najaemin92 / covi

Official code for the ECCV 2022 paper "Contrastive Vicinal Space for Unsupervised Domain Adaptation"

CoVi: Contrastive Vicinal Space for Unsupervised Domain Adaptation

Contrastive Vicinal Space for Unsupervised Domain Adaptation
Jaemin Na, Dongyoon Han, Hyung Jin Chang, Wonjun Hwang
In ECCV 2022.

Abstract: Recent unsupervised domain adaptation methods have utilized vicinal space between the source and target domains. However, the equilibrium collapse of labels, a problem where the source labels are dom- inant over the target labels in the predictions of vicinal instances, has never been addressed. In this paper, we propose an instance-wise minimax strategy that minimizes the entropy of high uncertainty instances in the vicinal space to tackle the stated problem. We divide the vicinal space into two subspaces through the solution of the minimax prob- lem: contrastive space and consensus space. In the contrastive space, inter-domain discrepancy is mitigated by constraining instances to have contrastive views and labels, and the consensus space reduces the confusion between intra domain categories. The effectiveness of our method is demonstrated on public benchmarks, including Office-31, Office-Home, and VisDA-C, achieving state-of-the-art performances. We further show that our method outperforms the current state-of-the-art methods on PACS, which indicates that our instance-wise approach works well for multi-source domain adaptation as well.

Introduction

Video: Click the figure to watch the explanation video.

Requirements

• Linux
• Python >= 3.7
• PyTorch == 1.7.1
• CUDA (must be a version supported by the pytorch version)

Getting Started

Training process.

Below we provide an example for training a CoVi on Office-31.

python main.py \
-gpu 0
-source amazon \
-target dslr \
-db_path $DATASET_PATH \
-baseline_path $BASELINE_PATH

• $DATASET_PATH denotes the location where datasets are installed.
• $BASELINE_PATH requires the path where pretrained models (DANN, MSTN, etc.) are stored.
• For DANN, the following code may be used: pytorch-DANN

Contact

For questions, please contact: [email protected]

Citation

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

@article{na2021contrastive,
  title={Contrastive Vicinal Space for Unsupervised Domain Adaptation},
  author={Na, Jaemin and Han, Dongyoon and Chang, Hyung Jin and Hwang, Wonjun},
  journal={arXiv preprint arXiv:2111.13353},
  year={2021}
}

covi's Issues

What is the `network` version？

Traceback (most recent call last):
File "main.py", line 9, in
import utils
File "/home/zhouwenshuo/workspace/domain_shift/nature/CoVi-master/utils.py", line 7, in
from network import models
ImportError: cannot import name 'models' from 'network' (/home/zhouwenshuo/anaconda3/envs/p38t/lib/python3.8/site-packages/network.py)`

Pre-training models

What datasets are used to train the pre-training models of DANN and MSTN? Can you provide a pre-trained model or code? Thank you very much!

code problem

        swap_src_labels = src_labels[mask_idx]
        pure_tgt_top1 = top1_label[mask_idx]
        swap_ul_loss, swap_lu_loss = 0, 0
        for i in range(num_of_mask):
            swap_ul_loss += sample_wise_loss(upper_out[i], swap_src_labels[i], top1_lower[i], upper_ratio[i])
            swap_lu_loss += sample_wise_loss(lower_out[i], pure_tgt_top1[i], top1_upper[i].cuda(), lower_ratio[i])

Is the swap_src_labels equal to the top1_lower? As the paper saying that there should be the top2 and top1 labels here, is the code correct?