Variational Auto Encoder

Project description:

Input: MNIST dataset

Encoder: It maps an input image to a mean and variance of a normal distribution in the image latent space with low dimension.

Decoder: Generate a sample drawn from the normal distribution with given mean and variance to reconstruct the original image.

Network model: For both encoder and decoder the models are shown as following:

After training the model, decoder is used to map the sampled points of latent space to images.

The grid of decoded images are shown as follows:

The source code of this project is available in my Github page.