FullyConnected::FullyConnected(OutputInfo info, int J, bool bias,
ActivationFunction act, double stdDev,
double maxSquaredWeightNorm)
: I(info.outputs()), J(J), bias(bias), act(act), stdDev(stdDev),
maxSquaredWeightNorm(maxSquaredWeightNorm), W(J, I), Wd(J, I),
b(J), bd(J), x(0), a(1, J), y(1, J), yd(1, J), deltas(1, J), e(1, I)
{
}
Subsampling::Subsampling(OutputInfo info, int kernelRows, int kernelCols,
bool bias, ActivationFunction act, double stdDev,
Regularization regularization)
: I(info.outputs()), fm(info.dimensions[0]), inRows(info.dimensions[1]),
inCols(info.dimensions[2]), kernelRows(kernelRows),
kernelCols(kernelCols), bias(bias), act(act), stdDev(stdDev), x(0),
e(1, I), fmInSize(-1), outRows(-1), outCols(-1), fmOutSize(-1),
maxRow(-1), maxCol(-1), regularization(regularization)
{
}
OutputInfo MaxPooling::initialize(std::vector<double*>& parameterPointers,
std::vector<double*>& parameterDerivativePointers)
{
OutputInfo info;
info.dimensions.push_back(fm);
outRows = inRows / kernelRows;
outCols = inCols / kernelCols;
fmOutSize = outRows * outCols;
info.dimensions.push_back(outRows);
info.dimensions.push_back(outCols);
fmInSize = inRows * inCols;
maxRow = inRows - kernelRows + 1;
maxCol = inCols - kernelCols + 1;
y.resize(1, info.outputs());
deltas.resize(1, info.outputs());
if(info.outputs() < 1)
throw OpenANNException("Number of outputs in max-pooling layer is below"
" 1. You should either choose a smaller filter"
" size or generate a bigger input.");
return info;
}
Dropout::Dropout(OutputInfo info, double dropoutProbability)
: info(info), I(info.outputs()),
dropoutProbability(dropoutProbability), y(1, I), dropoutMask(1, I), e(1, I)
{
}
OutputInfo Subsampling::initialize(std::vector<double*>& parameterPointers,
std::vector<double*>& parameterDerivativePointers)
{
OutputInfo info;
info.dimensions.push_back(fm);
outRows = inRows / kernelRows;
outCols = inCols / kernelCols;
fmOutSize = outRows * outCols;
info.dimensions.push_back(outRows);
info.dimensions.push_back(outCols);
fmInSize = inRows * inCols;
maxRow = inRows - kernelRows + 1;
maxCol = inCols - kernelCols + 1;
W.resize(fm, Eigen::MatrixXd(outRows, outCols));
Wd.resize(fm, Eigen::MatrixXd(outRows, outCols));
int numParams = fm * outRows * outCols * kernelRows * kernelCols;
if(bias)
{
Wb.resize(fm, Eigen::MatrixXd(outRows, outCols));
Wbd.resize(fm, Eigen::MatrixXd(outRows, outCols));
numParams += fm * outRows * outCols;
}
parameterPointers.reserve(parameterPointers.size() + numParams);
parameterDerivativePointers.reserve(parameterDerivativePointers.size() + numParams);
for(int fmo = 0; fmo < fm; fmo++)
{
for(int r = 0; r < outRows; r++)
{
for(int c = 0; c < outCols; c++)
{
parameterPointers.push_back(&W[fmo](r, c));
parameterDerivativePointers.push_back(&Wd[fmo](r, c));
if(bias)
{
parameterPointers.push_back(&Wb[fmo](r, c));
parameterDerivativePointers.push_back(&Wbd[fmo](r, c));
}
}
}
}
initializeParameters();
a.resize(1, info.outputs());
y.resize(1, info.outputs());
yd.resize(1, info.outputs());
deltas.resize(1, info.outputs());
if(info.outputs() < 1)
throw OpenANNException("Number of outputs in subsampling layer is below"
" 1. You should either choose a smaller filter"
" size or generate a bigger input.");
OPENANN_CHECK(fmInSize > 0);
OPENANN_CHECK(outRows > 0);
OPENANN_CHECK(outCols > 0);
OPENANN_CHECK(fmOutSize > 0);
OPENANN_CHECK(maxRow > 0);
OPENANN_CHECK(maxCol > 0);
return info;
}
MaxPooling::MaxPooling(OutputInfo info, int kernelRows, int kernelCols)
: I(info.outputs()), fm(info.dimensions[0]),
inRows(info.dimensions[1]), inCols(info.dimensions[2]),
kernelRows(kernelRows), kernelCols(kernelCols), x(0), e(1, I)
{
}