nll_loss 输入的是求过log_softmax之后的值，默认reduction=‘mean‘，计算的是平均loss，即将targets对应的log_prob相加再求均值：
>>> import torch >>> logits=torch.randn(2,3) >>> logits tensor([[-0.1818, -1.2657, 1.6381], [ 0.2038, 0.2661, -2.3768]]) >>> probs=torch.nn.functional.log_softmax(logits, dim=-1) >>> probs tensor([[-2.0162, -3.1000, -0.1963], [-0.7608, -0.6985, -3.3414]]) >>> targets=torch.tensor([0,1]) >>> targets tensor([0, 1]) >>> loss = torch.nn.functional.nll_loss(probs, targets) >>> loss tensor(1.3574) >>> loss = torch.nn.functional.nll_loss(probs, targets, reduction=‘none‘) >>> loss tensor([2.0162, 0.6985]) >>> loss = torch.nn.functional.nll_loss(probs, targets, reduction=‘mean‘) >>> loss tensor(1.3574) >>> loss = torch.nn.functional.nll_loss(probs, targets, reduction=‘sum‘) >>> loss tensor(2.7147) >>> loss = torch.nn.functional.nll_loss(probs, targets, reduction=‘none‘) >>> loss.mean() tensor(1.3574)
交叉熵Loss，输入是没有做softmax的logits，torch.nn.CrossEntropyLoss() 计算结果= log_softmax 结合 nll_loss
>>> ce = torch.nn.CrossEntropyLoss() >>> loss = ce(logits, targets) >>> loss tensor(1.3573)
pytorch nll_loss与CrossEntropyLoss示例
原文：https://www.cnblogs.com/AliceYing/p/14603346.html