doubt regarding the final loss about jule.torch HOT 2 OPEN

I am confused too. I have searched in the source code, but I found that the criterion function "TripletEmbeddingCriterion" computed rather using the distances based on the representations learned from the CNN than the affinity between two clusters. And why "TripletEmbeddingCriterion" is more similar to the loss defined in FaceNet which the paper referenced , but not same with the formula proposed in the paper? The Loss proposed in the paper actually uses the affinity between clusters, but why the source code seems to only use the affinity in agglomerative clustering and updating label except computing the triplet loss?

from jule.torch.

In general, the loss used to optimize DNN is >=0, and the purpose to optimize DNN can be expressed to making loss zero in a mathematical form .
But why the triplet loss defined in the paper(equation 7 or 11) is <=0 forever, so what's the destination/criterion of the loss to optimize the CNN? is zero? if it is zero, why add a minus sign?
Hoping anybody comment or answer my problem, thanks a lot.

from jule.torch.