noisylinear_impl.hpp

Go to the documentation of this file.

#ifndef MLPACK_METHODS_ANN_LAYER_NOISYLINEAR_IMPL_HPP
#define MLPACK_METHODS_ANN_LAYER_NOISYLINEAR_IMPL_HPP

// In case it hasn't yet been included.
#include "noisylinear.hpp"

namespace mlpack {
namespace ann  {

template<typename InputDataType, typename OutputDataType>
NoisyLinear<InputDataType, OutputDataType>::NoisyLinear() :
    inSize(0),
    outSize(0)
{
  // Nothing to do here.
}

template<typename InputDataType, typename OutputDataType>
NoisyLinear<InputDataType, OutputDataType>::NoisyLinear(
  const NoisyLinear& layer) :
    inSize(layer.inSize),
    outSize(layer.outSize),
    weights(layer.weights)
{
  Reset();
}

template<typename InputDataType, typename OutputDataType>
NoisyLinear<InputDataType, OutputDataType>::NoisyLinear(
    const size_t inSize,
    const size_t outSize) :
    inSize(inSize),
    outSize(outSize)
{
  weights.set_size(WeightSize(), 1);
  weightEpsilon.set_size(outSize, inSize);
  biasEpsilon.set_size(outSize, 1);
}

template<typename InputDataType, typename OutputDataType>
NoisyLinear<InputDataType, OutputDataType>::NoisyLinear(
    NoisyLinear&& layer) :
    inSize(std::move(layer.inSize)),
    outSize(std::move(layer.outSize)),
    weights(std::move(layer.weights))
{
  layer.inSize = 0;
  layer.outSize = 0;
  layer.weights = nullptr;
  Reset();
}

template<typename InputDataType, typename OutputDataType>
NoisyLinear<InputDataType, OutputDataType>&
NoisyLinear<InputDataType, OutputDataType>::operator=(const NoisyLinear& layer)
{
  if (this != &layer)
  {
    inSize = layer.inSize;
    outSize = layer.outSize;
    weights = layer.weights;
    Reset();
  }
  return *this;
}

template<typename InputDataType, typename OutputDataType>
NoisyLinear<InputDataType, OutputDataType>&
NoisyLinear<InputDataType, OutputDataType>::operator=(NoisyLinear&& layer)
{
  if (this != &layer)
  {
    inSize = std::move(layer.inSize);
    layer.inSize = 0;
    outSize = std::move(layer.outSize);
    layer.outSize = 0;
    weights = std::move(layer.weights);
    layer.weights = nullptr;
    Reset();
  }
  return *this;
}

template<typename InputDataType, typename OutputDataType>
void NoisyLinear<InputDataType, OutputDataType>::Reset()
{
  weightMu = arma::mat(weights.memptr(),
      outSize, inSize, false, false);
  biasMu = arma::mat(weights.memptr() + weightMu.n_elem,
      outSize, 1, false, false);
  weightSigma = arma::mat(weights.memptr() + weightMu.n_elem + biasMu.n_elem,
      outSize, inSize, false, false);
  biasSigma = arma::mat(weights.memptr() + weightMu.n_elem * 2 + biasMu.n_elem,
      outSize, 1, false, false);
  this->ResetNoise();
}

template<typename InputDataType, typename OutputDataType>
void NoisyLinear<InputDataType, OutputDataType>::ResetNoise()
{
  arma::mat epsilonIn = arma::randn<arma::mat>(inSize, 1);
  epsilonIn = arma::sign(epsilonIn) % arma::sqrt(arma::abs(epsilonIn));
  arma::mat epsilonOut = arma::randn<arma::mat>(outSize, 1);
  epsilonOut = arma::sign(epsilonOut) % arma::sqrt(arma::abs(epsilonOut));
  weightEpsilon = epsilonOut * epsilonIn.t();
  biasEpsilon = epsilonOut;
}

template<typename InputDataType, typename OutputDataType>
void NoisyLinear<InputDataType, OutputDataType>::ResetParameters()
{
  const double muRange = 1 / std::sqrt(inSize);
  weightMu.randu();
  weightMu = muRange * (weightMu * 2 - 1);
  biasMu.randu();
  biasMu = muRange * (biasMu * 2 - 1);
  weightSigma.fill(0.5 / std::sqrt(inSize));
  biasSigma.fill(0.5 / std::sqrt(outSize));
}

template<typename InputDataType, typename OutputDataType>
template<typename eT>
void NoisyLinear<InputDataType, OutputDataType>::Forward(
    const arma::Mat<eT>& input, arma::Mat<eT>& output)
{
  weight = weightMu + weightSigma % weightEpsilon;
  bias = biasMu + biasSigma % biasEpsilon;
  output = weight * input;
  output.each_col() += bias;
}

template<typename InputDataType, typename OutputDataType>
template<typename eT>
void NoisyLinear<InputDataType, OutputDataType>::Backward(
    const arma::Mat<eT>& /* input */, const arma::Mat<eT>& gy, arma::Mat<eT>& g)
{
  g = weight.t() * gy;
}

template<typename InputDataType, typename OutputDataType>
template<typename eT>
void NoisyLinear<InputDataType, OutputDataType>::Gradient(
    const arma::Mat<eT>& input,
    const arma::Mat<eT>& error,
    arma::Mat<eT>& gradient)
{
  // Locally stored to prevent multiplication twice.
  arma::mat weightGrad = error * input.t();

  // Gradients for mu values.
  gradient.rows(0, weight.n_elem - 1) = arma::vectorise(weightGrad);
  gradient.rows(weight.n_elem, weight.n_elem + bias.n_elem - 1)
      = arma::sum(error, 1);

  // Gradients for sigma values.
  gradient.rows(weight.n_elem + bias.n_elem, gradient.n_elem - bias.n_elem - 1)
      = arma::vectorise(weightGrad % weightEpsilon);
  gradient.rows(gradient.n_elem - bias.n_elem, gradient.n_elem - 1)
      = arma::sum(error, 1) % biasEpsilon;
}

template<typename InputDataType, typename OutputDataType>
template<typename Archive>
void NoisyLinear<InputDataType, OutputDataType>::serialize(
    Archive& ar, const uint32_t /* version */)
{
  ar(CEREAL_NVP(inSize));
  ar(CEREAL_NVP(outSize));

  // This is inefficient, but we have to allocate this memory so that
  // WeightSetVisitor gets the right size.
  if (cereal::is_loading<Archive>())
    weights.set_size((outSize * inSize + outSize) * 2, 1);
}

} // namespace ann
} // namespace mlpack

#endif