Yolov5 code reproduction and server operation

---

Preface

Mentioned in the previous article , There are two ways of target detection , One is one_stage（ Single stage ） Such as YOLO One is two_stage（ Two stage ） Such as Faster_Rcnn,Mask_Rcnn. As I mentioned earlier Faster_Rcnn, This article mainly introduces YOLOV5 Code reproduction process , And the modification of the configuration file .

---

One 、YOLO brief introduction

YOLO It is a network framework for target detection , The detection speed is very fast , Suitable for some real-time detection ,YOLO Now after several generations of versions , It's now V5 edition ,V5 Provides 5l,5m,5s,5x Several versions , This article uses 5s Model of .

Two 、 The code download

You can download to Baidu online disk by yourself , Link attached ：

link ： Baidu SkyDrive
Extraction code ：qnn5

3、 ... and 、 Data set preparation

In the last article , It introduces xml turn txt Code process of document , You can check it by yourself , Link attached ：
Code
After the transformation , It can also be divided into training set and test set

images Place the file
labels place txt file
What we should pay attention to here is , The folder name must be images Follow labels, The original code is as follows ：

def img2label_paths(img_paths):
    # Define label paths as a function of image paths
    sa, sb = os.sep + 'images' + os.sep, os.sep + 'labels' + os.sep  # /images/, /labels/ substrings
    return [sb.join(x.rsplit(sa, 1)).rsplit('.', 1)[0] + '.txt' for x in img_paths]

Four 、 Modification of configuration file

1.data Under the yaml

modify data Under directory VOCyaml file ,

 Before the change ：
path: ../datasets/VOC
train: # train images (relative to 'path') 16551 images
  - images/train2012
  - images/train2007
  - images/val2012
  - images/val2007
val: # val images (relative to 'path') 4952 images
  - images/test2007
test: # test images (optional)
  - images/test2007
 After modification 
train: # train images (relative to 'path') 16551 images
  - /mnt/wu/images/train
val: # val images (relative to 'path') 4952 images
  - /mnt/wu/images/val

Modify the data path here

 Before the change ：
nc: 20  # number of classes
names: ['aeroplane', 'bicycle', 'bird', 'boat', 'bottle', 'bus', 'car', 'cat', 'chair', 'cow', 'diningtable', 'dog',
        'horse', 'motorbike', 'person', 'pottedplant', 'sheep', 'sofa', 'train', 'tvmonitor']  # class names
 After modification ：
nc: 4  # number of classes
names: ["Traffic_Light_go","Traffic_Light_stop","Traffic_Light_warning",'Traffic_Light_ambiguous'] # class names

Change here nc Follow the label
nc： It indicates how many categories of data you train .

2.models Under the yaml

What we use here is 5s Model of , change yolov5s.yaml file

It just needs to be changed nc Sure , How many categories of data represent your training .

3. Training train

1. change YOLOV5s The path of the model

parser.add_argument('--weights', type=str, default=ROOT / 'yolov5s.pt', help='initial weights path')

2. Change the above modification models Of yaml File path

parser.add_argument('--cfg', type=str, default='models/traffic.yaml', help='model.yaml path')

3. Change the above modification data Of yaml File path

parser.add_argument('--data', type=str, default=ROOT / 'data/traffic.yaml', help='dataset.yaml path')

4. Training batch

parser.add_argument('--epochs', type=int, default=50)

5. Batch size ,batch_size If there is insufficient video memory during training, it can be reduced batch_size Size

parser.add_argument('--batch-size', type=int, default=64, help='total batch size for all GPUs, -1 for autobatch')

5、 ... and 、 Training with server

image YOLO This depth frame , It's best to train with a server , Bloggers use moment pool cloud , Link attached Quechi cloud
Attach invitation code ：BQlYmrQhDiex1lZ

1. Upload data

Before renting a server , Upload data first , If the data is large , Uploading after renting a server will waste time , It costs money .
Enter my online disk , Upload data .

2. Rent servers

Choose the model of the graphics card here , Rent .

Environmental choice yolov5 Environmental Science , Omit the link of matching the environment .

3.pycharm Connect to server

After renting , Enter the rental interface , Display the details of the leased server .

Use here ssh Connect .

1. add to ssh

According to the ssh link , Input host , user name , port .

2. Input password

3. Configure the server environment path and code mapping path

Open the link , Access terminal .

Input conda env list View the environment path

Configure the environment path , Code mapping path
mnt The path below is your own network disk path .

As shown in the figure , Configuration complete .

4. Decompress data

Open the terminal ,cd To the compressed data set directory ,unzip file name unpack , As shown in the figure below ：

5. Start training

Wait for the code to be uploaded to the server , Follow the server dataset path , modify data Below yaml Data set path in file , You can train .

As shown in the figure below ： It is to train with a server .


As shown in the figure , Start training .


End of training .

Under this path is the weight file generated by training .


best.ph Is the best weight file .
last.ph It is the weight file of the last training .

6. Download the file locally

We can go through winscp This software , A good folder for training , Download to local
1. Enter host name , Port number , user name , password , Connect to server

As shown in the figure , Successful connection

2. Find the folder , Download to local .

6、 ... and 、 test

We can test according to the trained weight .
Test code detect.py

Weight file path

 parser.add_argument('--weights', nargs='+', type=str, default=ROOT / 'runs/train/exp6/weights/best.pt', help='model path(s)')

The test image , Video path ,0 Indicates that the camera is turned on .

parser.add_argument('--source', type=str, default="traffic.mp4", help='file/dir/URL/glob, 0 for webcam')

After setting, you can test .



Testing is completed , The save path... Will be displayed .
Open folder , See the effect .

summary

That's all YOLOV5 The whole process of Server Replication , Everyone is in the process of reproduction , Problems encountered , You can feed back to the comment area .