/IMPLEMENTATION_Variational-Auto-Encoder

Simple implementation of Variational Autoencoder

Primary LanguagePythonMIT LicenseMIT

Variational Autoencoder

This is a enhanced implementation of Variational Autoencoder. Both fully connected and convolutional encoder/decoder are built in this model. Please star if you like this implementation.

Use

$python vae_train_amine.py # for training
$python sample.py # for sampling

Update

  1. Removed standard derivation learning on Gaussian observation decoder.
  2. Set the standard derivation of observation to hyper-parameter.
  3. Add deconvolution CNN support for the Anime dataset.
  4. Remove Anime dataset itself to avoid legal issues.

Pre-Trained Models

There are two pretrained models

  1. Anime
  2. MNIST

The weights of pretrained models are locaded in weights folder

Samples

ANIME

MNIST

Latent Space Distribution