第一周学习总结

• 《Maximum Classififier Discrepancy for Unsupervised Domain Adaptation》

论文小结

一、《Maximum Classififier Discrepancy for Unsupervised Domain Adaptation》

1、Method

• 使用两个分类器（F1，F2)，去寻找在Target中判定错误的区域。
  

2、Training Steps

• Step A

  • train both classififiers and generator to classify the source samples correctly.

  • 损失函数

    • L ( X S , Y S ) = − E ( x S , y s ) − ( X S , Y S ) ∑ k = 1 K 1 [ k = y S ] l o g p ( y ∣ x S ) L(X_S,Y_S) = -E_{(x_S,y_s)-(X_S,Y_S)}\sum^{K}_{k=1}{1_{[k=y_S]}logp(y|x_S)} L(XS​,YS​)=−E(xS​,ys​)−(XS​,YS​)​k=1∑K​1[k=yS​]​logp(y∣xS​)

    • m i n G , F 1 , F 2 L ( X S , Y S ) min_{G,F1,F2} L(X_S,Y_S) minG,F1,F2​L(XS​,YS​)

• Step B

• 

• 损失函数

  • m i n F 1 , F 2 L ( X S , Y S ) − L a d v ( X t ) min_{F1,F2} L(X_S,Y_S) - L_{adv}(X_t) minF1,F2​L(XS​,YS​)−Ladv​(Xt​)

  • L a d v ( X t ) = E x t − X t [ d ( p 1 ( y ∣ x t ) , p 2 ( y ∣ x t ) ) ] L_{adv}(X_t)=E_{x_t-X_t}{[d{(p_1(y|x_t),p_2(y|x_t))}]} Ladv​(Xt​)=Ext​−Xt​​[d(p1​(y∣xt​),p2​(y∣xt​))]

• Step C

  -

  • 损失函数

    • m i n G L a d v ( X t ) min_{G} L_{adv}(X_t) minG​Ladv​(Xt​)

3、Experiments on Classifification

• 结果图

• 参数设置

  • Optim： Adam

  • Learning rate: 0.0002

  • batch size : 128

  • hyper-parameter : num_k (2 - 4)

论文复现

• 《Maximum Classififier Discrepancy for Unsupervised Domain Adaptation》

一、源码分析

1. 网络模型

• svhn to mnist（model）

  • G： conv1 => bn1 => max_pool

    ​ => conv2 => bn2 => max_pool

    ​ => conv3 => bn3 =>

    ​ => fc1 => bn1_fc => dropout

    ​ => out

  • F：fc1 = > bn1_fc => fc2 = > bn2_fc => fc3 = > bn3_fc =>out

  • 不太明白的地方

    • 在F结构中的forward加了一个reverse的选项。

- grad_reverse函数

• syn to gtsrb

  • G： conv1 => bn1 => max_pool

    ​ => conv2 => bn2 => max_pool

    ​ => conv3 => bn3 => max_pool

    ​ => view（拉平）

    ​ => out

  • F： => fc2 = > bn2_fc => fc3 = > bn3_fc => out

• usps

  • G： conv1 => bn1 => max_pool

    ​ => conv2 => bn2 => max_pool

    ​ => view（拉平）

    ​ => out

  • F： fc1 = > bn1_fc => fc2 = > bn2_fc => fc3 = > bn3_fc => out

二、实验结果

METHOD	SVHN to MNIST	SYNSIG to GTSRB	MNIST to USPS	USPS to MNIST
Source Only	70.1	92.5	68.6	58.3
MCD(n = 4)	95.52 ± 0.58	95.75 ± 0.46	94.06 ± 0.36	96.79 ± 0.31