An implementation of cycle-gan that trains on celebA dataset. Using wgan with gradient penalty to stabilise training and avoid model collapse.

Run code

install requirements:

git clone https://github.com/MorvanZhou/celebA-cyclegan
cd celebA-cyclegan
pip3 install -r requirements.txt

Download CelebA img_align_celeba.zip (~1.4GB) from https://drive.google.com/file/d/0B7EVK8r0v71pZjFTYXZWM3FlRnM/view?usp=sharing, and list_attr_celeba.txt (25MB) from https://drive.google.com/file/d/0B7EVK8r0v71pblRyaVFSWGxPY0U/view?usp=sharing.

Parse data

python dataset.py --data_dir ~/data/celebA_img_align/

Training

python train.py --data_dir ~/data/celebA_img_align/ --soft_gpu -b 32 --epoch 51 --cycle_lambda 10 --gp_lambda 10 -lr 0.0005 -b1 0.01 -b2 0.99

Test

python restore.py --model_path visual\2020-12-17_16-06-29\model\cp-0020-00002000.ckpt -t f2m --image_path demo/female.jpg