Awesome

Variational Auto-Encoder for MNIST

An implementation of variational auto-encoder (VAE) for MNIST descripbed in the paper:

Auto-Encoding Variational Bayes by Kingma et al.

Results

Reproduce

Well trained VAE must be able to reproduce input image.
Figure 5 in the paper shows reproduce performance of learned generative models for different dimensionalities.
The following results can be reproduced with command:

python run_main.py --dim_z <each value> --num_epochs 60

<table align='center'> <tr align='center'> <td> Input image </td> <td> 2-D latent space </td> <td> 5-D latent space </td> <td> 10-D latent space </td> <td> 20-D latent space </td> </tr> <tr> <td><img src = 'results/input.jpg' height = '150px'> <td><img src = 'results/dim_z_2.jpg' height = '150px'> <td><img src = 'results/dim_z_5.jpg' height = '150px'> <td><img src = 'results/dim_z_10.jpg' height = '150px'> <td><img src = 'results/dim_z_20.jpg' height = '150px'> </tr> </table>

Denoising

When training, salt & pepper noise is added to input image, so that VAE can reduce noise and restore original input image.
The following results can be reproduced with command:

python run_main.py --dim_z 20 --add_noise True --num_epochs 40

<table align='center'> <tr align='center'> <td> Original input image </td> <td> Input image with noise </td> <td> Restored image via VAE </td> </tr> <tr> <td><img src = 'results/input.jpg' height = '300px'> <td><img src = 'results/input_noise.jpg' height = '300px'> <td><img src = 'results/denoising.jpg' height = '300px'> </tr> </table>

Learned MNIST manifold

Visualizations of learned data manifold for generative models with 2-dim. latent space are given in Figure. 4 in the paper.
The following results can be reproduced with command:

python run_main.py --dim_z 2 --num_epochs 60 --PMLR True

<table align='center'> <tr align='center'> <td> Learned MNIST manifold </td> <td> Distribution of labeled data </td> </tr> <tr> <td><img src = 'results/PMLR.jpg' height = '400px'> <td><img src = 'results/PMLR_map.jpg' height = '400px'> </tr> </table>

Usage

Prerequisites

Tensorflow
Python packages : numpy, scipy, PIL(or Pillow), matplotlib

Command

python run_main.py --dim_z <latent vector dimension>

Example: python run_main.py --dim_z 20

Arguments

Required :

--dim_z: Dimension of latent vector. Default: 20

Optional :

--results_path: File path of output images. Default: results
--add_noise: Boolean for adding salt & pepper noise to input image. Default: False
--n_hidden: Number of hidden units in MLP. Default: 500
--learn_rate: Learning rate for Adam optimizer. Default: 1e-3
--num_epochs: The number of epochs to run. Default: 20
--batch_size: Batch size. Default: 128
--PRR: Boolean for plot-reproduce-result. Default: True
--PRR_n_img_x: Number of images along x-axis. Default: 10
--PRR_n_img_y: Number of images along y-axis. Default: 10
--PRR_resize_factor: Resize factor for each displayed image. Default: 1.0
--PMLR: Boolean for plot-manifold-learning-result. Default: False
--PMLR_n_img_x: Number of images along x-axis. Default: 20
--PMLR_n_img_y: Number of images along y-axis. Default: 20
--PMLR_resize_factor: Resize factor for each displayed image. Default: 1.0
--PMLR_n_samples: Number of samples in order to get distribution of labeled data. Default: 5000

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

Acknowledgements

This implementation has been tested with Tensorflow r0.12 on Windows 10.