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.
[Generated MNIST images at epoch 50.]
[Loss at every epoch for 50 epochs.]