morvanzhou / celeba-gan

This experiment is an implementation of combining ACGAN and WGAN with gradient penalty. Generating images that look like the people in CelebA dataset.

This experiment is an implementation of combining ACGAN and WGAN with gradient penalty. Generating images that look like the people in CelebA dataset.

After 1 day training, some of OK results:

female, no smile	male, no smile	female, smile	male, smile
random seed 1	random seed 1	random seed 1	random seed 1
random seed 2	random seed 2	random seed 2	random seed 2
random seed 3	random seed 3	random seed 3	random seed 3

Some of weird results:

Run code

install requirements:

pip3 install -r requirements.txt

parse data

python dataset.py --data_dir "D:/data/celebA_img_align"

train model

# acgan with gradients penalty
python train.py --latent_dim 128 --data_dir "D:/data/celebA_img_align/" --label_dim 2 --batch_size 8 --epoch 101 --lambda 10 --d_loop 1 -lr 0.0002 --beta1 0. --beta2 0.9 --net resnet --soft_gpu --norm batch

tensorborad --logdir visual

After some runs, I find enlarging the network size would help generating better images. But restricted by my GPU and memory, I didn't try out all network structures.