deep_network/src/layers/layer.hpp

#ifndef LAYER_HPP
#define LAYER_HPP

#include "debug.hpp"
#include "vector.hpp"
#include "shape.hpp"
#include <cmath>

namespace Layer{
  /**
  * An abstract class representing a layer of a deep network.
  * Such a layer can be seen as a map
  \f[
  \begin{array}{rcl}
  f:\mathbb{R}^n&\mapsto&\mathbb{R}^m\\
  x&\to&y
  \end{array}
  \f].
  */
  class Layer{
  public:
    /** The name of the layer. Used for debugging.*/
    string name;

    /** Size of the input vector.*/
    size_t n;

    /** Size of the output vector.*/
    size_t m;

    /** A reference to the input vector.*/
    Vector x;

    /** Computed output vector. Owned by the layer.*/
    Vector y;

    /** Computed input delta vector computed by back propagation algorithm. Owned by the layer.*/
    Vector d;

    Layer(size_t n,size_t m);

    ~Layer();

    /** Return the input size.*/
    size_t get_input_size() const;

    /** Return the output size.*/
    size_t get_output_size() const;

    /** Return a reference to the computed output vector.*/
    Vector get_output() const;

    /** Apply the layer to the input vector \c x. Vectors \c x_in_ref and \c x_out must be updated in consequence. Return a reference to x_out.*/
    virtual Vector feed_forward(Vector x)=0;

    /** Initialize nabla vectors which are used during gradient descent.*/
    virtual void init_nabla()=0;

    /** Apply back propagation algorithm on the delta output vector d. Used the input vector stored in x_in_ref during feedforward.
    Return a reference to the computed (and stored) input delta vector. Nabla vectors must be computed here.*/
    virtual Vector back_propagation(Vector e)=0;

    /** Update layer parameters using gradient descent algorithm with learning rate eta. */
    virtual void update(Real eta)=0;
  };

  inline
  Layer::Layer(size_t n_,size_t m_){
    n=n_;
    m=m_;
    y=init_vector(m);
    d=init_vector(n);
  }

  inline
  Layer::~Layer(){
    delete_vector(y);
    delete_vector(d);
  }

  inline size_t
  Layer::get_input_size() const{
    return n;
  }

  inline size_t
  Layer::get_output_size() const{
    return m;
  }

  inline Vector
  Layer::get_output() const{
    return y;
  }
}

#endif