pytorch版Loss实现苏剑林版【将“softmax+交叉熵”推广到多标签分类问题】
from torch import nn
import numpy as np
import torch
class MultiLabelCircleLoss(nn.Module):
def __init__(self, reduction="mean", inf=1e12):
"""CircleLoss of MultiLabel, 多个目标类的多标签分类场景，希望“每个目标类得分都不小于每个非目标类的得分”
多标签分类的交叉熵(softmax+crossentropy推广, N选K问题), LSE函数的梯度恰好是softmax函数
让同类相似度与非同类相似度之间拉开一定的margin。
- 使同类相似度比最大的非同类相似度更大。
- 使最小的同类相似度比最大的非同类相似度更大。
- 所有同类相似度都比所有非同类相似度更大。
urls: [将“softmax+交叉熵”推广到多标签分类问题](https://spaces.ac.cn/archives/7359)
args:
reduction: str, Specifies the reduction to apply to the output, 输出形式.
eg.``'none'`` | ``'mean'`` | ``'sum'``. ``'none'``
inf: float, Minimum of maths, 无穷大. eg. 1e12
returns:
Tensor of loss.
examples:
>>> label, logits = [[1, 1, 1, 1], [0, 0, 0, 1]], [[0, 1, 1, 0], [1, 0, 0, 1],]
>>> label, logits = torch.tensor(label).float(), torch.tensor(logits).float()
>>> loss = MultiLabelCircleLoss()(logits, label)
"""
super(MultiLabelCircleLoss, self).__init__()
self.reduction = reduction
self.inf = inf # 无穷大
def forward(self, logits, labels):
logits = (1 - 2 * labels) * logits # <3, 4>
logits_neg = logits - labels * self.inf # <3, 4>
logits_pos = logits - (1 - labels) * self.inf # <3, 4>
zeros = torch.zeros_like(logits[..., :1]) # <3, 1>
logits_neg = torch.cat([logits_neg, zeros], dim=-1) # <3, 5>
logits_pos = torch.cat([logits_pos, zeros], dim=-1) # <3, 5>
neg_loss = torch.logsumexp(logits_neg, dim=-1) # <3, >
pos_loss = torch.logsumexp(logits_pos, dim=-1) # <3, >
loss = neg_loss + pos_loss
if "mean" == self.reduction:
loss = loss.mean()
else:
loss = loss.sum()
return loss