naivoder/DCVAE

Keras implementation of a deep convolutional variational autoencoder (DCVAE) evaluated on the MNIST dataset

Deep Convolutional Variational Autoencoder

This repository contains a Keras implementation using the MNIST digits dataset.
It improves upon the results demonstrated in the Keras blog.
To achieve better performance, the model employs:

• deeper convolutional architecture
• batch normalization layers
• max pooling rather than strided convolutions

Latent Space Embedding:

---

Installation

A requirements.txt file has been provided.
Once the repository has been cloned, create a new conda environment:

cd DeepConvolutionalVariationalAutoencoder
conda create env --name DCVAE --file requirements.txt
conda activate DCVAE

---

Usage

Run training from the command line:

cd code  
python train.py

Run training with jupyter notebook:

cd notebook
jupyter notebook