Divergence Optimization for Noisy UniDA

This is a PyTorch implementation of Divergence Optimization for Noisy Universal Domain Adaptation.

Requirements

• Python 3.8
• PyTorch 1.6.0
• torchvision 0.7.0
• matplotlib
• numpy
• scikit-learn

Preparation

Downlaod following data:

Office

OfficeHome

VisDA

and put them in data directory as follows:

Divergence-Optimization
│   README.md
│   train.py
│   run.py
│   ...
│   
└───data
    └───amazon
    |   └───images
    └───dslr
    └───webcam
    └───Art
    |   └───Alarm_Clock
    └───Clipart
    └───Product
    └───Real
    └───visda
        └───train
        └───validation
        

Usage

Train the network with Office dataset under Noisy UniDA setting having pairflip noise (noise rate = 0.2):

python run.py --gpu 0 --dataset office --noise-type pairflip --percent 0.2

The trained model and output will be saved at result/pairflip_0.2/configs/office-train-config_opda.

For more details and parameters, please refer to --help option.

Reference codes

• https://github.com/VisionLearningGroup/DANCE

References

• [1]: Qing Yu, Atsushi Hashimoto and Yoshitaka Ushiku. "Divergence Optimization for Noisy Universal Domain Adaptation", in CVPR, 2021.