This repository contains the official code of OSTAR in "Mapping Conditional Distributions for Domain Adaptation Under Generalized Target Shift" (ICLR 2022).

• Install the requirements pip install -r requirements.txt
• Run training. ex: python run.py -t 000000000001 -d digits -i 1 -g 0 -s 10
• Results are logged in ./results/run_id where run_id is the id of the run.

python run.py [-t MODEL] [-d DATASET] [-i RUN_ITERATIONS] [-g GPUID] [-s SETTING]

• Choose the model (see Section 5 of the paper for more details):
  • -t 100000000000: Source
  • -t 010000000000: DANN
  • -t 001000000000: WD_beta for beta = 0
  • -t 000111100000: WD_beta for beta in {1, 2, 3, 4}
  • -t 000000011000: MARSg / MARSc
  • -t 000000000100: IW-WD
  • -t 000000000010: WD_gt with true class-rations
  • -t 000000000001: OSTAR
• Choose the dataset:
  • -d digits: Digits
  • -d office: Office31 and OfficeHome. Requires downloading pre-computed features at https://github.com/jindongwang/transferlearning/blob/master/data/dataset.md
  • -d visda: VisDA12. Requires downloading pre-computed features at http://csr.bu.edu/ftp/visda17/clf/ and preprocessing downloaded file with prepare_data_visda12.py
• Choose the number of runs (e.g. 1 for a single run)
• Choose the gpu id (e.g. 0)
• Choose the label shift setting defined in compare_digits_setting.py, compare_office_setting.py, compare_visda_setting.py

@inproceedings{Kirchmeyer2022,
title={Mapping conditional distributions for domain adaptation under generalized target shift},
author={Matthieu Kirchmeyer and Alain Rakotomamonjy and Emmanuel de Bezenac and patrick gallinari},
booktitle={International Conference on Learning Representations},
year={2022},
url={https://openreview.net/forum?id=sPfB2PI87BZ}
}