/Keras-MNIST-GAN

Simple Generative Adversarial Networks for MNIST data with Keras.

Primary LanguagePythonMIT LicenseMIT

Keras GAN for MNIST

Simple and straightforward Generative Adverserial Network (GAN) implementations using the Keras library.
Several of the tricks from ganhacks have already been implemented.

mnist_dcgan.py: a Deep Convolutional Generative Adverserial Network (DCGAN) implementation.
Each epoch takes approx. 1 minute on a NVIDIA Tesla K80 GPU (using Amazon EC2).
Generated images after 50 epochs can be seen below.

mnist_gan.py: a standard GAN using fully connected layers. Each epoch takes ~10 seconds on a NVIDIA Tesla K80 GPU.
Generated images after 200 epochs can be seen below.

DCGAN

Generated MNIST images at epoch 50 with a DCGAN
[Generated MNIST images at epoch 50.]

Loss at every epoch for 50 epochs with a DCGAN
[Loss at every epoch for 50 epochs.]

Deep GAN

Generated MNIST images at epoch 200 with a GAN
[Generated MNIST images at epoch 200.]

Loss at every epoch for 200 epochs with a GAN
[Loss at every epoch for 200 epochs.]