Is it right to use `torch.mean(s_tgt)` when C < B？ about bnm HOT 1 OPEN

I admit I am a little careless about the weight.
In the equation, batch B is divided.
In the code, I achieve it with min(B, C). L_bnm can be combined with a hyperparameter \lambda, and in this case the actual value of \lambda is changed.
In other practices, I find the value might be better to set as \sqrt(B * C), and the performance can be better with different values of \lambda.

from bnm.