Home

Awesome

Adversarial AutoEncoder


Adversarial Autoencoder [arXiv:1511.05644] implemented with MXNet.

Requirements

Unsupervised Adversarial Autoencoder

Please run aae_unsupervised.py for model training. Set task to unsupervised in visualize.ipynb to display the results. Notice the desired prior distribution of the 2-d latent variable can be one of {gaussian, gaussian mixture, swiss roll or uniform}. In this case, no label info is being used during the training process.

Some results:

p(z) and q(z) with z_prior set to gaussian distribution.

p(z) gaussian q(z) gaussian

p(z) and q(z) with z_prior set to 10 gaussian mixture distribution.

p(z) gaussian q(z) gaussian

p(z) and q(z) with z_prior set to swiss roll distribution.

p(z) gaussian q(z) gaussian

Supervised Adversarial Autoencoder

Please run aae_supervised.py for model training. Set task to supervised in visualize.ipynb to display the results. Notice the desired prior distribution of the 2-d latent variable can be one of {gaussian mixture, swiss roll or uniform}. In this case, label info of both real and fake data is being used during the training process.

Some results:

p(z), q(z) and output images from fake data with z_prior set to 10 gaussian mixture distribution.

p(z) gaussian q(z) gaussian output images from gaussian fake data

p(z) and q(z) with z_prior set to swiss roll distribution.

p(z) gaussian q(z) gaussian

p(z) and q(z) with z_prior set to 10 uniform distribution.

p(z) gaussian q(z) gaussian

Semi-Supervised Adversarial Autoencoder

Not implemented yet.