Autoencoder Reconstruction Visualizer

Visualizes the gradual process where autoencoder encodes and decodes images using MNIST hand-written digit dataset as an example.

Dependencies

• TensorFlow
  • Tested with TF 1.7, however, tf.contrib.learn.datasets is deprecated and will be removed in a future version
  • Not included in requirements.txt because there are multiple versions available (cpu and gpu), install manually
• Numpy
• Matplotlib
• Imagemagick (used for creating gif from individual images)

Installation

You can install dependencies with make init, which uses pip to install Numpy and Matplotlib. Virtualenv is always recommended.

Usage

You can run the script with python autoencoder_visualizer.py. Alternatively, make run does the same thing. This saves the images into figures folder (from different steps of reconstruction).

When you want to combine images into a gif, just run make gif.

Example output gif

Here is an example gif generated by the script visualizing the first 200 iterations of autoencoder reconstructing one single image from the MNIST datset.