Image Generation with Tensorflow

Image Generation with Tensorﬂow
Cenk Bircano˘glu
Boyner Group/Bah¸ce¸sehir Uni
cenk.bircanoglu@gmail.com
April 21, 2018
Cenk Bircano˘glu (Boyner Group/Bah¸ce¸sehir Uni) GANs April 21, 2018 1 / 44

Overview
1 Convolutional Neural Networks
2 What are Generative Adverserial Networks?
3 Common GAN Architectures and Applications
4 Implementing GANs with Tensorﬂow
5 Question&Answer

Convolutional Neural Networks
Source: https://www.mathworks.com/discovery/convolutional-neural-network.html

Deceiving the CNN
Source: https://blog.openai.com/adversarial-example-research/

What are Generative Adverserial Networks?
Source: https://deeplearning4j.org/generative-adversarial-network

GAN Training
Source: Goodfellow, 20017

Generator
Source: http://arxiv.org/abs/1511.06434

Disriminator
Source: http://arxiv.org/abs/1511.06434

Common GAN Applications
Generating High-Quality
Images
Semi-supervised learning
Ensembles
Clustering
Image blending
Image Inpainting
Joint Probability
Super-Resolution
De-occlusion
Improving Classiﬁcation &
Recognition
Semantic Segmentation
Object Detection
RNN-GANs
Conditional Adversarial Nets
Video Prediction & Generation
Texture Synthesis & Style
Transfer
Image Translation
GAN Theory
3-Dimensional GANs
Music Face Generation &
Editing

GAN Results
Source: https://kaonashi-tyc.github.io/2017/04/06/zi2zi.html, https://github.com/hanzhanggit/StackGAN,
https://arxiv.org/pdf/1708.05509.pdf, https://arxiv.org/pdf/1705.09368.pdf, https://github.com/david-gpu/srez,
http://research.nvidia.com/sites/default/ﬁles/pubs/2017-10 Progressive-Growing-of/karras2018iclr-paper.pdf,
https://github.com/Guim3/IcGAN

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

Conditional Generative Adversarial Nets - 2014
Source: http://guimperarnau.com/blog/2017/03/Fantastic-GANs-and-where-to-ﬁnd-them
https://arxiv.org/pdf/1511.06434.pdf

cGAN Results
Source: https://arxiv.org/pdf/1411.1784.pdf

Unsupervised Representation Learning with Deep
Convolutional Generative Adversarial Networks - 2015
https://arxiv.org/pdf/1511.06434.pdfCenk Bircano˘glu (Boyner Group/Bah¸ce¸sehir Uni) GANs April 21, 2018 14 / 44

StackGAN: Text to Photo-realistic Image Synthesis with
Stacked Generative Adversarial Networks - 2016

StackGAN Results

InfoGAN: Interpretable Representation Learning by
Information Maximizing Generative Adversarial Nets - 2016
Source: https://towardsdatascience.com/infogan-generative-adversarial-networks-part-iii-380c0c6712cd

InfoGAN Results

Wasserstein GAN - 2017

Improved Training of Wasserstein GANs - 2017

WGAN-GP Results - 2017

BEGAN: Boundary Equilibrium Generative Adversarial
Networks - 2017

BEGAN Results

Progressive Growing of GANs for Improved Quality,
Stability, and Variation - 2017

ProGAN Results

Unpaired Image-to-Image Translation using
Cycle-Consistent Adversarial Networks - 2017

CycleGAN Results

Implementing GANs with Tensorﬂow
Generator Network
Discriminator Network
Initialize Graph
Loss, Optimizer, Weights
Training

Discriminator in Tensorﬂow

First Convolutional Layer in Discriminator

Second Convolutional Layer in Discriminator

First Fully Connected Layer in Discriminator

Second Fully Connected Layer in Discriminator

Generator in Tensorﬂow

First Convolutional Layer in Generator

Second Convolutional Layer in Generator

Third Convolutional Layer in Generator

Fourth Convolutional Layer in Generator

Initialize Graph

Loss, Optimizer, Weights in Tensorﬂow

Training

Results

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

Image Generation with Tensorflow

More Related Content

Recently uploaded (20)

Featured (20)

Image Generation with Tensorflow