[论文笔记] Domain-Adaptive Few-Shot Learning

原文链接：https://openaccess.thecvf.com/content/WACV2021/papers/Zhao_Domain-Adaptive_Few-Shot_Learning_WACV_2021_paper.pdf

github链接：[GitHub - dingmyu/DAPN: A pytorch implementation of "Domain-Adaptive Few-Shot Learning"](https://github.com/dingmyu/DAPN "GitHub - dingmyu/DAPN: A pytorch implementation of "Domain-Adaptive Few-Shot Learning"") （有bug和文件缺失，无法运行）

---

Methodology (论文原图，侵删)

Problem Definition

given - a large sample set from a set of source classes in a source domain

- a few-shot sample set from a set of target classes in the target domain

- a test set from another set of target classes in the target domain,

where 1) , ,

2. data distributions on and are also different

Objective: training a model with and and then evaluating its generalization
ability on

---

Episode Training

To form a training episode :

1. randomly choose classes from

2. build two sets of samples from the classes: the support set consists of samples (k samples per class), and the query set is composed of samples from the same classes.

Also build training episodes from the few-shot sample set Dd (data augmentation method needed).

To form a training episode :

1. randomly choose classes from

2. build two sets of samples from the classes: the support set consists of samples (k samples per class), and the query set is composed of samples from the same classes.

---

Feature Extractor

原文代码的 Feature Extractor 使用的是ResNet18

---

Prototypical Learning

基本思路是基于Prototypical network: Learn a prototype of each class in the support set , and classifies each sample in the query set based on the distances between each sample and different prototypes.

sample embedding 和 class prototype 的距离可以用以下公式计算(dist 是欧氏距离)

loss function over each episode 就是基于每一个query sample的negative log-probability:

同理，loss function over each episode 就是基于每一个query sample的negative log-probability:

---

Domain Adversarial Adaptation Module

1. Domain Adaptive Embedding

The embedding module consists of an autoencoder and an attention submodule.

目的： 为了使得到的feature尽可能domain-confused

2. Domain Adaptive Loss

根据文章Conditional Domain Adversarial Network (CDAN) 额外加入一个domain
discriminator 来处理 source distribution 和 target distribution 之间的 domain adversarial loss：

---

Domain Discriminative Loss

The features before and after the embedding layer with self-attention are distinguished and confused, respectively:

---

Adaptive Reweighting Module

用自适应模块将四个loss进行累加

---

Experiments

[略]在miniImageNet，tieredImageNet和DomainNet 上进行训练