Awesome

SAMix -- Spectral Adversarial MixUp for Few-Shot UDA

This repository includes open-source codes, detailed experimental results and full references of our MICCAI 2023 paper Spectral Adversarial MixUp for Few-Shot Unsupervised Domain Adaptation.

Overview

The figure above summarizes our algorithm comparisons framework, Spectral Adversarial MixUp.

• Our method first computes DoDiSS map, which is a variant of our AAAI-23 paper.
• Then, we run SAMix to generate the augmented data and train the model with the ERM loss, normal UDA loss and the Jenssen-Shannon divergence regularization.

Prerequisites

• Python 3.8
• PyTorch 1.8.1+
• A computing device with GPU

Getting started

Installation

• (Not necessary) Install Anaconda3
• Install CUDA 11.0+
• Install PyTorch 1.12.1+

Noted that our code is tested based on PyTorch 1.12.1

Dataset & Preparation

The two medical image datasets used in our work are publicly available.

The Fundus datasets include the REFUGE and the RIM. We follow the DoFE to preprocess the datasets.

The Camelyon dataset is a subset from the WILDS. This dataset is already preprocessed by the challenge.

Measure the sensitivity map of an ERM model

Train and evaluate an ERM model by

python ./AAAIcodeSubmissoin__model_sensitivity_map/train_ERM.py

The model will be saved in ./AAAIcodeSubmissoin__model_sensitivity_map/save_dir.

Measure the DoDiSS map of an ERM model by

python ./AAAIcodeSubmissoin__model_sensitivity_map/model_sensitivity_map.py

Train model with Spectral Adversarial Data Mixup (SAMix)

Train and evaluate the models with SADA by

python ./AAAIcodeSubmissoin__SADA/train_SADA.py

The key SAMix data augmentation module is in

./AAAIcodeSubmissoin__SADA/sada.py

Augmentation settings for all datasets:

• --epsilon iteration of the checkpoint to load. #Default: 0.2
• --step_size step size of the adversarial attack on the amplitude spectrum. #Default: 0.08
• --num_steps batch size of the attack steps. #Default: 5
• --tau settings for the early stop acceleration. #Default: 1
• --random_init if or not initializing amplitude spertrums with random perturbations. #Default: True
• --randominit_type type of random init type. #Default: 'normal'
• --criterion loss functions to attack. #Default: 'ce', choices=['ce', 'kl']
• --direction neg: standard attack || pos:inverse attack. #Default: neg, choices=['pos','neg']

Citation

Please cite these papers in your publications if it helps your research:

@article{zhang2022spectral,
  title={Spectral Adversarial MixUp for Few-Shot Unsupervised Domain Adaptation},
  author={Zhang, Jiajin and Chao, Hanqing and Dhurandhar, Amit and Chen, Pin-Yu and Tajer, Ali and Xu, Yangyang and Yan, Pingkun},
  journal={https://arxiv.org/abs/2309.01207},
  year={2023}
}

Acknowledgement

We would like to thank the authors we cited in our paper for sharing their codes.