dlwbm123/ggf

Gradual Domain Adaptation via Gradient Flow

This repository is the official implementation of Gradual Domain Adaptation via Gradient Flow (GGF).

Requirements

To install requirements:

pip install -r requirements.txt

Data Preparation

Download the Portraits dataset from here.

Run dataset/make_mnist.ipynb and dataset/make_portraits.ipynb to build the UMAP features of source and target domains.

Training

First, to train the modules (score network, rectified flow, and initial classifier) in the paper, run this command:

python train_init.py --task portraits --class_num 2 --save_path save/ --gpu_id 0
python train_init.py --task mnist45 --class_num 10 --save_path save/ --gpu_id 0
python train_init.py --task mnist60 --class_num 10 --save_path save/ --gpu_id 0

Second, to gradual generate intermediate domains and update the classifier, run this command:

python train_ggf.py --task portraits --class_num 2 
python train_ggf.py --task mnist45 --class_num 10 --alpha --iterations --lambda --eta1 --eta2 --eta3 --confidence
python train_ggf.py --task mnist60 --class_num 10 --alpha --iterations --lambda --eta1 --eta2 --eta3 --confidence

Reference

The "Denoise score matching" algorithm is built upon the implementation from https://github.com/Ending2015a/toy_gradlogp.

The codes related to "Rectified Flow" are built upon the tutorial code of the official implementation from https://github.com/gnobitab/RectifiedFlow.

Citation

@inproceedings{
    zhuang2024gradual,
    title={Gradual Domain Adaptation via Gradient Flow},
    author={Zhan Zhuang and Yu Zhang and Ying Wei},
    booktitle={The Twelfth International Conference on Learning Representations},
    year={2024},
    url={https://openreview.net/forum?id=iTTZFKrlGV}
}