Release: v1.1.0 – Updated Report and Final Project

Release: v1.1.0 – Final Submission

Overview

This is the first official release of the MNIST Optimization Project. It includes a complete end-to-end implementation of a logistic regression classifier on the binary MNIST dataset (0 vs. not-0), with manual implementations of key optimization algorithms and training workflows.

Included Features

• Binary MNIST classifier using NumPy (flattened images, custom logistic regression model)
• Custom optimization algorithms:
  • Stochastic Gradient Descent (SGD) with fixed learning rates
  • SGD with decreasing learning rate
  • SGD with Momentum
  • Adam optimizer (with and without bias correction)
• Training utilities:
  • Loss tracking during training
  • Smoothed loss curve visualization using Matplotlib
  • Final loss values printed in terminal

Outputs

• Two plots saved to the plots/ folder:
  • sgd_momentum.png: Loss comparison of SGD variants and Momentum
  • adam_comparison.png: Loss comparison of Adam (with vs. without bias correction)
• Final loss values displayed in terminal

Documentation

• README.md with run instructions, project description and report

Getting Started

pip install torch torchvision matplotlib numpy
python main.py

Contributors

• Rizik Haddad (SGD and project report)
• Sam Buttonow (Momentum and Adam)