BEGAN in CNTK

Microsoft Cognitive Toolkit (CNTK) implementation of BEGAN: Boundary Equilibrium Generative Adversarial Networks

Prerequisites

• Python 2.7 or Python 3.5+
• CNTK==2.5.1+

Note: I found that lower versions of CNTK caused abnormal training results of the GAN.

Usage

First, download MNIST dataset:

$ python install_mnist.py

Second, train the GAN:

$ python train.py

Third, generate digits with the trained generator:

$ python test.py --num_imgs=1000 --checkpointPath=models/BEGAN_G_274100.dnn

Result on MNIST

From scratch to 300k (frames captured every 1000 iter.)

gamma=1.0

Generated images with a trained model