[ICML 2025] Official Code of SYNC
Zhuo He, Shuang Li, Wenze Song, Jian Liang, Han Li, Kun Gai
- By taking a novel causal perspective towards EDG problem, we design a time-aware SCM that enables the refined modeling of both dynamic causal factors and causal mechanism drifts. After that, we propose Static-DYNamic Causal Representation Learning (SYNC), an approach for effectively learning time-aware causal representations, thereby mitigating spurious correlations.
- Theoretically, we show that SYNC can build the optimal causal predictor for each time domain, resulting in improved model generalization.
- Results on both synthetic and real-world datasets, along with extensive analytic experiments demonstrate the effectiveness of proposed approach.
The code is implemented with Python 3.8.18 and run on NVIDIA GeForce RTX 4090. To try out this project, it is recommended to set up a virtual environment first.
# Step-by-step installation
conda create -n sync python=3.8.18
conda activate sync
# install torch, torchvision and torchaudio
conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 cudatoolkit=11.6 -c pytorch -c conda-forge
# this installs required packages
pip install -r requirements.txt
All datasets are available for download here.
# running for Circle dataset:
bash scripts/train_circle_sync.sh
# running for Sine dataset:
bash scripts/train_sine_sync.sh
# running for RMNIST dataset:
bash scripts/train_rmnist_sync.sh
# running for Portraits dataset:
bash scripts/train_portraits_sync.sh
# running for Caltran dataset:
bash scripts/train_caltran_sync.sh
# running for PowerSupply dataset:
bash scripts/train_power_sync.sh
# running for ONP dataset:
bash scripts/train_onp_sync.sh
This project is mainly based on the open-source project: DomainBed and LSSAE. We thank the authors for making the source code publicly available.