Awesome

Learning from Extrinsic and Intrinsic Supervisions for Domain Generalization

by Shujun Wang, Lequan Yu, Caizi Li, Chi-Wing Fu, and Pheng-Ann Heng.

Introduction

This repository is for our ECCV2020 paper 'Learning from Extrinsic and Intrinsic Supervisions for Domain Generalization'. The framework of the proposed EISNet for domain generalization. We train a feature Encoder $f$ for discriminative and transferable feature extraction and a classifier for object recognition. Two complementary tasks, a momentum metric learning task and a self-supervised auxiliary task, are introduced to prompt general feature learning. We maintain a momentum updated Encoder (MuEncoder) to generate momentum updated embeddings stored in a large memory bank. Also, we design a $K$-hard negative selector to locate the informative hard triplets from the memory bank to calculate the triplet loss. The auxiliary self-supervised task predicts the order of patches within an image.

Requirements

• python 3.6.8

conda create -n EISNet python=3.6.8

• PyTorch 1.4.0

  source activate EISNet 
  conda install pytorch==1.4.0 torchvision cudatoolkit=9.2 -c pytorch 
  

• Other packages in requirements.txt

Usage

1. Clone the repository and download the dataset PACS and VLCS into folder EISNet/Dataset.

If you have already put the dataset into other paths, you need to change path in txt files in EISNet/code/data/txt_lists/*.txt.

2. Train the model.

   sh run_PACS_photo_Resnet50.sh
   

Citation

If EISNet is useful for your research, please consider citing:

@inproceedings{wang2020learning,
  title={Learning from Extrinsic and IntrinsicSupervisions for Domain Generalization},
  author={Wang, Shujun and Yu, Lequan and Li, Caizi and Fu, Chi-Wing and Heng, Pheng-Ann},
  booktitle={ECCV}, 
  year={2020}
}