ckb's Introduction

Conditional Bures Metric for Domain Adaptation

This is the Pytorch implementation for Conditional Bures Metric for Domain Adaptation (CKB) (CVPR 2021).

Overview

"Conditional Kernel Bures (CKB) is a conditional distribution adaptation model, which explores Wasserstein-Bures geometry and learns conditional invariant representations for knowledge transfer."

Insight

Network Architectures

Environments

Ubuntu 18.04
python 3.6
PyTorch 1.0

Dataset

The datasets should be placed in ./dataset, e.g.,

./dataset/OfficeHome
The structure of the datasets should be like

OfficeHome (Dataset)
|- Art (Domain)
|  |- Alarm_Clock (Class)
|     |- XXXX.jpg (Sample) 
|     |- ...
|  |- Backpack (Class)
|  |- ...
|- Clipart (Domain)
|- Product (Domain)
|- Real_World (Domain)

Train & Test

For OfficeHome dataset with SGD or Adam optimizer, please run

python main.py --dataset OfficeHome --exp_times 10 --batch_size 40 --CKB_lambda 1e-1 -- CKB_type hard --inv_epsilon 1e-2 --lr 1e-3 --optim_param GD
python main.py --dataset OfficeHome --exp_times 10 --batch_size 40 --CKB_lambda 1e-1 -- CKB_type hard --inv_epsilon 1e-2 --lr 3e-4

For ImageCLEF dataset with SGD or Adam optimizer, please run

python main.py --dataset ImageCLEF --exp_times 10 --batch_size 40 --CKB_lambda 1e0 --inv_epsilon 1e-1 --lr 1e-3 --optim_param GD
python main.py --dataset ImageCLEF --exp_times 10 --batch_size 40 --CKB_lambda 1e0 --inv_epsilon 1e-1 --lr 3e-4

Citation

If this repository is helpful for you, please cite our paper:

@inproceedings{luo2021conditional,
  title={Conditional Bures Metric for Domain Adaptation},
  author={Luo, You Wei and Ren, Chuan Xian},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={13989--13998},
  year={2021}
}

Contact

If you have any questions, please feel free contact me via [email protected].

ckb's People

Contributors

Stargazers

Watchers

ckb's Issues

When I want to port CKB to my project, some errors occur

Traceback (most recent call last):

train(epoch)

CKB_loss = CKButils.CKB_Metric(source, target, s_label, plab_t, t_pred, n_classes, inv_epsilon,CKB_type)

U_n, S_n, V_n = torch.svd(Nuclear)

RuntimeError: svd_cuda: For batch 0: U(33,33) is zero, singular U.

Can anyone solve the above problem, I would appreciate any help

Recommend Projects