A simple C++11 library for basic neural network operations on CPU.
- Feed-forward Neural Network: Supports basic feed-forward neural networks.
- Learning Algorithm: Implements a simple backpropagation algorithm for training.
- Customizable Activation Functions: Allows custom activation functions and their derivatives.
- Serialization: Supports saving and loading neural network architectures and weights.
- Template-based Design: Highly customizable with C++ templates.
To use the Neuronov Net library, simply include the neuronov_net.hpp header file in your project.
#include "neuronov_net.hpp"To create a neural network, you need to specify the architecture (number of neurons in each layer), activation functions, and a random number generator for initializing weights.
#include "neuronov_net.hpp"
#include <cmath>
#include <random>
// Define activation function and its derivative
float sigmoid(float x) {
return 1.0 / (1.0 + std::exp(-x));
}
float sigmoid_derivative(float x) {
const float s = sigmoid(x);
return s * (1.0 - s);
}
// Random number generator
std::random_device rd;
std::mt19937 gen(rd());
std::uniform_real_distribution<> dis(-1.0, 1.0);
auto random_number_generator = [&]() { return dis(gen); };
int main() {
// Define the architecture of the neural network
std::vector<std::size_t> architecture = {2, 3, 1}; // Input layer: 2 neurons, Hidden layer: 3 neurons, Output layer: 1 neuron
// Create the neural network
neuronov_net::perseptron net(architecture, sigmoid, sigmoid_derivative, random_number_generator);
// ...
}To perform a feed-forward pass, set the input values and call the feed_forward method.
// Set input values
auto input = net.get_input();
input[0].value = 0.5;
input[1].value = 0.2;
// Perform feed-forward pass
net.feed_forward();
// Get output values
const auto output = net.get_output();
for (const auto& neuron : output)
std::cout << "Output: " << neuron.value << '\n';To train the neural network, provide the correct results and a learning rate, then call the learn method.
// Correct results for the output layer
std::vector<float> correct_results = {0.8};
// Learning rate
float learn_rate = 0.1;
// Perform learning step
net.learn(correct_results, learn_rate);You can save the neural network architecture and weights to a stream and load them back.
// Save the neural network to a stream
{
std::ofstream output_stream("network.dat");
net.safe(output_stream);
}
// Load the neural network from a stream
{
std::ifstream input_stream("network.dat");
net.load(input_stream);
}Here is a simple example demonstrating the basic usage of the Neuronov Net library:
#include "neuronov_net.hpp"
#include <cmath>
#include <random>
#include <iostream>
#include <fstream>
// Define activation function and its derivative
float sigmoid(float x) {
return 1.0 / (1.0 + std::exp(-x));
}
float sigmoid_derivative(float x) {
float s = sigmoid(x);
return s * (1.0 - s);
}
// Random number generator
std::random_device rd;
std::mt19937 gen(rd());
std::uniform_real_distribution<> dis(-1.0, 1.0);
auto random_number_generator = [&]() { return dis(gen); };
int main() {
// Define the architecture of the neural network
std::vector<std::size_t> architecture = {2, 3, 1}; // Input layer: 2 neurons, Hidden layer: 3 neurons, Output layer: 1 neuron
// Create the neural network
neuronov_net::perseptron net(architecture, sigmoid, sigmoid_derivative, random_number_generator);
// Set input values
auto input = net.get_input();
input[0].value = 0.5;
input[1].value = 0.2;
// Perform feed-forward pass
net.feed_forward();
// Get output values
const auto& output = net.get_output();
for (const auto& neuron : output)
std::cout << "Output: " << neuron.value << '\n';
// Correct results for the output layer
std::vector<float> correct_results = {0.8};
// Learning rate
float learn_rate = 0.1;
// Perform learning step
net.learn(correct_results, learn_rate);
// Save the neural network to a stream
{
std::ofstream output_stream("network.dat");
net.safe(output_stream);
}
// Load the neural network from a stream
{
std::ifstream input_stream("network.dat");
net.load(input_stream);
}
return 0;
}This project is licensed under the MIT License. See the LICENSE file for details.