The code for EDIT: Exemplar-Domain Aware Image-to-Image Translation

Introduction

This is the code for paper "EDIT: Exemplar-Domain Aware Image-to-Image Translation" （https://arxiv.org/abs/1911.10520 ）, which is based on pyTorch implementation of paper "meta network for neural style transfer" (from https://ypw.io/style-transfer/ ) and cycleGAN (from https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix ). We design a novel generative adversarial network, namely exemplar-domain aware image-to-image translator (EDIT for short). The principle behind is that, for images from multiple domains, the content features can be obtained by a uniform extractor, while (re-)stylization is achieved by mapping the extracted features specifically to different purposes (domains and exemplars). The generator of our EDIT comprises of a part of blocks configured by shared parameters, and the rest by varied parameters exported by an exemplar-domain aware parameter network. In addition, a discriminator is equipped during the training phase to guarantee the output satisfying the distribution of the target domain. Our EDIT can flexibly and effectively work on multiple domains and arbitrary exemplars in a unified neat model.

Network Architecture

Dependnecy

python 3.5, pyTorch >= 1.4.0 (from https://pytorch.org/), numpy, Pillow.

Usage

Training

Download the dataset you want to use and change the dataset directory. More datatsets can be found https://people.eecs.berkeley.edu/~taesung_park/CycleGAN/datasets/
Starting training using the following command

python train.py

If one wants to use the exemplar image which doesn't belong to target domain, just starting training using the following command

python train_CrossDomain.py

Testing

Put the pre-trained model in your own path and change the checkpoint path of the code
Starting testing using the following command

python test.py

Results

More Results can be found in our website: https://forawardstar.github.io/EDIT-Project-Page/

Implementation Details

When translating shoes/handbags to edges or translating facades (buildings) to semantic maps, style losses are not needed because edges or semantic maps are exemplar images. Thus, our code in 'train.py' use a if statement to distinguish shoes/handbags to edges and facades to maps from the other domains. Our code sets label = 0 and label = 1 to represent shoes to edges and facades to maps respectively, please change the following code in 'train.py' if necessary according to your own settings.

if label == 3:
    loss_style_AB = criterion_style(fakeB_mean_std, realB_mean_std)
else:
    loss_style_AB = 0

Citation

@misc{fu2019edit,
    title={EDIT: Exemplar-Domain Aware Image-to-Image Translation},
    author={Yuanbin Fu and Jiayi Ma and Lin Ma and Xiaojie Guo},
    year={2019},
    eprint={1911.10520},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}

ForawardStar/EDIT