Neural Networks in the C Language

This repository was created to deepen my understanding of neural networks, enhance my programming skills, and gain experience in building libraries in C.

Description

This project implements a library named Synapse for constructing various MLP (Multilayer Perceptron) architectures in C/C++. The library is optimized for fast training of large models and offers flexibility for easily adjusting hyperparameters. By adhering to encapsulation within the project’s modules, the library simplifies usage while maintaining full control over the training process.

Installation

To clone this repository, execute the following bash command:

git clone https://github.com/profjuvii/neural-net-from-scratch.git

Navigation

The repository is organized into several folders, each serving a specific purpose:

• datasets: Stores raw dataset archives before preprocessing. These files are used by the preparation scripts to generate CSV inputs for training.
• mnist_preparation: Includes a Python script for preparing handwritten digit data from the MNIST dataset in CSV format.
• mnist_training: Contains code for training a model on pre-processed MNIST data.
• models: Stores pre-trained models.
• synapse: Contains the core neural network library.

To train your own model, ensure that the necessary compiler extensions are included as follows:

-I../synapse/include -L../synapse/lib -lsynapse

Note

This depends on the location of the project files in your environment.

Requirements for Training the MNIST Model

To run the code for training a neural network on handwritten MNIST data, follow these steps:

1. Prepare MNIST Data:

   • In the mnist_preparation/ folder, you need set up a Python virtual environment:

     python -m venv .venv

   • Activate the virtual environment:

     source .venv/bin/activate

   • Install the necessary library:

     pip install idx2numpy

   • After this, you can run the MNIST data preparation script:

     python mnist2csv.py

2. Build and Run the Training Code:

   • In the mnist_training/ folder, run the following command to compile the code:

     make

   • Once the build is complete, execute the program using:

     ./train

Once these steps are completed, the training of the neural network on the MNIST dataset will begin as intended.

Thanks to the hyperparameter settings defined in the code, I achieved an accuracy of 94.2%. The trained model has been saved in the models/ folder.

Note

You can also find all the library functions in the header files located in synapse/include/.

Author

Created by Denys Bondarchuk. Feel free to reach out or contribute to the project!