Auto-Encoders for MNIST: TF 2.0

Implementations of AE(Auto-Encoder), VAE(Variational AE) and CVAE(Conditional VAE)

Results

AE(Auto-Encoder)

python run_main.py --ae_type AE --num_epochs 150 --latent_dim 2

Reconsructed Image
Distribution of z

VAE(Variational Auto-Encoder)

python run_main.py --ae_type VAE --num_epochs 150 --latent_dim 2

Reconsructed Image
Distribution of z

CVAE(Conditional Variational Auto-Encoder)

python run_main.py --ae_type CVAE --num_epochs 150 --latent_dim 2

Reconsructed Image
Distribution of z

More Conceptual Plots

VAE

Visualizations of generated images from (z_1, z_2) ∈ [-2, 2] uniform distribution.

CVAE

Visualizations of generated images from z_1=0, and z_2 ∈ [-3, 3] uniform distribution.

Usage

Prerequisites

1. tensorflow 2.0.0-alpha0 or higher
2. python packages: numpy, matplotlib

Command

python run_main.py --ae_type <autoencoder_type>

Example: python run main.py --ae_type AE

Arguments

Required:

• --ae_type: Type of autoencoder, must be in ['AE', 'VAE', 'CVAE']

Optional:

• --latent_dim: Dimension of latent vector(z). Default: 2
• --num_epochs: The number of epochs to run. Default: 100
• --batch_size: The size of batch. Default: 1000
• --learn_rate: Learning rate of Adam optimizer. Default: 1e-4

References

[1] https://github.com/hwalsuklee/tensorflow-mnist-VAE
[2] https://www.slideshare.net/NaverEngineering/ss-96581209
[3] https://www.tensorflow.org/alpha/tutorials/generative/cvae