Practical blog Guide ( Ongoing update …)：

1. Construction of actual combat environment
2. Custom datasets
3. model training
4. Model testing and evaluation
5. YOLOv5 Integrate PyQt5

One 、 Parameter description

In the last step , We trained the model , The screenshot of the training is as follows ：

Explain the meaning of each field here ：

• Epoch: num0/num1,num0 Indicates the current number of rounds ,num1 Indicates the total number of rounds . The picture shows 0/19, To express a total of 20 round , The present is the 0 round .
• gpu_mem: GPU Memory usage .
• The middle ones will be explained in the training results .
• The rightmost digits indicate the current training status .8/57 This time epoch Need to carry out 5 Sub iteration , The current number of iterations is 8.
• 00:11<01:09, This time epoch Has run 11s, It is estimated that it needs to be operated 1m9s.1.43s/it Indicates that each iteration probably requires 1.43s.( It can be calculated simply ,1.43*57=81.51≈1m9s+11s=80s). With the help of these parameters, you can judge your every epoch About how long it will take , Then multiply it by yours epoch, We can roughly infer the training time .

Two 、 Evaluation of results

After training , It will be in your corresponding result directory (runs/train/expi,i For your number of training ) Next . For example, I have been trained three times before to test whether there is a problem , Without deleting the relevant files , Then complete the training for the fourth time , So the results are saved in exp3 Next , Pictured ：

weights/ The catalog saves the best and last training models in your training process .
exp3/ Contains many files and pictures , Here's a brief introduction result.png, Briefly introduce other documents ：

2.1 results.png

results.png as follows ：

• Box: YOLO V5 Use GIOU Loss As bounding box The loss of ,Box Presumably GIoU The loss function means , The smaller the box, the more accurate , That is, the coincidence degree between the prediction box and the real value box ;
• Objectness: It's supposed to be target detection loss mean value , The smaller the target, the more accurate the detection ;
• Classification： It is presumed to be classified loss mean value , The smaller the size, the more accurate the classification ;

• Precision： Precision rate （ Find the right positive class / All positive classes found ）. Its calculation formula is ：$Precision=\frac{TP}{TP+FP}$, Indicates the probability that the sample you think is true is actually true , Indicates your accuracy .
• Recall: Recall rate （ Find the right positive class / All the positive classes that should have been found right ）. Its calculation formula is ：$Recall=\frac{TP}{TP+FN}$, It means the probability of finding the right one in all the samples that are true .
• [email protected] & [email protected]:0.95： Namely mAP Yes, it is Precision and Recall As the area enclosed by two axes ,m Means average ,@ The following number denotes a decision iou Is the threshold of positive and negative samples ,@0.5:0.95 The threshold value is taken as 0.5:0.05:0.95 Take the mean value after .

General training results mainly observe the fluctuation of accuracy and recall rate （ If the fluctuation is not great, the training effect is better ）. Then observe [email protected] & [email protected]:0.95 Evaluate the training results , Generally speaking, the larger the two values, the better , It means your model is better .

2.2 Screenshots of other files

1. confusion_matrix.png( Confusion matrix )
   Confusion matrix can summarize the prediction results of classification problems , It shows which part of the classification model will be confused when making predictions . The vertical axis is the predicted value , The horizontal axis is the true value .
   

2. F1_curve：
   F1 The relationship between score and confidence .F1 fraction （F1-score） It's a measure of classification , It's accuracy precision And recall rate recall Harmonic mean of , The maximum is 1, The minimum is 0, 1 It's the best ,0 Is the worst
   

3. labels.jpg
   The first picture classes: The amount of data in each category
   The second picture labels： label
   The third picture center xy
   The fourth picture labels The length and width of the label
   

4. P_curve.png :
   Accuracy rate precision And confidence confidence Diagram for
   

5. PR_curve.png:
   PR In the curve P It stands for precision（ Accuracy ）,R It stands for recall（ Recall rate ）, It represents the relationship between accuracy and recall , In general , take recall Set to abscissa ,precision Set to ordinate .PR The area enclosed under the curve is AP, All categories AP The average value is Map.
   If PR One of the curves in Figure A Completely enclose the curve of another learner B, It can be asserted that A Better performance than B, When A and B When crossing occurs , It can be compared according to the area under the curve . General training results mainly observe the fluctuation of accuracy and recall rate （ If the fluctuation is not great, the training effect is better ）.
   

6. R_curve.png : The relationship between recall rate and confidence
   

Observe the screenshot above and you can find , Blogger's model is spicy chicken ( After all, the equipment doesn't work , Fewer datasets are used , Resulting in low accuracy ).

3、 ... and 、 Model test

python detect.py --weights weights/best.pt --source data/images/test --img-size 640 --device 0

The above command specifies that the weight of the model used in the test is the trained model best.pt, The test data set is test/ Catalog , It can also be specified as a single picture , The process of testing is as follows ：

The results are stored in runs/detect/exp/ Under the table of contents . An example is shown in the figure ：