/DUC

[ICLR 2023 Spotlight] Code release for "Dirichlet-based Uncertainty Calibration for Active Domain Adaptation"

Primary LanguagePython

Dirichlet-based Uncertainty Calibration for Active Domain Adaptation [ICLR 2023 Spotlight]

by Mixue Xie, Shuang Li, Rui Zhang and Chi Harold Liu

arXiv  

Overview

We propose a Dirichlet-based Uncertainty Calibration (DUC) approach for active domain adaptation (DA). It provides a novel perspective for active DA by introducing the Dirichlet-based evidential model and designing an uncertainty origin-aware selection strategy to comprehensively evaluate the value of samples.

image

Prerequisites Installation

  • For cross-domain image classification tasks, this code is implemented with Python 3.7.5, CUDA 11.4 on NVIDIA GeForce RTX 2080 Ti. To try out this project, it is recommended to set up a virtual environment first.

    # Step-by-step installation
conda create --name DUC_cls python=3.7.5
conda activate DUC_cls

# this installs the right pip and dependencies for the fresh python
conda install -y ipython pip

# this installs required packages
pip install -r requirements_cls.txt

  • For cross-domain semantic segmentation tasks, this code is implemented with Python 3.7.5, CUDA 11.2 on NVIDIA GeForce RTX 3090. To try out this project, it is recommended to set up a virtual environment first.

    # Step-by-step installation
conda create --name DUC_seg python=3.7.5
conda activate DUC_seg

# this installs the right pip and dependencies for the fresh python
conda install -y ipython pip

# this installs required packages
python3 -m pip install torch==1.7.0+cu110 torchvision==0.8.1+cu110 torchaudio===0.7.0 -f https://download.pytorch.org/whl/torch_stable.html
pip install -r requirements_seg.txt

Datasets Preparation

  • Image Classification Datasets

    Symlink the required dataset by running

    ln -s /path_to_home_dataset data/home
ln -s /path_to_visda2017_dataset data/visda2017
ln -s /path_to_domainnet_dataset data/domainnet

    The data folder should be structured as follows:

    ├── data/
│   ├── home/     
|   |   ├── Art/
|   |   ├── Clipart/
|   |   ├── Product/
|   |   ├── RealWorld/
│   ├── visda2017/
|   |   ├── train/
|   |   ├── validation/
│   ├── domainnet/	
|   |   ├── clipart/
|   |   |—— infograph/
|   |   ├── painting/
|   |   |—— quickdraw/
|   |   ├── real/	
|   |   ├── sketch/	
|   |——

  • Semantic Segmentation Datasets

    Symlink the required dataset by running

    ln -s /path_to_cityscapes_dataset datasets/cityscapes
ln -s /path_to_gtav_dataset datasets/gtav
ln -s /path_to_synthia_dataset datasets/synthia

    Generate the label static files for GTAV/SYNTHIA Datasets by running

    python datasets/generate_gtav_label_info.py -d datasets/gtav -o datasets/gtav/
python datasets/generate_synthia_label_info.py -d datasets/synthia -o datasets/synthia/

    The data folder should be structured as follows:

    ├── datasets/
│   ├── cityscapes/     
|   |   ├── gtFine/
|   |   ├── leftImg8bit/
│   ├── gtav/
|   |   ├── images/
|   |   ├── labels/
|   |   ├── gtav_label_info.p
│   └──	synthia
|   |   ├── RAND_CITYSCAPES/
|   |   ├── synthia_label_info.p
│   └──

Code Running

  • for cross-domain image classification:

    # running for cross-domain image classification:
sh script_cls.sh

  • for cross-domain semantic segmentation:

    # go to the directory of semantic segmentation tasks
cd ./segmentation

# running for GTAV to Cityscapes
sh script_seg_gtav.sh

# running for SYNTHIA to Cityscapes
sh script_seg_syn.sh

Acknowledgments

This project is based on the following open-source projects. We thank their authors for making the source code publicly available.

Citation

If you find this work helpful to your research, please consider citing the paper:

@inproceedings{xie2023DUC,
  title={Dirichlet-based Uncertainty Calibration for Active Domain Adaptation},
  author={Xie, Mixue and Li, Shuang and Zhang, Rui and Liu, Chi Harold},
  booktitle={International Conference on Learning Representations (ICLR)},
  year={2023}
}

Contact

If you have any problem about our code, feel free to contact mxxie@bit.edu.cn or describe your problem in Issues.