Scribd
Upload
392 views16 pages

DCGAN Presentation

This was a presentation for Multi-modal Perception Lab @ IIIT-B. Link to more of them: https://github.com/Nikunj-Gupta/Presentations

Uploaded by

Nikunj Gupta
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
392 views16 pages

DCGAN Presentation

This was a presentation for Multi-modal Perception Lab @ IIIT-B. Link to more of them: https://github.com/Nikunj-Gupta/Presentations

Uploaded by

Nikunj Gupta
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 16

Unsupervised Representation

Learning Through Deep


Convolutional Generative
Adversarial Networks (DCGANs)

-Alec Radford & Luke Metz


indico Research
Boston, MA
-Soumith Chintala
Facebook AI Research
New York, NY

Presented by
Nikunj Gupta
Contributions

● Stabilizing the training of GANs in most settings.


○ By proposing and evaluating a set of constraints on the architecture of GANs.

● Usage of trained discriminators for image classification tasks.

● Visualization of the filters learnt by GANs


○ And empirically show that specific filters have learned to draw specific
objects.

● Showing that generators have interesting vector arithmetic properties


○ For easy manipulation of many semantic qualities of generated samples.
Related Work

● Representation Learning from Unlabelled Data


○ Clustering (K-means, etc.)
○ Hierarchical Clustering
○ Auto encoders
○ Deep Belief Networks

● Generating Natural Images


○ Variational Sampling Approach in GANs
○ Laplacian Pyramid Extension to GANs

● Visualizing the Internals of CNN


○ Using deconvolutions and filtering the maximal activations (Zieler et. al.)
GAN Architecture
Inside GAN Architecture
Approach and Model Architecture

● Replace any pooling layers with


○ strided convolutions (discriminator), and
○ fractional-strided convolutions (generator).

● Use Batch Normalisation in both the generator and the discriminator.

● Remove fully connected hidden layers for deeper architectures.

● Use ReLU activation in generator for all layers except for the output, which uses
Tanh.

● Use LeakyReLU activation in the discriminator for all layers.


Generator Network
Discriminator Network
Goals

● Generator Goal: D(G(z)) = 1


● Discriminator Goal: D(G(z)) = 0,

● Both goals are conflicting and unsupervised.


Details of Adversarial Training

● The training is done on 3 datasets: LSUN, ImageNet-1K, Faces Dataset

● Preprocessing: Scale images between -1 and +1 (tanh range)

● Mini batch SGD (m = 128)

● Weight Initialisation: zero centered normal distribution (std dev = 0.02)

● Leaky ReLU slope: 0.2

● Adam Optimizer

● Learning Rate: 0.0002

● Momentum Term = 0.5, to stabilize training


Results

Classification using the Discriminator Network


Bedroom Generation from LSUN Dataset
Face Generation
Interesting Vector Arithmetic
Representations
One more example…
Thank You

You might also like