Domain Whitening Transform for Unsupervised Domain Adaptation (CVPR 2019)

Official PyTorch github repository for the paper Unsupervised Domain Adaptation using Feature-Whitening and Consensus Loss published in The Conference on Computer Vision and Pattern Recognition (CVPR) held at Long Beach, California in June, 2019.

Prerequisites

• Pytorch 1.0
• Python 3.5

Usage

• Office-Home: To run the experiments on the OfficeHome dataset first you need to download the dataset from this page. Following this step, you would need to download the ResNet50 pre-trained checkpoint, trained on ImageNet with the BatchNorm layers (in the first conv layer and the first Res block) replaced by whitening normalization layers. The pre-trained weights is available here.

python resnet50_dwt_mec_officehome.py --s_dset_path path-to-source-dataset-folder --t_dset_path path-to-target-dataset       folder --resnet_path path-to-pre-trained-resnet50-weights

• USPS -> MNIST:

python usps_mnist.py --group_size 4 --source 'usps' --target 'mnist'

If you find this code useful for your research, please cite our paper:

@article{roy2019unsupervised,
  title={Unsupervised Domain Adaptation using Feature-Whitening and Consensus Loss},
  author={Roy, Subhankar and Siarohin, Aliaksandr and Sangineto, Enver and Bulo, Samuel Rota and Sebe, Nicu and Ricci, Elisa},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  year={2019}
}