HAHA-DL / Episodic-DG

This is the repo for the paper "Episodic Training for Domain Generalization" https://arxiv.org/abs/1902.00113

Episodic-DG

This is the repo for reproducing the results in the paper Episodic Training for Domain Generalization.

Data

Please download the data from https://drive.google.com/drive/folders/0B6x7gtvErXgfUU1WcGY5SzdwZVk?resourcekey=0-2fvpQY_QSyJf2uIECzqPuQ&usp=sharing and use the official train/val split.

ImageNet pretrained model

We use the pytorch pretrained ResNet-18 model from https://download.pytorch.org/models/resnet18-5c106cde.pth

Enviroments

verified on

GPU GeForce RTX 2080 Ti
pytorch 1.0.0
Python 3.7.3
Ubuntu 16.04.6

Method	Art	Cartoon	Photo	Sketch	Ave.
AGG	76.1	75.2	94.9	69.7	79.0
Epi-FCR	79.6	76.8	93.7	77.1	81.8

and

GPU TITAN X (Pascal)
pytorch 0.4.1
Python 2.7
Scientific Linux 7.6

Method	Art	Cartoon	Photo	Sketch	Ave.
AGG	77.6	73.9	94.4	70.3	79.1
Epi-FCR	82.1	77.0	93.9	73.0	81.5

Run

sh run_main_epi_fcr.sh #data_folder #model_path
sh run_main_agg.sh #data_folder #model_path

Sensitivity of loss weights

each point is the average performance of 20 runs on VLCS

lambda_1	lambda_2	lambda_3

Reference

If you consider using this code or its derivatives, please consider citing:

@InProceedings{Li_2019_ICCV,
author = {Li, Da and Zhang, Jianshu and Yang, Yongxin and Liu, Cong and Song, Yi-Zhe and Hospedales, Timothy M.},
title = {Episodic Training for Domain Generalization},
booktitle = {The IEEE International Conference on Computer Vision (ICCV)},
month = {October},
year = {2019}
}

Note

When working with a different enviroment, you can get different results and need to tune the hyper parameters yourself.