tencentyouturesearch / activelearning-sdm

Implementation of CVPR 2022 paper "Learning Distinctive Margin toward Active Domain Adaptation”

Learning Distinctive Margin toward Active Domain Adaptation

• A Pytorch implementation of our CVPR 2022 paper "Learning Distinctive Margin toward Active Domain Adaptation"
• arXiv

Installation

• Python 3.7
• Pytorch 1.8.0
• torchvision 0.9
• Numpy 1.20

Run the code

Preliminaries

• Prepare dataset: OfficeHome, Office31 and VisDA
  • We have provided text index files.

Training

• Setting

Modify the configuration in SDM_code/config/ini.config

Arg:
[data]
name : dataset
path = dataset location
source = the initial of certain scenario 
target = the initial of certain scenario
class = number of categories
[sample]
strategy = certain sample strategy
[param]
epoch : we set it to 40 in our experiments
lr : learning rate
batch : batch size
sdm_lambda : default value is 0.01
sdm_margin : default value is 1.0

• Usage

After modify setting, just run the code:

python3 run.py

• Log

We also provide our experiment logs saved in SDM_code/log/{dataset}_{source}{target}.log. For example, officehome_AC.log

Acknowledgement

This codebase is built upon TQS.

Citation

If you find our work helps your research, please kindly consider citing our paper in your publications.

@article{xie2022sdm
	title={Learning Distinctive Margin toward Active Domain Adaptation},
    author={Xie, Ming and Li, Yuxi and Wang, Yabiao and Luo, Zekun and Gan, Zhenye and Sun, Zhongyi and Chi, Mingmin and Wang, Chengjie and Wang, Pei},
    booktitle={IEEE/CVF International Conference on Computer Vision and Pattern Recognition},
    year={2022}
}

activelearning-sdm's Issues

Using the same target dataset for both training and testing in Office31

Hi, I realized that you are testing the model on the same target dataset that was used during trainng and active sampling ('tr_or_te' doesn't exist in the filename). Here:

ActiveLearning-SDM/run.py

Lines 29 to 32 in 0ee700e

	train_target_data = loader.get_data(cfg.DATA_NAME,os.path.join(cfg.DATA_PATH,cfg.DATA_TARGET),
	transform.test_transform,tr_or_te='train',n_views=cfg.N_VIEWS)
	test_target_data = loader.get_data(cfg.DATA_NAME,os.path.join(cfg.DATA_PATH,cfg.DATA_TARGET),
	transform.test_transform,tr_or_te='test',n_views=cfg.N_VIEWS)

Could you explain why you did that? Thanks