Smoking-detection-YOLO26s

🚬 Smoking Detection — YOLOv8s

A fine-tuned YOLOv8s object detection model trained to detect cigarettes in images and video streams. Designed for real-world deployment with crash-resilient training and automatic checkpoint uploads to Hugging Face.

---

📊 Model Performance

YOLO Validation Metrics

Image-Level Classification Report (307 validation images)

Model Evaluation Metrics

Confusion Matrix Summary

---

🏗️ Training Setup

Dataset

• Source: Roboflow — pro-hiyaw / cigarettes-reality-2-trnm2 (Version 1)
• Format: YOLOv8 (yolo26)

Model

• Base weights: yolo26s.pt
• Architecture: YOLOv8 Small

Hyperparameters

Data Augmentation

---

🔄 Crash-Resilient Training

The training pipeline includes automatic crash recovery:

• At the end of every epoch, a training_status.json file is written with the current epoch, path to last.pt, and path to best.pt.
• On restart, if a status file is found with "status": "training", training automatically resumes from the last saved weights.
• Every 10 epochs, checkpoints are uploaded to Hugging Face:
  • weights/epoch_{N}_last.pt — rolling checkpoint
  • weights/best.pt — best model so far (overwritten each time)
  • training_status.json — current training state

---

🤗 Hugging Face Integration

Model checkpoints are automatically uploaded to a private Hugging Face repository during training.

• Repo: <your-username>/Smoking-detection-YOLO26s
• Visibility: Private
• Upload frequency: Every 10 epochs

Authentication uses a Hugging Face token stored as a Kaggle secret (HF_TOKEN).

---

🚀 Quick Start

Install Dependencies

pip install ultralytics roboflow huggingface_hub opencv-python

Run Inference

from ultralytics import YOLO

model = YOLO("weights/best.pt")
results = model.predict(source="your_image.jpg", imgsz=800, conf=0.5)
results[0].show()

Resume Training

Simply re-run the training script. If training_status.json exists and points to valid weights, training will resume automatically from the last checkpoint.

---

📁 Project Structure

.
├── training_status.json        # Auto-generated crash recovery state
├── runs/
│   └── detect/
│       └── train/
│           └── weights/
│               ├── best.pt     # Best checkpoint
│               └── last.pt     # Latest checkpoint
└── <dataset>/
    └── data.yaml               # Dataset config (auto-downloaded from Roboflow)

---

🔐 Required Secrets (Kaggle)

---

📄 License

This project is for research and educational purposes. Dataset sourced from Roboflow Universe under its respective license.