ELU.hpp

Go to the documentation of this file.

#ifndef SRC_MLNN_ELU_HPP_
#define SRC_MLNN_ELU_HPP_

#include <mlnn/layer/Layer.hpp>

namespace mic {
namespace mlnn {
namespace activation_function {

template <typename eT=float>
class ELU : public mic::mlnn::Layer<eT> {
public:

    ELU(size_t size_,std::string name_ = "ELU") :
        ELU(size_, 1, 1, name_)
    {

    }


    ELU(size_t height_, size_t width_, size_t depth_, std::string name_ = "ELU") :
        Layer<eT>::Layer(height_, width_, depth_,
                height_, width_, depth_,
                LayerTypes::ELU, name_)
    {

    }

    virtual ~ELU() {};

    void forward(bool test = false) {
        // Access the data of both matrices.
        eT* x = s['x']->data();
        eT* y = s['y']->data();

        // Iterate through elements.
        size_t size = (size_t) s['x']->rows() * s['x']->cols();
        for (size_t i = 0; i < size;  i++) {
            y[i] = x[i] > 0.0f ? x[i] : (expf(x[i]) - 1.0f);
        }//: for
    }

    void backward() {
        // Access the data of matrices.
        eT* gx = g['x']->data();
        eT* gy = g['y']->data();
        eT* y = s['y']->data();

        // Iterate through elements.
        size_t size = (size_t) g['x']->rows() * g['x']->cols();
        for (size_t i = 0; i < size;  i++) {
            // Calculate the ELU y derivative.
            eT dy = y[i] > 0.0f ? 1.0f : exp(y[i]);
            // Pass the gradient.
            gx[i] = dy * gy[i];

        }//: for
    }

    virtual void update(eT alpha_, eT decay_  = 0.0f) { };

protected:
    // Unhiding the template inherited fields via "using" statement.
    using Layer<eT>::g;
    using Layer<eT>::s;

private:
    // Friend class - required for using boost serialization.
    template<typename tmp> friend class mic::mlnn::MultiLayerNeuralNetwork;

    ELU<eT>() : Layer<eT> () { }

};

} /* activation_function */
} /* namespace mlnn */
} /* namespace mic */

#endif /* SRC_MLNN_ELU_HPP_ */