Another implementation of DRAW with deepy framework

Paper: http://arxiv.org/pdf/1502.04623

Note

Core components in the implementation are copied from https://github.com/jbornschein/draw, thanks for the great work of https://github.com/jbornschein.

Tricky parts in the model

• Differential filter functions that can zoom in and zoom out an image to get a glimpse
• Q Sampler
• Differential sampling function to get a sample from distributions

What does the sampler do, an analysis from computation graph

• Prior distribution P_z
• This distribution generates latent variables used in image generation
• mean and deviation transform networks
• These two networks transform inputs to mean and deviation vectors to form a distribution Q(z|x)
• The goal for training these two network is to make Q(z|x) close to prior P_z
• KL(Q(z|x) || P_z) is to evaluate how close these two distributions are
• But if we sample a latent variable from Q(z|x), another goal is to restore the original input x from z
• So this means the model try to map any input to a similar latent vector
• but it has to maintain a tiny difference so as to decode it back to x
• Then if we sample from P_z, we are sampling from Q(z|x) which we don't know what the x is

Experiment on MNIST

python mnist_training.py

MNIST Animation (work in progress)