Implementation of Various Autoencoders

Contains tensorflow code to run

• Autoencoder
• Convolutional Autoencoder
• Variational Autoencoder
• Convolutional Variational Autoencoder

All models are trained on the famous and boring MNIST dataset (suboptimal choice for the bVAE)

Requirements

Python 3.6 and tensorflow 1.0

Structure

Everything you need is contained in the folder code

1. Model Selection
   Change the flags in main.py

2. Model Modification
   Each model consists of a specification file (e.g. this one for the AE) and a runner runXY.py (e.g. this one for the AE). Change the model architecture in the corresponding model.py file. Some auxiliar functions are defined in the nntools subfolder.

3. Training a model
   Call the respective runXY.py

4. Evaluating a model
   Change the training flag in main.py to False and call the respective runXY.py

Monitoring

The code creates scalar and image summaries during training.

Local

to start a monitoring session:

 tensorboard --logdir=path/to/log-directory --port=6006

then, navigate in browser to:

0.0.0.0:6006

to visualize the summaries and monitor the training process

Remote

to start monitoring session:

ssh -L 16006:127.0.0.1:6006 user@remote.machine
tensorboard --logdir=path/to/log-directory --port=6006

then, navigate in browser of local machine (whilst ssh tunnel is open) to:

127.0.0.1:16006