Generalizing to Evolving Domains with Latent Structure-Aware Sequential Autoencoder

---

This repository provides the official implementations and experiments for our research in evolving domian generalization, including LSSAE and MMD-LSAE. Both of them employ a sequential autoencoder architecture that can be implemented within a same framework, the core codes for implementing these models can be found in network/vae_algorithms.py. Please follow the steps below to prepare the datasets, install the required packages, and run the code.

LSSAE (ICML 2022)

---

The network architecture for LSSAE.

Generalizing to Evolving Domains with Latent Structure-Aware Sequential Autoencoder
Tiexin Qin, Shiqi Wang and Haoliang Li
Paper: https://proceedings.mlr.press/v162/qin22a/qin22a.pdf

LSSAE is a VAE-based probabilistic framework which incorporates variational inference to identify the continuous latent structures of covariate shift and concept shift in latent space separately and simultaneously for the problem of non-stationary evolving domain generalization.

MMD-LSAE (TPAMI 2023)

---

The network architecture for MMD-LSAE.

Evolving Domain Generalization Via Latent Structure-Aware Sequential Autoencoder
Tiexin Qin, Shiqi Wang and Haoliang Li
Paper: https://ieeexplore.ieee.org/abstract/document/10268347

MMD-LSAE is built on LSSAE. Different from LSSAE utilizing KL divergence to align the individual posterior distributions with the corresponding prior distributions for the latent codes, MMD-LSAE aligns aggregated posteriors with the priors via minimizing MMD that can yield a tighter lower bound for optimization, better representation learning and stability.

Datasets

• Circle/Circle-C
• Sine/Sine-C
• RMNIST
• Portraits
• Caltran
• Airline 🔗
• ONP
• PowerSupply

We provide the Google Drive links here, so you can download these datasets directly and move them to your own file path for storage.

Requirements

• python 3.8
• Pytorch 1.10 or above
• Pyyaml
• tqdm

All the required packages can be installed via conda.

🚀 Quick Start

1. Toy Circle/-C

cd ./LSSAE
chmod +x ./scripts/*

# 1. Specify the path for Circle
    --data_path "/data/Toy_Circle/half-circle.pkl" 

# 2. Run script for LSSAE
./scripts/train_circle_lssae.sh or ./scripts/train_circle_c_lssae.sh

# 2. Run script for MMD-LSAE
./scripts/train_circle_mmd.sh or ./scripts/train_circle_c_mmd.sh

2. Other Datasets

To apply our method to other datasets, we can first copy the script of train_circle_*.sh directly, then specify the dataset information and employed network architectures. Here, an example for RMNIST is provided :

# 1. Specify the dataset info for RMNIST
    --data_path "/data/DataSets" 
    --num_classes 10 
    --data_size '[1, 28, 28]' 
    --source-domains 10 
    --intermediate-domains 3 
    --target-domains 6 

# 2. Specify the feature extractor and classifier
    --model-func MNIST_CNN
    --cla-func Linear_Cla

# 3. Run script for LSSAE
./scripts/train_rmnist_lssae.sh

# 3. Run script for MMD-LSAE
./scripts/train_rmnist_mmd.sh

For different datasets, the feature_extractor (model_func in our implementation), classifier (cla_func in our implementation) and hyper-parameters need to be specified. We provide the detailed description of network architectures and most of the hyper-parameters in our Appendix. As this is a reproduced version that implements LSSAE and MMD-LSAE into one unified framework, the results could be a little different. See ./logs for running records.

⭐ Notice

The originally released code for LSSAE only can be found in the ori-lssae branch.

Citations

If you find this repo useful for your research, please cite the following papers:

@inproceedings{Qin2022LSSAE,
title={Generalizing to Evolving Domains with Latent Structure-Aware Sequential Autoencoder},
author={Tiexin Qin and Shiqi Wang and Haoliang Li},
booktitle={ICML},
year={2022}
}

@article{Qin2023MMDLSAE,
author={Qin, Tiexin and Wang, Shiqi and Li, Haoliang},
journal={IEEE Transactions on Pattern Analysis and Machine Intelligence}, 
title={Evolving Domain Generalization Via Latent Structure-Aware Sequential Autoencoder}, 
year={2023},
pages={1-14},
doi={10.1109/TPAMI.2023.3319984}
}

---

Acknowledgments

Our codes are influenced by the following repos: DomainBed and Disentangled Sequential Autoencoder.