Scribd
Upload
3 views

PPT Script

The document outlines a project focused on utilizing Generative Adversarial Networks (GANs) to generate Islamic art patterns, addressing the challenges of traditional methods that are time-consuming. It includes sections on problem statement, objectives, significance, related works, requirements analysis, proposed solutions, market research, and proof of concept. The project aims to raise awareness of Islamic patterns, facilitate their use in architecture and branding, and provide a dataset for research purposes.

Uploaded by

itsrbm1811
Copyright
© © All Rights Reserved
Available Formats
Download as DOCX, PDF, TXT or read online on Scribd
3 views

PPT Script

The document outlines a project focused on utilizing Generative Adversarial Networks (GANs) to generate Islamic art patterns, addressing the challenges of traditional methods that are time-consuming. It includes sections on problem statement, objectives, significance, related works, requirements analysis, proposed solutions, market research, and proof of concept. The project aims to raise awareness of Islamic patterns, facilitate their use in architecture and branding, and provide a dataset for research purposes.

Uploaded by

itsrbm1811
Copyright
© © All Rights Reserved
Available Formats
Download as DOCX, PDF, TXT or read online on Scribd
You are on page 1/ 10

Outline

1. IPattern (15 - 20 sec)


2. Introduction (5 – 10 sec)
a. Problem statement (1 min)
b. Objectives (1min)
c. Significance (40 sec)
3. Related works (4min)
a. DC-Art GAN
b. CANS
c. Navigating the latent
d. StyleGANs - Jill Shah
4. Requirements Analysis (3min)
a. Development Process
b. Functional requirements (use case)
c. Non-Function requirements
5. Proposed Solution (4min)
a. Alternative solutions and trade-offs
b. Selected Solution
c. High level Architecture
d. Software Used
6. Market Research (4min)
a. Pie Charts (9-slides)
b. Anticipated risks

7. Proof of Concept
PPT Script

Problem Statement

Throughout history, art has been used as a medium to express thoughts, beliefs, and creativity.
The same holds true for Islamic art, at first glance it looks very complex and intricate, but one
soon realizes that it's a beautifully organized arrangement based on geometry and
mathematical equations. However, one cannot deny that it requires dedicating long hours to
practice and perfect. Not to mention, that people nowadays gravitate towards computer-
generated content, so it is feared that people will lose interest Islamic patterns due to their
time-consuming nature. Making them more at risk of gaining negligence over time. But with the
development of technology and emergence of machine learning, we seek to close the gap
between technology and Islamic patterns by using the most recent deep learning models, which
are Generative Adversarial Networks (GANs).

GANs are a deep learning model, that primarily consists of two networks, the generator and the
discriminator. The networks are in constant adversary, each with their own objective.

Simply put, the generator tried to generate fake data that resembles the real data, and gives it
to the discriminator network, in hopes of fooling the discriminator

Whereas the discriminators job is to classify whether the data it was given is real or fake.

If the generator fools the discriminator it advances, and the discriminator gets feedback to
improve its guessing abilities

If the opposite happened, and the discriminator classifies correctly and is not fooled then the
generator gets the feed to improve its fooling abilities
Problem Objectives - Read from PPT

Problem Significance and Benefits - Read from PPT

Here we mention four benefits that could arise from this project

Awareness:

Raise awareness about the importance of Islamic pattern

and inspire new geometric possibilities; as machines are not restricted by rules of creation,
which could open new doors.

Architecture & Branding:

Facilitate the use generated patterns in buildings, which could be utilized by architects

as well as branding purposes, as people might use them as for designs and logos for their
brands

As for Research:

We are utilizing GANs with purposes related to our cultural and tradition, which is a benefit
with intrinsic value.

Last but certainly not least, is providing the research community with a dataset of Islamic
patterns.
*IMPORTANT: you might start with theory part of the Train the network (Why stylegan2-ada?
and the model architecture)

PPT – Model chosen (why STYLEGAN2-ADA)


PPT – Model Architecture

Demo - Below

Results obtained
-Slides

Human Subject Evaluation (Train the network)


-Slides
______________________________________________________________________________

POC Theory (Slides)


______________________________________________________________________________

Model Chosen (Why StyleGAN2-ADA?)

Model Architecture

Results

- Show the images


- Human subject evaluation
Action – Red
Script - Blue

______________________________________________________________________________

POC Demo
______________________________________________________________________________

Image Gathering:
[Safari => uncurated dataset]

Goal: We started with the first use-case in our use case diagram which was the curate and
process images use case, which firstly involved collecting a large dataset set comprising of
pattern.

The question in how large?


There weren’t many papers discussing the issue, but we found one paper that concluded that
1300 was the way to go

However, keeping in mind that the conclusion came about by showing the generated results to
human subjects, and no quantitative accounts were considered.
We sought to do the experiments by ourselves. Basically, we trained the model various times
with different sizes of dataset

Process: -

Problems: Firstly we found that the best solution was screenshotting as it maintained the
photos quality. but you see screen shotting the images came with its problems, like:

- Time consuming
- Licensing issue

Solution: We found a website that had unlimited possibilities for pattern creation, which was
the Repper website. Which also came with some problem that we will discuss when viewing the
results

Image Processing:

[imageProcessing.ipynb]

Environment:
Why Google Colab:
- Provides access to free cloud GPUs, which is essential for us in the testing phase.
- Also, with subscribing to the pro plan, we can execute the code in the background. And
gives access to better performing GPUs such as the A100 and V100
- Easily Shareable between teammates
- Ease of syncing with google drive

Goal: To achieve uniformity before feeding the images to the networks, basically the network
only accepts, images power 2 dimensions, and requires us to standardize the color mode

Process:
-First, we mount google drive to gain full access of the drives content
-Then we crop the images at the center
-Scale them to our desired dimension
-Standardize the color mode, in our case the color mode is RGB
Problems: -

Training the Network:

[copy-stylegan2-ada. ipynb]

- Explain the steps briefly in Colab


- The attempts

Process: just read from Colab

What is happening in the training code:

Results:
Attempt:

Test01: Screen-Shotting method => 500 images => mode collapse

Test02: Screen-Shotting method => 800 images => mode collapse

Test03: Found Repper => 2300images => Good results

Test03: Found Repper => 3000images =>

Test03: Found Repper => 3000images => Good results


StyleGAN2-ADA is a variant of the StyleGAN2 model, which is a type of
generative adversarial network (GAN). GANs consist of two neural networks,
a generator and a discriminator, that are trained to work against each other
in a minimax game. The generator tries to produce fake data that is
indistinguishable from real data, while the discriminator tries to accurately
identify whether each piece of data is real or fake. Through this process, the
generator learns to create more realistic data.

StyleGAN2-ADA is an extension of the StyleGAN2 model that was specifically


designed for generating high-resolution images of human faces. It is an
"adaption" of the original StyleGAN2 model, which means that it has been
fine-tuned for this specific task.

The architecture of the StyleGAN2-ADA model consists of several


components:

1. The generator network: This network is responsible for generating


images. It consists of several layers of convolutional neural networks
(CNNs) and residual blocks. The output of the generator network is a
high-resolution image.
2. The discriminator network: This network is responsible for
distinguishing real images from fake ones. It also consists of several
layers of CNNs and is trained to classify images as either real or fake.
3. The mapping network: This network maps the input latent code (a
random noise vector) to a higher-dimensional latent space, which the
generator network uses to create the output image.
4. The style modulation module: This module modulates the style of the
generated image by applying style vectors to the output of the
mapping network. This allows the model to generate images with
different styles and variations.
5. The progressive growing module: This module gradually increases the
resolution of the generated image by adding new layers to the
generator network during training. This allows the model to generate
high-resolution images while still being able to learn effectively.

1. Convolutional layers: These layers apply a convolution operation to the


input data, which helps the network learn spatial hierarchies of
features. The convolutional layers in the generator network and
discriminator network of StyleGAN2-ADA are made up of several
stacked convolutional blocks, each of which consists of a convolutional
layer, a normalization layer, and an activation function.
2. Residual blocks: These blocks are used in the generator network to
allow the network to learn complex, non-linear relationships in the
data. Each residual block consists of two convolutional layers, a
normalization layer, and a skip connection that adds the input of the
block to its output.
3. Mapping network: The mapping network consists of a series of fully
connected (FC) layers that map the input latent code to a higher-
dimensional latent space.
4. Style modulation module: The style modulation module consists of
several fully connected (FC) layers that apply style vectors to the
output of the mapping network. These style vectors modulate the style
of the generated image.
5. Progressive growing module: The progressive growing module
gradually increases the resolution of the generated image by adding
new layers to the generator network during training. This allows the
model to generate high-resolution images while still being able to learn
effectively.

You might also like