22 lines
608 B
Python
22 lines
608 B
Python
|
import torch
|
||
|
|
import torch.nn.functional as F
|
||
|
|
|
||
|
|
def cal_loss(pred, gold, smoothing=True):
|
||
|
|
''' Calculate cross entropy loss, apply label smoothing if needed. '''
|
||
|
|
|
||
|
|
gold = gold.contiguous().view(-1)
|
||
|
|
|
||
|
|
if smoothing:
|
||
|
|
eps = 0.2
|
||
|
|
n_class = pred.size(1)
|
||
|
|
|
||
|
|
one_hot = torch.zeros_like(pred).scatter(1, gold.view(-1, 1), 1)
|
||
|
|
one_hot = one_hot * (1 - eps) + (1 - one_hot) * eps / (n_class - 1)
|
||
|
|
log_prb = F.log_softmax(pred, dim=1)
|
||
|
|
|
||
|
|
loss = -(one_hot * log_prb).sum(dim=1).mean()
|
||
|
|
else:
|
||
|
|
loss = F.cross_entropy(pred, gold, reduction='mean')
|
||
|
|
|
||
|
|
return loss
|