KDLoss and KLDiscretLoss difference , KDLoss mistake

[CatDroid]

in KLDiscretLoss (classification_loss.py)
log_pt's or labels's shape is [BK, width] or [BK, height]
width or height is ‘number of categories'

def criterion(self, dec_outs, labels):
   ......
  loss = torch.mean(self.kl_loss(log_pt, labels), dim=1)  
   ......

it will calculate the mean KL divergence of 'a key point x coordinate' on ' all categories'(along the dim=1 dimension)

but in KDLoss (kd.py)

def loss(self, logit_s, logit_t, beta, weight):
    ......
    loss_all = torch.sum(self.kl_loss(s_i, t_i), dim=1)
    ......

it will calculate the sum of KL divergence along dim=1 dimension .

does doing this have a different meaning?

Beside, in 'def loss(self, logit_s, logit_t, beta, weight):' if 'if len(logit_s.shape) == 3:' not true, K is not defined , maybe a mistake ?