├── README.md
└── multilabelceloss.py


/README.md:
--------------------------------------------------------------------------------
1 | # MultilabelCrossEntropyLoss-Pytorch
2 | multilabel categorical crossentropy
3 | 
4 | This is a Pytorch implementation of multilabel crossentropy loss, which is modified from Keras version here: 
5 | 
6 | 苏剑林. (2020, Apr 25). 《将“softmax+交叉熵”推广到多标签分类问题 》[Blog post]. Retrieved from https://kexue.fm/archives/7359
7 | 


--------------------------------------------------------------------------------
/multilabelceloss.py:
--------------------------------------------------------------------------------
 1 | import torch
 2 | def multilabel_categorical_crossentropy(y_true, y_pred):
 3 |     """多标签分类的交叉熵
 4 |     说明：y_true和y_pred的shape一致，y_true的元素非0即1，
 5 |          1表示对应的类为目标类，0表示对应的类为非目标类。
 6 |     """
 7 |     y_pred = (1 - 2 * y_true) * y_pred
 8 |     y_pred_neg = y_pred - y_true * 1e12
 9 |     y_pred_pos = y_pred - (1 - y_true) * 1e12
10 |     zeros = torch.zeros_like(y_pred[..., :1])
11 |     y_pred_neg = torch.cat([y_pred_neg, zeros], dim=-1)
12 |     y_pred_pos = torch.cat([y_pred_pos, zeros], dim=-1)
13 |     neg_loss = torch.logsumexp(y_pred_neg, dim=-1)
14 |     pos_loss = torch.logsumexp(y_pred_pos, dim=-1)
15 |     return neg_loss + pos_loss


--------------------------------------------------------------------------------