Tensorflow implementation of StarGAN: Unified Generative Adversarial Networks for Multi-Domain Image-to-Image Translation
The author's torch code can be found here
- Python 3.5
- Tensorflow 1.3.0
- Scipy
So far I'm only using celebA dataset
First, download dataset with:
$ python download.py
To train a model:
$ python main.py --phase=train --image_size=64 --batch_size=16
To test the model:
$ python main.py --phase=test --image_size=64 --binary_attrs=1001110
Bianry attributes are now set up with the following sequence:
'Black_Hair', 'Blond_Hair', 'Brown_Hair', 'Male', 'Young', 'Mustache', 'Pale_Skin'
All the available attributes are:
'5_o_Clock_Shadow', 'Arched_Eyebrows', 'Attractive', 'Bags_Under_Eyes', 'Bald', 'Bangs', 'Big_Lips',
'Big_Nose', 'Black_Hair', 'Blond_Hair', 'Blurry', 'Brown_Hair', 'Bushy_Eyebrows', 'Chubby',
'Double_Chin', 'Eyeglasses', 'Goatee', 'Gray_Hair', 'Heavy_Makeup', 'High_Cheekbones', 'Male',
'Mouth_Slightly_Open', 'Mustache', 'Narrow_Eyes', 'No_Beard', 'Oval_Face', 'Pale_Skin', 'Pointy_Nose',
'Receding_Hairline', 'Rosy_Cheeks', 'Sideburns', 'Smiling', 'Straight_Hair', 'Wavy_Hair', 'Wearing_Earrings',
'Wearing_Hat', 'Wearing_Lipstick', 'Wearing_Necklace', 'Wearing_Necktie', 'Young'
which are the attributes provided in celebA dataset. you can change the attribute set in model.py
if you change the attribute set, you should also change the n_label argument. for example:
$ python main.py --phase=train --n_label=10
Details of the loss of Discriminator and Generator