This code is the official implementation of the following paper: Semi-Supervised Domain Generalization with Stochastic StyleMatch. The paper addresses a practical and yet under-studied setting for domain generalization: one needs to use limited labeled data along with abundant unlabeled data gathered from multiple distinct domains to learn a generalizable model. This setting greatly challenges existing domain generalization methods, which are not designed to deal with unlabeled data and are thus less scalable in practice. Our approach, StyleMatch, extends the pseudo-labeling-based FixMatch—a state-of-the-art semi-supervised learning framework—in two crucial ways: 1) a stochastic classifier is designed to reduce overfitting and 2) the two-view consistency learning paradigm in FixMatch is upgraded to a multi-view version with style augmentation as the third complementary view. Two benchmarks are constructed for evaluation. Please see the paper at https://arxiv.org/abs/2106.00592 for more details.
25-08-2022: In case you are using the latest version of office_home_dg in which the folder names do not contain labels, e.g., 000_Alarm_Clock has been changed to Alarm_Clock, you can use update_oh_path.py to correct the image paths.
23-10-2021: StyleMatch has been accepted to NeurIPS 2021 Workshop on Distribution Shifts: Connecting Methods and Applications.
This code is built on top of Dassl.pytorch. Please follow the instructions provided in https://github.com/KaiyangZhou/Dassl.pytorch to install the dassl environment, as well as to prepare the datasets, PACS and OfficeHome. The five random labeled-unlabeled splits can be downloaded at the following links: pacs, officehome. The splits need to be extracted to the two datasets' folders. Assume you put the datasets under the directory $DATA, the structure should look like
$DATA/
|–– pacs/
| |–– images/
| |–– splits/
| |–– splits_ssdg/
|–– office_home_dg/
| |–– art/
| |–– clipart/
| |–– product/
| |–– real_world/
| |–– splits_ssdg/
The style augmentation is based on AdaIN and the implementation is based on this code https://github.com/naoto0804/pytorch-AdaIN. Please download the weights of the decoder and the VGG from https://github.com/naoto0804/pytorch-AdaIN and put them under a new folder ssdg-benchmark/weights.
The script is provided in ssdg-benchmark/scripts/StyleMatch/run_ssdg.sh. You need to update the DATA variable that points to the directory where you put the datasets. There are three input arguments: DATASET, NLAB (total number of labels), and CFG. See the tables below regarding how to set the values for these variables.
Dataset |
NLAB |
|---|---|
ssdg_pacs |
210 or 105 |
ssdg_officehome |
1950 or 975 |
CFG |
Description |
|---|---|
v1 |
FixMatch + stochastic classifier + T_style |
v2 |
FixMatch + stochastic classifier + T_style-only (i.e. no T_strong) |
v3 |
FixMatch + stochastic classifier |
v4 |
FixMatch |
v1 refers to StyleMatch, which is our final model. See the config files in configs/trainers/StyleMatch for the detailed settings.
Here we give an example. Say you want to run StyleMatch on PACS under the 10-labels-per-class setting (i.e. 210 labels in total), simply run the following commands in your terminal,
conda activate dassl
cd ssdg-benchmark/scripts/StyleMatch
bash run_ssdg.sh ssdg_pacs 210 v1In this case, the code will run StyleMatch in four different setups (four target domains), each for five times (five random seeds). You can modify the code to run a single experiment instead of all at once if you have multiple GPUs.
At the end of training, you will have
output/
|–– ssdg_pacs/
| |–– nlab_210/
| | |–– StyleMatch/
| | | |–– resnet18/
| | | | |–– v1/ # contains results on four target domains
| | | | | |–– art_painting/ # contains five folders: seed1-5
| | | | | |–– cartoon/
| | | | | |–– photo/
| | | | | |–– sketch/
To show the results, simply do
python parse_test_res.py output/ssdg_pacs/nlab_210/StyleMatch/resnet18/v1 --multi-expIf you use this code in your research, please cite our paper
@article{zhou2021stylematch,
title={Semi-Supervised Domain Generalization with Stochastic StyleMatch},
author={Zhou, Kaiyang and Loy, Chen Change and Liu, Ziwei},
journal={arXiv preprint arXiv:2106.00592},
year={2021}
}