Loss function - divide by N?

[simran-modi]

In the original paper, the sum of confidence loss and localization loss is divided by "N", where N is the number of matched default boxes. I fail to see where that division is implemented in the code. Without that my loss is very high...