This program presents a Neural Network capable of recognizing hand written digits. The project was developed for one of my student projects namely Parallel computing, along with Michael. The goal is to create 2 version of a neural network, one that is trained serially (written in C++) and one that is trained using one or more GPGPUs (written in CUDA), then prove the accerelation that HPC offers. The CUDA version can be found on the respective branch.
This repo contains 2 branches, one for each version of the code. The serial version exists primarily for benchmarking against the CUDA code. I am aware of the many design and performance flaws of the current project and do not intend to spend more time on it for now. However i will maintain this repo primarily for personal reasons (this is the first serious programming project i developed as a student, i would like to remember it).The training set was downloaded from the MNIST database. It consists of 60.000 sample images and their respective labels. Each image shows a digit ranging from 0-9.
The network is capable of correctly predicting more than 90% even on unknown data (image samples that were not used during training phase) when training with more than 20000 images. This should take at most 1 minute on any modern CPU. The CUDA version can be trained with the same dataset and therefore achieve the same results, in a fraction of that time, namely less than 10 seconds on a desktop GTX NVIDIA. A CMake file is provided for automating the build process. All you need to do is download CMake, navigate to the build folder and build:cd build && cmake && make
You can then run the application providing your image data and their labels for training. Sample data are included in this repo, you can run it like this:
./Main ../resources/train-images.idx3-ubyte ../resources/train-labels.idx1-ubyte