Yolov7 model reasoning and training its own data set

        Some time ago YOLOv7 Release , There has been no time to reproduce , It's just time today , Simple reproduction .

1. The necessary environment

         My environment is ：Windows10+python3.9.7+CUDA11.1+CUDNN8.1.1+torch1.8. I won't say more about this environment configuration , There are many tutorials on the Internet . This article mainly records YOLOv7 Use .

2.YOLOv7 install

2.1 Get into github Download the source code and weight file on the official website ：

https://github.com/WongKinYiu/yolov7

2.2 Open the project , Enter the command at the terminal , Install dependency packages

pip install -r requirements.txt

         There is no error in installing the dependent package , If an error occurs , Please Baidu

3. Use YOLOv7 Official model for reasoning

          open detect.py file , Modify the path of the weight file , And modify the path of the reasoning picture . Right click to run ok, The result will be runs/detect Show

          The reasoning result is excellent , Far exceed YOLOX Performance of . 

4. Train your own dataset

First, set the project folder ：

           Put the picture in images in ,xml File put in Annotations.

          First use txt_write.py File partition dataset . The content of this document refers to another article of myself YOLOX In the article .

         Next use txt_yolo_label.py file , Data sets are converted to yolo Dataset format , After conversion, you can see VOCData/labels Generate the of each graph txt file .

          Then modify cfg/training In the corresponding model you downloaded yaml file , I downloaded it here yolov7.pt file , Change here yolov7.yaml file , take nc The quantity becomes the corresponding number of categories in your dataset .

          And then data Backup a copy in the folder coco.yaml file , Rename it to myvocdataset.yaml, Change the content as shown in the figure , Remember to change the path to your own path .

train: E:/python_practice/object_detection/yolov7-main/VOC/train.txt
val:  E:/python_practice/object_detection/yolov7-main/VOC/val.txt

nc: 2  # Category , Change to your own 
names: ['screen','sign']# Category name 

         Then open train.py file , Modify the following items ：

  From top to bottom , In turn, the weight file ,cfg The configuration file ,voc Data files , Super parameter file , And training rounds . The super parameter can not be modified , Others are changed to their corresponding files .batch-size Change according to computer strength , My configuration is general ,batch smaller , It is suggested that dataloader workers Default to 0, Turn off its multithreading . Right click to run the file , The program runs perfectly .

  If the following error occurs ：

Error #15: Initializing libiomp5md.dll, but found libiomp5md.dll already initialized.

 stay train.py Add the following code to ：

import os
os.environ['KMP_DUPLICATE_LIB_OK'] = 'TRUE'

         To demonstrate yolov7 Algorithm , After two rounds of training, I stopped training , Use the model of two rounds of training for reasoning . The result is shown in Fig. .

Because I only trained for two rounds , The effect is very poor , Only the box is recognized .

Here we are ,YOLOv7 Reproduction complete .