Why the KL Divergence Loss needs to be multiplied by number of classes?

[tim5go]

Confident_classifier/src/run_joint_confidence.py

Line 156 in 462db01

KL_loss_fake = F.kl_div(KL_fake_output, uniform_dist)*args.num_classes

I'm bit confused about why you needa multiply the KL divergence loss by the number of classes.
I can't find it from the definition of your paper, could you briefly explain it?