StarGAN

Tensorflow implementation of StarGAN: Unified Generative Adversarial Networks for Multi-Domain Image-to-Image Translation

Prerequisites

• Python 3.6
• Tensorflow 1.6

Results for Facial Attribute Transfer on CelebA

The images are generated by StarGAN trained on the CelebA dataset.

Usage

1. Clone the repository

$ git clone https://github.com/ioanachelu/StarGAN.git
$ cd StarGAN/

2. Download the dataset into ./data. It should have the following substructure:

    .data
        celebA
        list_attr_celebs.txt

3. Train StarGAN with CelebA

$ python train.py --train=True --logdir=./logdir --config=default --dataset_path=../data

4. Resume StarGAN training with CelebA

$ python train.py --train=True --resume=True --load_from=./logdir/0-default --config=default --dataset_path=../data

4. Test StarGAN

$ python run_wild.py --logdir=./logdir --config=default --load_from=./logdir/0-default --dataset_type=test