MKRoughDiamond / AwesomeGenerativeModels

For collecting useful algorithms of generative ML

Awesome Generative Models

• For collecting useful algorithms of generative ML
• Contains wiki and link of example codes(repositories)

Simple Overview

Generative Models (especially ML)

• Variational Autoencoder(VAE)
• Generative Adversarial Network(GAN)
• Generative Flow
• etc.

Variational Autoencoder

• Stochastic Backpropagation and Approximate Inference in Deep Generative Models
• Auto-Encoding Variational Bayes
• Independent gaussians : VAE's latent vectors are means and stddevs of independent gaussian distributions
• Random sampling : in decoding, VAE generates samples from gaussian distributions

DRAW

• VAE+ RNN
• serial data
• DRAW: A Recurrent Neural Network For Image Generation

Generative Query Network

• evidences(Images + positions) + (target position) -> (image)

• Neural scene representation and rendering

• Learning models for visual 3D localization with implicit mapping

• Sample codes

• GQN world model

  • evidences(Image + (positions, time)) + (action) -> (Image) + (next action)
  • GQN + RL
  • World Models
  • Sample codes

Generative Adversarial Networks

• Generator vs discriminator : Generator generates "fake results" to deceive discriminator. Discriminator finds "fake data" from whole data.
• Generative Adversarial Nets

DC-GAN

• Uses convolution(transposed convolution) layers
• Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks
• Sample codes

PG-GAN

• Generates images from low resolution to high resolution
• Learns to generate bigger data from smaller ones
• Progressive Growing of GANs for Improved Quality, Stability, and Variation
• Sample codes

Conditional-GAN

• Z layer : Noise + Y(condition)
• GAN : random generate
• Conditional-GAN : pseudo-random generate (random generate with specific characteristics)
• Conditional Generative Adversarial Nets

multi-modal GANs

• Cycle-GAN
  • Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks
  • Sample codes
• DISCO-GAN
  • Learning to Discover Cross-Domain Relations with Generative Adversarial Networks
  • Sample codes(pytorch)
• STAR-GAN
  • StarGAN: Unified Generative Adversarial Networks for Multi-Domain Image-to-Image Translation
  • Sample codes

Adversarial Variational Bayes

• Adversarial Variational Bayes: Unifying Variational Autoencoders and Generative Adversarial Networks
• VAE <-- GAN
• Sample codes

BEGAN

• BEGAN: Boundary Equilibrium Generative Adversarial Networks
• GAN <-- AE
• Sample codes

Generative FLOW (GLOW)

• Glow: Generative Flow with Invertible 1x1 Convolutions