CUDA Neural Network

This is a simple neural network built using CUDA for matrix operations.

Features

• Fully connected (MLP) network
• Convolutional network
• Optimisers
  • Stochastic Gradient Descent (SGD)
  • Gradient normalised clipping
  • SGD with momentum
  • Adam
• 2D matrix operations
• 3D matrix operations (to include batch)

Dependencies

The library has been designed to have as few dependencies as possible with the only dependencies for the core library:

• cuda (build and runtime)
• spdlog (build)
• MNIST (runtime)
• Compiler with C++20 support
• CMake 3.22 or newer

MNIST and spdlog are automatically fetched via CMake.

Building

Build via CMake

cmake --preset release
cmake --build --preset release --target cuda-nn --parallel 8