Practical Deep Learning

This repository documents my personal journey into practical deep learning.
It contains hands-on notebooks, experiments, and notes as I learn and apply deep learning concepts through practice.

What you’ll find

• Jupyter Notebooks: Experiments, exercises, and model training logs.
• Implementations: Practical code inspired by real-world use cases.
• Learning Notes: Observations and "gotchas" encountered while building intuition around DL frameworks.

Purpose

The goal of this repository is learning by doing:

• Understanding concepts through experimentation.
• Making mistakes and improving over time.
• Building intuition around deep learning workflows and data pipelines.

Datasets

This repository utilizes primary datasets for various machine learning and deep learning tasks:

1. Microsoft Cats vs Dogs Dataset

• Source: Kaggle - Microsoft Cats vs Dogs Dataset
• Categories: Cat and Dog.
• Important Note: This dataset contains corrupted images (notably 666.jpg and 11702.jpg) and non-image system files like Thumbs.db. A validation step is mandatory during data loading to prevent training crashes.

2. UTKFace Dataset

• Source: Kaggle - UTKFace (New)
• Categories: Human faces with diverse demographics (over 20,000 images).
• Labels: Age, Gender, and Ethnicity.
• Filename Format: [age]_[gender]_[race]_[date&time].jpg

Label	Description	Mapping
Age	Integer	0 to 116
Gender	Binary	0 (Male), 1 (Female)
Race	Categorical	0 (White), 1 (Black), 2 (Asian), 3 (Indian), 4 (Others)

Tip for Implementation: Since the UTKFace dataset encodes labels in the filename, you will need a custom parser (e.g., using Python's os.listdir and string.split('_')) to extract the target variables.

3. Coin Image Dataset

• Source: Kaggle - Coin Image Dataset
• Categories: US Coins (Pennies, Dimes, Nickels, Quarters).
• Size: 750 pictures.
• Task: Object detection using YOLO.
• Annotation Format: Regardless of the specific type of data, the images need to be structured and separated for annotation using Label Studio.

4. Tatoeba French-English Corpus

• Source: ManyThings.org (Tatoeba project)
• Files: fra-eng.zip (extracts to fra.txt)
• Format: Tab-separated text file structured as: [English sentence] \t [French sentence] \t [Attribution]
• Task: Used for building and training a Transformer model from scratch.

---

Articles, Blogs & Theoretical Deep Dives

Long Short-Term Memory (LSTMs)

• Understanding LSTMs: A visual and intuitive explanation of Long Short-Term Memory networks and how they solve the vanishing gradient problem.

Recurrent Neural Networks (RNNs)

• Backpropagation in RNNs: A deep dive into how gradients flow through time.
• RNN Architectures & Use Cases: A guide to choosing the right structure for sequential data.

Convolutional Neural Networks (CNNs)

• Convolutions and Backpropagation: Visualizing the math behind the spatial feature extraction.
• Padding and Strides in CNN: Understanding how these parameters affect the output feature map size.
• Practical Guide to Data Augmentation: Techniques to improve model generalization (crucial for datasets like UTKFace).

Optimization & Initialization

• Types of Optimizers: Comparison between SGD, Adam, RMSprop, and more.
• Weight Initialization Techniques: Why Xavier and He initialization are vital for training deep architectures.

Hardware & Productivity

• How to use Kaggle GPU: Maximizing free compute resources for training your models.

---

This repo reflects my progress step by step as I grow my skills in deep learning.