Note:[ wechat:Y466551 | 付费咨询,非诚勿扰 ]
论文信息
论文标题:Cluster Alignment with a Teacher for Unsupervised Domain Adaptation
论文作者:Zhijie Deng, Yucen Luo, Jun Zhu
论文来源:2020 ICCV
论文地址:download
论文代码:download
视屏讲解:click
1 介绍
2 方法
2.1 模型框架
2.2 Cluster Alignment with a Teacher
目标:discriminative learning 和 class-conditional alignment between domains?
$\min _{\theta} \mathcal{L}_{y}+\alpha\left(\mathcal{L}_{c}+\mathcal{L}_{a}\right) \quad(1)$
2.2.1 Discriminative clustering with a teacher
目标函数:
$\mathcal{L}_{c}\left(\mathcal{X}_{s}, \mathcal{X}_{t}\right)=\mathcal{L}_{c}\left(\mathcal{X}_{s}\right)+\mathcal{L}_{c}\left(\mathcal{X}_{t}\right)$
$\begin{aligned}\mathcal{L}_{c}(\mathcal{X})= \frac{1}{|\mathcal{X}|^{2}} \sum_{i=1}^{|\mathcal{X}|} \sum_{j=1}^{|\mathcal{X}|}\left[\delta_{i j} d\left(f\left(x^{i}\right), f\left(x^{j}\right)\right)+\right.\left.\left(1-\delta_{i j}\right) \max \left(0, m-d\left(f\left(x^{i}\right), f\left(x^{j}\right)\right)\right)\right]\end{aligned}$
其中 ,$\delta_{i j}$ 代表样本 $x_i$ 和 样本 $x_j$ 是不是同一类;
Note:目标域样本的标签(伪)由 教师分类器给出;
Note:可能会怀疑,教师分类器的错误预测是否会破坏训练的动态。然而,先前关于半监督学习[17,43]的研究已经验证了这种训练总是能导致良好的收敛性,并证明了对不正确标签的鲁棒性。
2.2.2 Cluster alignment via conditional feature matching
类条件特征对齐:
$\min _{\theta} \mathcal{D}\left(\mathcal{F}_{s, k} \| \mathcal{F}_{t, k}\right)$
其中,$\mathcal{F}_{s, k}\left(\mathcal{F}_{t, k}\right) $ 表示由属于源域(目标域)的类 $k$ 的所有特征组成的集合。
Cluster alignment loss 如下:$\mathcal{L}_{a}\left(\mathcal{X}_{s}, \mathcal{Y}_{s}, \mathcal{X}_{t}\right)=\frac{1}{K} \sum_{k=1}^{K}\left\|\lambda_{s, k}-\lambda_{t, k}\right\|_{2}^{2}$
其中: $\lambda_{s, k}=\frac{1}{\left|\mathcal{X}_{s, k}\right|} \sum_{x_{s}^{i} \in \mathcal{X}_{s, k}} f\left(x_{s}^{i}\right)$$\lambda_{t, k}=\frac{1}{\left|\mathcal{X}_{t, k}\right|} \sum_{x_{t}^{i} \in \mathcal{X}_{t, k}} f\left(x_{t}^{i}\right)$
2.3 Improved marginal distribution alignment
最后作者还做了一些提高,这是因为实验观察到:一开始训练的时候,teacher 对于目标域的判断并不果断,即分类结果更多聚集在分类边界附近,而不是类别中心。
目标函数:
$\begin{array}{c}\min _{\theta} \max _{\phi} \mathcal{L}_{d}\left(\mathcal{X}_{s}, \mathcal{X}_{t}\right)=\frac{1}{N} \sum_{i=1}^{N}\left[\log c\left(f\left(x_{s}^{i} ; \theta\right) ; \phi\right)\right]+ \frac{1}{\tilde{M}} \sum_{i=1}^{\tilde{M}}\left[\log \left(1-c\left(f\left(x_{t}^{i} ; \theta\right) ; \phi\right)\right) \gamma_{i}\right]\end{array}$
3 实验
标签:Domain,right,frac,Unsupervised,sum,Cluster,mathcal,left From: https://www.cnblogs.com/BlairGrowing/p/17596624.html