Discover Cross-Domain Relations with GAN (DiscoGAN) with TensorFlow & slim

implementation based on http://arxiv.org/abs/1703.05192

"Learning to Discover Cross-Domain Relations with Generative Adversarial Networks" Taeksoo Kim, Moonsu Cha, Hyunsoo Kim, Jung Kwon Lee, and Jiwon Kim

Network architecture

Requirements

• TensorFlow 1.1.0 or greater (?)
• opencv (for image_translate.py)
• numpy
• pandas

Training

dataset download (celebA dataset)

• celebA
• list_attr_celeba.txt

make datasets folder and unzip

$ mkdir datasets/celebA
$ mv ${YOUR_DOWNLOAD_DIR}/img_align_celeba.zip datasets/celebA
$ mv ${YOUR_DOWNLOAD_DIR}/list_attr_celeba.txt datasets/celebA
$ cd datasets/celebA
$ unzip img_align_celeba.zip

editing train.sh

• Set the "TRAIN_DIR" (where to save checkpoint) to path you want
• Set the hyper-parameters

train.sh

ROOT_DIR=$HOME/projects

# Where the training (fine-tuned) checkpoint and logs will be saved to.
TRAIN_DIR=$ROOT_DIR/discoGAN.tensorflow.slim/exp1

CUDA_VISIBLE_DEVICES=0 \
python train.py \
    --train_dir=${TRAIN_DIR} \
    --batch_size=256 \
    --initial_learning_rate=0.0002 \
    --max_steps=80000 \
    --save_steps=4000 \
    --adam_beta1=0.5 \
    --style_A='Male' \
    #--style_A='Blond_Hair' \
    #--style_B='Black_Hair' \
    #--constraint='Male' \
    #--constraint_type='1' \

run ./train.sh

$ ./train.sh

• You can use tensorboard for monitoring loss and generated images

$ tensorboard --logdir exp1

Image translate

image_translate.sh

ROOT_DIR=$HOME/projects

# Where the training (fine-tuned) checkpoint and logs will be saved to.
TRAIN_DIR=$ROOT_DIR/discoGAN.tensorflow.slim/exp1

BATCH_SIZE=$1

CUDA_VISIBLE_DEVICES=0 \
python image_translate.py \
    --checkpoint_path=${TRAIN_DIR} \
    --is_all_checkpoints=True \
    --checkpoint_step=-1 \
    --batch_size=$BATCH_SIZE \
    --style_A='Male' \
    #--style_A='Blond_Hair' \
    #--style_B='Black_Hair' \
    #--constraint='Male' \
    #--constraint_type='1' \

run ./image_translate.py

$ ./image_translate.sh batch_size (the number of images you want)

Results

celebA datasets

Male -> Female -> Male

Female -> Male -> Female

Reference codes

SK T-Brain official code (pytorch)

• https://github.com/SKTBrain/DiscoGAN

carpedm20 code (pytorch)

• https://github.com/carpedm20/DiscoGAN-pytorch

Author

Il Gu Yi