MatrixFlow C++ Deep Learning Library (TensorFlow from scratch)

MatrixFlow is a lightweight C++ library designed to build and train neural networks from scratch. It implements core deep learning concepts like matrix operations, layers stacking, forward propagation, and backward propagation.

✨ Features

• Matrix Operations: A robust Matrix class with support for:
  • Dot products, transpose, and element-wise operations.
  • Scalar mathematics and randomization.
• Layers:
  • Dense: Fully connected layers with learnable weights and biases.
  • Activation: Base class for activation functions.
• Activation Functions:
  • ReLU: Rectified Linear Unit.
  • Sigmoid: Standard sigmoid activation.
• Model API:
  • Sequential: A container to stack layers and manage forward/backward passes.
• Training:
  • Trainer: Handles batch generation, loss calculation (MSE), and the training loop.

🚀 Getting Started

Prerequisites

• A C++ compiler (e.g., g++ or clang++).
• Standard Library support (C++11 or later).

Compilation

You can compile the library with your main.cpp using the following command:

g++ main.cpp -o main.exe

Running

To run the compiled executable:

./main.exe

📚 Usage Example

Here is a quick example of defining a XOR model using MatrixFlow:

#include "MatrixFlow.h"
#include "Matrix.h"

int main() {
    // 1. Prepare Data (XOR)
    Matrix X(4, 2);
    X.data = {0,0, 1,0, 0,1, 1,1};

    Matrix Y(4, 1);
    Y.data = {0, 1, 1, 0};

    // 2. Define Architecture
    Sequential model({
        new Dense(2, 10),
        new ReLU(),
        new Dense(10, 1),
        new Sigmoid()
    });

    // or
    // Sequential model;
    // model.add(new Dense(2, 10));
    // model.add(new ReLU());
    // model.add(new Dense(10, 1));
    // model.add(new Sigmoid());

    // 3. Setup Trainer
    Trainer trainer(model, X, Y, 10000, 4, 0.1, 1000);

    // 4. Train
    trainer.train();

    // 5. Predict
    model.forward(X).print();

    return 0;
}

📁 File Structure

• Matrix.h: Core matrix manipulation library.
• MatrixFlow.h: Neural network components (Layers, Model, Trainer).
• main.cpp: Example entry point demonstrating an XOR implementation.

🛠️ Built With

• Language: C++
• Focus: Object-Oriented Programming (OOP)
• Domain: Deep Learning

To Do

• Add more layers (Flaten, Softmax, etc.).
• Add more activation functions (Tanh, Leaky ReLU, etc.).
• Add optimizers (SGD, Adam, etc.).
• Add custom loss functions options.