Skip to content

NikolasMarkou/multiscale_variational_autoencoder

master
Switch branches/tags

Name already in use

A tag already exists with the provided branch name. Many Git commands accept both tag and branch names, so creating this branch may cause unexpected behavior. Are you sure you want to create this branch?
Code

Latest commit

 

Git stats

Files

Permalink
Failed to load latest commit information.
Type
Name
Latest commit message
Commit time
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Multiscale Variational Autoencoder

Building a multiscale variational autoencoder (m-vae).

It is similar to a wavelet decomposition with a learnable encoding in the middle.

It creates different scale representation of an image and then encodes it into z-domain.

The downscaled versions are used to learn high level features and the higher level versions encode minute details.

The model allows to create as much as log_2(input_size) numbers of levels.

GitHub Logo

I intend to use this m-vae as a building block for Classifiers, Fuzzers, Anomaly Detection and more.

Tasks

  • Build basic model
  • Abstract Encoding and Decoding
  • CIFAR10 notebook
  • MNIST notebook
  • Generator model
  • Classifier model
  • Anomaly detection

CIFAR10 Autoencoder

Trained on the CIFAR10 dataset to recreate images for 10 epochs.

GitHub Logo

Original / Recreation comparison

Cifar10 - Epoch 1

GitHub Logo

Cifar10 - Epoch 150

GitHub Logo