Domain Adaptation 論文閱讀筆記

文章目錄

• Domain Adaptation 論文閱讀筆記
• 一、Method Summary
• • Unsupervised Domain Adaptation by Backpropagation
  • Learning Transferable Features with Deep Adaptation Networks
  • Coupled Generative Adversarial Networks
  • Domain Separation Networks
  • DiDA: Disentangled Synthesis for Domain Adaptation
  • Unsupervised Domain Adaptation in the Wild via Disentangling Representation Learning
  • Unsupervised Domain Adaptation via Disentangled Representations: Application to Cross-Modality Liver Segmentation
  • Learning from Synthetic Data: Addressing Domain Shift for Semantic Segmentation
  • Contrastive Adaptation Network for Unsupervised Domain Adaptation (CVPR 2019)
  • MME: Semi-supervised Domain Adaptation via Minimax Entropy
  • PAC: Surprisingly Simple Semi-Supervised Domain Adaptation with Pretraining and Consistency (BMVC 2021)
  • Learning Invariant Representations and Risks for Semi-supervised Domain Adaptation
• 2. Experiment part
• • 1. (Unsupervised) Domain Adaptation
  • 2. Joint-Domain Learning
• 3. Analysis part

---

What is Domain Adaptation(DA)? — attempt to map representations between the two domains or learn to extract features that are domain–invariant.
source有label，target沒有

一、Method Summary

Unsupervised Domain Adaptation by Backpropagation

domian classifier部分的梯度需要通過gradient reverse layer，使encoder提到的信息不利于domain 分類，也就是domain-invariant feature

Learning Transferable Features with Deep Adaptation Networks

(https://blog.csdn.net/weixin_40526176/article/details/79065861)

• 多層適配
  • 適配最后3層——認為（AlexNet）最后3層是task-specific，對于其他網絡要另外計算相似度
• Multi kernel-MMD(Maximum Mean Discrepancy)
  • 可以用來計算不同域feature的距離，相當于把1中的maximize domain error換成這邊的minimize MMD

Coupled Generative Adversarial Networks

• 即便沒有labeled cross-domain pair，也可以通過weight sharing和adversarial learning學習到2個domain的joint distribution——相當于輸入同一個vector z，2個generator的輸出是一對語義相關但是各有特點的pair。
• weight sharing如highlight部分所示，其中z是random vector，因為有了weight sharing，可以保證對應高層語義信息的layer，其處理信息的方式是一致的。
• 這似乎不是DA，但是這個框架可以用在DA上，效果似乎很不錯——因為雖然target沒有label，但是source有label，并且有weight sharing機制，使得2個generator得到的圖像g(z)理論上是同一個數字。

Domain Separation Networks

構建一個直接提取domain-invariant的框架，會導致 these representations might trivially include noise that is highly correlated with the shared representation.

• Overall Loss:
  
• Reconstruction Loss:
  • 用scale mse，因為普通mse penalizes predictions that are correct up to a scaling term.，而scale msepenalizes differences between pairs of pixels. This allows the model to learn to reproduce the overall shape of the objects being modeled without expending modeling power on the absolute color or intensity of the inputs. （為什么scale會導致model分心？）
    
• L_dif: （可以encourages orthogonality，why？）
  
• L_sim:
  • domain classfier(gradient reverse layer)+CE loss
  • MMD loss

DiDA: Disentangled Synthesis for Domain Adaptation

通過交替進行domain adaptation和disentangle synthesis這兩步，逐漸得到更好的labeled target data

• DA：訓domain-invariant common feature
• Disentangle：在DA的基礎上，訓specific feature，要讓common和specific的combination可以reconstruct input，但是這個specific feature得對分類不利（這邊可能有個GRL？）

Unsupervised Domain Adaptation in the Wild via Disentangling Representation Learning

• As the category information between the source and the target domains can be imbalanced, directly aligning latent feature representation may lead to negative transfer.
• So they disentangle the latent feature to category related code (global code) as well as style related code (local code).

Unsupervised Domain Adaptation via Disentangled Representations: Application to Cross-Modality Liver Segmentation

• 對每個domain分別提取style code $s_i$和content code $c_i$，然后把這些code輸入G中（怎么輸？），得到對應的img（要得到content-only img，必須解耦才行嗎？）
• 通過這樣的訓練，可以得到content-only img
• 然后用得到的content-only img來訓練一個新的模型
• 這個方法可以用來做domain adaptation，也可以做joint-domain learning

Learning from Synthetic Data: Addressing Domain Shift for Semantic Segmentation

(https://github.com/swamiviv/LSD-seg)

• Discriminator: Patch discriminator的變形，each pixel in the output map indicates real/fake probabilities across source and target domains hence resulting in four classes per pixel: src-real, src-fake, tgt-real, tgt-fake.
• Auxiliary Classifier GAN (ACGAN)思想：by conditioning G during training and adding an auxiliary classification loss to D, they can realize a more stable GAN training and even generate large scale images. —— 或許可以用來reconstruct full img

Iteratively update:

• F的訓練和cross-domain有關系：To update F, we use the gradients from D that lead to a reversal in domain classification, i.e. for source embeddings, we use gradients from D corresponding to classifying those embeddings as from target domain.

Contrastive Adaptation Network for Unsupervised Domain Adaptation (CVPR 2019)

• 類似MMD，提出了個CDD，用來拉近fc層處target和source的距離
• Alternative optimization：先cluster，得到pseudo target label，然后根據這些label去用CDD算intra-class、inter-class discrepancy，再回去更好的cluster

MME: Semi-supervised Domain Adaptation via Minimax Entropy

• 先用labeled訓F+C，F提feature，C包含一組weight (the weight vectors can be regarded as estimated prototypes for
  each class.)，將feature轉換為domain-invariant prototype
• 然后對F minimize entropy——得到discriminative feature
• 對C maximize entropy (similarity)——讓每類prototype (C的weight) 都和unlabeled target feature相近

PAC: Surprisingly Simple Semi-Supervised Domain Adaptation with Pretraining and Consistency (BMVC 2021)

先用rotation pretext task pretraining，然后再做domain adaptation fine-tuning，對labeled img要滿足分類正確，對unlabeled img要讓perturb版本輸出和原來差不多

Learning Invariant Representations and Risks for Semi-supervised Domain Adaptation

• 挺理論的一篇，核心在于同時 learn invariant representation 和 invariant risk (data are collected from multiple envrionments with different distributions where spurious correlations are due to dataset biases. This part of spurious correlation will confuse model to build predictions on unrelated correlations rather than true causal relations.)
• 讓他們的optimal predictor對齊？

2. Experiment part

1. (Unsupervised) Domain Adaptation

• Train:source
• Test: target

methodexp setup

Unsupervised Domain Adaptation by Backpropagation	如果source比target更復雜，則還行；source比target簡單，就不太行
Learning Transferable Features with Deep Adaptation Networks	1. Unsupervised adaptation -> use all source examples with labels and all target examples without labels 2.semi-supervised adaptation -> randomly down-sample the source examples, and further require 3 labeled target examples per category.
Domain-Adversarial Training of Neural Networks
Coupled Generative Adversarial Networks
Domain Separation Networks	用2個baseline作為lower bound和upper bound（不用DA，只在source或只在target上訓練）

2. Joint-Domain Learning

• 多個domain的數據混一起train
• 目標：得到的結果比只在單個domain上train的好

3. Analysis part

Visualization(t-SNE): 證明在Target域模型得到的feature是：

easier to discriminate

more align with source

總結

以上是生活随笔為你收集整理的Domain Adaptation 论文笔记的全部內容，希望文章能夠幫你解決所遇到的問題。

如果覺得生活随笔網站內容還不錯，歡迎將生活随笔推薦給好友。