Official Open

VGAP Assignment

Visual Generative AI Applications

Specialist Diploma in Generative AI
Oct 2024 Semester

ASSIGNMENT

Submission Deadline:
Presentation: 5-6th Dec 2024 11:59PM
Report and Code: 7th Dec 2024 11:59PM

Students Name :

Student Number :

Penalty for late submission:

10% of the marks will be deducted every calendar day after the deadline.
NO submission will be accepted after 14th Dec 2024, 11:59PM.

1
SDGAI Oct Semester 2024/25
Official Open

VGAP Assignment
1. Assignment Objectives
This assignment aims to assess your competencies in the following areas:

- Understanding Generative AI Models

Explain the core principles and applications of generative AI models, including Generative
Adversarial Networks (GANs), and/or diffusion models.

- Building and Implementing Generative Models

To build generative AI models for image generation, with a focus on both theoretical
understanding and practical implementation, including the fine-tuning of model hyper-
parameters to optimize performance and improve output quality.

- Controlled Generation
Implement controlled generative processes to influence and guide the output of AI models.

2. Assignment Overview
You will use the Fashion MNIST dataset in two parts:

Part 1: Build and train unconditional generative models to generate images of fashion
items.

Implement two types of models:

• Unconditional Generative Adversarial Networks (GANs)
• Unconditional Diffusion Models

Part 2: Build and train conditional generative models to control the image generation
process.

Implement two types of models:

• Conditional GANs
• Conditional Diffusion Models

3. Dataset
Download the dataset from PoliteMall, which contains a collection of images on fashion items
from 10 different categories:
• T-shirt/Top
• Trouser
• Pullover
• Dress

2
SDGAI Oct Semester 2024/25
Official Open

VGAP Assignment
• Coat
• Sandal
• Shirt
• Sneaker
• Bag
• Ankle Boot

The dataset consists of 60,000 training images across 10 categories. Each image is grayscale
with a resolution of 28x28 pixels, where the pixel intensity values range from 0 to 255. Figure
1 displays 25 random examples from this dataset.

Figure 1 Visualization of the Fashion MNIST Dataset

3
SDGAI Oct Semester 2024/25
Official Open

VGAP Assignment
4. Building Generative Models
4.1. Part 1: Unconditional Generative Model

Design and implement a generative model capable of creating images of fashion items from
10 distinct categories (classes). Your model should generate visually diverse and accurate
representations for each class.

The Jupyter notebook starter file (ASG_starter.ipynb) downloads the MNIST fashion item
dataset into a dataloader.

Your task is to complete this notebook by implementing your various generative models. It is
recommended that you follow the suggested structure provided below to tackle the problem
effectively.

You will need to develop 2 models for unconditional image generation following the
recommended steps below:

4.1.1. Unconditional Generative Adversarial Network (Vanilla GAN

or Deep Convolutional GAN)

Step 1 – Load and explore your Dataset.

- Load your data and visualize the data.
- Demonstrate a clear understanding of the MNIST fashion item dataset.

Step 2 – Build an unconditional GAN Model

- Develop an unconditional GAN model to generate images of fashion items.
- You may use vanilla GAN or Deep Convolutional GAN to achieve your desired
outcome.
- Document the following:
o Start with a baseline GAN model.
o Considerations for hyperparameters.
- Performance optimization strategies.
- Analyze the model’s performance and adjust hyperparameters during training.
- Ensure to record performance curves for reporting purposes.

4
SDGAI Oct Semester 2024/25
Official Open

VGAP Assignment
Step 3 – Evaluate your model.
- Use the model to generate new images.
- Provide a random (noise) tensor as input to generate your desired image.

Step 4 – Save your notebook.

- Save your notebook for submission as

ASG1_unGAN_[your_name].ipynb

4.1.2. Unconditional Diffusion Model

Step 1 – Load your Dataset.

Step 2 – Build an unconditional Diffusion Model

- Develop a diffusion model to generate images of fashion items.
- Document the following:
o Start with a baseline diffusion model.
o Considerations for hyperparameters.
- Performance optimization strategies.
- Analyze the model’s performance and adjust hyperparameters during training.
- Ensure to record performance curves for reporting purposes.

Step 3 – Evaluate your model.

- Use the model to generate new images.
- Provide a random (noise) tensor as input to generate your desired image.

Step 4 – Save your notebook.

- Save your notebook for submission as

ASG1_unDiffusion _[your_name].ipynb.

5
SDGAI Oct Semester 2024/25
Official Open

VGAP Assignment
4.2. Part 2: Conditional Generative Model

Design and implement a generative model capable of creating images of fashion items from
10 distinct categories (classes). Your model should generate visually diverse and accurate
representations for each class.

You will need to develop 2 models for conditional image generation following the
recommended steps:

4.2.1. Conditional Generative Adversarial Network (Vanilla GAN or Deep

Convolutional GAN)

Step 1 – Load your Dataset.

Step 2 – Build a conditional GAN Model

- Develop a conditional GAN model to generate images of fashion items.
- You may use conditional GAN or Deep Convolutional GAN to achieve your desired
outcome.
- Document the following:
o Start with a baseline GAN model.
o Considerations for hyperparameters.
- Performance optimization strategies.
- Analyze the model’s performance and adjust hyperparameters during training.
- Ensure to record performance curves for reporting purposes.

Step 3 – Evaluate your model.

- Use the model to generate new images.
- Provide a random (noise) tensor and your image generation condition as input to
generate your desired image.

Step 4 – Save your notebook.

- Save your notebook for submission as

ASG1_cGAN_[your_name].ipynb.

6
SDGAI Oct Semester 2024/25
Official Open

VGAP Assignment
4.2.2. Conditional Diffusion Model

Step 1 – Load your Dataset.

Step 2 – Build a conditional Diffusion Model

- Develop a conditional diffusion model to generate images of fashion items.
- Document the following:
o Start with a baseline diffusion model.
o Considerations for hyperparameters.
- Performance optimization strategies.
- Analyze the model’s performance and adjust hyperparameters during training.
- Ensure to record performance curves for reporting purposes.

Step 3 – Evaluate your model.

- Use the model to generate new images.
- Provide a random (noise) tensor and your image generation condition as input to
generate your desired image.

Step 4 – Save your notebook.

- Save your notebook for submission as

ASG1_cDiffusion _[your_name].ipynb.

7
SDGAI Oct Semester 2024/25
Official Open

VGAP Assignment
5. Deliverables
The deliverable for this assignment includes:
- Jupyter Notebook files
- Report
- Online live presentation.

5.1. Jupyter Notebook Files Submission

Complete your notebook files.
- ASG1_unGAN_[your_name].ipynb
- ASG1_unDiffusion _[your_name].ipynb,
- ASG1_cGAN_[your_name].ipynb,
- ASG1_cDiffusion _[your_name].ipynb.)

Zip the 4 files into one file, rename the file as ASG1 _[your_name].zip. Submit your zip file
on PoliteMall by 7th Dec 2024 11:59PM.

Late submissions of assignment-based coursework component without leave of

absence (LOA) for the module will be subjected to the late penalty. Note that only the
most recent submission before the deadline will be considered for grading.

DO NOT PLAGIARIZE (please refer to Ngee Ann Polytechnic Plagiarism Policy webpage
for more information)

5.2. Report Submission

Write an INDIVIDUAL report with the following sections (see Table below) and submit your
report on PoliteMall by 7th Dec 2024 11:59PM.

Suggested content description and word count are provided for each section. You are free to
include other relevant information you deem necessary in the sections.

(Note: For a page with 1-inch margins, 12-point Arial font, and minimal spacing elements, a good rule
of thumb is 500 words for a single-spaced page)

8
SDGAI Oct Semester 2024/25
Official Open

VGAP Assignment
Suggested Report Sections & Content Guidelines Word Count
1. Table of Contents NA

2. Overview
• Describe the problem, the objective and the approach. Min: 300 words
Max: 1000 words

3. Data Preprocessing and Data Loading

• Describe how you preprocess the data and load data into Min: 200 words
Jupyter Notebook Max: 500 words

4. Develop the 2 unconditional generative models (GAN and diffusion

model) Min: 1000 words
• Describe how you build & train the models, tuning model Max: 2000 words
hyperparameters.
• Analyze the model performance.
• Compare your models, recommend the best model and
explain why.

5. Develop the 2 conditional models (GAN and diffusion model)

• Describe how you build & train the models, tuning model Min: 1000 words
hyperparameters. Max: 2000 words
• Analyze the model’s performance.

6. Summary
• Summarize your model performance and provide Min: 100 words
suggestions for further improvements Max: 500 words

7 Reflect on the learning outcomes of the course outlined below, and

write a reflection on how this course has enhanced your
understanding of generative AI:
• Explain the core principles and applications of generative AI
models, including Variational Autoencoders (VAEs),
Generative Adversarial Networks (GANs), and diffusion
models.
• To build generative AI models for image generation, with a
focus on both theoretical understanding and practical
implementation, including the fine-tuning of model hyper-
parameters to optimize performance and improve output
quality.
• Implement controlled generative processes to influence and
guide the output of AI models.

9
SDGAI Oct Semester 2024/25
Official Open

VGAP Assignment
• Evaluate AI ethics and apply ethical considerations in the
design and deployment of generative AI models,
emphasizing the role of prompt engineering in shaping ethical
outcomes.

5.3. Presentation
You are required to do an online live presentation and share your findings. The presentation
should not exceed 15 minutes, followed by a Question-and-Answer session.

You are encouraged to use your Jupyter notebook for your presentation. The Online
Presentation is scheduled in the week starting 2nd Dec 2024. The exact timing will be arranged
by your tutor. If class is affected by public holiday, alternative arrangements will be made.

6. Grading Criteria
Component
Grading Criteria
Weightage

a) Quality of work
b) Flow of presentation based on content guidelines
(see section 5.2)
Presentation 20%
c) Presentation and articulation skills
d) Clear comprehension of code and fundamental
concepts.

a) Quality of work
b) Completeness of report based on suggested report
Final Report sections and content guidelines (see section 5.2) 80%
c) Clarity of report, Quality of analysis and discussions
d) Use of proper visual aids and Use of proper grammar

10
SDGAI Oct Semester 2024/25