Yolov6:yolov6+win10--- train your own dataset

yolov6 I also ran at the meeting when I just came out , There are too many problems , All kinds of bug, Just stop for a while , There's time later. Just look at it and start again git Run down . It is said that YoloV7 It also comes out AB The great God team produced , It's not what someone calls YoloV7....

GitHub - meituan/YOLOv6: YOLOv6: a single-stage object detection framework dedicated to industrial applications.YOLOv6: a single-stage object detection framework dedicated to industrial applications. - GitHub - meituan/YOLOv6: YOLOv6: a single-stage object detection framework dedicated to industrial applications.https://github.com/meituan/YOLOv6

1、 Data sets

I use my own dataset , Let's take a look at the data set format first .

val and train Store pictures ,labels Deposit is train  val Folder , It's in there txt

  At the same time, check the configuration file , My dataset There are only two categories

  Here we need to change dataset.py There must be a place inside ;

 

2、 Training

Training orders :

python tools/train.py --batch 6 --conf configs/yolov6s_finetune.py --data data/MyDataCoCo.yaml --device 0  --epochs 800 --workers 0

I trained on my own machine , therefore bs Set it down

End of training , We see ,map Not for 0 了 , That's what we use yolov6s_finetune.py  If you use yolov6s.py, Equivalent to no loading and training model , You can go to Yolov6 Of github  issues Go to see .

https://github.com/meituan/YOLOv6/issues/205#issuecomment-1173409870

 800 Round training time is approaching 19 Hours .........

3、 test

Test code :

python tools/infer.py --weights runs/train/exp/weights/best_ckpt.pt --source CiWaData/val --yaml data/MyDataCoCo.yaml

test result :

 

 

4、ONNX export

Export command ：

export_onnx.py --weights runs/train/exp/weights/best_ckpt.pt --img 640 --batch 1