Adversarial-Learned Loss for Domain Adaptation

By Minghao Chen, Shuai Zhao, Haifeng Liu, Deng Cai.

Introduction

A PyTorch implementation for our AAAI 2020 paper "Adversarial-Learned Loss for Domain Adaptation" (ALDA). In ALDA, we use a domain discriminator to correct the noise in the pseudo-label. ALDA outperforms state-of-the-art approaches in four standard unsupervised domain adaptation datasets.

Requirements

The code is implemented with Python(3.6) and Pytorch(1.0.0).

Install the newest Pytorch from https://pytorch.org/.

To install the required python packages, run

pip install -r requirements.txt

Setup

Digits：

Download SVHN dataset and unzip it at data/svhn2mnist.

Office-31

Download Office-31 dataset and unzip it at data/office.

Office-Home

Download Office-Home dataset and unzip it at data/office-home.

VisDA-2017

Download VisDA-2017 dataset

Training

Digits：

SVHN->MNIST
python train_svhnmnist.py ALDA --gpu_id 0 --epochs 50 --loss_type all --start_epoch 2 --threshold 0.6

USPS->MNIST
python train_uspsmnist.py ALDA --gpu_id 0 --epochs 50 --task USPS2MNIST --loss_type all --start_epoch 2 --threshold 0.6

MNIST->USPS
python train_uspsmnist.py ALDA --gpu_id 0 --epochs 50 --task MNIST2USPS --loss_type all --start_epoch 2 --threshold 0.6

Office-31:

Amazon->Webcam
python  train.py ALDA --gpu_id 0 --net ResNet50 --dset office --test_interval 500 --s_dset_path ./data/office/amazon_list.txt --t_dset_path ./data/office/webcam_list.txt --batch_size 36 --trade_off 1 --output_dir "A2W_ALDA_all_thresh=0.9_test" --loss_type all --threshold 0.9

We provide a shell file to train all six adaptation tasks at once.

sh train.sh

Office-Home

Train all twelve adaptation tasks at once:

sh train_home.sh

VisDA-2017

The code of VisDA-2017 dataset is still processing.

Results

The code is tested on GTX 1080 with cuda-9.0.

The results presented in the paper:

Citation

If you use this code in your research, please cite:

@article{chen2020adversariallearned,
    title={Adversarial-Learned Loss for Domain Adaptation},
    author={Minghao Chen and Shuai Zhao and Haifeng Liu and Deng Cai},
    journal={arXiv},
    year={2020},
    volume={abs/2001.01046}
}

Acknowledgment

The structure of this code is largely based on CDAN. We are very grateful for their open source.