Official implementation of "Learning to Discover Cross-Domain Relations with Generative Adversarial Networks"
Python Shell
Switch branches/tags
Nothing to show
Clone or download
jazzsaxmafia Merge pull request #8 from TengdaHan/master
Fix some typos in image_translation.py
Latest commit f044178 Apr 12, 2017
Permalink
Failed to load latest commit information.
assets Add files via upload Mar 18, 2017
datasets First Commit Mar 18, 2017
discogan Fix some typos in image_translation.py Apr 12, 2017
.gitignore First Commit Mar 18, 2017
README.md Update README.md Apr 12, 2017
script.sh Update script.sh Mar 18, 2017

README.md

DiscoGAN

Official PyTorch implementation of Learning to Discover Cross-Domain Relations with Generative Adversarial Networks.

Prerequisites

  • Python 2.7
  • PyTorch
  • Numpy/Scipy/Pandas
  • Progressbar
  • OpenCV

Training DiscoGAN

CelebA

Download CelebA dataset using

$ python ./datasets/download.py celebA 

(Currently, the link for downloading CelebA dataset is not available).

To train gender conversion,

$ python ./discogan/image_translation.py --task_name='celebA' --style_A='Male'

To train hair color conversion

$ python ./discogan/image_translation.py --task_name='celebA' --style_A='Blond_Hair' --style_B='Black_Hair' --constraint='Male'

Handbags / Shoes

Download Edges2Handbags dataset using

$ python ./datasets/download.py edges2handbags

Download Edges2Shoes dataset using

$ python ./datasets/download.py edges2shoes

To train Edges2Handbags,

$ python ./discogan/image_translation.py --task_name='edges2handbags'

To train Edges2Shoes,

$ python ./discogan/image_translation.py --task_name='edges2shoes' 

To train Handbags2Shoes,

$ python ./discogan/image_translation.py --task_name='Handbags2Shoes' --starting_rate=0.5

Facescrub

Download Facescrub dataset using

$ python ./datasets/download.py facescrub

To train gender conversion,

$ python ./discogan/image_translation.py --task_name='facescrub'

Car, Face

Download 3D car dataset used in Deep Visual Analogy-Making, and 3D face dataset into ./datasets folder and extract them.

To train Car2Car translation,

$ python ./discogan/angle_pairing.py --task_name='car2car' 

To train Car2Face translation,

$ python ./discogan/angle_pairing.py --task_name='car2face'

Run script.sh in order to train a model using other datasaet, after uncommenting corresponding line.

Results

All example results show x_A, x_AB, x_ABA and x_B, x_BA, x_BAB

Example results of hair color conversion

Example results of gender conversion (CelebA)

Example results of Edges2Handbags

Example results of Handbags2Shoes

Example results of gender conversion (Facescrub)

Example results of Car2Face