Reproducing quantitative results

Adversarial autoencoders

This repository contains code to implement adversarial autoencoder using Tensorflow.

Medium posts:

A Wizard's guide to Adversarial Autoencoders: Part 1. Autoencoders?
A Wizard's guide to Adversarial Autoencoders: Part 2. Exploring the latent space with Adversarial Autoencoders.
A Wizard's guide to Adversarial Autoencoders: Part 3. Disentanglement of style and content.
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:

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.

Matching prior and posterior distributions.

Distribution of digits in the latent space.

Supervised Adversarial Autoencoder:

Architecture:

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:

Semi-Supervised Adversarial Autoencoder:

Architecture:

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:

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.

	dc_c_loss_fake = tf.reduce_mean(
	tf.nn.sigmoid_cross_entropy_with_logits(labels=tf.ones_like(d_c_fake), logits=d_c_fake))

naresh1318 / adversarial_autoencoder Goto Github PK

adversarial_autoencoder's Introduction

Adversarial autoencoders

Installing the dependencies

Dataset

Training!

Autoencoder:

Architecture:

Adversarial Autoencoder:

Architecture:

Supervised Adversarial Autoencoder:

Architecture:

Semi-Supervised Adversarial Autoencoder:

Architecture:

Thank You

adversarial_autoencoder's People

Contributors

Stargazers

Watchers

Forkers

adversarial_autoencoder's Issues

Recommend Projects

Recommend Topics

Recommend Org