1、介绍
Bi-Directional Generation for Unsupervised Domain Adaptation结合了双向对抗生成网络和多类别领域适应的方法实现了源域到目标域的知识迁移适应,论文在office数据集上测试正确率较高,基本网络结构图如下所示:

Bi-Directional Generation for Unsupervised Domain Adaptation(Bi-Directional Generation & class MMD)

2、数学原理
该方法总共结合了5个损失函数,在多批次训练后达到收敛,各损失函数介绍如下:

2.1. Gs与Cs结合,达到生成器Gs和Cs在训练中收敛:

Bi-Directional Generation for Unsupervised Domain Adaptation(Bi-Directional Generation & class MMD)

2.2. Gt与Ct结合,达到生成器Gt和Ct在训练中收敛:

Bi-Directional Generation for Unsupervised Domain Adaptation(Bi-Directional Generation & class MMD)

2.3. 对每个类别计算损失函数,采用class MMD技术:

Bi-Directional Generation for Unsupervised Domain Adaptation(Bi-Directional Generation & class MMD)

2.4. 计算全局MMD:

Bi-Directional Generation for Unsupervised Domain Adaptation(Bi-Directional Generation & class MMD)

2.5. 按经验参数整合全局MMD和class MMD:

Bi-Directional Generation for Unsupervised Domain Adaptation(Bi-Directional Generation & class MMD)

2.6. 计算分类器分类差别损失:

Bi-Directional Generation for Unsupervised Domain Adaptation(Bi-Directional Generation & class MMD)

2.7. 整合所有损失函数:

Bi-Directional Generation for Unsupervised Domain Adaptation(Bi-Directional Generation & class MMD)

相关文章:

  • 2021-12-02
  • 2021-05-28
  • 2021-10-11
  • 2021-09-16
  • 2021-11-17
  • 2021-07-25
  • 2021-07-04
猜你喜欢
  • 2021-05-27
  • 2021-04-07
  • 2021-06-05
  • 2021-12-14
  • 2021-10-23
  • 2021-08-23
  • 2021-08-08
相关资源
相似解决方案