Variational Auto Encoding

An implementation of variational auto-encoding (VAE) for MNIST represented by Kingma et al. Paper(2013)

• Auto-Encoding Variational Bayes by Kingma et al.
• This implementation is referred from hwalsuklee’s tutorial

Different from the hwalsuklee’s program, this tutorial is not required to use other class file or functional set file.
Within this IPython notebook, all operations in VAE are performed.

Results

Reproduce

Given training set data, VAE architecture can generate the similar images without any label.
In this example, we introduce 2-Dimensional latent space unit.

• Input Image
  

• Recounstruction Image
  

Reconstruction

For exploiting latent space in VAE, we generate and put synthetic mesh-grid data (-4:4, -4:4) into the VAE model as hidden variable, Z instead of training set data.

• Reconstruction Image
  

Learned MNIST manifold

Visualizations of learned data manifold for generative models with 2-dim.
Latent space are given in Figure. 4 in the paper.

• Manifold Image
  

References

The implementation is based on the projects:
[1] https://github.com/oduerr/dl_tutorial/tree/master/tensorflow/vae
[2] https://github.com/fastforwardlabs/vae-tf/tree/master
[3] https://github.com/kvfrans/variational-autoencoder
[4] https://github.com/altosaar/vae
[5] https://github.com/hwalsuklee/tensorflow-mnist-VAE

Acknowledgements

This implementation has been performed on Tensorflow 1.7 in Ubuntu 16.04.