👁️ Vision Transformer

This repository contains an implementation of the Vision Transformer (ViT) architecture built from scratch.

📄 Reference Paper

The architecture is based on the paper:

"An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale"
Authors: Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, et al.
Published by Google Research, Brain Team.
Paper link

🛠️ Features of This Implementation

• Tokenization: Patch-based image splitting to create input tokens.
• Transformer Encoder: Multi-head self-attention and feed-forward layers.
• Positional Encoding: Integration of positional information into input tokens.
• Classifier Head: Fully connected layer for classification tasks.
• Modular Design: Clean and reusable code for core components like tokenizers, encoders, and heads.

🚀 Getting Started

Prerequisites

Ensure you have Python 3.8+ installed and install the requirements.txt using

pip install -r requirements.txt

Running the Code

1. Clone this repository:

   git clone https://github.com/Ruhaan838/Vision-Transformer
   cd Vision-Transformer

2. Train the ViT on a dataset

   python train.py 

3. Evaluate the model:

   python eval.py

📁 Project Structure

vision-transformer/  
├── dataset/              # Dataset Classes  
├── models/               # Vision Transformer model components  
│   ├── vit.py            # Main ViT implementation  
│   ├── attention.py      # Scripts for SelfAttention and MultiHeadAttention
│   ├── Encoder.py        # Scripts for Encoder for Vit.
│   ├── get_model.py      # retrun the full model with Config set on config.py
│   ├── embedding.py      # Scripts for patch Embedding 
├── notebooks/
    ├── vit-scratch.ipynb # Jupyter notebook for model training and more ... 
├── config.py             # Configurations of full model
├── train.py              # Training script  
├── eval.py               # Evaluation script  
├── requirements.txt      # Required libraries  
└── README.md             # Project documentation  

Let me know if you’d like assistance with any part of the implementation!