tf-exercise-gan

Tensorflow implementation of different GANs and their comparisions

GAN implementations

• DCGAN from 'Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks' (https://arxiv.org/abs/1511.06434)
• WGAN from 'Wasserstein GAN' (https://arxiv.org/abs/1701.07875)
• BEGAN from 'BEGAN: Boundary Equilibrium Generative Adversarial Networks' (https://arxiv.org/abs/1703.10717)
• MAD-GAN from 'Multi-Agent Diverse Generative Adversarial Networks' (https://arxiv.org/abs/1704.02906)
• GoGAN from 'Gang of GANs: Generative Adversarial Networks with Maximum Margin Ranking' (https://arxiv.org/abs/1704.04865)
• ... (To be added)

Tasks

• Impl. DCGAN, GoGAN, WGAN
• Impl. BEGAN, MAD-GAN
• Reproduce GANs on MNIST and CelebA datasets
• Impl. training & evaluation on synthetic datasets
• Add sample results
• Impl. inference-only code for GANs (may require refactoring)
• Impl. better evaluation function for real images (e.g. IvOM, energy dist., ...)
• Impl. a result logger
• Compare GANs (synthetic)
• Compare GANs (MNIST and CelebA dataset)
• Add quantitative comparisons
• Add more GAN implementations

Experiments & Benchmarks

170718 / Comparison of different GAN models on synthetic datasets

• Done without any hyper-parameter search.
• MAD-GAN worked best in the tested datasets.
• 

170718 / Sample results on MNIST dataset

170809 / Sample results on CelebA dataset

Other References