Online Adaptation to Label Distribution Shift

1. Dependencies and Data Preparation

1.1 Dependencies

derivatives, tqdm, torch

1.2 Data Preparation for Simulated Shift

The processed data are

./data_preparation/val_cifar10_resnet18_cal.pt
./data_preparation/test_cifar10_resnet18_cal.pt

We also provide the code to generate them from scratch: first download the resnet18 pretrained on CIFAR10 from https://drive.google.com/file/d/1gkPX4HKhwyXl820yUF6LEjXnhNzWTh99/view?usp=sharing. Then run the following script:

cd data_preparation
python cifar10_prepare.py
cd -

1.3 Data Preparation for Shift in Arxiv

The processed data are

./data_preparation/val_arxiv_linear_cal.pt
./data_preparation/test_arxiv_linear_cal.pt

We also provide the code to generate them from scratch: download arxiv-metadata-oai-snapshot.json at data_preparation/data/ from https://www.kaggle.com/Cornell-University/arxiv. Then run the following two jupyternotebooks sequentially:

Arxiv Data Process.ipynb
Arxiv Training.ipynb

2. Experiment with Simulated Shift

python simulation_main.py --shift_process ${shift_process} --algo {algo} --seed {seed}

• shift_process is constant_shift, monotone_shift, period_shift_{T_p} or exp_period_shift_{k} if we'd like to run the simulation with constant shift, monotone shift, periodic shift with period T_p, exponential periodic shift with exponential parameter k.
• algo is const, opt_const, fth, ftfwh_{w}, ogd if we choose the method as base classifier, opt fixed classifier, Follow The HIstory, Follow The Fixed Window History with window size w, Online Gradient Descent.
• When we choose online gradient descent, run the above command with --conf_type zero_one if using finite difference to approximate the gradient or run it with --conf_type prob

3. Experiment with Shift in Arxiv

python arxiv_main.py --algo {algo}

• algo is const, opt_const, fth, ftfwh_{w}, ogd if we choose the method as base classifier, opt fixed classifier, Follow The HIstory, Follow The Fixed Window History with window size w, Online Gradient Descent.
• When we choose online gradient descent, run the above command with --conf_type zero_one if using finite difference to approximate the gradient or run it with --conf_type prob

Reference

This code corresponds to the following paper:

Ruihan Wu, Chuan Guo, Yi Su, and Kilian Q. Weinberger. Online Adaption to Label Distribution Shift. NeurIPS 2021.

@inproceedings{
wu2021online,
title={Online Adaptation to Label Distribution Shift},
author={Ruihan Wu and Chuan Guo and Yi Su and Kilian Q Weinberger},
booktitle={Advances in Neural Information Processing Systems},
editor={A. Beygelzimer and Y. Dauphin and P. Liang and J. Wortman Vaughan},
year={2021}
}