changwxx / UniOT-for-UniDA

Source code for the NeurIPS 2022 Spotlight paper: "Unified Optimal Transport Framework for Universal Domain Adaptation"

[NeurIPS 2022 Spotlight] Unified Optimal Transport Framework for Universal Domain Adaptation

Code release for Unified Optimal Transport Framework for Universal Domain Adaptation (NeurIPS 2022 Spotlight).

[paper] [project page]

Requirements

• Python 3.7+
• PyTorch 1.8.0
• GPU Memory 12 GB

To install requirements:

pip install -r requirements.txt

Preparation

• Download the dataset: Office31, OfficeHome, VisDA and DomainNet (real, painting and sketch).

• Prepare dataset in data directory as follows

  /path/to/your/dataset/images/amazon/      # Office
  /path/to/your/dataset/RealWorld/          # OfficeHome 
  /path/to/your/dataset/train/              # VisDA synthetic images
  /path/to/your/dataset/test/               # VisDA real images
  /path/to/your/dataset/sketch/             # DomainNet
  

• For OfficeHome dataset, make sure that your folder name is RealWorld instead of Real World.

• Modify root_path with /path/to/your/dataset/ in config files ./config/<dataset>-config.yaml.

• Make a log directory by mkdir ./log.

• Make a model directory by mkdir ./model. Download ImageNet pretrained model from Google Drive, then put the downloaded model into ./model.

Getting started

• Train with command line (take office for example)

  python main.py --gpu 0 --exp office31 --dataset office31 --source amazon --target dslr
  

• Train with script

  Modify ./config/<dataset>.sh:

  • delete the lines which begin with #SBATCH
  • specify $CUDA_VISIBLE_DEVICES

  then

  cd ./script
  sh office31.sh          # or officehome/visda/domainnet
  

• Train with Slurm script

  Modify ./config/<dataset>.sh:

  • #SBATCH --mail-user=[e-mail]
  • #SBATCH -p YOUR_partition

  then

  cd ./script
  mkdir output
  sbatch office31.sh      # or officehome/visda/domainnet
  

• Monitor (TensorBoard required)

  tensorboard --logdir=./log --port xxxx
  

• Test with command line (take office for example)

  python eval.py --gpu 0 --dataset office31 --source amazon --target dslr --model_path /path/to/your/model/final.pkl
  

Checkpoints

We provide the checkpoints for Office, OfficeHome, VisDA and DomainNet at Google Drive.

Citation

If you find this repository useful in your research, please consider citing:

@inproceedings{
chang2022unified,
title={Unified Optimal Transport Framework for Universal Domain Adaptation},
author={Wanxing Chang and Ye Shi and Hoang Duong Tuan and Jingya Wang},
booktitle={Advances in Neural Information Processing Systems},
editor={Alice H. Oh and Alekh Agarwal and Danielle Belgrave and Kyunghyun Cho},
year={2022},
url={https://openreview.net/forum?id=RTan64GlCLV}
}