# Pastebin hiuZXY8l


class NormalSampler
{
  public:
  template<typename InputVecType, typename OutputVecType>
  static void Fn(const InputVecType& x, OutputVecType& y)
  {
    for (std::size_t i = 0; i < y.n_elem; i++)
      y[i] = math::RandNormal(x[2 * i], x[2 * i + 1);
  }
};

class BinomialSampler
{
  public:
  template<typename InputVecType, typename OutputVecType>
  static void Fn(const InputVecType& x, OutputVecType& y)
  {
    // this should be similar to core/math/random
    std::random_device rd;
    std::mt19937 gen(rd());
    for (std::size_t i = 0; i < y.n_elem; i++)
    {
      std::binomial_distribution dist(1, x[i]);
      y[i] = dist(gen);
    }
  }
};

/*
 * You can use the BaseLayer class if you define ActivationFunction::Deriv
 */
template <
    class ActivationFunction = NormalSampler,
    typename InputDataType = arma::mat,
    typename OutputDataType = arma::mat
>
class BaseProbabilisticLayer
{
  template<typename InputType, typename OutputType>
  void Forward(const InputType&& input, OutputType&& output)
  {
    ActivationFunction::Fn(input, output);
  }
};

template <
    class ActivationFunction = BinomialSampler,
    typename InputDataType = arma::mat,
    typename OutputDataType = arma::mat
>
using BinomialLayer = BaseProbabilisticLayer<
    ActivationFunction, InputDataType, OutputDataType>;

/*
 * In such a way Gibbs sampling (for a binary RBM) is just a cyclic forward propogation over the network 
 * VisibleLayer -> Linear -> SigmoidLayer -> BinomialSampler -> HiddenLayer -> Linear -> SigmoidLayer -> BinomialSampler
 * 
 * Of course the implementation of the ssRBM in this manner is a little tricky but this is possible.
 * 
 * Actually BinomialSampler is similar to your ConditionalDistribution and BinomialLayer is similar to the Try class.
 * But I think it is better to use the LayerType notation for that.
 */