Assignment 3: Deep Learning for Medical Image Classification

Due June 3rd, 2026 at 11:59pm
github details

Overview

You will design and train a convolutional neural network (CNN) to classify breast lesions as benign or malignant using the BreastMNIST dataset — 28 × 28 grayscale images, binary classification.

Your goal is to maximise test-set accuracy. Grading is out of 15 points:

Test-set accuracy Score
≤ reference accuracy (~0.73) 5 / 15
1.0 (perfect) 15 / 15
anything in between linearly interpolated

Estimated Time: 8–12 hours

Learning Objectives

By completing this assignment, you will:

  • Implement and train a CNN for binary medical image classification
  • Design and iterate on neural network architectures
  • Tune hyperparameters (learning rate, batch size, number of epochs)
  • Evaluate model performance on a held-out test set
  • Understand the grading workflow via GitHub Actions

Getting Started

Prerequisites

  • Python 3.11 or higher
  • Git
  • uv (recommended) — fast Python package manager. Install once with:
curl -LsSf https://astral.sh/uv/install.sh | sh

Setup

git clone <your-repo-url>
cd <name-of-repo>
make setup
source .venv/bin/activate

Accept the assignment on GitHub Classroom using the link above, then clone your personal repository.

Assignment Tasks

You may only edit two files. Everything else is part of the grading scaffold — do not modify it.

File What to change
model.py CNN architecture inside BreastLesionCNN
config.yaml Training hyperparameters: batch_size, learning_rate, num_epochs

Step 1 — See the reference model

make baseline

Runs inference with the pre-trained reference CNN (models/baseline_cnn.pt) and prints its test-set accuracy. This is the score you need to beat.

Step 2 — Edit model.py and config.yaml

Change the architecture and/or hyperparameters. See the comments inside each file for guidance.

Keep num_epochs ≤ 100 in config.yaml. Submissions with a saved model file larger than 10 MB or inference time exceeding 30 seconds on the test set will fail the corresponding grading tests.

Step 3 — Train your model

make train

Trains BreastLesionCNN using your config.yaml settings and saves weights to models/submission_cnn.pt. Per-epoch accuracy is printed so you can monitor progress.

Step 4 — Evaluate your model

make evaluate

Loads models/submission_cnn.pt — no retraining — runs inference on the test set, and prints your estimated score out of 15.

Step 5 — Run the grading tests locally

make test

Runs the same tests that will run on GitHub when you push. Each of the 30 threshold tests is worth 0.5 points.

Grading

Grading is fully automated via GitHub Actions.

When you push, GitHub Actions checks out your repository, installs dependencies, and runs pytest tests/ -v. The tests load models/submission_cnn.pt, reconstruct your model using the BreastLesionCNN class from model.py, and run inference on the BreastMNIST test set. No training happens on the server.

Submission

Commit and push your model weights along with your code changes:

git add model.py config.yaml models/submission_cnn.pt
git commit -m "Improve CNN architecture and hyperparameters"
git push

models/submission_cnn.pt must be committed. GitHub grading loads the saved file and runs inference — it does not retrain. If the file is missing, all grading tests will fail.

Check the Actions tab in your repository to see results after each push.