Project: MNIST Classifiers

This repository implements several simple classifiers for MNIST provided as raw PNG images in the MNIST/ folder. Implementations include:

• K-Nearest Neighbors (NumPy)
• Bernoulli Naive Bayes (NumPy)
• Linear classifier (L2 loss) (PyTorch)
• Multilayer Perceptron (PyTorch)
• Convolutional Neural Network (PyTorch)

Quick environment setup

python -m pip install --upgrade pip
python -m pip install numpy pillow matplotlib scikit-learn tqdm torch

How to run

• RUN ALL:

  bash run_all.sh

• MLP (PyTorch):

  python multilayer_perception.py

• CNN (PyTorch):

  python convolutional_network.py

• Linear classifier (PyTorch):

  python linear_classifier.py

• Naive Bayes (NumPy):

  python naive_bayes.py

• KNN (NumPy):

  python k_nearest_neighbors.py

What the scripts do

• Each training script saves results under results/<model>/<timestamp>/
  • Classification report (text)
  • Confusion matrix (CSV)
  • Optional visualizations (PNG)

Notes

• The repository uses preprocessing.py helpers to load PNG images and return either NumPy arrays (collect_np_dataset) or PyTorch tensors (collect_torch_dataset).
• Labels are integer class indices (0–9). PyTorch models expect targets as torch.long when using CrossEntropyLoss.
• tqdm is only needed to run k_nearest_neighbors.py as it helps determine the length of time to finish processing.
• Code is only tested using python 3.13.7