PCL: Proxy-based Contrastive Learning for Domain Generalization (CVPR'22)

Official PyTorch implementation of PCL: Proxy-based Contrastive Learning in Domain Generalization.

Xufeng Yao, Yang Bai, Xinyun Zhang, Yuechen Zhang, Qi Sun, Ran Chen, Ruiyu Li, Bei Yu

Note that this project is built upon SWAD and DomainBed@3fe9d7.

Preparation

Dependencies

pip install -r requirements.txt

Datasets

python -m domainbed.scripts.download --data_dir=/my/datasets/path

Environments

Environment details used for our study.

Python: 3.8.6
PyTorch: 1.7.0+cu92
Torchvision: 0.8.1+cu92
CUDA: 9.2
CUDNN: 7603
NumPy: 1.19.4
PIL: 8.0.1

How to Run

train_all.py script conducts multiple leave-one-out cross-validations for all target domain.

# OfficeHome
python train_all.py OH0 --dataset OfficeHome --deterministic \
--trial_seed 0 --steps 3000 --checkpoint_freq 300 --data_dir your_data_dir

# PACS
python train_all.py PACS0 --dataset PACS --deterministic \
--trial_seed 0 --checkpoint_freq 300 --steps 5000 --data_dir your_data_dir

# TerraIncognita
python train_all.py TR0 --dataset TerraIncognita --deterministic \
--trial_seed 0 --checkpoint_freq 1000 --steps 5000 --data_dir your_data_dir

Experiment results are reported as a table. In the table, the row SWAD indicates out-of-domain accuracy from SWAD. The row SWAD (inD) indicates in-domain validation accuracy.

Example results:

# OfficeHome
+------------+---------+---------+---------+------------+---------+
| Selection  |   Art   | Clipart | Product | Real_World |   Avg.  |
+------------+---------+---------+---------+------------+---------+
|   oracle   | 64.882% | 55.842% | 76.267% |  78.600%   | 68.898% |
|    iid     | 64.882% | 53.265% | 76.267% |  78.371%   | 68.196% |
|    last    | 60.968% | 50.773% | 74.240% |  76.363%   | 65.586% |
| last (inD) | 82.440% | 80.166% | 76.457% |  79.601%   | 79.666% |
| iid (inD)  | 83.816% | 84.081% | 79.826% |  81.133%   | 82.214% |
|    SWAD    | 67.302% | 58.763% | 79.448% |  81.440%   | 71.738% |
| SWAD (inD) | 86.022% | 86.993% | 82.424% |  83.130%   | 84.642% |
+------------+---------+---------+---------+------------+---------+

Citation

@inproceedings{yao2022pcl,
  title={PCL: Proxy-based Contrastive Learning for Domain Generalization},
  author={Yao, Xufeng and Bai, Yang and Zhang, Xinyun and Zhang, Yuechen and Sun, Qi and Chen, Ran and Li, Ruiyu and Yu, Bei},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={7097--7107},
  year={2022}
}