ViT-Implementation

AI Club DC Mini Project

This project is a step-by-step journey into modern deep learning architectures, with a special focus on understanding and implementing Vision Transformers (ViT) using PyTorch. The tasks below are designed to build foundational knowledge and practical coding skills.

---

🚀 Tasks Overview

✅ Task 1: Convolutional Neural Networks (CNNs)

• Objective: Understand and implement a basic CNN.
• What to do:
  • Study CNN architecture fundamentals (convolutional layers, pooling, activation functions).
  • Implement a simple CNN from scratch in PyTorch (e.g., on MNIST or CIFAR-10).

---

✅ Task 2: Attention Mechanism & Transformer Encoder-Decoder

1. Paper Reading:

   • Read and understand the landmark paper: Attention Is All You Need

2. Practical Exploration:

   • Go through a blog/tutorial where an encoder-decoder transformer model is implemented from scratch.
   • Suggested blog: The Annotated Transformer (or choose your preferred one).

---

✅ Task 3: Vision Transformer (ViT)

1. Paper Reading:

   • Dive into the original Vision Transformer paper: An Image is Worth 16x16 Words

2. Implementation:

   • Code the ViT architecture from scratch using only PyTorch (no high-level transformer libraries like Hugging Face).
   • Understand and build:
     • Patch embeddings
     • Positional encodings
     • Transformer encoder blocks
     • Classification head

---

📁 Repository Structure (Implemented)

ViT-Implementation/
├── task1_cnn/
│   └── cnn_model.ipynb
├── task3_vit/
│   └── vit_from_scratch.ipynb
└── README.md