我们从Python开源项目中，提取了以下2个代码示例，用于说明如何使用torch.nn.KLDivLoss()。
def __init__(self, generator, tgt_vocab, label_smoothing=0.0): super(NMTLossCompute, self).__init__(generator, tgt_vocab) # CHECK assert (label_smoothing >= 0.0 and label_smoothing <= 1.0) # END CHECK if label_smoothing > 0: # When label smoothing is turned on, # KL-divergence between q_{smoothed ground truth prob.}(w) # and p_{prob. computed by model}(w) is minimized. # If label smoothing value is set to zero, the loss # is equivalent to NLLLoss or CrossEntropyLoss. # All non-true labels are uniformly set to low-confidence. self.criterion = nn.KLDivLoss(size_average=False) one_hot = torch.randn(1, len(tgt_vocab)) one_hot.fill_(label_smoothing / (len(tgt_vocab) - 2)) one_hot[0][self.padding_idx] = 0 self.register_buffer('one_hot', one_hot) else: weight = torch.ones(len(tgt_vocab)) weight[self.padding_idx] = 0 self.criterion = nn.NLLLoss(weight, size_average=False) self.confidence = 1.0 - label_smoothing
def kldiv_loss(loss_weight=None, size_ave=True): return nn.KLDivLoss(weight=loss_weight,size_average=size_ave)
