Building a denoising autoencoder

The network architecture is very simple. An input image, of size 784 pixels, is stochastically corrupted, and then it is dimensionally reduced by an encoding network layer. The reduction step is from 784 to 256 pixels.

In the decoding phase, we prepare the network for output, re-changing the original image size from 256 to 784 pixels.

As usual, we start loading all the necessary libraries to our implementation:

import numpy as np 
import tensorflow as tf 
import matplotlib.pyplot as plt     
from tensorflow.examples.tutorials.mnist import input_data 

Set the basic network parameters:

n_input    = 784  
n_hidden_1 = 256  
n_hidden_2 = 256  
n_output   = 784 

We also set the session's parameters:

epochs     = 110 
batch_size = 100 
disp_step  = 10 

We build the training and test sets. We again use the input_data feature imported from the tensorflow.examples.tutorials.mnist library...