Deep learning -yolov5 introduction to actual combat click data set training

One 、 Environmental preparation

• yolov5 Is the main role ： Given a data set that has been marked 》 run yolov5 The code gets a target detection model 》 Choose a picture or video and give it to yolov5 Code ,yolov5 The code can automatically recognize and distinguish the target body

• It is recommended that the blogger's Video Script install the environment hand in hand to train recognition , Or look directly at The blogger's blog has tutorials

• Native configuration ： Win11,Anaconda(python3.7),GTX3050,cuda11.1,pytorch1.8, yolov5, It is recommended to check your computer first cuda Version information ： open cmd Window type nvidia-smi see , Or by NVIDIA Control panel see
  

• pytorch Environmental preparation ： First of all, there are anconda, Then open the cmd Execute the order as follows （ Corresponding to the configuration above ）, For example, search # CUDA 11.1 open cuda Each version corresponds to the installation command search , If you encounter problems, you can see the installation steps in this article

  conda create --p=D:\Programmer\Captcha_env python=3.7
  activate D:\Programmer\Captcha_env
  conda install pytorch==1.8.0 torchvision==0.9.0 torchaudio==0.8.0 cudatoolkit=11.1 -c pytorch -c conda-forge
  

• download yolov5, Download the compressed package address , Then decompress it
  

• yolov5 relevant python Dependent package installation ： stay yolov5-master Open in the current directory cmd window , Then enter the previously built virtual environment, such as activate D:\Programmer\Captcha_env, Then install the dependencies pip install -r requirements.txt -i https://pypi.tuna.tsinghua.edu.cn/simple/

  activate D:\Programmer\Captcha_env
  pip install -r requirements.txt -i https://pypi.tuna.tsinghua.edu.cn/simple/ --trusted-host pypi.tuna.tsinghua.edu.cn
  

  

Two 、yolov5 A brief introduction to the code

• See here for a more detailed introduction , Or this article
• Recommend 2 Half an hour video nanny introduces the meaning of code parameters

1、yolov5 In the code train.py

• stay yolov5 In the code , Our training mainly runs train.py file , The test mainly runs detect.py file ; So as long as Prepare a data set , Then change the main configuration of these two files as shown in the figure, such as which data set path to train or test , Picture size, etc , Then the process is finished .
  
• yolov5 Code train.py Introduction to important parameters of ：
  
• train.py Parameters of weights Brief introduction ： Specify the weight model , Such as ：yolov5s.pt / yolov5m / yolov5l/ yolov5x（ It can be understood as s Small ,m Medium ,l Large ,xl Big size , The corresponding detection speed is from fast to slow , But the accuracy is from low to high ）
  
  

2、yolov5 In the code detect.py

• yolov5 Code detect.py Introduction to important parameters of ：
  

3、yolov5 Data set file format for code preparation

• yolov5 The custom preparation format of the data set to be trained is generally as follows ： Three parts Training set + Verification set +.yaml To configure , among images There are marked pictures under the folder ,labels All are yolo Format .txt file ,images And labels There is one-to-one correspondence in , Only the suffix is different
  
• yolo The format label is txt File format ： The file name is the same as the corresponding picture name , Except that the suffix is changed to .txt

  • One line for each goal , If the whole picture has no goal, there is no need to have txt file , Or empty file

  • ` The format of each line is 5 Parameters ：class_num（ Classified index , such as 0 On behalf of wearing masks ,1 The representative doesn't wear a mask ）, after 4 The column is x_center/image_width、y_center/image_height、width/image_width、height/image_height, The value range is 0 ~ 1
    

  • among class_num The value is 0 to total_class - 1, As shown in the following figure, there are four values x_center Relative to the picture axis is 0.48,y_center Relative to the picture axis is 0.63,width The width of the picture is 0.69, height The height of the picture is 0.71; The reason why image values are normalized , Because there are many sizes of training pictures , In this way, it can be unified through normalization
    

• .yaml Configuration file for ： It mainly specifies the image path for training and verification , And classification quantity and name
  

3、 ... and 、 A tiny test identification case

• Test with the existing model on the official website yolov5 Is it working ： open pychram Direct operation detect.py file （detect.py file ： Use the pre trained model for prediction and recognition test , Such as yolov5s.pt Model )
• During operation, it will Automatically download one yolov5s.pt Model , If you can't download it , You can connect directly to this connection Weight model download , and train.py In the same directory ; then Automatic identification data/images Two pictures in the directory , And then Automatically generate a new folder runs>detect>exp Used to store the identified results .
  
• Here is runs>detect>exp Identification results under the folder , The target has been automatically identified as bus still person, And marked
  

Four 、 Open data sets 【 masks 】 Train and identify cases

• Open data sets roboflow network , There are many data sets , I downloaded the data set of masks here , The picture is about 149 Zhang ; Other public datasets DataCastle network , Jishi network , Others, etc
  

• First of all, we Download the mask data set A marked data set is decompressed and stored in I created a new folder MaskWearing, The directory of this folder
  

• MaskWearing Data set introduction ： The main Yes 3 individual ： Data set of training pictures , Verify the data set of the picture , as well as data.yaml To configure . among images Folders are for pictures ,labels Under the folder is yolo Format .txt Label file ,images Folders and labels The contents of the folder are one-to-one correspondence , The file names are all the same , The suffix is different .data.yaml The configuration file is described in detail below
  

• change yaml File configuration ： Pictured data.yaml Correct the storage path of data set , among train It refers to the path of the image file to be trained , and val It refers to the image path to be verified , and nc representative 2 A classification ,names Represents the name corresponding to the two classifications ; be based on yolov5s.pt The training will be modified yolov5s.yaml Inside nc classification , according to data.yaml Of nc Modify the classification quantity
  

• modify train.py Parameters and then start training ： The training mainly modifies the following parameters , After running, it will automatically generate runs>train>exp>weights Folder , One will be generated under the folder best.pt Weight model

  • weights： Select the model weight , Such as ：yolov5s.pt / yolov5m / yolov5l（ It can be understood as s Small ,m Medium ,l Large , The corresponding detection speed is from fast to slow , But the accuracy is from low to high ）
  • cfg： Specify the location of the model configuration file, such as models/yolov5s.yaml
  • data： Specify your own dataset configuration file location, such as ./MaskWearing/data.yaml
  • epochs: Specify the number of training rounds , The default initial setting is 300, For example, the data set of this mask 149 A picture , I started setting 3 The trained model is not effective , The test did not frame the mask correctly , Set to 10 Then the target is recognized normally , So adjust according to the actual situation
  • batchsize： according to GPU Set the size of the video memory , It's usually 2 Multiple , I set up here 4
  • img-size： The size of the image resolution reduced or enlarged , Default 640, Adjust according to the world

  

• If you make a mistake OSError: [WinError 1455] The page file is too small , Unable to complete operation ., be Look at this article to solve , Search directly to view advanced system settings , Then process according to the picture , Restart the computer after saving ; If this setting still reports an error and does not solve the problem , First restore it to the previous default state of automatically managing all drives , After saving the application , Reopen it again and press the image again .
  
  

• modify detect.py Parameters and then start testing recognition , It will be generated automatically after training runs>detect>exp Folder , There are recognition results under this folder , So far, training to test the small recognition demo Completed

  • weight： Specify the location of the previously trained weight model file
  • source： Specify a single picture recognized by the detection , Or picture folder path , Or video path , Or the camera

  
  

5、 ... and 、 Label the dataset by yourself 【 Point selection 】 Train and identify cases

1、 utilize labelimg Annotation dataset generation yolo Format

• labelimg Is a data annotation tool , The following three formats can be marked ：① VOC Label format , Save as xml file ,② yolo Label format , Save as txt file ,③ createML Label format , Save as json Format ; This annotation selects yolob Label format
• labelimg Installation ：pip install labelimg -i https://pypi.tuna.tsinghua.edu.cn/simple

  activate D:\Programmer\Captcha_env\
  pip install labelimg -i https://pypi.tuna.tsinghua.edu.cn/simple
  

  
• newly build 2 A folder ： One is for storing pictures images, One is for storing labels labels,cmd Continue to input labelimg You can enter the home page of the label

  activate D:\Programmer\Captcha_env\
  labelimg
  

  
• Set according to the following picture , Remember to choose first yolo Format , Then start marking
  
• Be careful ： Whether it's a labeled picture , Or the path , Or tag name , Try not to write Chinese , Otherwise, there will be unexpected mistakes , For example, it may not work all the time and report errors OSError: [WinError 1455] The page file is too small , even to the extent that pycharm Flashback, etc
• Mark the prepared pictures as shown in the figure , Press w key You can start marking , Then customize the label classification , You can also set the default classification on the right , After saving, the folder where you save the tag will Automatic generation yolo Format .txt file , There is a general classes.txt This will count all your category names according to the order in which the labels appear ; Then mark it
  
  

2、 Standardization of data set format

• First, create an empty folder , New inside make_file.py as follows , After running, it will automatically generate some necessary folder paths, such as images,labels, .yaml; Then copy the marked pictures and labels into the corresponding folders ;.yaml The file configuration is set according to your specific situation
  
  

  import os
  def make_data_dir_yaml(yaml_name, data_dir_name, nums, class_names):
      #  Generate the necessary image and label folders for the dataset 
      for dirs in ['train/images', 'train/labels', 'valid/images', 'valid/labels']:
          try:
              os.makedirs(dirs)
          except FileExistsError:
              print(f'{
          dirs}  directory already exists ')
      #  Generate dataset configuration .yaml
      with open(yaml_name, "w", encoding='utf-8') as f:
          f.write(f"train: ./{
          data_dir_name}/train/images\n")
          f.write(f"val: ./{
          data_dir_name}/valid/images\n")
          f.write("\n")
          f.write(f"nc: {
          nums}\n")
          f.write(f"names: {
          class_names}\n")
  
  
  make_data_dir_yaml('test.yaml', 'ClickData', 3, ['dog', 'cat', 'pig'])
  

3、 Training and test identification

• train.py The training is changed as follows , Focus on --data, The path is changed to your own data set .yaml file , Other configurations will be adjusted according to the actual situation , See Table 2 above for the specific parameter meanings ; And then run train.py Started training
  
  

• After the training, the results will be automatically generated to runs/train/exp Under the folder , The video 7 minute ~10 Minutes away Will introduce the folder where the training results are generated runs/train/exp The meaning of each document
  

• detect.py Identify the picture to be tested , Change several configurations as follows , Focus on --weights and --source These two configurations , Choose a well trained weight model , Choose a picture to be tested , And then run , Eventually it will be yolov5-master/runs/detect/exp Test results will be automatically generated under the path
  
  

• val.py： Verify the weight file separately , The time-consuming display of pictures and the accuracy of recognition are here 0.187 A relatively low
  

• So far, the whole process is almost over , You can also continue to optimize, such as the output coordinates of the recognition results