Overview

Deep neural network implementation from scratch in C++ with CUDA for GPU acceleration. This project focuses on creating a neural network to classify digits using the MNIST dataset.

The main features include:

Matrix and vector operations optimized for GPU
Feedforward neural network with customizable architecture
Training using backpropagation and gradient descent
MNIST digit classification

System Requirements

This project was developed and tested on a system with the following specifications:

OS: Arch Linux x86_64
CPU: Intel i7-9750H (12) @ 4.500GHz
GPU: NVIDIA GeForce GTX 1650 Mobile / Max-Q
Memory: 16GB
NVIDIA Driver Version: 560.35.03
CUDA Version: 12.6
CUDA Compilation Tools: Release 12.6, V12.6.77

Please ensure your system meets these requirements or adjust accordingly.

Installation

Clone the repository:
git clone https://github.com/juan-esteban-berger/cuda_networks.git

cd cuda_networks
Ensure you have the CUDA Toolkit (version 12.6 or compatible) installed on your system.
Install the Google Test library if not already present:
sudo pacman -S gtest
Build the project:
make

Usage

To run the MNIST classifier:

make run

This will execute the main program, which trains the neural network on the MNIST dataset and displays the training progress and final accuracy.

Demo of CUDA Neural Networks

Running Tests

To run the test suite:

make test

This will execute all the unit tests for the project.

Memory Check

To run the compute sanitizer memory check:

make memcheck

This will use NVIDIA's Compute Sanitizer to check for memory errors in the CUDA code.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Modifications

If you encounter any issues running the project, you may need to:

Adjust the NVCCFLAGS in the Makefile to match your GPU architecture.
Ensure the paths to the CUDA Toolkit and Google Test library are correct for your system.
If you're using a different CUDA version, you might need to update some CUDA function calls or syntax.