yolo_series：Integration of yolo target detection

​ The purpose of our creation of yolo_series is to better learn the algorithms of the YOLO series, better train and deploy models.

At present, the YOLO series algorithms mainly included in YOLOU are:

Anchor-base: YOLOv3, YOLOv4, YOLOv5, YOLOv5-Lite, YOLOv7

Anchor-Free: YOLOv6, YOLOX, YOLOX-Lite

Comparison of ablation experiment results

Model	size(pixels)	[email protected]	[email protected]:95	Parameters(M)	GFLOPs	TensorRT-FP32(b16) ms/fps	TensorRT-FP16(b16) ms/fps
YOLOv5n	640	45.7	28.0	1.9	4.5	0.95/1054.64	0.61/1631.64
YOLOv5s	640	56.8	37.4	7.2	16.5	1.7/586.8	0.84/1186.42
YOLOv5m	640	64.1	45.4	21.2	49.0	4.03/248.12	1.42/704.20
YOLOv5l	640	67.3	49.0	46.5	109.1
YOLOv5x	640	68.9	50.7	86.7	205.7
YOLOv6-T	640
YOLOv6-n	640
YOLOv6	640	58.4	39.8	20.4	28.8	3.06/326.93	1.27/789.51
YOLOv7	640	69.7	51.4	37.6	53.1	8.18/113.88	1.97/507.55
YOLOv7-X	640	71.2	53.7	71.3	95.1
YOLOv7-W6	640	72.6	54.9
YOLOv7-E6	640	73.5	56.0
YOLOv7-D6	640	74.0	56.6
YOLOv7-E6E	640	74.4	56.8
YOLOX-s	640	59.0	39.2	8.1	10.8	2.11/473.78	0.89/1127.67
YOLOX-m	640	63.8	44.5	23.3	31.2	4.94/202.43	1.58/632.48
YOLOX-l	640			54.1	77.7
YOLOX-x	640			104.5	156.2
v5-Lite-e	320	35.1		0.78	0.73	0.55/1816.10	0.49/2048.47
v5-Lite-s	416	42.0	25.2	1.64	1.66	0.72/1384.76	0.64/1567.36
v5-Lite-c	512	50.9	32.5	4.57	5.92	1.18/850.03	0.80/1244.20
v5-Lite-g	640	57.6	39.1	5.39	15.6	1.85/540.90	1.09/916.69
X-Lite-e	320	36.4	21.2	2.53	1.58	0.65/1547.58	0.46/2156.38
X-Lite-s	416	Training…	Training…	3.36	2.90
X-Lite-c	512	Training…	Training…	6.25	5.92
X-Lite-g	640	58.3	40.7	7.30	12.91	2.15/465.19	1.01/990.69

You can download all pretrained weights of yolo_series with Baidu Drive (YOLO)

How to use

Install

git clone https://github.com/Mobu59/yolo_series.git
cd yolo_series
pip install -r requirements.txt

Training

Single GPU

# train yolov5-lite
python train.py --mode yolo --data data/head_det.yaml --cfg models/yolov5-lite/v5Lite-c.yaml --weights weights/yolov5-lite/v5lite-c.pt --batch-size 32

Multi GPUs(DDP mode recommended)

# train yolox
python -m torch.distributed.launch --nproc_per_node 4 train.py --mode yolox --data data/head_det.yaml --cfg models/yolox/yoloxs.yaml --weights weights/yolox/yolox-s.pt --batch-size 32
# train yolov6
python -m torch.distributed.launch --nproc_per_node 4 train.py --mode yolov6 --data data/head_det.yaml --cfg models/yolov6/yolov6s.yaml --weights "" --batch-size 16

Note:

• For models deployed on mobile_platform like yolox_tiny, yolox_nano etc, you should set "self.mobile_platform == True"(common.py line720)

Detect

python detect.py --source 0  # webcam
                            file.jpg  # image 
                            file.mp4  # video
                            path/  # directory
                            path/*.jpg  # glob
                            'https://youtu.be/NUsoVlDFqZg'  # YouTube
                            'rtsp://example.com/media.mp4'  # RTSP, RTMP, HTTP stream

DataSet

For data in yolo format
    train: ../coco/images/train2017/
    val: ../coco/images/val2017/
For data in json format
    train:../head_detection/train.json
    val:../head_detection/val.json

├── images            # xx.jpg example
│   ├── train2017        
│   │   ├── 000001.jpg
│   │   ├── 000002.jpg
│   │   └── 000003.jpg
│   └── val2017         
│       ├── 100001.jpg
│       ├── 100002.jpg
│       └── 100003.jpg
└── labels             # xx.txt example      
    ├── train2017       
    │   ├── 000001.txt
    │   ├── 000002.txt
    │   └── 000003.txt
    └── val2017         
        ├── 100001.txt
        ├── 100002.txt
        └── 100003.txt
#xx.json example        
data/0002902.jpg [{"ymax": 658.9170170117557, "xmax": 709.5, "xmin": 460.3, "ymin": 282.4999999999999, "name": 0}, {"ymax": 736, "xmax": 1020.4, "xmin": 728.2, "ymin": 368.6, "name": 0}]        

Export ONNX

python export.py --weights ./weights/yolov6/yolov6s.pt --data data/coco.yaml --include onnx

Note:

• When export yolox.pt model to onnx, you should set "self.decode_in_inference == True"(yolox.py line36), and if your model will be deployed on mobile platform, you should set "mobile_platform == True"(yolox.py line168).Finally, you should also pay attention to "self.hw"(yolox.py line169), and you should modify it according to your input imgsz.
• When export yolov6.pt model to onnx, if your model will be deployed on mobile platform, you should set "mobile_platform == True"(yolov6.py line123).

Reference

https://github.com/ultralytics/yolov5

https://github.com/WongKinYiu/yolor

https://github.com/ppogg/YOLOv5-Lite

https://github.com/WongKinYiu/yolov7

https://github.com/meituan/YOLOv6

https://github.com/ultralytics/yolov3

https://github.com/Megvii-BaseDetection/YOLOX

https://github.com/WongKinYiu/ScaledYOLOv4

https://github.com/WongKinYiu/PyTorch_YOLOv4

https://github.com/WongKinYiu/yolor

https://github.com/shouxieai/tensorRT_Pro

https://github.com/Tencent/ncnn

https://github.com/Gumpest/YOLOv5-Multibackbone-Compression

https://github.com/positive666/yolov5_research

https://github.com/cmdbug/YOLOv5_NCNN

https://github.com/OAID/Tengine

https://github.com/jizhishutong/YOLOU