tensorflow T-GAN

The code is tested on Linux operating system with Python 2.7 or 3.x, TensorFlow 1.4.0+

Keras Address

Paper Address:Training Generative Adversarial Networks Via Turing Test

Run the model using this command:

if you want to run t-sgan_64:

python main.py --gan_type TGAN_64 --loss_type sgan --dataset celebA

t-wgan_64:

python main.py --gan_type TGAN_64 --loss_type wgan --dataset celebA

t-sgan_128:

python main.py --gan_type TGAN_128 --loss_type sgan --dataset celebA

t-wgan_128:

python main.py --gan_type TGAN_128 --loss_type wgan --dataset celebA

you also can try lsgan loss.

Please kindly look at the file main.py for hyperparameter arguments.

Tensorflow Version

if your tensorflow version >= 1.8, you can set self.custom_dataset = True else self.custom_dataset = False

Results

Random generation for 64x64 images

All results are randomly sampled.

Name	Result-64x64
T-SGAN
T-WGAN

Random generation for 128x128 images

All results are randomly sampled.

Name	Result-128x128
T-SGAN
T-WGAN