Awesome
Divide to Adapt: Mitigating Confirmation Bias for Domain Adaptation of Black-Box Predictors
This repo is the official implementation of "Divide to Adapt: Mitigating Confirmation Bias for Domain Adaptation of Black-Box Predictors". Our method is termed as BETA. To cite this work:
@article{yang2022divide,
title={Divide to Adapt: Mitigating Confirmation Bias for Domain Adaptation of Black-Box Predictors},
author={Yang, Jianfei and Peng, Xiangyu and Wang, Kai and Zhu, Zheng and Feng, Jiashi and Xie, Lihua and You, Yang},
journal={arXiv preprint arXiv:2205.14467},
year={2022}
}
News
2023-01-21: The paper was accepted by ICLR 2023 as spotlight (notable-top-25%)!
Environment
- Install
pytorch
andtorchvision
(we usepytorch==1.9.1
andtorchvision==0.10.1
). pip install -r requirements.txt
Datasets
Please download and organize the datasets
as the following structure, where DN_domain
for DomainNet is clipart/infograph/painting/quickdraw/real/sketch
BETA/
├── data/
├── office_home/
│ ├── Art
│ ├── Clipart
│ ├── Product
│ ├── Real World
├── office31/
│ ├── amazon
│ ├── dslr
│ ├── webcam
├── visda17/
│ ├── train
│ ├── validation
├── DomainNet/
│ ├── ${DN_domain}/
│ ├── ${DN_domain}_train.txt
│ ├── ${DN_domain}_test.txt
Then generate info files with the following commands:
python dev/generate_infos.py --ds office_home
python dev/generate_infos.py --ds office31
python dev/generate_infos.py --ds visda17
Train on Office-Home
# train black-box source model on domain A
python train_src_v1.py configs/office_home/src_A/train_src_A.py
# adapt with BETA, from A to C
python train_BETA.py configs/office_home/src_A/BETA_C.py
# finetune on C
python finetune.py configs/office_home/src_A/finetune_C.py
Train on Office-31
# train black-box source model on domain a
python train_src_v1.py configs/office31/src_a/train_src_a.py
# adapt with BETA, from a to d
python train_BETA.py configs/office31/src_a/BETA_d.py
# finetune on d
python finetune.py configs/office31/src_a/finetune_d.py
Train on VisDA-2017
# train black-box source model
python train_src_v2.py configs/visda17/train_src.py
# adapt with BETA
python train_BETA.py configs/visda17/BETA.py
Train on DomainNet
# C/I/P/Q/R/S stands for clipart/infograph/painting/quickdraw/real/sketch respectively
SRC=C
TGT=I
# train source model
python train_src_v2.py configs/DomainNet/src_${SRC}/train_src_${SRC}.py
# adapt with BETA
python train_BETA.py configs/DomainNet/src_${SRC}/BETA_${TGT}.py
Easy-hard domain division
Here we show an example of the easy-hard target domain division (Office-Home: Art -> Clipart).
<img src="resources/division.png">