Adversarial Reweighting with α-Power Maximization for Partial Domain Adaptation

Code for paper "Xiang Gu, Xi Yu, Yan Yang, Jian Sun, Zongben Xu, Adversarial Reweighting with α-Power Maximization for Domain Adaptation, IJCV, 2024". [Official Version]

This is the extended version of the conference paper "Xiang Gu, Xi Yu, Yan Yang, Jian Sun, Zongben Xu, Adversarial Reweighting with α-Power Maximization for Partial Domain Adaptation. NeurIPS, 2021". The extended version is more stable and more effective.

We have extended our method to open-set domain adaptation (OSDA), universal domain adaptation (UniDA), and test-time adaptation (TTA). Please refer to them.

Prerequisites:

python==3.9.12
torch==2.0.1
torchaudio==2.0.2
torchvision==0.15.2
numpy==1.25.2
cvxpy==1.3.2
tqdm==4.66.1
Pillow==10.0.0
scikit-learn==1.3.0
torch_ema==0.3

Datasets:

Download the datasets of
VisDA-2017
DomainNet
Office-Home
Office
ImageNet
Caltech-256
and put them into the folder "./data/" and modify the path of images in each '.txt' under the folder './data/'. Note the full list of ImageNet (imagenet.txt) is too big. Please download it here and put it into './data/imagenet_caltech/'.

Domain ID:

VisDA-2017: train (synthetic), validation (real) ==> 0,1
DomainNet: clipart, painting, real, sketch ==> 0,1,2,3
Office-Home: art, clipart, product, real_world ==> 0,1,2,3
Office: amazon, dslr, webcam ==> 0,1,2
ImageNet-Caltech: imagenet, caltech ==> 0,1

Training

VisDA-2017:

python main.py --dset visda-2017 --s 0 --t 1

DomainNet:

python main.py --dset domainnet --s 0 --t 1

Office-Home:

python main.py --dset office_home --s 0 --t 1

Office:

python main.py --dset office --s 0 --t 1

ImageNet-Caltech:

python main.py --dset imagenet_caltech --s 0 --t 1

Results

We run the code on a single Tesla V-100 GPU. The results are as follows.

Office:

seed	A2D	A2W	D2A	D2W	W2A	W2D	Avg
2019	99.4	99.3	96.7	100.0	96.8	100.0	98.7
2021	100.0	99.7	96.4	100.0	96.6	100.0	98.8
2023	99.4	99.3	96.6	99.7	96.9	100.0	98.7
Avg	99.6	99.4	96.6	99.9	96.8	100.0	98.7

Office-Home:

seed	A2C	A2P	A2R	C2A	C2P	C2R	P2A	P2C	P2R	R2A	R2C	R2P	Avg
2019	67.9	90.1	92.8	78.4	81.2	85.0	80.0	67.5	89.5	87.3	72.5	89.2	81.8
2021	67.4	85.7	90.8	76.2	85.8	88.5	81.5	71.0	88.8	85.3	69.2	88.9	81.6
2023	69.5	87.7	93.0	78.8	86.8	85.4	81.7	69.1	90.2	86.1	68.2	89.2	82.1
Avg	68.3	87.8	92.2	77.8	84.6	86.3	81.1	69.2	89.5	86.2	70.0	89.1	81.8

VisDA-2017:

seed	2019	2021	2023	Avg
S2R	92.2	93.9	93.6	93.2

ImageNet-Caltech:

seed	I2C	C2I	Avg
2019	84.3	87.1	85.7
2021	84.6	87.0	85.8
2023	85.0	87.2	86.1
Avg	84.6	87.1	85.9

DomainNet:

seed	C2P	C2R	C2S	P2C	P2R	P2S	R2C	R2P	R2S	S2C	S2P	S2R	Avg
2019	65.3	77.0	66.5	78.3	84.3	81.9	86.7	77.9	78.5	66.8	63.8	70.9	74.8
2021	69.1	81.0	65.9	77.4	84.4	82.2	86.5	78.1	77.7	60.1	65.7	72.5	75.1
2023	69.2	81.4	66.6	79.4	83.7	81.5	86.2	78.0	79.6	60.7	65.0	71.8	75.3
Avg	67.9	79.8	66.3	78.4	84.1	81.9	86.5	78.0	78.6	62.5	64.8	71.7	75.0

Reference code:

https://github.com/thuml/CDAN
https://github.com/tim-learn/BA3US
https://github.com/XJTU-XGU/RSDA

Contact：

If you have any problem, feel free to contect [email protected].