Name		Name	Last commit message	Last commit date
parent directory ..
assets		assets
.gitignore		.gitignore
README.md		README.md
main.py		main.py
models.py		models.py
utils.py		utils.py

README.md

Adversarial Autoencoder

Usage

> python main.py --help

The arguments are as follows-

usage: main.py [-h] [--directory DIRECTORY] [--epochs EPOCHS]
               [--batch_size BATCH_SIZE] [--gen_lr GEN_LR] [--dis_lr DIS_LR]
               [--download DOWNLOAD]

optional arguments:
  -h, --help            show this help message and exit
  --directory DIRECTORY
                        directory of dataset
  --epochs EPOCHS       total number of epochs you want to run. Default: 100
  --batch_size BATCH_SIZE
                        Batch size for dataset
  --gen_lr GEN_LR       generator learning rate
  --dis_lr DIS_LR       discriminator learning rate
  --download DOWNLOAD   Argument to download dataset. Set to True.

Contributed by-

Som Tambe

References

Adversarial Autoencoders Alireza Makhzani, Jonathon Shlens, Navdeep Jaitly, Ian Goodfellow, Brendan Frey

ICLR 2016 / ArXiv

This blog helped me with a lot with the learning rates for the Adam optimizer, AAE with PyTorch, great stuff!!

Summary

Introduction

Adversarial Autoencoders help us in mapping the hidden latent space vector to a given prior distribution defined by the user.

There can be several choices for the encoder we take, like-

Deterministic - This is the normal encoder used in a standard autoencoder setup. The only stocasticity lies in the given dataset distribution.
Gaussian Posterior - This is the one taken from the VAE setup, where the encoder network predicts the mean and variance. A sample is then sampled from $normal distribution$ . Thus the stochasticity lies in the randomness of the gaussian distribution and the data distribution here.

I have simply used the standard Deterministic network, as suggested by the experiments they carried on the MNIST dataset.

Losses and the objective function

I have taken the losses as defined as in the paper, with help from the forementioned blog.

There are 3 objectives that should come into picture in the Adversarial Autoencoder setup.

Reconstruction loss, follows the same from the standard autoencoder setup, I have used Binary Crossentropy Loss here.
Discriminator loss, which discriminates the real samples(defined by the prior distribution) from the fake samples(generated by the encoder).
Generator loss, we use this to update the generator network (encoder network here), by making sure generated output lies in the given prior distribution.