README.md

Implementation for CoSDA: Continual Source-Free Domain Adaptation

README | 中文文档

Here is the code for our work CoSDA:Continual Source-Free Domain Adaptation. To ensure fair comparison, we build a unified codebase for the methods of source-free domain adaptation and continual DA, as shown in supported methods.

• Implementation for CoSDA: Continual Source-Free Domain Adaptation
  • Introduction
  • Installation and Usage
    • Installation of Datasets.
    • Source-free Domain Adaptation
  • Supported Methods
  • Citation

Introduction

Continual source-free domain adaptation is a new and practical task in the field of domain adaptation, which seeks to preserve the performance of a model across all domains encountered during the adaptation process while also protecting the privacy of private data, as illustrated in the following figure:

CoSDA is a continual source-free domain adaptation approach that employs a dual-speed optimized teacher-student model pair and is equipped with consistency learning, as shown in the following figure. The implementaion details of CoSDA are shown in [train/cosda/cosda.py].

Installation and Usage

Installation of Datasets

First, download the datasets from the following links:

• DomainNet
• OfficeHome: (1) image_list (2) Art (3) Clipart (4) Product (5) Real_World
• Office31: (1) image_list (2) amazon (3) dslr (4) webcam
• VisDA17

Next, select a base_path, create a dataset folder within it and place the downloaded files in this folder, as shown below:

base_path
├── dataset
    ├── DomainNet
    │   ├── splits
    │   ├── clipart
    │   ├── infograph
    │   ├── painting
    │   ├── quickdraw
    │   ├── real
    │   ├── sketch
    ├── Office31
    │   ├── image_list
    │   ├── amazon
    │   ├── dslr
    │   └── webcam
    ├── OfficeHome
    │   ├── image_list
    │   ├── Art
    │   ├── Clipart
    │   ├── Product
    │   ├── Real_World
    └── Visda2017
        ├── image_list
        ├── train
        └── validation

Source-free Domain Adaptation

• Dependencies and environment setup.

  pip install -r requirements.txt
  

• Pretrain.

  python pretrain.py -bp [base_path] --config [config_file] 
  

  The base_path is the selected location where the dataset will be installed. The config_file is stored in the [pretrain/config/backup] directory. We have designed specific configurations for GSFDA, and SHOT++, while for other methods, we use the same pretrain configuration as SHOT method used.

  Once the pretraining process is complete, the model will be saved within the base_path directory. Below is an example of the resulting file structure for DomainNet:

  base_path
  ├── DomainNet
      ├── pretrain_parameters_shot
      │   ├── source_{}_backbone_{}.pth.tar
  

  • We supply pretrained parameters for all datasets using the SHOT, SHOT++, and GSFDA methods. Download Links: DomainNet, OfficeHome, Office31, VisDA2017.

• Single-target adaptation.

  python single_tar.py -bp [base_path] --config [config_file] --writer [tensorboard / wandb]

  We have created separate configuration files for each method, which can be found in the [adaptationcfg/backup)] directory. The source_domain and target_domain can be manually specified under the DAConfig key in the configuration file.

  We provide support for two methods of recording the training process: tensorboard and wandb. To use tensorboard, you need to specify the log_path to store the event files locally using -lp [log_path]. To use wandb, you need to specify the entity using -e [entity].

• Multi-targets adaptation.

  python multi_tar.py --config [config_file] -bp [base_path] --writer [tensorboard / wandb] (-lp [log_path] or -e [entity])

  The settings for multi-target domain adaptation are the same as those for single-target adaptation. For DomainNet, the sequential adaptation order is Real → Infograph → Clipart → Painting → Sketch → Quickdraw. For OfficeHome, the sequential adaptation order is Art → Clipart → Product → Real-world.

Supported Methods

Apart from CoSDA, we also support the following methods.

• SHOT - Do We Really Need to Access the Source Data? Source Hypothesis Transfer for Unsupervised Domain Adaptation. (ICML'20) [train/shot/shot_plus.py]

• SHOT++ - Source Data-absent Unsupervised Domain Adaptation through Hypothesis Transfer and Labeling Transfer. (TPAMI) [train/shot/shot_plus.py]

• G-SFDA - Generalized Source-free Domain Adaptation. (ICCV'21) [train/gsfda/gsfda.py]

• NRC - Exploiting the Intrinsic Neighborhood Structure for Source-free Domain Adaptation. (NeurIPS'21) [train/nrc/nrc.py]

• AaD - Attracting and Dispersing: A Simple Approach for Source-free Domain Adaptation. (NeurIPS'22) [train/aad/aad.py]

• DaC - Divide and Contrast: Source-free Domain Adaptation via Adaptive Contrastive Learning. (NeurIPS'22) [train/dac/dac.py]

• Edgemix - Balancing Discriminability and Transferability for Source-Free Domain Adaptation. (ICML'22) [train/dataaug/edgemix.py]

• CoTTA - Continual Test-Time Domain Adaptation. (CVPR'22) [train/cotta/cotta.py]

Citation

If you find our paper helpful, consider citing us via:

@article{feng2023cosda,
  title={CoSDA: Continual Source-Free Domain Adaptation},
  author={Feng, Haozhe and Yang, Zhaorui and Chen, Hesun and Pang, Tianyu and Du, Chao and Zhu, Minfeng and Chen, Wei and Yan, Shuicheng},
  journal={arXiv preprint arXiv:2304.06627},
  year={2023}
}