xmed-lab / TripleE-DG

TripleE: Easy Domain Generalization via Episodic Replay

By Xiaomeng Li, Hongyu Ren, Huifeng Yao and Ziwei Liu.

This repository is for our paper TripleE: Easy Domain Generalization via Episodic Replay.

Installation

conda env create -f IDC_environment.yml

Data Preparation

• Download PACS dataset, Digits-DG dataset, Office-Home dataset, VLCS dataset, TerraInc dataset, DomainNet dataset. You can download these datasets using this script.

• Put data under ./DATASET/

• The correct path should be

./DATASET/PACS/kfold/sketch/...
./DATASET/digits_dg/mnist/...
./DATASET/OfficeHome/Art/...  
./DATASET/VLCS/Caltech101...
./DATASET/terra_incognita/L38...

Training

• Check training command in ./scripts/.

• Change times to 1,2,3 to run the experiment for 3 times.

• Set target to get results on different target domains.

• By averaging three results on each domain, you will get the following results:

Sketch	Cartoon	Art	Photo	Average
84.45%	80.84%	85.30%	96.27%	86.72%

Citation

If this code is useful for your research, please consider citing: