Skip to content
No description, website, or topics provided.
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
figs
.gitignore
README.md
dvae.py
main_freyface.py
main_mnist.py
optimize.py
stochastic_layer.py
utils.py

README.md

Denoising Criterion for Variational Auto-encoding Framework

Python (Theano) implementation of Denoising Criterion for Variational Auto-encoding Framework code provided by Daniel Jiwoong Im, Sungjin Ahn, Roland Memisevic, and Yoshua Bengio. Denoising criterion injects noise in input and attempts to generate the original data. This is shown to be advantageous. The codes include training criterion which corresponds to a tractable bound when input is corrupted. For more information, see

@article{Im2016dvae,
    title={Denoising Criterion for Variational Framework},
    author={Im, Daniel Jiwoong and Ahn,Sungjin and Memisevic, Roland and Bengio, Yoshua},
    journal={http://arxiv.org/abs/1511.06406},
    year={2016}
}

If you use this in your research, we kindly ask that you cite the above arxiv paper

Dependencies

Packages

How to run

Entry code for one-bit flip and factored minimum probability flow for mnist data are

    - /main_mnist.py
    - /main_freyface.py

Here are some samples generated from trained Denoising Variational Auto-encoder (DVAE):

Image of Freyface

Traversing over 2D latent space of trained DVAE on Freyface dataset.

Image of Freyface

Here are some samples generated from trained Denoising Variational Auto-encoder (DVAE):

Image of MNIST

VAE encoder filter versus DVAE encoder filter

VAE ENC FILTERDVAE ENC FILTER

You can’t perform that action at this time.