Classification Project

Initial project documentation for the EDK classification assignment.

Status

• MNIST: implemented for Part 1 and Part 2.
• Iris: under construction.

Project Structure

Key folders/files:

• mnist/: MNIST implementation scripts and tests.
• data/mnist/: MNIST dataset files (.gz).
• utils.py: shared confusion-matrix and error-rate helpers.
• report/figures/: report-ready saved figures.

Environment

Recommended Python version: 3.10+.

Required packages:

• numpy
• scipy
• matplotlib
• scikit-learn (required for Part 2 k-means)

Install example:

pip install numpy scipy matplotlib scikit-learn

MNIST

Dataset Placement

Place the following files under data/mnist/:

• train-images-idx3-ubyte.gz
• train-labels-idx1-ubyte.gz
• t10k-images-idx3-ubyte.gz
• t10k-labels-idx1-ubyte.gz

Expected layout:

classification_project/
	data/
		mnist/
			train-images-idx3-ubyte.gz
			train-labels-idx1-ubyte.gz
			t10k-images-idx3-ubyte.gz
			t10k-labels-idx1-ubyte.gz

Part 1: NN with Full Training Templates

Script: mnist/part1_nn.py

Run from mnist/:

python part1_nn.py --lite
python part1_nn.py --full
python part1_nn.py --full --quiet

Flags:

• --lite: quick synthetic smoke run (no file outputs).
• --full: full MNIST run (default behavior if no mode flag is provided).
• --quiet: suppress metric/confusion-matrix printing.

Notes:

• Full mode creates confusion matrices and image grids under mnist/results/ and report/figures/.
• If plot windows block terminal completion, use:

MPLBACKEND=Agg python part1_nn.py --full

Part 2: K-Means Templates + NN/KNN

Script: mnist/part2_kmeans.py

Run from mnist/:

python part2_kmeans.py --lite
python part2_kmeans.py --full --clusters 64 --k 7
python part2_kmeans.py --full --clusters 64 --k 7 --quiet

Flags:

• --lite: quick synthetic smoke run (no file outputs).
• --full: full MNIST run (default behavior if no mode flag is provided).
• --quiet: suppress metric/confusion-matrix printing.
• --clusters: templates per class for k-means (default: 64).
• --k: KNN neighbor count for Part 2c (default: 7).

Smoke Tests

Run from mnist/:

python tests/test_part1.py
python tests/test_part2.py

What these tests do:

• Validate core logic on small synthetic data.
• Verify chunked processing and output shape constraints.
• Confirm assignment-specific behavior (Euclidean distance, template usage, and pipeline side-effect controls).
• Avoid heavy full-dataset runtime.

Output Files

Generated outputs are saved under:

• mnist/results/confusion_matrices/
• mnist/results/misclassified_images/
• report/figures/

Iris

Iris section is intentionally minimal for now and will be expanded once the implementation is complete.