This repository contains the PyTorch implementation of the paper Consistency Regularization for Domain Generalization with Logit Attribution Matching, accepted at UAI 2024.
Here, you will find code to execute ERM training, ERM with Copy-Paste augmentation, and LAM (Logit Attribution Matching) training with Copy-Paste augmentation on the iWildCam2020-WILDS dataset (iWildCam)
We have incorporated parts of the code for Copy-Paste augmentation from the Targeted Augmentation project, and data handling code from the WILDS repository.
The iWildCam dataset, which includes dataset splits, segmentation masks, and bounding boxes essential for the Copy-Paste augmentation, can be downloaded from this link: iwildcam_data.zip.
Please download and extract the contents to retrieve the iwildcam_data
folder.
To install dependencies, run
pip install -r requirements.txt
python train_erm.py \
--config=configs/lam_wildcam.yaml \
--log_dir=erm_log \
--root_dir=iwildcam_data\
--gpu_id=0
python train_erm_cp.py \
--config=configs/lam_wildcam.yaml \
--log_dir=erm_cp_log \
--root_dir=iwildcam_data\
--gpu_id=0
python train_lam_cp.py \
--config=configs/lam_wildcam.yaml \
--log_dir=lam_cp_log \
--root_dir=iwildcam_data\
--gpu_id=0
The result on the ID and OOD testing set is shown below.
Model | ID Macro F1 score | OOD Macro F1 score | Model Checkpoint |
---|---|---|---|
ERM | 48.1 | 30.0 | ERM |
ERM+DA | 53.8 | 36.0 | ERM+DA |
LAM | 52.6 | 42.3 | LAM |
python evaluation.py \
--config=configs/lam_wildcam.yaml \
--log_dir=evaluation_log \
--root_dir=iwildcam_data\
--gpu_id=0\
--checkpoint_path=lam_checkpoints/ERM
If this codebase / these models are useful in your work, please consider citing our paper.
@inproceedings{
gao2024consistency,
title={Consistency Regularization for Domain Generalization with Logit Attribution Matching},
author={Han Gao and Kaican Li and Weiyan Xie and Zhi LIN and Yongxiang Huang and Luning Wang and Caleb Chen Cao and Nevin L. Zhang},
booktitle={The 40th Conference on Uncertainty in Artificial Intelligence},
year={2024},
url={https://openreview.net/forum?id=WNy1ooHYHx}
}
Weiyan Xie via [email protected]