Week 1 — MNIST Baseline (LogReg vs MLP)

Week 1 Deliverables

This release contains the MNIST baseline project for Week 1 of learning NN.
Two models are implemented and compared:

• Logistic Regression (Linear baseline)

  • Achieves ~90–91% accuracy after 10 epochs
  • Represents the simplest linear classifier

• 2-layer MLP (Nonlinear)

  • Achieves ~97–98% accuracy after 10 epochs
  • Demonstrates the power of adding hidden layers + ReLU

---

Contents

• train.py — training entry point (LogReg / MLP selectable)
• predict.py — inference on a single MNIST digit
• models/ — model definitions (logistic_regression.py, mnist_mlp.py)
• utils/ — dataloaders, metrics, plotting
• checkpoints/ — best.pt and last.pt saved per run
• results/2025-08-27/ — loss/accuracy curves, confusion matrices
• README.md — project documentation
• requirements.txt — pinned dependencies

---

Results

Model	Accuracy	Notes
Logistic Regression	~90–91%	Strong linear baseline
MLP (784→128→10+ReLU)	~97–98%	Captures nonlinearity

---

Highlights

• Pure PyTorch implementation (no sklearn)
• Confusion Matrix written from scratch
• Reproducible (set seeds + pinned versions)
• Ready for extension: Fashion-MNIST, other optimizers

---

Next Steps

• Apply the same pipeline to Fashion-MNIST
• Add EarlyStopping & LR scheduler
• Extend visualization (misclassified samples)

Full Changelog: https://github.com/Ryan0522/MNIST_PyTorch/commits/v0.1.0-mnist