Turing Generative Adversarial Network

I found Turing GAN a way to train GANs quickly! This excited me to write my own versions in PyTorch refering the original Keras code.

Experiments

So, following are my experiments' resulting image data.

Note

• For all the experiments the images shown below are sampled after 100K iterations of training the Turing GAN on various datasets.
• All the experiments used spectral normalization for 1-Lipschitz contraint enforcement.
• I trained all of the Turing GANs with both Jensen-Shannon and Wasserstein divergences.

CIFAR-10

Turing Standard GAN with Spectral Normalization

Turing Wasserstein GAN with Spectral Normalization

MNIST

Turing Standard GAN with Spectral Normalization

Turing Wasserstein GAN with Spectral Normalization

Fashion MNIST

Turing Standard GAN with Spectral Normalization

Turing Wasserstein GAN with Spectral Normalization

References

• Training Generative Adversarial Networks Via Turing Test [arXiv]
• Original T-GANs implementation
• Spectral Normalization for Generative Adversarial Networks [arXiv]
• Spectral Normalization implementation in PyTorch