Vision-Transformer-Image-Classification

A Vision Transformer (ViT) implementation for image classification using CIFAR-10 dataset, leveraging HuggingFace's Trainer API for computational efficiency

Vision Transformer for Image Classification

Overview

This repository contains an implementation of the Vision Transformer (ViT) model, a novel architecture leveraging self-attention mechanisms for image classification tasks. Unlike traditional CNNs, ViT splits images into patches and processes them as sequences, enabling the model to capture global context effectively.

Objective

• To explore the capabilities of Vision Transformer on the CIFAR-10 dataset.
• To compare its performance with traditional CNN models.
• To implement and evaluate using HuggingFace's Trainer API for improved computational efficiency.

Methodology

1. Dataset: CIFAR-10 (60,000 32x32 images across 10 classes).
2. Preprocessing: Data augmentation and patch embedding for input preparation.
3. Model Architecture: Implementation of Vision Transformer with patch encoding and positional encoding.
4. Training: Leveraged HuggingFace's Trainer API to streamline training and overcome computational limitations.
5. Evaluation: Achieved high accuracy through transfer learning and efficient training.

Results

• Accuracy: Reached 98.6% validation accuracy by epoch 3.
• Efficiency: Demonstrated the use of pre-trained weights and transfer learning for computationally constrained setups.

Challenges

Faced computational resource constraints but overcame them using HuggingFace’s Trainer API, reducing the training burden while maintaining accuracy.

Usage

1. Clone the repository:

   git clone https://github.com/KhushiRajurkar/Vision-Transformer-Image-Classification.git