VAE

Build a vanila VAE model. It is composed of 2 parts: 1.Encoder 2.Decoder

1.Encoder - a MLP with 2 hidden layers. Activation function: softplus

2.Decoder - a MLP with 2 hidden layers. Activation function: softplus

Thanks to [1] to deliver the inspiration of Bernoulli cross-entropy

Dependencies

1. Python - 3.6

2. Tensorflow - 1.0

Result

2-D Latent Code space Distribution:

Use 2-D latent codes to generate image. Latent codes are sampled from uniform distribution. Below is the generated images:

The dimension of latent codes has an impact on VAE performance. I trained 3 models with latent dimension: 2, 20, 100. 20 model renders the best performance while 100 model ranks the last

Reference

[1].http://jmetzen.github.io/2015-11-27/vae.html