This is a Pytorch implementation of Progressive GAN.

Borrow code from PyTorch-progressive_growing_of_gans.

Requirements

• Test on Mac, Linux and Windows 7.
• Python 3 (>=3.5).
• PyTorch 1.0, python-opencv3.

Dataset

Download the original CelebA dataset from here and additional deltas files from here. Unzip them, then you can generate CelebA-HQ dataset using dataset_tools.py:

python dataset_tools.py create_celeba_hq Celeba-HQ ~/celeba ~/celeba-hq-deltas

The CelebA-HQ dataset will be placed in folder Celeba-HQ.

To ease image reading, I generate a file list for all training images:

python dataset_tools.py generate_filelist Celeba-HQ/ data/

The file list will be generated in folder data.

Training

celeba_hq_dir=Celeba-HQ
g_lr=1e-3
d_lr=1e-3
resolution=4
epochs=40
gan_type=lsgan
norm=pixelnorm
output_act=tanh
start_idx=0  # start epoch
l_gp=1.
device_id=0  # GPU id
batch_size=16

phase=stabilize
# phase=fadein

ckpt_path=ckpt/reso-${resolution}x${resolution}/${phase}_${gan_type}_${norm}_${output_act}
result_path=result/reso-${resolution}x${resolution}/${phase}_${gan_type}_${norm}_${output_act}

python train.py --celeba_hq_dir ${celeba_hq_dir} \
                --g_lr ${g_lr} \
                --d_lr ${d_lr} \
                --batch_size ${batch_size} \
                --epochs ${epochs} \
                --gan_type ${gan_type} \
                --l_gp ${l_gp} \
                --device_id ${device_id} \
                --resolution ${resolution} \
                --norm ${norm} \
                --output_act ${output_act} \
                --start_idx ${start_idx} \
                --phase ${phase} \
                --ckpt_path ${ckpt_path} \
                --result_path ${result_path} 

Result

Currently, I have only run out some results using LsGAN. For WGAN-GP, the results are very bad and I still don't know why.