Optimized Vision Transformer Training using GPU and Multi-threading

This is the official repository for the paper titled, "Optimized Vision Transformer Training using GPU and Multi-threading," published at the IEEE Conference on Artificial Intelligence 2024 (IEEE CAI 2024). This repository contains an optimized implementation of Convolutional Neural Networks (CNN), Transformer, and Vision Transformer (ViT) models.

Authors

• Jonathan Ledet (@jonledet)
• Ashok Kumar
• Dominick Rizk
• Rodrigue Rizk
• KC Santosh

Overview

This project focuses on optimizing Vision Transformer training using GPU acceleration and multi-threading techniques. It provides implementations of popular deep learning models, including Convolutional Neural Networks (CNN), Transformer, and a customized version of Vision Transformer (ViT) tailored for improved performance.

Contents

• CNN: Implementation of Convolutional Neural Networks.
• Transformer: Implementation of the Transformer model.
• ViT: Customized version of the Vision Transformer (ViT) model, based on the vision-transformers-cifar10 repository.

Getting Started

Prerequisites

• Python (>=3.6)
• Anaconda 3
• PyTorch
• CUDA-enabled GPU (for GPU acceleration)

Installation

1. Clone this repository:

   git clone https://github.com/jonledet/vision-transformer.git

2. Create and activate a new Anaconda environment:

   conda create --name your-env-name python=3.6
   conda activate your-env-name

3. Install dependencies:

   pip install -r requirements.txt

Usage

• To run the models, execute the corresponding Python script:

python cnn.py

python transformer.py

python vit.py

Acknowledgments

• The Vision Transformer (ViT) model is based on the work from the vision-transformers-cifar10 repository.

License

This project is licensed under the MIT License.