Towards effective instance discrimination contrastive loss for unsupervised domain adaptation

This is the source code of our proposed method ICCV2023 paper "Towards effective instance discrimination contrastive loss for unsupervised domain adaptation".

Project structure

The project structure is presented as follows

|ICCV2023/
├──configs
| ├──_base_
| ├──fixmatch_srcmix_officehome
| ├──fixmatch_srcmix_gvb_officehome
├──data
├──runs
├──clsda
| ├──loader
| ├──models
| ├──runner
| ├──trainers
├──experiments

config: training configs files for different experiments

data: contain dataset images and labels

runs: automatically created which stores checkpoints, tensorboard and text logging files

clsda: source code of our method

experiments: training scripts

Below are the structure under data.

│officehome/
├──Art/
│  ├── Alarm_Clock
│  │   ├── 00001.jpg
│  │   ├── 00002.jpg
│  │   ├── ......
│  ├── Backpack
│  │   ├── 00001.jpg
│  │   ├── 00002.jpg
│  │   ├── ......
│  ├── ......
├──Clipart/
│  ├── Alarm_Clock
│  │   ├── 00001.jpg
│  │   ├── 00002.jpg
│  │   ├── ......
│  ├── Backpack
│  │   ├── 00001.jpg
│  │   ├── 00002.jpg
│  │   ├── ......
│  ├── ......
│txt/
├──officehome/
│  ├── labeled_source_images_Art.txt
│  ├── unlabeled_target_images_Clipart_0.txt
|  ├── unlabeled_target_images_Clipart_1.txt
|  ├── unlabeled_target_images_Clipart_3.txt

Core files

1. Model definition:

   ./clsda/models/cls_models/srcmix_contrastive_model.py (for SSDA)

   ./clsda/models/cls_models/gvb_srcmix_contrastive_model.py (for UDA)

2. Training process:

   clsda/trainers/trainer_fixmatch_srcmix.py (for SSDA)

   ./clsda/trainers/trainer_fixmatch_gvb_srcmix.py (for UDA)

3. Loss Definition

   Our contrastive loss is defined within each trainer, such as contrastive_loss in trainer_fixmatch_hda_srcmix.py file.

Training scripts

CUDA_VISIBLE_DEVICES=0,1 bash ./experiments/scripts/uda_fixmatch_gvb_srcmix_train.sh exp ./configs/gvb/gvb_officehome_A_C_fixmatch_nce.py

Citing EIDCo

@inproceedings{zhang2023eidco,
  title={Towards effective instance discrimination contrastive loss for unsupervised domain adaptation},
  author={Zhang, Yixin and Wang, Zilei and Li, Junjie and Zhuang, Jiafan and Lin, Zihan},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={11388--11399},
  year={2023}
}