void Subsampling::backpropagate(Eigen::MatrixXd* ein, Eigen::MatrixXd*& eout,
bool backpropToPrevious)
{
const int N = a.rows();
yd.conservativeResize(N, Eigen::NoChange);
e.conservativeResize(N, Eigen::NoChange);
// Derive activations
activationFunctionDerivative(act, y, yd);
deltas = yd.cwiseProduct(*ein);
e.setZero();
for(int fmo = 0; fmo < fm; fmo++)
{
Wd[fmo].setZero();
if(bias)
Wbd[fmo].setZero();
}
for(int n = 0; n < N; n++)
{
int outputIdx = 0;
for(int fmo = 0; fmo < fm; fmo++)
{
for(int ri = 0, ro = 0; ri < maxRow; ri += kernelRows, ro++)
{
int rowBase = fmo * fmInSize + ri * inCols;
for(int ci = 0, co = 0; ci < maxCol;
ci += kernelCols, co++, outputIdx++)
{
const double d = deltas(n, outputIdx);
for(int kr = 0; kr < kernelRows; kr++)
{
for(int kc = 0, inputIdx = rowBase + ci; kc < kernelCols;
kc++, inputIdx++)
{
e(n, inputIdx) += W[fmo](ro, co) * d;
Wd[fmo](ro, co) += d * (*x)(n, inputIdx);
}
}
if(bias)
Wbd[fmo](ro, co) += d;
}
}
}
}
if(regularization.l1Penalty > 0.0)
{
for(int fmo = 0; fmo < fm; fmo++)
Wd[fmo].array() += regularization.l1Penalty * W[fmo].array() / W[fmo].array().abs();
}
if(regularization.l2Penalty > 0.0)
{
for(int fmo = 0; fmo < fm; fmo++)
Wd[fmo] += regularization.l2Penalty * W[fmo];
}
eout = &e;
}
void FullyConnected::backpropagate(Eigen::MatrixXd* ein, Eigen::MatrixXd*& eout)
{
const int N = a.rows();
yd.conservativeResize(N, Eigen::NoChange);
// Derive activations
activationFunctionDerivative(act, y, yd);
deltas = yd.cwiseProduct(*ein);
// Weight derivatives
Wd = deltas.transpose() * *x;
if(bias)
bd = deltas.colwise().sum().transpose();
// Prepare error signals for previous layer
e = deltas * W;
eout = &e;
}
void RBM::backpropagate(Eigen::MatrixXd* ein, Eigen::MatrixXd*& eout,
bool backpropToPrevious)
{
const int N = ph.rows();
phd.conservativeResize(N, Eigen::NoChange);
// Derive activations
activationFunctionDerivative(LOGISTIC, ph, phd);
deltas = phd.cwiseProduct(*ein);
if(backprop)
{
Wd = deltas.transpose() * v;
bhd = deltas.colwise().sum().transpose();
}
// Prepare error signals for previous layer
if(backpropToPrevious)
e = deltas * W;
eout = &e;
}