Neural Network and Transfer Learning: Assignment 2

This repository contains solutions to Assignment 2 of the Neural Network and Transfer Learning course. The assignment is divided into two main exercises: implementing custom Dataset and DataLoader classes in PyTorch, and performing Principal Component Analysis (PCA) for image reconstruction using Eigen Decomposition (EVD) and Singular Value Decomposition (SVD).

Table of Contents

• Overview
• Exercise 2.2 - Datasets and Dataloaders
• Exercise 2.3 - Eigencats
• Getting Started
• Results

---

Overview

Exercise 2.2: Datasets and Dataloaders

This exercise focuses on creating custom PyTorch Dataset and DataLoader classes for two datasets. The implementation demonstrates:

1. Dataset1 - Tokenized SMS Messages:

   • Tokenized text data (SMS messages labeled as spam or not spam).
   • Processed to include padding for uniform input size.
   • Binary labels (0 for not spam, 1 for spam).
   • Implemented a custom Dataset class to preprocess the data and a DataLoader to batch and shuffle it.

2. Dataset2 - Vector-borne Disease Data:

   • Tabular data containing symptoms and prognoses.
   • Labels encoded using one-hot encoding.
   • Created a custom Dataset class to preprocess the tabular data, handle categorical encoding, and manage data splits.

Exercise 2.3: Eigencats

This exercise demonstrates Principal Component Analysis (PCA) applied to a dataset of cat images. Key aspects include:

1. Computing the Mean Cat:

   • Calculated the mean image by averaging pixel values across all samples.
   • Visualized the result to understand the average characteristics of the dataset.

2. Eigencats - Principal Components of Images:

   • Used Eigen Decomposition (EVD) to compute the principal components (eigencats).
   • Visualized the top principal components as "eigenfaces" for cats.

3. Reconstructing Images:

   • Reconstructed images using varying numbers of principal components.
   • Demonstrated image fidelity improvement with additional components.
   • Used Singular Value Decomposition (SVD) as an alternative method for PCA and compared its results to EVD.

---

Getting Started

Prerequisites

• Python 3.6 or higher
• PyTorch
• Numpy
• Matplotlib
• Pandas

Installation

1. Clone the repository:

   git clone https://github.com/username/NNTI_Assignment2.git

2. Navigate to the repository:

   cd NNTI_Assignment2

3. Install dependencies

Running the Code

• Dataset and Dataloader Implementation: Execute exercise_2_2.py to see the custom Dataset and DataLoader classes in action.
• Eigencats Analysis: Open exercise_2_3.ipynb and run the notebook to see PCA-based image reconstruction.

---

Results

Datasets and Dataloaders

• Dataset1: Successfully tokenized, padded, and batched SMS messages with their corresponding labels.
• Dataset2: Successfully preprocessed tabular data, applied one-hot encoding, and loaded data for analysis.

Eigencats

• Mean Cat: Visualized the average image of the dataset.
• Principal Components: Computed and visualized top eigencats using both EVD and SVD.
• Reconstructed Images:
  • Visualized reconstructed images with varying numbers of principal components (e.g., 10, 40, and 80 components).
  • Observed trade-offs between reconstruction quality and the number of components used.

---