Vision Transformer for Image Classification

Project Overview

This project implements a Vision Transformer (ViT) model for image classification, leveraging the power of transformer architectures that have revolutionized natural language processing for computer vision tasks.

Key Features

• Transfer Learning: Fine-tuned a pre-trained ViT model on a custom dataset
• Performance Optimization: Implemented techniques to reduce inference time while maintaining accuracy
• Interpretability: Added visualization tools to understand model decisions
• Deployment Pipeline: Created a streamlined pipeline for model deployment to Hugging Face Spaces
• Interactive Demo: Built a web interface for real-time image classification

Technologies Used

• PyTorch: Framework for model training and evaluation
• Hugging Face Transformers: For pre-trained model access and fine-tuning
• Weights & Biases: Experiment tracking and visualization
• Gradio: Web interface for the demo application
• Docker: Containerization for deployment

Results and Impact

The final model achieved 94.5% accuracy on the test set, with inference time reduced by 62% compared to the base model while maintaining performance within 1% of the original accuracy.