report_synthetic.md

Experiments on Synthetic Datasets

GAN Candidates

• DCGAN (or just GAN)
• MAD-GAN (an extension of DCGAN)
• WGAN
• GoGAN (an extension of WGAN)
• BEGAN

Datasets

• 2-dim MoG (Mixture of Gaussian) with 5 x 5 modes
• 2-dim spiral

Comments

• Among MAD-GAN generated the best samples among these GANs.
• Basic GAN worked reasonably well in these datasets. (better than WGAN or BEGAN)
• 'Mode collapse' occurs during BEGAN training.
• Disclaimer: no hyper-parameter search has been done yet

Results

Mixture of Gaussian (2-dim, 25 modes)

GAN Name	HQ samples (%)	Modes (%)
DCGAN	61.9	100
MAD-GAN	91.9	100
WGAN	58.4	100
GoGAN	48.24	100
BEGAN	62.0	24

• Sample results with a fixed LR (=1e-4) and 100k iterations.
• About the metrics, please refer to the VEEGAN paper (https://arxiv.org/abs/1705.07761)

DCGAN and MADGAN samples

DCGAN(left) and MAD-GAN(right)

WGAN and GoGAN samples

WGAN(left) and 2-stage GoGAN(right)

BEGAN samples

BEGAN with LR=1e-4 (left) and LR=1e-6 (right)

Spiral (2-dim, 100 modes)

GAN Name	HQ samples (%)	Modes (%)
DCGAN	100.0	100
MAD-GAN	99.6	100
WGAN	94.4	100
GoGAN	92.9	100
BEGAN	100.0	27

• Sample results with a fixed LR (=1e-4) and 100k iterations.

DCGAN and MADGAN samples

DCGAN(left) and MAD-GAN(right)

WGAN and GoGAN samples

WGAN(left) and 2-stage GoGAN(right)

BEGAN samples

BEGAN with LR=1e-4 (left) and LR=1e-6 (right)