Variational Auto Encoders (VAEs), Generative Adversarial Networks (GANs) and Generative Normalizing Flows (NFs) and are the most famous and powerful deep generative models. With this reposetory, I attempt to gather many deep generative model architectures, within a clean structured code enviroment. Lastly, I also attempt to analyzed both from theoretical and practical spectrum, with mathematical formulas and annimated pictures.
A VAE is a latent variable model that leverages the flexibility of Neural Networks (NN) in order to learn/specify a latent variable model.
Auto-Encoding Variational Bayes
Learning Structured Output Representation using Deep Conditional Generative Models
Generative Adversarial Networks (GAN) are a type of deep generative models. Similar to VAEs, GANs can generate images that mimick images from the dataset by sampling an encoding from a noise distribution. In constract to VAEs, in vanilla GANs there is no inference mechanism to determine an encoding or latent vector that corresponds to a given data point (or image).
Generative Adversarial Networks
From left to right; Vanilla VAE on 2-dimentional space, Conditional VAE on 20-dimentional space.
Vanilla GAN training progress.
All the results can be found the folder vae/logs (or gan/logs) with tensorboard:
tensorboard --logdir=vae/logs
python vae/main.py --model="cvae"
- Python 3.x: PyTorch, NumPy, Tensorboard
Copyright © 2019 Ioannis Gatopoulos.