Skip to content

A Unified Feature Disentangler for Multi-Domain Image Translation and Manipulation

License

Notifications You must be signed in to change notification settings

Alexander-H-Liu/UFDN

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

12 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

UFDN - Pytorch Implementation

Description

This is a PyTorch implementation of my paper A Unified Feature Disentangler for Multi-Domain Image Translation and Manipulation (UFDN) accepted by NIPS 2018. Please feel free to use/modify them, any bug report or improvement suggestion will be appreciated.

Model

For more detailed information, please refer to the paper.

Application

  • Image-to-Image translation / Cross-domain image generation

    With domain invariant representation, domain vector and unified decoder, UFDN is able to perform (a) continuous image translation across different domains and/or attributes at interest (b) image generation across different domains. Pre-trained model & dataset are available, see next section.

  • Unsupervised domain adaption (UDA)

    UDA can be addressed via the domain invariant representation learned by UFDN. As reported in the paper, UFDN's UDA performance on digit dataset is the current state-of-the-art. The figure visualizes the domain invariant representation learned by our UFDN, colored w.r.t. (a) dataset (b) digit.

    Model trained on MNIST and tested on SVHN is available here, it's logger is also available here. This model was trained with the example config identical to the paper, recording accuracy 0.9579 on SVHN testing setting.

Requirements

Execution Environment
  • Python 3
  • CUDA & GPU available device
Packages

Setup

Download pretrained model & dataset

Dataset is available here. To run the code, please download and place it under data/. Human face dataset is currently unavailible due to CelebA's prohibition of futher publication, we will make it availible if we get the premission in the future. Pretrained model is also available here, you can download it if you'd like to try.

Train your own model

To train UFDN, make sure all requirements are satisfied and run

python3 train_face.py <path/to/config>

See example config for more options avialible. Please refer to the paper's supplementary for config used in the main paper.

Training log & inference

To see learning curve and some translation/generation result, use tensorboard to access training log (location specified in config). E.g. tensorboard --logdir=log/

Notes

  • face_hd is the config for demo (image-to-image translation) only.
  • For UDA reproducibility, please refer to the supplemental of main paper and also issue #6 for more experiment config/settings.

Reference

Please cite the article:

"A Unified Feature Disentangler for Multi-Domain Image Translation and Manipulation" Alexander H. Liu, Yen-Cheng Liu, Yu-Ying Yeh, Yu-Chiang Frank Wang, NIPS'18

Please also cite the article if you find the face dataset helpful:

"Detach and Adapt: Learning Cross-Domain Disentangled Deep Representation" Yen-Cheng Liu, Yu-Ying Yeh, Tzu-Chien Fu, Sheng-De Wang, Wei-Chen Chiu, Yu-Chiang Frank Wang, CVPR'18 (spotlight)

About

A Unified Feature Disentangler for Multi-Domain Image Translation and Manipulation

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages