Generative Models

GAN

• http://kvfrans.com/generative-adversial-networks-explained/
• How to Train a GAN? Tips and tricks to make GANs work
• Generative Adversarial Networks (GANs) in 50 lines of code (PyTorch)
  • real/fake 두 데이터 셋이 있어야 한다. D를 먼저 훈련시키고 다음에 G를 훈련시킨다.
• least squres GAN
• Tagger: Deep Unsupervised Perceptual Grouping
  • f-GAN: Training Generative Neural Samplers using Variational Divergence Minimization Sebastian Nowozin, Botond Cseke, Ryota Tomioka. Jun 2016
• https://blogs.nvidia.com/blog/2017/05/17/generative-adversarial-network
• SimGAN 2017 CVPR best paper. by Apple.

VAE

• http://kvfrans.com/variational-autoencoders-explained/
  • autoencoder와 동일하나 latent vector를 생성할 때 (unit) gaussian으로만 생성하도록 constraint를 줌. 그래서 unit gaussian random variable로부터 generate.
  • In practice, there's a tradeoff between how accurate our network can be and how close its latent variables can match the unit gaussian distribution.
  • latent vector를 바로 만들지도 않고 mean, std만 만들어낸다.
  • we can compare generated images directly to the originals, which is not possible when using a GAN.
• VAE in tensorflow

etc

GAN, VAE, pixel-rnn (by OpenAI)

https://blog.openai.com/generative-models/

GAN vs VAE

https://www.reddit.com/r/MachineLearning/comments/4r3pjy/variational_autoencoders_vae_vs_generative/