NeuralNetwork Class Reference

Represents a simple feedforward neural network with one hidden layer. More...

#include <neural_network.h>

Public Member Functions
	NeuralNetwork (int input_size, int hidden_size, int output_size)
	Construct a new NeuralNetwork object. More...
	~NeuralNetwork ()
	Destroy the NeuralNetwork object. More...
void	initialize ()
	Initialize the neural network parameters. More...
void	forward (const Matrix &X)
	Perform forward propagation through the network. More...
void	backward (const Matrix &X, const Matrix &Y)
	Perform backward propagation through the network. More...
void	update_params (double learning_rate)
	Updates the network parameters based on computed gradients. More...
Vector	get_predictions () const
	Get predictions from the output layer (A2) More...
double	get_accuracy (const Matrix &Y) const
	Calculate the accuracy of predictions compared to true labels. More...
void	gradient_descent (const Matrix &X, const Matrix &Y, double learning_rate, int epochs)
	Perform gradient descent to train the neural network. More...
double *	get_W1_data () const
	Get the pointer to the W1 matrix data. More...
double *	get_W2_data () const
	Get the pointer to the W2 matrix data. More...
double *	get_b1_data () const
	Get the pointer to the b1 vector data. More...
double *	get_b2_data () const
	Get the pointer to the b2 vector data. More...
std::pair< int, int >	get_W1_dimensions () const
	Get the dimensions of the W1 matrix. More...
std::pair< int, int >	get_W2_dimensions () const
	Get the dimensions of the W2 matrix. More...
int	get_b1_size () const
	Get the size of the b1 vector. More...
int	get_b2_size () const
	Get the size of the b2 vector. More...
Matrix	get_DW1 () const
	Get the DW1 matrix. More...
double	get_db1 () const
	Get the db1 scalar. More...
Matrix	get_DW2 () const
	Get the DW2 matrix. More...
double	get_db2 () const
	Get the db2 scalar. More...
double *	get_A_data () const
	Get the pointer to the A matrix data (input matrix) More...
std::pair< int, int >	get_A_dimensions () const
	Get the dimensions of the A matrix. More...
double *	get_Z1_data () const
	Get the pointer to the Z1 matrix data (pre-activation of hidden layer) More...
std::pair< int, int >	get_Z1_dimensions () const
	Get the dimensions of the Z1 matrix. More...
double *	get_A1_data () const
	Get the pointer to the A1 matrix data (activation of hidden layer) More...
std::pair< int, int >	get_A1_dimensions () const
	Get the dimensions of the A1 matrix. More...
double *	get_Z2_data () const
	Get the pointer to the Z2 matrix data (pre-activation of output layer) More...
std::pair< int, int >	get_Z2_dimensions () const
	Get the dimensions of the Z2 matrix. More...
double *	get_A2_data () const
	Get the pointer to the A2 matrix data (activation of output layer) More...
std::pair< int, int >	get_A2_dimensions () const
	Get the dimensions of the A2 matrix. More...

Detailed Description

Represents a simple feedforward neural network with one hidden layer.

Definition at line 16 of file neural_network.h.

Constructor & Destructor Documentation

NeuralNetwork()

NeuralNetwork::NeuralNetwork	(	int	input_size,
		int	hidden_size,
		int	output_size
	)

Construct a new NeuralNetwork object.

Parameters

input_size	Number of input features
hidden_size	Number of neurons in the hidden layer
output_size	Number of output classes

Definition at line 8 of file neural_network_constructor.cu.

    : input_size(input_size),
      hidden_size(hidden_size),
      output_size(output_size),
      W1(hidden_size, input_size),
      b1(hidden_size),
      W2(output_size, hidden_size),
      b2(output_size),
      A(input_size, 1),
      Z1(hidden_size, 1),
      A1(hidden_size, 1),
      Z2(output_size, 1),
      A2(output_size, 1),
      DZ2(output_size, 1),
      DW2(output_size, hidden_size),
      db2(0.0),
      DZ1(hidden_size, 1),
      DW1(hidden_size, input_size),
      db1(0.0) {
    // Initialize the neural network parameters
    initialize();
}

~NeuralNetwork()

NeuralNetwork::~NeuralNetwork

(

)

Destroy the NeuralNetwork object.

Definition at line 8 of file neural_network_destructor.cu.

                              {
    // The destructor for Matrix and Vector objects will be called automatically
    // to free the GPU memory
}

Member Function Documentation

backward()

void NeuralNetwork::backward	(	const Matrix &	X,
		const Matrix &	Y
	)

Perform backward propagation through the network.

Parameters

X	Input data matrix
Y	True labels matrix

Definition at line 8 of file neural_network_backward.cu.

                                                             {
    // Get the number of training examples
    int m = X.get_cols();
 
    // Compute the gradient of the output layer
    // DZ2 = A2 - Y
    DZ2 = A2.subtract(Y);
    // std::cout << "Gradient of output layer DZ2:" << std::endl;
    // DZ2.print(4);
 
    // Compute gradient for W2
    // DW2 = 1/m * DZ2 * A1^T
    DW2 = DZ2.multiply(A1.transpose());
    DW2.divide_scalar(m);
    // std::cout << "Gradient for W2:" << std::endl;
    // DW2.print(4);
 
    // Compute gradient for b2
    // db2 = 1/m * sum(DZ2)
    db2 = DZ2.sum() / m;
    // std::cout << "Gradient for b2: " << db2 << std::endl;
 
    // Compute the gradient of the hidden layer
    // DZ1 = W2^T * DZ2 .* ReLU'(Z1)
    Matrix W2_transpose = W2.transpose();
    DZ1 = W2_transpose.multiply(DZ2);
    Matrix Z1_relu_derivative = Z1.relu_derivative();
    DZ1 = DZ1.multiply_elementwise(Z1_relu_derivative);
    // std::cout << "Gradient of hidden layer DZ1:" << std::endl;
    // DZ1.print(4);
 
    // Compute gradient for W1
    // DW1 = 1/m * DZ1 * X^T
    DW1 = DZ1.multiply(X.transpose());
    DW1.divide_scalar(m);
    // std::cout << "Gradient for W1:" << std::endl;
    // DW1.print(4);
 
    // Compute gradient for b1
    // db1 = 1/m * sum(DZ1)
    db1 = DZ1.sum() / m;
    // std::cout << "Gradient for b1: " << db1 << std::endl;
}

forward()

void NeuralNetwork::forward

(

const Matrix &

X

)

Perform forward propagation through the network.

Parameters

X	Input data matrix

Definition at line 8 of file neural_network_forward.cu.

                                           {
    // Store the input matrix
    A = X.copy();
    // std::cout << "Input matrix A:" << std::endl;
    // A.print(4);
 
    // Compute the pre-activation of the hidden layer
    Z1 = W1.multiply(A);
    // std::cout << "Pre-activation of hidden layer Z1:" << std::endl;
    // Z1.print(4);
 
    // Add biases to the pre-activation
    Z1.add_vector(b1);
    // std::cout << "Z1 after adding biases:" << std::endl;
    // Z1.print(4);
 
    // Apply ReLU activation to the hidden layer
    A1 = Z1.relu();
    // std::cout << "Activation of hidden layer A1:" << std::endl;
    // A1.print(4);
 
    // Compute the pre-activation of the output layer
    Z2 = W2.multiply(A1);
    // std::cout << "Pre-activation of output layer Z2:" << std::endl;
    // Z2.print(4);
 
    // Add biases to the pre-activation
    Z2.add_vector(b2);
    // std::cout << "Z2 after adding biases:" << std::endl;
    // Z2.print(4);
 
    // Apply softmax activation to the output layer
    A2 = Z2.softmax();
    // std::cout << "Activation of output layer A2:" << std::endl;
    // A2.print(4);
}

get_A1_data()

double* NeuralNetwork::get_A1_data ( ) const

inline

Get the pointer to the A1 matrix data (activation of hidden layer)

Returns

Pointer to the A1 matrix data on the device

Definition at line 177 of file neural_network.h.

{ return A1.get_data(); }

get_A1_dimensions()

std::pair<int, int> NeuralNetwork::get_A1_dimensions ( ) const

inline

Get the dimensions of the A1 matrix.

Returns

std::pair<int, int> containing rows and columns of A1

Definition at line 183 of file neural_network.h.

{ return {A1.get_rows(), A1.get_cols()}; }

get_A2_data()

double* NeuralNetwork::get_A2_data ( ) const

inline

Get the pointer to the A2 matrix data (activation of output layer)

Returns

Pointer to the A2 matrix data on the device

Definition at line 201 of file neural_network.h.

{ return A2.get_data(); }

get_A2_dimensions()

std::pair<int, int> NeuralNetwork::get_A2_dimensions ( ) const

inline

Get the dimensions of the A2 matrix.

Returns

std::pair<int, int> containing rows and columns of A2

Definition at line 207 of file neural_network.h.

{ return {A2.get_rows(), A2.get_cols()}; }

get_A_data()

double* NeuralNetwork::get_A_data ( ) const

inline

Get the pointer to the A matrix data (input matrix)

Returns

Pointer to the A matrix data on the device

Definition at line 153 of file neural_network.h.

{ return A.get_data(); }

get_A_dimensions()

std::pair<int, int> NeuralNetwork::get_A_dimensions ( ) const

inline

Get the dimensions of the A matrix.

Returns

std::pair<int, int> containing rows and columns of A

Definition at line 159 of file neural_network.h.

{ return {A.get_rows(), A.get_cols()}; }

get_accuracy()

double NeuralNetwork::get_accuracy

(

const Matrix &

Y

)

const

Calculate the accuracy of predictions compared to true labels.

Parameters

Y	True labels matrix

Returns

Accuracy as a fraction of correct predictions

Definition at line 22 of file neural_network_get_accuracy.cu.

                                                        {
    // Get predictions
    Vector predictions = get_predictions();
    // std::cout << "Predictions:" << std::endl;
    // predictions.print(0);
 
    // Convert Y matrix to argmax form for comparison
    Vector Y_argmax = Y.argmax();
    // std::cout << "True labels (argmax):" << std::endl;
    // Y_argmax.print(0);
 
    // Allocate device memory for correct count
    int* d_correct_count;
    cudaMalloc(&d_correct_count, sizeof(int));
    cudaMemset(d_correct_count, 0, sizeof(int));
 
    // Define block and grid dimensions
    int threadsPerBlock = 256;
    int blocksPerGrid = (Y.get_cols() + threadsPerBlock - 1) / threadsPerBlock;
 
    // Launch kernel to calculate accuracy
    calculate_accuracy_kernel<<<blocksPerGrid, threadsPerBlock>>>(
        predictions.get_data(), Y_argmax.get_data(), Y.get_cols(), d_correct_count
    );
 
    // Copy correct count from device to host
    int h_correct_count;
    cudaMemcpy(&h_correct_count, d_correct_count, sizeof(int), cudaMemcpyDeviceToHost);
 
    // Calculate accuracy
    double accuracy = static_cast<double>(h_correct_count) / Y.get_cols();
 
    // Free device memory
    cudaFree(d_correct_count);
 
    return accuracy;
}

get_b1_data()

double* NeuralNetwork::get_b1_data ( ) const

inline

Get the pointer to the b1 vector data.

Returns

Pointer to the b1 vector data on the device

Definition at line 93 of file neural_network.h.

{ return b1.get_data(); }

get_b1_size()

int NeuralNetwork::get_b1_size ( ) const

inline

Get the size of the b1 vector.

Returns

Size of the b1 vector

Definition at line 117 of file neural_network.h.

{ return b1.get_rows(); }

get_b2_data()

double* NeuralNetwork::get_b2_data ( ) const

inline

Get the pointer to the b2 vector data.

Returns

Pointer to the b2 vector data on the device

Definition at line 99 of file neural_network.h.

{ return b2.get_data(); }

get_b2_size()

int NeuralNetwork::get_b2_size ( ) const

inline

Get the size of the b2 vector.

Returns

Size of the b2 vector

Definition at line 123 of file neural_network.h.

{ return b2.get_rows(); }

get_db1()

double NeuralNetwork::get_db1 ( ) const

inline

Get the db1 scalar.

Returns

The db1 scalar

Definition at line 135 of file neural_network.h.

{ return db1; }

get_db2()

double NeuralNetwork::get_db2 ( ) const

inline

Get the db2 scalar.

Returns

The db2 scalar

Definition at line 147 of file neural_network.h.

{ return db2; }

get_DW1()

Matrix NeuralNetwork::get_DW1 ( ) const

inline

Get the DW1 matrix.

Returns

The DW1 matrix

Definition at line 129 of file neural_network.h.

{ return DW1; }

get_DW2()

Matrix NeuralNetwork::get_DW2 ( ) const

inline

Get the DW2 matrix.

Returns

The DW2 matrix

Definition at line 141 of file neural_network.h.

{ return DW2; }

get_predictions()

Vector NeuralNetwork::get_predictions

(

)

const

Get predictions from the output layer (A2)

Returns

Vector containing the predicted class indices

Definition at line 7 of file neural_network_get_predictions.cu.

                                            {
    // Get the argmax of A2 along axis 0 (column-wise)
    return A2.argmax();
}

get_W1_data()

double* NeuralNetwork::get_W1_data ( ) const

inline

Get the pointer to the W1 matrix data.

Returns

Pointer to the W1 matrix data on the device

Definition at line 81 of file neural_network.h.

{ return W1.get_data(); }

get_W1_dimensions()

std::pair<int, int> NeuralNetwork::get_W1_dimensions ( ) const

inline

Get the dimensions of the W1 matrix.

Returns

std::pair<int, int> containing rows and columns of W1

Definition at line 105 of file neural_network.h.

{ return {W1.get_rows(), W1.get_cols()}; }

get_W2_data()

double* NeuralNetwork::get_W2_data ( ) const

inline

Get the pointer to the W2 matrix data.

Returns

Pointer to the W2 matrix data on the device

Definition at line 87 of file neural_network.h.

{ return W2.get_data(); }

get_W2_dimensions()

std::pair<int, int> NeuralNetwork::get_W2_dimensions ( ) const

inline

Get the dimensions of the W2 matrix.

Returns

std::pair<int, int> containing rows and columns of W2

Definition at line 111 of file neural_network.h.

{ return {W2.get_rows(), W2.get_cols()}; }

get_Z1_data()

double* NeuralNetwork::get_Z1_data ( ) const

inline

Get the pointer to the Z1 matrix data (pre-activation of hidden layer)

Returns

Pointer to the Z1 matrix data on the device

Definition at line 165 of file neural_network.h.

{ return Z1.get_data(); }

get_Z1_dimensions()

std::pair<int, int> NeuralNetwork::get_Z1_dimensions ( ) const

inline

Get the dimensions of the Z1 matrix.

Returns

std::pair<int, int> containing rows and columns of Z1

Definition at line 171 of file neural_network.h.

{ return {Z1.get_rows(), Z1.get_cols()}; }

get_Z2_data()

double* NeuralNetwork::get_Z2_data ( ) const

inline

Get the pointer to the Z2 matrix data (pre-activation of output layer)

Returns

Pointer to the Z2 matrix data on the device

Definition at line 189 of file neural_network.h.

{ return Z2.get_data(); }

get_Z2_dimensions()

std::pair<int, int> NeuralNetwork::get_Z2_dimensions ( ) const

inline

Get the dimensions of the Z2 matrix.

Returns

std::pair<int, int> containing rows and columns of Z2

Definition at line 195 of file neural_network.h.

{ return {Z2.get_rows(), Z2.get_cols()}; }

gradient_descent()

void NeuralNetwork::gradient_descent	(	const Matrix &	X,
		const Matrix &	Y,
		double	learning_rate,
		int	epochs
	)

Perform gradient descent to train the neural network.

Parameters

X	Input data matrix
Y	True labels matrix
learning_rate	Learning rate for parameter updates
epochs	Number of training epochs

Definition at line 10 of file neural_network_gradient_descent.cu.

                                                                                                       {
    const int bar_width = 50;
    std::string bar;
 
    // Iterate through epochs
    for (int epoch = 0; epoch < epochs; ++epoch) {
        // Perform forward propagation
        forward(X);
 
        // Perform backward propagation
        backward(X, Y);
 
        // Update parameters
        update_params(learning_rate);
 
        // Calculate accuracy for this epoch
        double accuracy = get_accuracy(Y);
 
        // Calculate progress
        float progress = static_cast<float>(epoch + 1) / epochs;
        int pos = static_cast<int>(bar_width * progress);
 
        // Update progress bar
        bar = "[";
        for (int i = 0; i < bar_width; ++i) {
            if (i < pos) bar += "=";
            else if (i == pos) bar += ">";
            else bar += " ";
        }
        bar += "] ";
 
        // Print progress bar and accuracy
        std::cout << "\r" << std::setw(3) << static_cast<int>(progress * 100.0) << "% "
                  << bar << std::setw(3) << epoch + 1 << "/" << std::setw(3) << epochs
                  << " - Accuracy: " << std::fixed << std::setprecision(4) << accuracy << std::flush;
    }
    // Move to the next line after completion
    std::cout << std::endl;
}

initialize()

void NeuralNetwork::initialize

(

)

Initialize the neural network parameters.

Definition at line 9 of file neural_network_initialize.cu.

                               {
    // Initialize W1 with random values
    W1.randomize();
    // Scale W1 by sqrt(2.0 / input_size) for better initial performance
    W1.multiply_scalar(std::sqrt(2.0 / input_size));
 
    // Initialize b1 with random values
    b1.randomize();
    // Scale b1 by 0.01 to keep initial values small
    b1.multiply_scalar(0.01);
 
    // Initialize W2 with random values
    W2.randomize();
    // Scale W2 by sqrt(2.0 / hidden_size) for better initial performance
    W2.multiply_scalar(std::sqrt(2.0 / hidden_size));
 
    // Initialize b2 with random values
    b2.randomize();
    // Scale b2 by 0.01 to keep initial values small
    b2.multiply_scalar(0.01);
 
    // Initialize other matrices with zeros
    A.initialize();
    Z1.initialize();
    A1.initialize();
    Z2.initialize();
    A2.initialize();
    DZ2.initialize();
    DW2.initialize();
    DZ1.initialize();
    DW1.initialize();
 
    // Initialize scalar gradients to zero
    db1 = 0.0;
    db2 = 0.0;
}

update_params()

void NeuralNetwork::update_params

(

double

learning_rate

)

Updates the network parameters based on computed gradients.

Parameters

learning_rate

The learning rate for the parameter update.

Definition at line 8 of file neural_network_update_params.cu.

                                                      {
    // Update weights for the first layer (W1)
    DW1.multiply_scalar(learning_rate);
    W1 = W1.subtract(DW1);
    // std::cout << "Updated W1:" << std::endl;
    // W1.print(4);
 
    // Update bias for the first layer (b1)
    b1.subtract_scalar(learning_rate * db1);
    // std::cout << "Updated b1:" << std::endl;
    // b1.print(4);
 
    // Update weights for the second layer (W2)
    DW2.multiply_scalar(learning_rate);
    W2 = W2.subtract(DW2);
    // std::cout << "Updated W2:" << std::endl;
    // W2.print(4);
 
    // Update bias for the second layer (b2)
    b2.subtract_scalar(learning_rate * db2);
    // std::cout << "Updated b2:" << std::endl;
    // b2.print(4);
}

---

The documentation for this class was generated from the following files:

• src/neural_network/neural_network.h
• src/neural_network/neural_network_backward.cu
• src/neural_network/neural_network_constructor.cu
• src/neural_network/neural_network_destructor.cu
• src/neural_network/neural_network_forward.cu
• src/neural_network/neural_network_get_accuracy.cu
• src/neural_network/neural_network_get_predictions.cu
• src/neural_network/neural_network_gradient_descent.cu
• src/neural_network/neural_network_initialize.cu
• src/neural_network/neural_network_update_params.cu