NK-YOLO is a comprehensive object detection benchmark built on Jittor, providing state-of-the-art YOLO implementations with automatic compilation and optimized performance.

📖 Documentation | 🚀 Quick Start | 🎯 Models | 📊 Results

• 🔥 Automatic Compilation: No manual CUDA/C++ compilation required
• 🚀 Multi-Model Support: YOLOv5, YOLOv6, YOLOv7, YOLOv8, YOLOv10, YOLO-MS
• ⚡ High Performance: Optimized for speed and accuracy
• 🎯 Easy Training: Simple training pipeline with comprehensive configuration
• 📊 Rich Evaluation: Built-in validation and testing tools
• 🛠️ Cross-Platform: Support for Linux, macOS, and WSL

• OS: Linux (Ubuntu/CentOS/Arch), macOS, or Windows Subsystem for Linux (WSL)
• Python: >= 3.7
• CPU Compiler: g++ (>=5.4.0) or clang (>=8.0)
• GPU Compiler (Optional): nvcc (>=10.0 for g++ or >=10.2 for clang)
• GPU Library: cudnn-dev (recommended tar file installation)

git clone https://github.com/NK-JittorCV/nk-yolo.git
cd nk-yolo
python -m pip install -r requirements.txt

💡 Note: If you encounter Jittor installation issues, please refer to the Jittor Installation Guide.

# Option 1: Install in development mode
pip install -v -e .

# Option 2: Install with user permissions (if needed)
pip install -v -e . --user

python demo.py

This will:

1. Train a YOLOv8 model on COCO8 dataset for 2 epochs
2. Evaluate the model performance
3. Run inference on sample images

from nkyolo import YOLO

# Load a model
model = YOLO('yolov8n.yaml')

# Train the model
results = model.train(
    data='coco8.yaml',
    epochs=100,
    imgsz=640,
    batch=16,
    device='cpu'  # or '0' for GPU
)

# Validate the model
metrics = model.val()

# Run inference
results = model('path/to/image.jpg')
results[0].show()

Create a YAML configuration file for your dataset:

# dataset.yaml
path: ../datasets/coco  # dataset root directory
train: images/train2017  # train images (relative to 'path')
val: images/val2017      # val images (relative to 'path')

# Classes
names:
  0: person
  1: bicycle
  2: car
  # ... more classes

python -m nkyolo.models.yolo.detect.train \
    --model yolov8n.yaml \
    --data coco.yaml \
    --epochs 100 \
    --imgsz 640 \
    --batch 16 \
    --device cpu

python -m nkyolo.models.yolo.detect.train \
    --model yolov8n.yaml \
    --data coco.yaml \
    --epochs 100 \
    --imgsz 640 \
    --batch 16 \
    --device 0 \
    --workers 8 \
    --patience 50 \
    --save-period 10 \
    --cache ram

# YOLOv8
python -m nkyolo.models.yolo.detect.train --model yolov8n.yaml

# Validate a trained model
python -m nkyolo.models.yolo.detect.train \
    --model ${CKPT_PATH} \
    --data coco.yaml \
    --task val

from nkyolo import YOLO

# Load trained model
model = YOLO(${CKPT_PATH})

# Single image inference
results = model('path/to/image.jpg')
results[0].show()
results[0].save('output.jpg')

# Batch inference
results = model(['image1.jpg', 'image2.jpg', 'image3.jpg'])

# Video inference
results = model('video.mp4')

# Evaluate model performance
python -m nkyolo.models.yolo.detect.train \
    --model runs/train/exp/weights/best.pt \
    --data coco.yaml \
    --task val \
    --conf 0.001 \
    --iou 0.6

nk-yolo/
├── nkyolo/           # Main package
│   ├── cfg/             # Configuration files
│   │   ├── datasets/    # Dataset configurations
│   │   └── models/      # Model configurations
│   ├── data/            # Data loading and processing
│   ├── engine/          # Training engine
│   ├── models/          # Model implementations
│   ├── nn/              # Neural network modules
│   └── utils/           # Utility functions
├── datasets/            # Dataset storage
├── runs/               # Training outputs
├── demo.py             # Demo script
└── requirements.txt    # Dependencies

We welcome contributions! Please see our Contributing Guidelines for details.

This project is licensed under the AGPL-3.0 License - see the LICENSE file for details.

This project is built on top of these excellent open-source projects:

• Jittor - Deep learning framework
• ultralytics - YOLOv8 implementation
• yolov10 - YOLOv10 implementation

If you find this work useful for your research, please cite:

@article{hu2020jittor,
  title={Jittor: a novel deep learning framework with meta-operators and unified graph execution},
  author={Hu, Shi-Min and Liang, Dun and Yang, Guo-Ye and Yang, Guo-Wei and Zhou, Wen-Yang},
  journal={Science China Information Sciences},
  volume={63},
  number={222103},
  pages={1--21},
  year={2020}
}

@article{Chen2025,
  title = {YOLO-MS: Rethinking Multi-Scale Representation Learning for Real-time Object Detection},
  ISSN = {1939-3539},
  url = {http://dx.doi.org/10.1109/TPAMI.2025.3538473},
  DOI = {10.1109/tpami.2025.3538473},
  journal = {IEEE Transactions on Pattern Analysis and Machine Intelligence},
  publisher = {Institute of Electrical and Electronics Engineers (IEEE)},
  author = {Chen, Yuming and Yuan, Xinbin and Wang, Jiabao and Wu, Ruiqi and Li, Xiang and Hou, Qibin and Cheng, Ming-Ming},
  year = {2025},
  pages = {1–14}
}

@article{wang2024yolov10,
  title={YOLOv10: Real-Time End-to-End Object Detection},
  author={Wang, Ao and Chen, Hui and Liu, Lihao and Chen, Kai and Lin, Zijia and Han, Jungong and Ding, Guiguang},
  journal={arXiv preprint arXiv:2405.14458},
  year={2024}
}