Project 2 — Optimizing Deep Learning Pipelines

This repository contains my implementation for Project 2: Optimizing Deep Learning Pipelines. The goal of this project is to explore optimizer and regularization techniques to improve model accuracy and generalization on the Sign Language MNIST dataset.

Overview

This project involves building baseline and optimized neural network models and comparing their training behavior and evaluation metrics. Specifically, the project covers:

• Data loading and preprocessing for the Sign Language MNIST dataset (24 classes)
• Exploration and visualization (class distributions, sample images)
• A baseline dense model (3 layers) trained with Adam optimizer
• Optimized models using different optimizers (Adam, SGD, RMSProp) and regularization techniques (Dropout, BatchNorm, L2)
• Evaluation: accuracy/loss curves, confusion matrix, classification report
• Reflection and discussion of results and ethical considerations

Source dataset: https://www.kaggle.com/datasets/datamunge/sign-language-mnist

Folder structure

Path	Type	Purpose / What to store
data/raw/	directory	Original downloaded dataset files (CSV, ZIP). Keep raw files here and exclude from git via .gitignore.
data/processed/	directory	Preprocessed and cached artifacts (NumPy arrays, pickles, smaller CSVs). Safe to commit small caches; large caches can be excluded.
src/	package	Reusable project code and modules. Commit source files here.
src/main_data/	package	Helper functions for data loading, preprocessing, and plotting.
notebooks/	directory	Primary Jupyter notebook(s) (e.g., main.ipynb). Prefer cleared outputs before committing.
outputs/	directory	Model checkpoints, figures, logs, and exported results. Large artifacts should be stored externally or added to .gitignore.
outputs/models/	directory	Saved model weights and checkpoints (.h5, .ckpt).
tests/	directory	Unit and integration tests (pytest).
docs/	directory	Project documentation, reports, and submission artifacts.
environment.yml	file	Conda environment specification. Commit for reproducibility.
.gitignore	file	Exclude raw data, large model artifacts, and other local files.

Notes on storage

• Keep raw dataset files in data/raw/ and exclude them from Git (use .gitignore).
• Store processed, small cache files in data/processed/ for reproducible runs; consider excluding very large caches.
• Save large models/checkpoints and long-term artifacts in outputs/ or external storage (Google Drive, S3); keep lightweight checkpoints in the repo only if necessary.
• Clear notebook outputs before committing to keep repo size manageable.

Data & schema

• Source: Sign Language MNIST on Kaggle (https://www.kaggle.com/datasets/datamunge/sign-language-mnist). The raw CSVs included here are sign_mnist_train.csv and sign_mnist_test.csv.
• CSV format: first column is label (class index), remaining 784 columns are pixel values for a 28x28 grayscale image (row-major order). When converting to arrays we save as NumPy arrays of shape (N, 28, 28) or flattened (N, 784) depending on the pipeline step.
• Where to store: place original downloaded CSV/ZIP files in data/raw/. Put preprocessed NumPy arrays, pickles, and smaller cached artifacts in data/processed/.
• Download (example using Kaggle CLI):

# install and authenticate kaggle-cli first (https://github.com/Kaggle/kaggle-api)
kaggle datasets download -d datamunge/sign-language-mnist -p data/raw/ --unzip

If you cannot use the Kaggle CLI, download via the web UI and move the files into data/raw/.

Requirements & Environment

This project was designed to run within a Conda environment. An environment.yml file is provided for reproducibility. Required packages include:

• python (3.8+)
• numpy, pandas
• matplotlib, seaborn
• tensorflow (or keras)
• scikit-learn
• jupyter

How to run

Local Setup

1. Create and activate the conda environment:

conda env create -f environment.yml
conda activate project2

2. Open the primary notebook main.ipynb in the root directory. This notebook contains all the code, visualizations, and analysis.

3. Run all cells in sequence. The notebook includes:

   • Data loading and exploration
   • Baseline model implementation
   • Optimized models with various optimizers and regularization techniques
   • Data augmentation implementation
   • Performance evaluation and visualization

Google Colab Setup

The notebook is also compatible with Google Colab:

1. Upload main.ipynb to Google Colab
2. Upload the dataset files to the Colab session or mount Google Drive
3. Run all cells in sequence

Note: This project has been tested and runs successfully in both local and Colab environments.

Project Components

Data Loading & Preprocessing

• Loaded Sign Language MNIST dataset (24 classes, letters A-Z excluding J and Z)
• Implemented data normalization and preprocessing pipeline
• Handled train/test split

Data Augmentation

• Implemented augmentation techniques to improve model generalization
• Applied rotation, shifts, zoom, and other transformations
• Strategically applied to training data

Baseline Model

• Built 3-layer dense neural network
• Trained with Adam optimizer
• Evaluated performance metrics

Optimized Models

• Compared multiple optimizers (Adam, SGD, RMSProp)
• Applied regularization techniques (Dropout, Batch Normalization, L2)
• Conducted hyperparameter tuning and experimentation

Visualization & Analysis

• Generated training/validation accuracy and loss curves
• Created confusion matrices
• Produced classification reports
• Analyzed class distribution
• Visualized sample images

Environment Configuration

• Provided complete environment.yml with all required dependencies
• Tested compatibility with both local and Colab environments

Dataset Notes

• Source: Sign Language MNIST on Kaggle
• Classes: 24 (A-Y, excluding J and Z which require motion)
• Format: 28x28 grayscale images
• Storage: Keep raw dataset files in data/raw/ (excluded from git via .gitignore)

Notes

• Large dataset files are kept out of the repository (excluded via .gitignore)
• Model checkpoints and artifacts can be stored in outputs/ or externally (e.g., Google Drive)
• The main notebook (main.ipynb) is located in the root directory for easy access
• All outputs are saved in the notebook to show results without re-running computationally expensive training

AI Use

AI was used in the following workflows

• Suggesting ideas and flow of experiments
• Ensuring answers are accurate
• Generating Graphs and function options/syntax
• Formatting and review of Analysis
• Formatting of the project structure to aid in bulk edits

Troubleshooting

Import Errors: Ensure the conda environment is activated with conda activate project2

Missing Data: Download the Sign Language MNIST dataset from Kaggle and place CSV files in data/raw/

Memory Issues: If running locally with limited RAM, reduce batch sizes or use Colab with GPU runtime