in

A deep stable learning method for out-of-distribution generalization


StableNet is a deep stable learning method for out-of-distribution generalization.

This is the official repo for CVPR21 paper “Deep Stable Learning for Out-Of-Distribution Generalization” and the arXiv version can be found at https://arxiv.org/abs/2104.07876.

Introduction

Approaches based on deep neural networks have achieved striking performance when testing data and training data share similar distribution, but can significantly fail otherwise. Therefore, eliminating the impact of distribution shifts between training and testing data is crucial for building performance-promising deep models. Conventional methods assume either the known heterogeneity of training data (e.g. domain labels) or the approximately equal capacities of different domains. In this paper, we consider a more challenging case where neither of the above assumptions holds. We propose to address this problem by removing the dependencies between features via learning weights for training samples, which helps deep models get rid of spurious correlations and, in turn, concentrate more on the true connection between discriminative features and labels. Extensive experiments clearly demonstrate the effectiveness of our method on multiple distribution generalization benchmarks compared with state-of-the-art counterparts. Through extensive experiments on distribution generalization benchmarks including PACS, VLCS, MNIST-M, and NICO, we show the effectiveness of our method compared with state-of-the-art counterparts.

Installation

Requirements

  • Linux with Python >= 3.6
  • PyTorch >= 1.1.0
  • torchvision >= 0.3.0
  • tensorboard >= 1.14.0

Quick Start

Train StableNet

python main_stablenet.py --gpu 0

You can see more options from

python main_stablenet.py -h

Result files will be saved in results/.

Performance and trained models

setting dataset source domain target domain network dataset split accuracy trained model
unbalanced(5:1:1) PACS A,C,S photo ResNet18 split file 94.864 model file
unbalanced(5:1:1) PACS C,S,P art_painting ResNet18 split file 80.344 model file
unbalanced(5:1:1) PACS A,S,P cartoon ResNet18 split file 74.249 model file
unbalanced(5:1:1) PACS A,C,P sketch ResNet18 split file 71.046 model file
unbalanced(5:1:1) VLCS L,P,S caltech ResNet18 split file 88.776 model file
unbalanced(5:1:1) VLCS C,P,S labelme ResNet18 split file 63.243 model file
unbalanced(5:1:1) VLCS C,L,S pascal ResNet18 split file 66.383 model file
unbalanced(5:1:1) VLCS C,L,P sun ResNet18 split file 55.459 model file
flexible(5:1:1) PACS ResNet18 split file 45.964 model file
flexible(5:1:1) VLCS ResNet18 split file 81.157

Citing StableNet

If you find this repo useful for your research, please consider citing the paper.

@inproceedings{zhang2021deep,
  title={Deep Stable Learning for Out-Of-Distribution Generalization},
  author={Zhang, Xingxuan and Cui, Peng and Xu, Renzhe and Zhou, Linjun and He, Yue and Shen, Zheyan},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={5372--5382},
  year={2021}
}

GitHub

https://github.com/xxgege/StableNet


Leave a Reply

Your email address will not be published. Required fields are marked *

This site is protected by reCAPTCHA and the Google Privacy Policy and Terms of Service apply.

GIPHY App Key not set. Please check settings

The free from Validation Framework cookbook 🌽

VuePeople lists and connects Vue.JS developers around the world