As shown in the figure, during the training process, the verification set has the above problem. Not all vals are like this. I don’t understand if such an effect occurs, does it mean that the model training must have failed? According to previous experience, this phenomenon should not occur, but observations have found that this effect has a tendency to degrade. . Check it out and update again. . .
Update, the problem is finally solved, the network model and data set have been tested for a week. .
The conclusion reached so far is that if there is such a color abnormality in val, there must be a problem with the training. The solution is to check the data preprocessing part and look at the bit width, etc. .