VAE for Image Generation

Variational AutoEncoder - Keras implementation on mnist and cifar10 datasets

Dependencies

implementation Details

code is highly inspired from keras examples of vae : vae, vae_deconv
(source files contains some code duplication)

MNIST

network architecture
mnist_vae_architecture
src/mnist_train.py
src/mnist_2d_latent_space_and_generate.py
src/mnist_3d_latent_space_and_generate.py
src/mnist_general_latent_space_and_generate.py

results

2D latent space
latent space uniform sampling
2D 2D
3D latent space

3D

3D latent space results
uniform sampling random sampling
3D 3D

CIFAR10

encoder
cifar10_vae_encoder
decoder
cifar10_vae_decoder
src/cifar10_train.py , src/cifar10_generate.py

implementation structure is same as mnist files

result - latent dimensions 16

25 epochs 50 epochs 75 epochs
cifar10 cifar10 cifar10
600 epochs
cifar10

CALTECH101