[yolov5 practice 3] traffic sign recognition system based on yolov5 - model training

Practical blog Guide ( Ongoing update …)：

1. Construction of actual combat environment
2. Custom datasets
3. model training
4. Evaluation of results

One 、 Parameter description

After the first two steps , Start model training and testing . First, model training . The official training order is as follows ：

python train.py --data coco.yaml --cfg yolov5n.yaml --weights '' --batch-size 128
                                       yolov5s                                64
                                       yolov5m                                40
                                       yolov5l                                24
                                       yolov5x                                16

We can see train.py file , It sets the parameters we need to specify during training , As shown in the figure ( This parameter has been modified by me , The initial value is different from that shown in this figure , We should modify it according to our own situation )：

Briefly introduce the parameters ：

• weights: Specify the pre training model .yolov5 Several pre training models are officially provided to improve training efficiency , I'll talk about it later .
• data: Written in the last Practical Guide myvoc.yaml, Used to specify the path of training set and verification set , And the number and name you want to classify , Must be modified
• epochs: How many rounds of training , Suggest to modify
• batch-size: Batch size , Personal understanding is how many data are read and processed each time ( Modify according to your own configuration , If you report an error of insufficient memory, you can reduce the value , Suggest to modify )
• img-size: The size of the training set and the verification set , The original data set resize Set the size for . The default is [640, 640], It can be left unchanged
• resume: Break point continued training . This is mainly because if you interrupt during training , You can continue training after the last training result . The default is False
• device: Set use CPU still GPU( If GPU How many , Specify which piece to use ).
• workers: It seems to be the number of threads , The equipment can be adjusted if allowed , Because my device is too spicy, it is set to 0

Basically, the above parameters are commonly used , Other parameters can keep their default values , You can also make corresponding modifications according to your actual situation .

Two 、 Pre training model

Officials have given the following pre training models .

Commonly used n,s,m,l Four sizes , The specification gradually increases , The training accuracy is also gradually increasing , But the training time will also increase . As shown in the figure below ：

In this blog I will use yolov5s.pt Model training .

3、 ... and 、 model training

After the above preparation steps are completed , Start training model . Training parameters can be in train.py The document specifies , You can also explicitly specify , The priority of the command line is higher than that in the file . I explicitly specify the parameters here , You may not specify , That is to use the parameter information in the file . We need to pay attention to , If you didn't download it at first yolov5s.pt( namely –weights Specified file ), Then he will download the corresponding file .( It seems that the default download path is not weights/ Next , Personal habit moves him to weights/ Under the table of contents ). If the download speed is too slow, it is recommended to modify pip Source , This can be changed by reading others' blogs , Then the speed will be much faster .
Open the terminal , Enter the following command ：

python train.py --weights weights/yolov5s.pt --data data/myvoc.yaml --epochs 20 --batch-size 8

If no error is reported , You should get the following interface ：

The storage location of this result training will be displayed , Put it in runs/train/ Next , from exp Start , Every training will add one to the number , I've got results here before , Therefore exp5/.

You can check in runs/train/ Whether the directory has been generated ：

Four 、 Next up

The next blog will introduce how to use the trained model to verify and test your data set , And judge the quality of the model .