VAE

This repository is for sharing the scripts of VAE.

<p align="center"> <img src="image\latent_space_plot.png" alt="Latent space of the test-set data." title="Latent space of the test-set data." width=500px> </p>

Description

You can train MNIST with any number of latent variable dimensions, and visualize the latent space from multiple perspectives.

Requirement

• Python: 3.7
• fire: 0.4

Usage

You can use the following command to train and visualize VAE.

python main.py main --z_dim [The dimensions of latent variables]

You can start tensorboard with the following command. Note that The default URL is http://localhost:6006/.

tensorboard --logdir ./logs

Network

The configuration of VAE is shown as below.

<p align="center"> <img src="image\network.png" alt="The configuration of VAE." title="The configuration of VAE." width=500px> </p>

Output

You can get the following images.

Reconstruction

<p align="center"> <img src="image\reconstructions.png" alt="Reconstructions of the test-set data." title="Reconstructions of the test-set data." width=500px> </p>

Mappings of the latent space

<p align="center"> <img src="image\latent_space_scatter.png" alt="Mappings of the latent space in the test-set data." title="Mappings of the latent space in the test-set data." width=500px> </p>

Generation from lattice points

<p align="center"> <img src="image\lattice_point.png" alt="Artificially generated lattice points." title="Artificially generated lattice points." width=500px> </p> <br> <p align="center"> <img src="image\lattice_point_reconstruction.png" alt="Reconstructions generated from lattice points." title="Reconstructions generated from lattice points." width=500px> </p>

Walkthrough

<p align="center"> <img src="image\latent_space_walking.png" alt="Four-direction-walkings at a constant speed in potential space." title="Four-direction-walkings at a constant speed in potential space." width=500px> </p> <br> <p align="center"> <img src="image\linear_changes.gif" alt="Reconstructions obtained from a walkthrough of the latent space." title="Reconstructions obtained from a walkthrough of the latent space." width=500px> </p>