RelGAN (ICCV 2019)

(Official) Keras implementation of RelGAN: Multi-Domain Image-to-Image Translation via Relative Attributes

The paper is accepted to ICCV 2019. We also have the PyTorch version here.

Preparation

• Prerequisites
  • Python 3.5
  • Keras 2.2.4
• Dataset
  • Celeba-HQ
    • Please follow the instructions in celeba-hq-modified to prepare the dataset
• Pre-trained model
  • generator519.h5

Get Started

Preprocessing

In this step, we export annotations to a numpy file. You will get anno_dic.npy and imgIndex.npy after running the script

-n  :   number of attributes (5, 9, 17)
-o  :   target output file

python3 preprocessing.py [--number=17] [--output=anno_dic.npy]

Training

python3 train.py
    --path=<path to celeba-256>
    --device=<device number>
    [--growth=False]
    [--step=0]
    [--lr=1e-5]
    [--beta1=0.5]
    [--beta2=0.999]
    [--batch_size=4]
    [--sample_size=2]
    [--epochs=400000]
    [--lambda1=10]
    [--lambda2=10]
    [--lambda4=10]
    [--lambda5=10]
    [--lambda_gp=150]
    [--img_size=256]
    [--vec_size=17]     #if you change the number of attributes, change this number

Testing

python3 demo_translation.py --device=<device number>
python3 demo_interpolation.py --device=<device number>