitGradient,trans(hiddenWeights));
//add the effect of the current time step
std::size_t param = 0;
for(std::size_t i = 0; i != numNeurons; ++i) {
for(std::size_t j = 0; j != numUnits; ++j) {
if(mpe_structure->connection(i,j)) {
m_unitGradient(param,i) += m_lastActivation(j);
++param;
}
}
}
//multiply with outer derivative of the neurons
for(std::size_t i = 0; i != m_unitGradient.size1(); ++i) {
noalias(row(m_unitGradient,i)) = element_prod(row(m_unitGradient,i),neuronDerivatives);
}
//and formula 4 (the gradient itself)
noalias(gradient) = prod(
columns(m_unitGradient,numNeurons-outputSize(),numNeurons),
row(coefficients,0)
);
//sanity check
SIZE_CHECK(param == mpe_structure->parameters());
}
void CMACMap::eval(RealMatrix const& patterns,RealMatrix &output) const{
SIZE_CHECK(patterns.size2() == m_inputSize);
std::size_t numPatterns = patterns.size1();
output.resize(numPatterns, m_outputShape.numElements());
output.clear();
for(std::size_t i = 0; i != numPatterns; ++i){
std::vector<std::size_t> indizes = getIndizes(row(patterns,i));
for (std::size_t o = 0; o != output.size2(); ++o) {
for (std::size_t j = 0; j != m_tilings; ++j) {
output(i,o) += m_parameters(indizes[j] + o*m_parametersPerTiling);
}
}
}
}
// subtract the mean from each row
void ZeroSum(RealMatrix& mat)
{
RealVector sum(mat.size2(), 0.0);
for (size_t j=0; j<mat.size1(); j++) sum += row(mat, j);
RealVector mean = (1.0 / mat.size1()) * sum;
for (size_t j=0; j<mat.size1(); j++) row(mat, j) -= mean;
}
double evalDerivative(
RealMatrix const&,
RealMatrix const& prediction,
RealMatrix& gradient
) const {
gradient.resize(prediction.size1(),prediction.size2());
gradient.clear();
return 0;
}
void CMACMap::weightedParameterDerivative(
RealMatrix const& patterns,
RealMatrix const& coefficients,
State const&,//not needed
RealVector &gradient
) const {
SIZE_CHECK(patterns.size2() == m_inputSize);
SIZE_CHECK(coefficients.size2() == m_outputSize);
SIZE_CHECK(coefficients.size1() == patterns.size1());
std::size_t numPatterns = patterns.size1();
gradient.resize(m_parameters.size());
gradient.clear();
for(std::size_t i = 0; i != numPatterns; ++i) {
std::vector<std::size_t> indizes = getIndizes(row(patterns,i));
for (std::size_t o=0; o!=m_outputSize; ++o) {
for (std::size_t j=0; j != m_tilings; ++j) {
gradient(indizes[j] + o*m_parametersPerTiling) += coefficients(i,o);
}
}
}
}
void DiscreteLoss::defineCostMatrix(RealMatrix const& cost){
// check validity
std::size_t size = cost.size1();
SHARK_ASSERT(cost.size2() == size);
for (std::size_t i = 0; i != size; i++){
for (std::size_t j = 0; j != size; j++){
SHARK_ASSERT(cost(i, j) >= 0.0);
}
SHARK_ASSERT(cost(i, i) == 0.0);
}
m_cost = cost;
}
//! Returns a model mapping encoded data from the
//! m-dimensional PCA coordinate system back to the
//! n-dimensional original coordinate system.
void PCA::decoder(LinearModel<>& model, std::size_t m) {
if(!m) m = std::min(m_n,m_l);
if( m == m_n && !m_whitening){
model.setStructure(m_eigenvectors, m_mean);
}
RealMatrix A = columns(m_eigenvectors, 0, m);
if(m_whitening){
for(std::size_t i=0; i<A.size2(); i++) {
//take care of numerical difficulties for very small eigenvalues.
if(m_eigenvalues(i)/m_eigenvalues(0) < 1.e-15){
column(A,i).clear();
}
else{
column(A, i) = column(A, i) * std::sqrt(m_eigenvalues(i));
}
}
}
model.setStructure(A, m_mean);
}
RealMatrix NormalizeComponentsWhitening::createWhiteningMatrix(
RealMatrix& covariance
){
SIZE_CHECK(covariance.size1() == covariance.size2());
std::size_t m = covariance.size1();
//we use the inversed cholesky decomposition for whitening
//since we have to assume that covariance does not have full rank, we use
//the generalized decomposition
RealMatrix whiteningMatrix(m,m,0.0);
//do a pivoting cholesky decomposition
//this destroys the covariance matrix as it is not neeeded anymore afterwards.
PermutationMatrix permutation(m);
std::size_t rank = pivotingCholeskyDecompositionInPlace(covariance,permutation);
//only take the nonzero columns as C
auto C = columns(covariance,0,rank);
//full rank, means that we can use the typical cholesky inverse with pivoting
//so U is P C^-1 P^T
if(rank == m){
noalias(whiteningMatrix) = identity_matrix<double>( m );
solveTriangularSystemInPlace<SolveXAB,upper>(trans(C),whiteningMatrix);
swap_full_inverted(permutation,whiteningMatrix);
return whiteningMatrix;
}
//complex case.
//A' = P C(C^TC)^-1(C^TC)^-1 C^T P^T
//=> P^T U P = C(C^TC)^-1
//<=> P^T U P (C^TC) = C
RealMatrix CTC = prod(trans(C),C);
auto submat = columns(whiteningMatrix,0,rank);
solveSymmPosDefSystem<SolveXAB>(CTC,submat,C);
swap_full_inverted(permutation,whiteningMatrix);
return whiteningMatrix;
}
void OnlineRNNet::eval(RealMatrix const& pattern, RealMatrix& output){
SIZE_CHECK(pattern.size1()==1);//we can only process a single input at a time.
SIZE_CHECK(pattern.size2() == inputSize());
std::size_t numUnits = mpe_structure->numberOfUnits();
if(m_lastActivation.size() != numUnits){
m_activation.resize(numUnits);
m_lastActivation.resize(numUnits);
zero(m_activation);
zero(m_lastActivation);
}
swap(m_lastActivation,m_activation);
//we want to treat input and bias neurons exactly as hidden or output neurons, so we copy the current
//pattern at the beginning of the the last activation pattern aand set the bias neuron to 1
////so m_lastActivation has the format (input|1|lastNeuronActivation)
noalias(subrange(m_lastActivation,0,mpe_structure->inputs())) = row(pattern,0);
m_lastActivation(mpe_structure->bias())=1;
m_activation(mpe_structure->bias())=1;
//activation of the hidden neurons is now just a matrix vector multiplication
fast_prod(
mpe_structure->weights(),
m_lastActivation,
subrange(m_activation,inputSize()+1,numUnits)
);
//now apply the sigmoid function
for (std::size_t i = inputSize()+1;i != numUnits;i++){
m_activation(i) = mpe_structure->neuron(m_activation(i));
}
//copy the result to the output
output.resize(1,outputSize());
noalias(row(output,0)) = subrange(m_activation,numUnits-outputSize(),numUnits);
}
