Skip to content
A wizard's guide to Adversarial Autoencoders
Branch: master
Clone or download

README.md

Adversarial autoencoders

Cover

This repository contains code to implement adversarial autoencoder using Tensorflow.

Medium posts:

  1. A Wizard's guide to Adversarial Autoencoders: Part 1. Autoencoders?

  2. A Wizard's guide to Adversarial Autoencoders: Part 2. Exploring the latent space with Adversarial Autoencoders.

  3. A Wizard's guide to Adversarial Autoencoders: Part 3. Disentanglement of style and content.

  4. A Wizard's guide to Adversarial Autoencoders: Part 4. Classify MNIST using 1000 labels.

Installing the dependencies

Install virtualenv and creating a new virtual environment:

pip install virtualenv
virtualenv -p /usr/bin/python3 aa

Install dependencies

pip3 install -r requirements.txt

Note:

  • I'd highly recommend using your GPU during training.
  • tf.nn.sigmoid_cross_entropy_with_logits has a targets parameter which has been changed to labels for tensorflow version > r0.12.

Dataset

The MNIST dataset will be downloaded automatically and will be made available in ./Data directory.

Training!

Autoencoder:

Architecture:

To train a basic autoencoder run:

    python3 autoencoder.py --train True
  • This trains an autoencoder and saves the trained model once every epoch in the ./Results/Autoencoder directory.

To load the trained model and generate images passing inputs to the decoder run:

    python3 autoencoder.py --train False

Adversarial Autoencoder:

Architecture:

Cover

Training:

    python3 adversarial_autoencoder.py --train True

Load model and explore the latent space:

    python3 adversarial_autoencoder.py --train False

Example of adversarial autoencoder output when the encoder is constrained to have a stddev of 5.

Cover

Matching prior and posterior distributions.

Adversarial_autoencoder Distribution of digits in the latent space.

Supervised Adversarial Autoencoder:

Architecture:

Cover

Training:

    python3 supervised_adversarial_autoencoder.py --train True

Load model and explore the latent space:

    python3 supervised_adversarial_autoencoder.py --train False

Example of disentanglement of style and content: Cover

Semi-Supervised Adversarial Autoencoder:

Architecture:

Cover

Training:

    python3 semi_supervised_adversarial_autoencoder.py --train True

Load model and explore the latent space:

    python3 semi_supervised_adversarial_autoencoder.py --train False

Classification accuracy for 1000 labeled images:

Cover

Cover

Note:

  • Each run generates a required tensorboard files under ./Results/<model>/<time_stamp_and_parameters>/Tensorboard directory.
  • Use tensorboard --logdir <tensorboard_dir> to look at loss variations and distributions of latent code.
  • Windows gives an error when : is used during folder naming (this is produced during the folder creation for each run).I would suggest you to remove the time stamp from folder_name variable in the form_results() function. Or, just dual boot linux!

Thank You

Please share this repo if you find it helpful.

You can’t perform that action at this time.