lzhangbj / D3GU

Official implementation of paper "D^3GU: Multi-target Active Domain Adaptation via Enhancing Domain Alignment" at WACV 2024

D³GU: Multi-target Active Domain Adaptation via Enhancing Domain Alignment (WACV 2024)

---

Overview

---

We propose a Multi-Target Active Domain Adaptation (MT-ADA) framework for image classification, named D³GU. It is designed to align source and multiple target domains via

• Decomposed Domain Discrimination (D³) during training to achieve both source-target and target-target domain alignments.
• GU-KMeans during active selection to sample informative images for annotation.

Installation

---

Install conda or miniconda then create new environment with the provided d3gu.yml file:

conda env create -f d3gu.yml

Dataset Preparation

---

We use Office31, OfficeHome, and DomainNet in experiments. Please create a datasets folder and download the images under it. The structure of datasets directory should be:

|-- datasets
|   |-- office31
|   |   |-- amazon
|   |   |-- dslr
|   |   |-- webcam
|   |-- office-home
|   |   |-- art
|   |   |-- clipart
|   |   |-- product
|   |   |-- real
|   |-- domain-net
|   |   |-- clipart
|   |   |-- infograph
|   |   |-- painting
|   |   |-- quickdraw
|   |   |-- real
|   |   |-- sketch

Training and Evaluation

---

Training is composed of two stages: unsupervised pretraining stage and active learning stage. Evaluation is automatically applied after training. In parallel, there are 3 domain discrimination methods {disc}:

• binary domain discrimination (bin)
• all-way domain discrimination (aw)
• docomposed domain discrimination (d3)

Unsupervised pretraining

Select discrimination method {disc}, dataset {dataset}, and find corresponding config file under config/pretrain_dann_{disc}/{dataset}. Then train on a single gpu with command:

python pretrain_dann_{disc}.py --config_file config/pretrain_dann_{disc}/{dataset}/xxx.yaml

Active training

Select discrimination method {disc}, dataset {dataset}, and active selectino algorithm {alg}. Find corresponding config file under config/active_dann_{disc}/{dataset}/target_combined/{alg}/xxx.yaml and train on a single gpu with command:

python active_dann_{disc}.py --config_file config/active_dann_{disc}/{dataset}/target_combined/alg/xxx.yaml

Pretrained checkpoints

---

We provide unsupervised pretrained checkpoints and actively trained checkpoints here:

binary domain discrimination

Method	Dataset	Link
Unsupervised Pretrain	Office31 OfficeHome DomainNet	Google Drive Google Drive Google Drive
GU-KMeans	Office31 OfficeHome DomainNet	Google Drive Google Drive Google Drive

all-way domain discrimination

Method	Dataset	Link
Unsupervised Pretrain	Office31 OfficeHome DomainNet	Google Drive Google Drive Google Drive
GU-KMeans	Office31 OfficeHome DomainNet	Google Drive Google Drive Google Drive

decomposed domain discrimination

Method	Dataset	Link
Unsupervised Pretrain	Office31 OfficeHome DomainNet	Google Drive Google Drive Google Drive
GU-KMeans	Office31 OfficeHome DomainNet	Google Drive Google Drive Google Drive

Citation

---

If you find this repo useful, please cite:

@inproceedings{zhang2024d3gu,
  title={D3GU: Multi-target Active Domain Adaptation via Enhancing Domain Alignment},
  author={Zhang, Lin and Xu, Linghan and Motamed, Saman and Chakraborty, Shayok and De la Torre, Fernando},
  booktitle = {WACV},
  year={2024}
}