jiwoongim / dvae

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

• numpy
• Theano

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):

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

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

VAE encoder filter versus DVAE encoder filter