Cross-Domain Feature Augmentation for Domain Generalization

This repository contains the implementation of the XDomainMix algorithm and empirical studies in Cross-Domain Feature Augmentation for Domain Generalization in IJCAI 2024. The arXiv version of the paper and the supplementary materials can be found at https://arxiv.org/abs/2405.08586.

Overview

XDomainMix performs feature augmentation to increase sample diversity while emphasizing the learning of invariant representations to achieve domain generalization. We decompose features into class-generic, class-specific, domain-generic, and domain-specific components, based on feature semantics' correlation with class and domain. Feature augmentation is performed by changing domain-specific components of a feature while preserving class-specific components.

On benchmark datasets, we achieve the following results:

Method	Camelyon17	FMoW	PACS	TerraIncognita	DomainNet
ERM	70.3±6.4	32.3±1.3	85.5±0.2	46.1±1.8	43.8±0.1
GroupDRO	68.4±7.3	30.8±0.8	84.4±0.8	43.2±1.1	33.3±0.2
RSC	77.0±4.9	32.6±0.5	85.2±0.9	46.6±1.0	38.9±0.5
MixStyle	62.6±6.3	32.9±0.5	85.2±0.3	44.0±0.7	34.0±0.1
DSU	69.6±6.3	32.5±0.6	85.5±0.6	41.5±0.9	42.6±0.2
LISA	77.1±6.5	35.5±0.7	83.1±0.2	47.2±1.1	42.3±0.3
Fish	74.7±7.1	34.6±0.2	85.5±0.3	45.1±1.3	42.7±0.2
XDomainMix	80.9±3.2	35.9±0.8	86.4±0.4	48.2±1.3	44.0±0.2

Implementation

The code is based on DomainBed by Gulrajani and Lopez-Paz, 2020, and LISA by Yao et al., 2022.

The dependencies in requirments.txt should be installed to run the experiments.

Training to get domain generalization performance

Camelyon and FMoW dataset

• under wilds/ directory, run

python main.py --dataset {dataset_name} --algorithm xdomain_mix --data-dir '/my/datasets/path'

PACS, TerraIncognita and DomainNet dataset (based on the README of DomainBed)

• download data by running

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

• train a model

python -m domainbed.scripts.train --algorithm xdomain_mix --data_dir /my/datasets/path --dataset {dataset_name} --test_envs 0 

• alternatively, launch a sweep that trains multiple models

python -m domainbed.scripts.sweep launch\
       --data_dir=/my/datasets/path\
       --output_dir=/my/sweep/output/path\
       --command_launcher multi_gpu\
       --algorithms xdomain_mix\
       --datasets PACS TerraIncognita DomainNet\
       --n_hparams 1\
       --n_trials 3\
       --single_test_envs

• collect result from a sweep

python -m domainbed.scripts.collect_results --input_dir=/my/sweep/output/path

Model invariance

Representation invariance

python -m feature_invariance.invariance_feature \
       --dataset {dataset_name} \
       --input_dir /my/sweep/path \
       --result_file /path/to/store/result/file

Predictor invariance

python -m feature_invariance.invariance_predictor \
       --dataset {dataset_name} \
       --input_dir /my/sweep/path \
       --result_file /path/to/store/result/file 

Feature divergence

python -m feature_divergence.augmentation_divergence \
       --dataset {dataset_name} \
       --data_dir /my/dataset/path \
       --output_dir /path/to/store/result/file \
       --algorithm xdomain_mix

Feature importance

python -m feature_importance.eval_feature_importance \
       --dataset {dataset_name} \
       --data_dir /my/dataset/path  \
       --output_dir /saved/model/path

Feature visualization

Specify data dir (path to dataset) and model dir (path to saved model) in the notebooks

Citation

If you find our paper useful, you are welcome to cite it as

@inproceedings{liu2024cross,
 title     = {Cross-Domain Feature Augmentation for Domain Generalization},
 author    = {Liu, Yingnan and Zou, Yingtian and Qiao, Rui and Liu, Fusheng and Lee, Mong Li and Hsu, Wynne},
 booktitle = {Proceedings of the Thirty-Third International Joint Conference on
              Artificial Intelligence, {IJCAI-24}},
 publisher = {International Joint Conferences on Artificial Intelligence Organization},
 editor    = {Kate Larson},
 pages     = {1146--1154},
 year      = {2024},
 month     = {8},
 note      = {Main Track},
 doi       = {10.24963/ijcai.2024/127},
 url       = {https://doi.org/10.24963/ijcai.2024/127},
}