Potential bug in the loss function

[yygrechka]

Why is Le being subtracted in the definition of Lc?
Is this a bug or just some clever optimization?

PENCIL/pencil_train.py

Line 244 in a1b65d2

Lc = -torch.mean(torch.sum(F.log_softmax(labels_update, dim=1) * pred, dim=1)) - Le

[MohammedAlkhrashi]

I think this is the definition of the KL Divergence loss function: CrossEntropy (x,y)- Entropy(y), which is, I believe, what the author of the code is trying to do here