In this repository, we provide the implementation of Hyper-YOLO v1.1, which intergrates the advantages of YOLOv9 and Hyper-YOLO, achieving the state-of-the-art performance on MS COCO dataset.
We replace the neck of YOLOv9 with the proposed HyperC2Net of Hyper-YOLO, termed Hyper-YOLOv1.1. Clearly, in each scale, the Hyper-YOLOv1.1 outperforms the YOLOv9, which demonstrates the effectiveness of our HyperC2Net in capturing high-order feature correlations. The comparison of four scale models are provided in the following table
| Model | Test Size | Params | FLOPs | ||
|---|---|---|---|---|---|
| YOLOv9-T | 640 | 38.3 | 53.1 | 2.0M | 7.7G |
| YOLOv9-S | 640 | 46.8 | 63.4 | 7.1M | 26.4G |
| YOLOv9-M | 640 | 51.4 | 68.1 | 20.0M | 76.3G |
| YOLOv9-C | 640 | 53.0 | 70.2 | 25.3M | 102.1G |
| Hyper-YOLOv1.1-T | 640 | 40.3 | 55.6 | 2.5M | 10.8G |
| Hyper-YOLOv1.1-S | 640 | 48.0 | 64.5 | 7.6M | 29.9G |
| Hyper-YOLOv1.1-M | 640 | 51.9 | 69.1 | 21.2M | 87.4G |
| Hyper-YOLOv1.1-C | 640 | 53.2 | 70.4 | 29.8M | 115.5G |
Clone repo and create conda environment (recommended). Then install requirements.txt in a Python>=3.8.0 environment, including PyTorch>=1.8.
git clone https://github.com/iMoonLab/Hyper-YOLOv1.1.git # clone
cd Hyper-YOLOv1.1
conda create -n Hyper-YOLOv1.1 python=3.8
conda activate Hyper-YOLOv1.1
pip install -r requirements.txt # installYou can also use the environment.yaml file and the conda command to install the required environment.
conda env create -f environment.yamlData Preparation: Download the MS COCO dataset images (training, validation, and test sets) and corresponding labels, or prepare your custom dataset as shown below. Additionally, modify the dataset path in data/coco.yaml to reflect the location of your data.
coco
--images
--train2017
--val2017
--labels
--train2017
--val2017Training configurations can be modified within the argument parser of “train.py” or “train_dual.py”.
The key factors are model, data, img, epoches, batch, device and training hyperparameters.
You can adjust the training hyperparameters in the data/hyps/hyp.scratch-XXX.yaml file. Here, XXX can be set to low, med, or high, which correspond to low, medium, and high levels of data augmentation, respectively.
python train.py --hyp hyp.scratch-low.yaml
hyp.scratch-med.yaml
hyp.scratch-high.yamlFor instance, you can employ “yolov9-s-hyper.yaml” to train the “HyperYOLOv1.1-S” object detection model, and subsequently use “convert.py” along with “gelan-s-hyper.yaml” to remove the Auxiliary Reversible Branch.
# train yolov9-s-hyper models
python train_dual.py --workers 8 --device 0 --batch 16 --data data/coco.yaml --img 640 --cfg models/detect/yolov9-s-hyper.yaml --weights '' --name yolov9-s-hyper --hyp hyp.scratch-low.yaml --epochs 500
# train gelan-s-hyper models
# python train.py --workers 8 --device 0 --batch 32 --data data/coco.yaml --img 640 --cfg models/detect/gelan-s-hyper.yaml --weights '' --name gelan-s-hyper --hyp hyp.scratch-low.yaml --epochs 500# train yolov9-s-hyper models
python -m torch.distributed.run --nproc_per_node 8 --master_port 9527 train_dual.py --workers 8 --device 0,1,2,3,4,5,6,7 --sync-bn --batch 128 --data data/coco.yaml --img 640 --cfg models/detect/yolov9-s-hyper.yaml --weights '' --name yolov9-s-hyper --hyp hyp.scratch-low.yaml --epochs 500
# train gelan-s-hyper models
# python -m torch.distributed.run --nproc_per_node 4 --master_port 9527 train.py --workers 8 --device 0,1,2,3 --sync-bn --batch 128 --data data/coco.yaml --img 640 --cfg models/detect/gelan-s-hyper.yaml --weights '' --name gelan-s-hyper --hyp hyp.scratch-low.yaml --epochs 500The key factors are model(weight), data, img, batch, conf, iou, half.
# evaluate converted yolov9-s-hyper models
python val.py --data data/coco.yaml --img 640 --batch 32 --conf 0.001 --iou 0.7 --device 0 --weights './yolov9-s-hyper-converted.pt' --save-json --name yolov9_s_hyper_c_640_val
# evaluate yolov9-s-hyper models
# python val_dual.py --data data/coco.yaml --img 640 --batch 32 --conf 0.001 --iou 0.7 --device 0 --weights './yolov9-s-hyper.pt' --save-json --name yolov9_s_hyper_640_val
# evaluate gelan-s-hyper models
# python val.py --data data/coco.yaml --img 640 --batch 32 --conf 0.001 --iou 0.7 --device 0 --weights './gelan-s-hyper.pt' --save-json --name gelan_s_hyper_640_valThe key factors are model(weight), source, img, conf, iou.
# inference converted yolov9-s-hyper models
python detect.py --source './data/images/horses.jpg' --img 640 --device 0 --weights './yolov9-s-hyper-converted.pt' --name yolov9_s_hyper_c_640_detect
# inference yolov9 model
# python detect_dual.py --source './data/images/horses.jpg' --img 640 --device 0 --weights './yolov9-s-hyper.pt' --name yolov9_s_hyper_640_detect
# inference gelan models
# python detect.py --source './data/images/horses.jpg' --img 640 --device 0 --weights './gelan-s-hyper.pt' --name gelan_s_hyper_640_detect
Please refer to YOLOv8 or YOLOv9.
If you find our work useful in your research, please consider citing:
@article{feng2024hyper,
title={Hyper-YOLO: When Visual Object Detection Meets Hypergraph Computation},
author={Feng, Yifan and Huang, Jiangang and Du, Shaoyi and Ying, Shihui and Yong, Jun-Hai and Li, Yipeng and Ding, Guiguang and Ji, Rongrong and Gao, Yue},
journal={arXiv preprint arXiv:2408.04804},
year={2024}
}Hypergraph computation is a powerful tool to capture high-order correlations among visual features. Compared with graphs, each hyperedge in a hypergraph can connect more than two vertices, which is more flexible to model complex correlations. Now, learning with high-order correlations still remains a under-explored area in computer vision. We hope our work can inspire more research in this direction. If you are interested in hypergraph computation, please refer to our series of works on hypergraph computation in the follows:
- Hypergraph Learning: Methods and Practices
- Hypergraph Nerual Networks
- HGNN+: General Hypergraph Nerual Networks
- Hypergraph Isomorphism Computation
Hyper-YOLO is maintained by iMoon-Lab, Tsinghua University. If you have any questions, please feel free to contact us via email: Yifan Feng and Jiangang Huang.