Apppublic

gphua1/rklb_materials_deploy

sourceHugging Faceupdated 1y agoView on Hugging Face

0likes

App README

🔍 Defect Detection with Vision Transformer

Production-grade defect detection system using state-of-the-art Vision Transformers.

🚀 Quick Start

Installation

bash

pip install -r requirements.txt

Training

bash

# Full training
python train.py --epochs 50 --model_type efficient_vit --pretrained

# Quick test (toy mode)
python train.py --toy --epochs 3

Inference

Web Interface (Streamlit)

bash

python app.py --mode web
# Open browser to http://localhost:8501

API Server

bash

python app.py --mode api
# API at http://localhost:8000
# Web UI at http://localhost:8000/interface

Command Line

bash

# Single image
python app.py --mode cli --image path/to/image.png

# Batch processing
python app.py --mode cli --directory path/to/images/ --output results.json

📁 Project Structure

.
├── app.py              # Main application (Web/API/CLI)
├── train.py            # Training pipeline
├── models/
│   └── vision_transformer.py  # Model architectures
├── data/               # Dataset directory
├── requirements.txt    # Dependencies
└── vercel.json        # Deployment config

🌐 Deployment

Vercel (Serverless)

bash

vercel --prod

Docker

bash

docker build -t defect-detection .
docker run -p 8000:8000 defect-detection

🧪 Model Options

—efficient_vit: Best accuracy (default)
—hybrid_vit: CNN + Transformer hybrid
—vit: Standard Vision Transformer
—dual_attention_vit: Dual attention mechanism

📊 Performance

—Accuracy: 95%+ on industrial defect datasets
—Inference: <100ms per image
—Model size: ~350MB (efficient_vit)

📝 License

MIT