Questions about calculating accuracy in training

[plastictime]

In the paper, it is mentioned that adding a sigmoid to the feature map output by the network can convert the feature map into a crack prediction graph. Then I looked at the code. The calculation of loss uses BCEWithLogitsLoss, that is, the network output is added to the sigmoid operation and then calculated. The visualisation also puts the output after the sigmoid operation as a prediction graph. However, when the accuracy is calculated, the output of the network was not processed with sigmoid. What's the reason?