neural_network

Neural Network Class Documentation

The neural_network class represents a feedforward neural network. It consists of multiple layers (neural_layer) and provides functionalities for forward propagation, backward propagation, and serialization/deserialization.

Class Structure

The neural_network class has the following structure:

class neural_network {
private:
    std::vector<neural_layer<double>> layers;
    double learning_rate = 0.4;

public:
    neural_network() = default;

    neural_network(std::initializer_list<neural_layer<double>> list);

    void add_layer(neural_layer<double>& layer);
    void add_layer(size_t input_size, size_t output_size);

    Matrix<double> inference(const Matrix<double>& input);
    Matrix<double> cost(const Matrix<double>& target, Matrix<double>& prediction);

    std::vector<std::pair<Matrix<double>, Matrix<double>>> forward(const Matrix<double>& input);
    Matrix<double> backward(const Matrix<double>& input, const Matrix<double>& label);

    std::vector<neural_layer<double>>& get_layers();
    neural_layer<double>& operator[](size_t pos);
    size_t size();
    double get_learning_rate() const;

    static std::string trim(std::string& str);

    void serialize(const std::string& filename);
    void deserialize(const std::string& filename);
};

classDiagram
    class neural_network {
        - std::vector<neural_layer<double>> layers
        - double learning_rate
        + neural_network()
        + neural_network(std::initializer_list<neural_layer<double>> list)
        + void add_layer(neural_layer<double>& layer)
        + void add_layer(size_t input_size, size_t output_size)
        + Matrix<double> inference(const Matrix<double>& input)
        + Matrix<double> cost(const Matrix<double>& target, Matrix<double>& prediction)
        + std::vector<std::pair<Matrix<double>,Matrix<double>>> forward(const Matrix<double>& input)
        + Matrix<double> backward(const Matrix<double>& input, const Matrix<double>& label)
        + std::vector<neural_layer<double>>& get_layers()
        + neural_layer<double>& operator[](size_t pos)
        + size_t size()
        + double get_learning_rate() const
        + static std::string trim(std::string &str)
        + void serialize(const string& filename)
        + void deserialize(const string& filename)
    }

Loading

Functionality

The neural_network class provides the following functionality:

• Constructors:

  • neural_network(): Default constructor.
  • neural_network(std::initializer_list<neural_layer<double>> list): Constructor that takes an initializer list of neural_layer objects and initializes the network with those layers.

• Layer Management:

  • void add_layer(neural_layer<double>& layer): Adds a neural_layer object to the network.
  • void add_layer(size_t input_size, size_t output_size): Creates a new neural_layer object with the specified input and output sizes and adds it to the network.

• Forward Propagation:

  • Matrix<double> inference(const Matrix<double>& input): Performs forward propagation on the network with the given input and returns the output prediction.
  • Matrix<double> cost(const Matrix<double>& target, Matrix<double>& prediction): Calculates the cost between the target output and the predicted output.

• Backward Propagation:

  • std::vector<std::pair<Matrix<double>, Matrix<double>>> forward(const Matrix<double>& input): Performs forward propagation on the network with the given input and returns a vector of pairs containing the weighted sum (Z) and activation (A) matrices for each layer.
  • Matrix<double> backward(const Matrix<double>& input, const Matrix<double>& label): Performs backward propagation on the network with the given input and label. Updates the weights and biases of each layer and returns the loss.

• Layer Access:

  • std::vector<neural_layer<double>>& get_layers(): Returns a reference to the vector of neural_layer objects in the network.
  • neural_layer<double>& operator[](size_t pos): Returns a reference to the neural_layer object at the specified position in the network.
  • size_t size(): Returns the number of layers in the network.

• Learning Rate:

  • double get_learning_rate() const: Returns the learning rate of the network.

• Serialization/Deserialization:

  • static std::string trim(std::string& str): Trims leading and trailing spaces from a string.
  • void serialize(const std::string& filename): Serializes the network to a file in a custom format.
  • void deserialize(const std::string& filename): Deserializes the network from a file in the custom format.

Implementation Details

• The neural_network class uses a vector of neural_layer objects to represent the layers of the network.
• The learning rate is set to a default value of 0.4 but can be modified.
• The inference function performs forward propagation on the network and returns the output prediction.
• The cost function calculates the cost between the target output and the predicted output.
• The forward function performs forward propagation on the network and returns a vector of pairs containing the weighted sum (Z) and activation (A) matrices for each layer.
• The backward function performs backward propagation on the network, updates the weights and biases of each layer, and returns the loss.
• The get_layers function returns a reference to the vector of neural_layer objects in the network.
• The operator[] function allows accessing a specific layer in the network.
• The size function returns the number of layers in the network.
• The serialize function serializes the network to a file in a custom format.
• The deserialize function deserializes the network from a file in the custom format.

Conclusion

The neural_network class provides a high-level interface for creating, training, and using feedforward neural networks. It encapsulates the layers, forward and backward propagation, and serialization/deserialization functionality. Developers can use this class to easily create and manipulate neural networks for various machine learning tasks.