Generative-Adversarial-Nets

Different GAN (Generative Adversarial Network) architectures in TensorFlow

w gan (./w_gan/)

Trained on Digimon images:

Trained on the original 150 Pokemon:

Wasserstein GAN

https://arxiv.org/abs/1701.07875

Generator

tf.layers.conv2d_transpose tf.contrib.layers.batch_norm

Discriminator

tf.layers.conv2d tf.contrib.layers.batch_norm

def leaky_relu(input, name, leak=0.2): return tf.maximum(input, leak * input, name=name)

w- gan

GAN (/vaniall_gan/)

A Generative Adversarial Net implemented with TensorFlow using the MNIST data set.

Generator:

• Input: 100
• Output: 784
• Purpose: Will learn to output images that look like a real image from random input.

Discriminator:

• Input: 784
• Output: 1
• Purpose: Will learn to tell a real ("looks like it could be a real image in MNIST dataset") image(784) from a fake one.

Notes and Outputs

A problem with the way that I built this is that I used the same architecture for both the generator and discriminator. Although I thought this save me, the developer, a lot of time it actually caused a lot of problems with trying to pigeonhole that architecture to work with a smaller input (Discriminator: 28x28 vs 10x10 : Generator).

Architecture

• conv1 -> relu -> pool ->
• conv2 -> relu -> pool ->
• conv3 -> relu -> pool ->
• fullyConnected1 -> relu ->
• fullyConnected2 -> relu ->
• fullyConnected3 ->

100 random numbers -> Generator -> ImageOutput -> Discriminator -> (Real|Fake)

ColorGan