Conditional Vision Transformer (CViT) for Traffic Sign Classification 🚦

This repository contains the implementation of a Conditional Vision Transformer (CViT) designed for accurate and robust traffic sign classification using the GTSRB (German Traffic Sign Recognition Benchmark) dataset. This model introduces a novel conditional attention mechanism that dynamically adapts the attention weights based on input context, significantly improving performance on misclassification-prone classes.

---

🚀 Key Features

• ✅ Conditional Attention Mechanism
• ✅ Custom Patch Embedding and Tokenization
• ✅ Gating Network for Adaptive QKV Generation
• ✅ Superior Accuracy and Generalization
• ✅ Well-structured and clean code, ready for training and evaluation

---

🧠 Model Overview

The Conditional ViT architecture is composed of three core phases:

1. Preprocessing Phase

   • Patch extraction and embedding
   • Positional encoding
   • Token sequence formation

2. Feature Extraction Phase

   • Conditional attention blocks with gating mechanisms
   • Adaptive generation of Query, Key, and Value matrices

3. Classification Phase

   • Fully connected layers
   • Softmax output for multi-class classification

---

📊 Results

Model	Accuracy (%)
Simple ViT	94.25
Conditional ViT (Proposed)	99.87

The proposed model significantly reduces misclassification in challenging traffic sign classes such as T21 and T2, ensuring more stable and interpretable predictions.

---