The document discusses image generation using generative adversarial networks (GANs) with Tensorflow. It provides an overview of convolutional neural networks, GANs and common GAN architectures. It then demonstrates how to implement GANs with Tensorflow by defining the generator and discriminator networks, initializing the graph, and training the model. The document concludes with examples of GAN results and tips for training GANs.
11. Generative Adversarial Networks - 2014
The first paper
Two-player game that the generator
is trained to generate images from
inputed noises to fool the
discriminator while the discriminator
is trained to well discriminate real
samples and fake samples
E real(log(D)) + E fake(log(1-D))
https://papers.nips.cc/paper/5423-generative-adversarial-nets.pdf
Cenk Bircano˘glu (Boyner Group/Bah¸ce¸sehir Uni) GANs April 21, 2018 11 / 44
43. GAN Training Tips & Tricks
Normalize the inputs
A modified loss function
Use a spherical Z
BatchNorm
Avoid Sparse Gradients:
ReLU, MaxPool
Use Soft and Noisy Labels
DCGAN / Hybrid Models
Use stability tricks from RL
Use the ADAM Optimizer
Track failures early
Dont balance loss via statistics
If you have labels, use them
Add noise to inputs, decay over
time
Train discriminator more
Batch Discrimination
Discrete variables in Conditional
GANs
Use Dropouts in G in both train
and test phase
https://github.com/soumith/ganhacks
Cenk Bircano˘glu (Boyner Group/Bah¸ce¸sehir Uni) GANs April 21, 2018 43 / 44