ffn_gradient_function

ffn_gradient_function.hpp

// Current implementation.
template<typename OutputLayerType, typename InitializationRuleType>
void FFN<OutputLayerType, InitializationRuleType>::Gradient()
{
  boost::apply_visitor(GradientVisitor(std::move(currentInput), std::move(
      boost::apply_visitor(deltaVisitor, network[1]))), network.front());

  for (size_t i = 1; i < network.size() - 1; ++i)
  {
    boost::apply_visitor(GradientVisitor(std::move(boost::apply_visitor(
        outputParameterVisitor, network[i - 1])), std::move(
        boost::apply_visitor(deltaVisitor, network[i + 1]))), network[i]);
  }

  boost::apply_visitor(GradientVisitor(std::move(boost::apply_visitor(
      outputParameterVisitor, network[network.size() - 2])), std::move(error)),
      network[network.size() - 1]);
}

// Slightly modified implementation to only update a single layer.
// Suppose we have the index of the layer we like to update,
// that could be set before training.
template<typename OutputLayerType, typename InitializationRuleType>
void FFN<OutputLayerType, InitializationRuleType>::Gradient()
{
  if (index == 0) // The first layer.
  {
    boost::apply_visitor(GradientVisitor(std::move(currentInput), std::move(
      boost::apply_visitor(deltaVisitor, network[1]))), network.front());
  }
  else if (index == network.size() - 1) // The last layer.
  {
    boost::apply_visitor(GradientVisitor(std::move(boost::apply_visitor(
      outputParameterVisitor, network[network.size() - 2])), std::move(error)),
      network[network.size() - 1]);
  }
  else
  {
    boost::apply_visitor(GradientVisitor(std::move(boost::apply_visitor(
        outputParameterVisitor, network[index - 1])), std::move(
        boost::apply_visitor(deltaVisitor, network[index + 1]))), network[index]);    
  }
}