Optimal Transport for Class-Conditional and Target Shift

This is a Pytorch implementation of the paper Optimal Transport for Class-Conditional and Target Shift that is available here https://arxiv.org/abs/2006.08161

Requirements

• python3.8
• pytorch 1.3
• torchvision
• numpy
• tqdm
• POT
• sklearn

The organisation of the files are

• ClassDANN.py ClassWDGRL.py ClassWDTS_Adv.py

  provide respectively implementations of a DANN, a weighted Wasserstein distance based approach as defined in Wu 2019 and our approach.

• compare_toy_one_iter.py compare_toy_setting.py allows to run one iteration of those algorithms on the toy problem defined in the setting file. It executes one run of DANN, a W_beta=1, and MARSg. (takes few minutes on a core I7 laptop)

• compare_visda_one_iter.py run the Source only, WD based DA, and our approach on a VisDA problem. Default dataset setting works on a 3-class problem. We use a subpart of the original 12-class problem

• compare_visda_setting.py compare_visda_models.py define the dataset and their imbalance and the network models used.

• utils_local.py contains some utility functions

• visda_shift.py defines our dataset loader function. It needs the downloading of the pre-computed data features available at https://figshare.com/articles/visda-train0411_npz/11865915

Some latent features obtained on the VisDA 3-class problem using agglomerative clustering

after training on the source domain only

after adaptation