YOLOv9

Implementation of paper - YOLOv9: Learning What You Want to Learn Using Programmable Gradient Information

Performance

MS COCO

Model	Test Size	AP^val	AP₅₀^val	AP₇₅^val	Param.	FLOPs
YOLOv9-T	640	38.3%	53.1%	41.3%	2.0M	7.7G
YOLOv9-S	640	46.8%	63.4%	50.7%	7.1M	26.4G
YOLOv9-M	640	51.4%	68.1%	56.1%	20.0M	76.3G
YOLOv9-C	640	53.0%	70.2%	57.8%	25.3M	102.1G
YOLOv9-E	640	55.6%	72.8%	60.6%	57.3M	189.0G

Useful Links

Expand

Custom training: WongKinYiu#30 (comment)

ONNX export: WongKinYiu#2 (comment) WongKinYiu#40 (comment) WongKinYiu#130 (comment)

ONNX export for segmentation: WongKinYiu#260 (comment)

TensorRT inference: WongKinYiu#143 (comment) WongKinYiu#34 (comment) WongKinYiu#79 (comment) WongKinYiu#143 (comment)

QAT TensorRT: WongKinYiu#327 (comment) WongKinYiu#253 (comment)

TFLite: WongKinYiu#374 (comment)

OpenVINO: WongKinYiu#164 (comment)

C# ONNX inference: WongKinYiu#95 (comment)

C# OpenVINO inference: WongKinYiu#95 (comment)

OpenCV: WongKinYiu#113 (comment)

Hugging Face demo: WongKinYiu#45 (comment)

CoLab demo: WongKinYiu#18

ONNXSlim export: WongKinYiu#37

YOLOv9 ROS: WongKinYiu#144 (comment)

YOLOv9 ROS TensorRT: WongKinYiu#145 (comment)

YOLOv9 Julia: WongKinYiu#141 (comment)

YOLOv9 MLX: WongKinYiu#258 (comment)

YOLOv9 StrongSORT with OSNet: WongKinYiu#299 (comment)

YOLOv9 ByteTrack: WongKinYiu#78 (comment)

YOLOv9 DeepSORT: WongKinYiu#98 (comment)

YOLOv9 counting: WongKinYiu#84 (comment)

YOLOv9 face detection: WongKinYiu#121 (comment)

YOLOv9 segmentation onnxruntime: WongKinYiu#151 (comment)

Comet logging: WongKinYiu#110

MLflow logging: WongKinYiu#87

AnyLabeling tool: WongKinYiu#48 (comment)

AX650N deploy: WongKinYiu#96 (comment)

Conda environment: WongKinYiu#93

AutoDL docker environment: WongKinYiu#112 (comment)

Installation

Docker environment (recommended)

Expand

# create the docker container, you can change the share memory size if you have more.
nvidia-docker run --name yolov9 -it -v your_coco_path/:/coco/ -v your_code_path/:/yolov9 --shm-size=64g nvcr.io/nvidia/pytorch:21.11-py3

# apt install required packages
apt update
apt install -y zip htop screen libgl1-mesa-glx

# pip install required packages
pip install seaborn thop

# go to code folder
cd /yolov9

Evaluation

yolov9-c-converted.pt yolov9-e-converted.pt yolov9-c.pt yolov9-e.pt gelan-c.pt gelan-e.pt

# evaluate converted yolov9 models
python val.py --data data/coco.yaml --img 640 --batch 32 --conf 0.001 --iou 0.7 --device 0 --weights './yolov9-c-converted.pt' --save-json --name yolov9_c_c_640_val

# evaluate yolov9 models
# python val_dual.py --data data/coco.yaml --img 640 --batch 32 --conf 0.001 --iou 0.7 --device 0 --weights './yolov9-c.pt' --save-json --name yolov9_c_640_val

# evaluate gelan models
# python val.py --data data/coco.yaml --img 640 --batch 32 --conf 0.001 --iou 0.7 --device 0 --weights './gelan-c.pt' --save-json --name gelan_c_640_val

You will get the results:

 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.530
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.702
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.578
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.362
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.585
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.693
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.392
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.652
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.702
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.541
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.760
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.844

Training

Data preparation

bash scripts/get_coco.sh

Download MS COCO dataset images (train, val, test) and labels. If you have previously used a different version of YOLO, we strongly recommend that you delete train2017.cache and val2017.cache files, and redownload labels

Single GPU training

# train yolov9 models
python train_dual.py --workers 8 --device 0 --batch 16 --data data/coco.yaml --img 640 --cfg models/detect/yolov9-c.yaml --weights '' --name yolov9-c --hyp hyp.scratch-high.yaml --min-items 0 --epochs 500 --close-mosaic 15

# train gelan models
# python train.py --workers 8 --device 0 --batch 32 --data data/coco.yaml --img 640 --cfg models/detect/gelan-c.yaml --weights '' --name gelan-c --hyp hyp.scratch-high.yaml --min-items 0 --epochs 500 --close-mosaic 15

Multiple GPU training

# train yolov9 models
python -m torch.distributed.launch --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-c.yaml --weights '' --name yolov9-c --hyp hyp.scratch-high.yaml --min-items 0 --epochs 500 --close-mosaic 15

# train gelan models
# python -m torch.distributed.launch --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-c.yaml --weights '' --name gelan-c --hyp hyp.scratch-high.yaml --min-items 0 --epochs 500 --close-mosaic 15

Re-parameterization

See reparameterization.ipynb.

Inference

# inference converted yolov9 models
python detect.py --source './data/images/horses.jpg' --img 640 --device 0 --weights './yolov9-c-converted.pt' --name yolov9_c_c_640_detect

# inference yolov9 models
# python detect_dual.py --source './data/images/horses.jpg' --img 640 --device 0 --weights './yolov9-c.pt' --name yolov9_c_640_detect

# inference gelan models
# python detect.py --source './data/images/horses.jpg' --img 640 --device 0 --weights './gelan-c.pt' --name gelan_c_c_640_detect

Citation

@article{wang2024yolov9,
  title={{YOLOv9}: Learning What You Want to Learn Using Programmable Gradient Information},
  author={Wang, Chien-Yao  and Liao, Hong-Yuan Mark},
  booktitle={arXiv preprint arXiv:2402.13616},
  year={2024}
}

@article{chang2023yolor,
  title={{YOLOR}-Based Multi-Task Learning},
  author={Chang, Hung-Shuo and Wang, Chien-Yao and Wang, Richard Robert and Chou, Gene and Liao, Hong-Yuan Mark},
  journal={arXiv preprint arXiv:2309.16921},
  year={2023}
}

Teaser

Parts of code of YOLOR-Based Multi-Task Learning are released in the repository.

Object Detection

gelan-c-det.pt

object detection

# coco/labels/{split}/*.txt
# bbox or polygon (1 instance 1 line)
python train.py --workers 8 --device 0 --batch 32 --data data/coco.yaml --img 640 --cfg models/detect/gelan-c.yaml --weights '' --name gelan-c-det --hyp hyp.scratch-high.yaml --min-items 0 --epochs 300 --close-mosaic 10

Model	Test Size	Param.	FLOPs	AP^box
GELAN-C-DET	640	25.3M	102.1G	52.3%
YOLOv9-C-DET	640	25.3M	102.1G	53.0%

Instance Segmentation

gelan-c-seg.pt

object detection instance segmentation

# coco/labels/{split}/*.txt
# polygon (1 instance 1 line)
python segment/train.py --workers 8 --device 0 --batch 32  --data coco.yaml --img 640 --cfg models/segment/gelan-c-seg.yaml --weights '' --name gelan-c-seg --hyp hyp.scratch-high.yaml --no-overlap --epochs 300 --close-mosaic 10

Model	Test Size	Param.	FLOPs	AP^box	AP^mask
GELAN-C-SEG	640	27.4M	144.6G	52.3%	42.4%
YOLOv9-C-SEG	640	27.4M	145.5G	53.3%	43.5%

Panoptic Segmentation

gelan-c-pan.pt

object detection instance segmentation semantic segmentation stuff segmentation panoptic segmentation

# coco/labels/{split}/*.txt
# polygon (1 instance 1 line)
# coco/stuff/{split}/*.txt
# polygon (1 semantic 1 line)
python panoptic/train.py --workers 8 --device 0 --batch 32  --data coco.yaml --img 640 --cfg models/panoptic/gelan-c-pan.yaml --weights '' --name gelan-c-pan --hyp hyp.scratch-high.yaml --no-overlap --epochs 300 --close-mosaic 10

Model	Test Size	Param.	FLOPs	AP^box	AP^mask	mIoU_164k/10k^semantic	mIoU^stuff	PQ^panoptic
GELAN-C-PAN	640	27.6M	146.7G	52.6%	42.5%	39.0%/48.3%	52.7%	39.4%
YOLOv9-C-PAN	640	28.8M	187.0G	52.7%	43.0%	39.8%/-	52.2%	40.5%

Image Captioning (not yet released)

object detection instance segmentation semantic segmentation stuff segmentation panoptic segmentation image captioning

# coco/labels/{split}/*.txt
# polygon (1 instance 1 line)
# coco/stuff/{split}/*.txt
# polygon (1 semantic 1 line)
# coco/annotations/*.json
# json (1 split 1 file)
python caption/train.py --workers 8 --device 0 --batch 32  --data coco.yaml --img 640 --cfg models/caption/gelan-c-cap.yaml --weights '' --name gelan-c-cap --hyp hyp.scratch-high.yaml --no-overlap --epochs 300 --close-mosaic 10

Model	Test Size	Param.	FLOPs	AP^box	AP^mask	mIoU_164k/10k^semantic	mIoU^stuff	PQ^panoptic	BLEU@4^caption	CIDEr^caption
GELAN-C-CAP	640	47.5M	-	51.9%	42.6%	42.5%/-	56.5%	41.7%	38.8	122.3
YOLOv9-C-CAP	640	47.5M	-	52.1%	42.6%	43.0%/-	56.4%	42.1%	39.1	122.0

Acknowledgements

Expand

starshineangu/yolov9

YOLOv9

Performance

Useful Links

Installation

Evaluation

Training

Re-parameterization

Inference

Citation

Teaser

Object Detection

Instance Segmentation

Panoptic Segmentation

Image Captioning (not yet released)

Acknowledgements