Enhanced Transport Distance for Unsupervised Domain Adaptation (ETD)

March 20, 2020 ยท View on GitHub

This is the pytorch demo code for Enhanced Transport Distance for Unsupervised Domain Adaptation (ETD) (CVPR 2020)

Requirements

  • python 3.7
  • torch 1.2.0
  • torchvision 0.4.0
  • pandas 0.24.2
  • numpy 1.17.3

Dataset

  • The structure of the datasets should be like
OfficeHome (Dataset)
|- Art (Domain)
|  |- Alarm_Clock (Class)
|     |- 00001.jpg (Sample) 
|     |- ...
|  |- Backpack (Class)
|  |- ...
|- Clipart
|- Product 
|- Real_World
  • The srtucture of all the code should be like
|- OfficeHome
|- UV_code

Usage

  • Download the OfficeHome dataset from Google Drive.

  • Set experiment configures in a csv file.

    • It is UV.csv in this code.
    • The csv file includes: epochs, Pretrain_Epoch, train_batch_size, lr lr_feature, lr_fc, beta1, beta2, lambda_1, lambda_2, source_domain, target_domain, class_num, resnet_name, fc_in_features, bottleneck_dim, dropout_p, and network_name.
    • An example is shown as following (configures in this figure may not be the best choices and this figure is just to explain the configure file more clear):

  • Set saving path.

    • The saving path is ./UV_code/UV in this code and the corresponding code is shown as following:
     file_path = '.'+os.path.sep+'UV' 
 if not os.path.exists(file_path):
    os.mkdir(file_path)
 experiment_base_path = '.'+os.path.sep+'UV'+os.path.sep+experiment_name        
 if not os.path.exists(experiment_base_path):
    os.mkdir(experiment_base_path)

  • Training with main.py.

  • The loss, acc, best acc and best model can be found in ./UV_code/UV/test1(in this code).

Note

This code is correspongding to the dual formulation of the reweighed OT problem. And we will introduce the semi-dual version later.

Citation

If this reposity is helpful for you, please cite our paper:
@inproceedings{Li2020ETD,
  title={Enhanced Transport Distance for Unsupervised Domain Adaptation},
  author={Mengxue Li, and Yi-Ming Zhai, and You-Wei Luo, and Peng-Fei Ge, and Chuan-Xian Ren},
  booktitle={2020 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2020},
}

Contact

If you have any questions, please feel free to contact me via zhaiym3@mail2.sysu.edu.cn.