H-SRDC

Code release for Towards Uncovering the Intrinsic Data Structures for Unsupervised Domain Adaptation using Structurally Regularized Deep Clustering, which is published in IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE in 2022.

Project Page $\cdot$ PDF Download

The paper is available here or at the arXiv archive.

Requirements

• python 3.6.4
• pytorch 1.4.0
• torchvision 0.5.0

Data preparation

The structure of the used datasets is shown in the folder ./data/datasets/.

For each adaptation task in an inductive setting, we use all the data on the source domain as the training ones, and make a random, half-half splitting of training and test data for samples of each class on the target domain; the data settings are fixed once prepared.

The lists of image names for the training and test sets of each target domain are provided in corresponding files, e.g., ./data/datasets/Office31/amazon_half/amazon_half.txt.

The original datasets can be downloaded here.

Model training

1. Replace paths and domains in run.sh with those in one's own system.
2. Install necessary python packages.
3. Run command sh run.sh.

The results are saved in the folder ./checkpoints/.

Article citation

@article{tang2021towards,
  author={Tang, Hui and Zhu, Xiatian and Chen, Ke and Jia, Kui and Chen, C. L. Philip},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence}, 
  title={Towards Uncovering the Intrinsic Data Structures for Unsupervised Domain Adaptation Using Structurally Regularized Deep Clustering}, 
  year={2022},
  volume={44},
  number={10},
  pages={6517-6533},
  doi={10.1109/TPAMI.2021.3087830}
}