Homework 1 — The ML Framework

This homework accompanies Session 02: The ML Framework. You will move from a Python warm-up to building logistic regression from scratch in PyTorch, and finish with a prediction contest.

📘 The assignment

Do the work in the notebook homework1.ipynb — that's the assignment. A read-only homework1.pdf copy is included for easy reading/printing, but you submit/run the notebook.

---

Setup

# from this folder (final_lectures/hw1)
python -m venv .venv
source .venv/bin/activate        # Windows: .venv\Scripts\activate
pip install -r requirements.txt
jupyter notebook homework1.ipynb

You do not need a GPU. Everything runs on CPU in seconds.

Dataset (Part 2). The MAGIC Gamma Telescope set (19,020 events, 10 numeric features, binary gamma-vs-hadron label), bundled as data/magic04.data — everything runs offline. The Part 3 contest uses a separate dataset (TBD); its files are added to data/ when released.

---

Structure

Part	Topic	What you do
Part 1	PyTorch basics	What PyTorch adds over plain Python: tensors, vectorization, autograd. Assumes you can already program (new to Python? see the Google Python tutorial).
Part 2	Logistic regression from scratch	Build a LinearClassifier (nn.Module), implement sigmoid, BCE loss, accuracy, evaluate, and the gradient-descent loop, then run a learning-rate ablation (loss curves for different learning rates).
Part 3	Contest	Train your best model and produce a submission.csv we will score.

The folder is intentionally minimal — the notebook is self-contained (it defines its own data/plotting helpers, so there is no utils.py to import):

hw1/
├── homework1.ipynb           <- the assignment (work here)
├── homework1.pdf             <- read-only PDF copy of the assignment
├── hw1_tests.py              <- the only support file: self-checks, an
│                                environment check, and the contest validator
├── requirements.txt          <- packages to install
├── reading/                  <- start here: Session 02 summary + logistic-regression reference
├── data/
│   ├── magic04.data          <- Part 2 dataset (MAGIC), bundled
│   ├── contest_train.csv     <- Part 3 training data (added when released)
│   └── contest_test.csv      <- Part 3 test set, labels hidden (added when released)
└── contest/
    └── sample_submission.csv <- example of the expected submission format

---

Self-checking your work

Throughout Part 2 you will see cells like:

from hw1_tests import check_sigmoid
check_sigmoid(sigmoid)        # pass YOUR function

These run lightweight unit tests against your implementation and print PASS / FAIL with a hint. You can also run everything at once at the end:

from hw1_tests import run_all
run_all(globals())

The tests check correctness of the building blocks (shapes, known values, edge cases). Passing them does not guarantee a high contest score — that depends on your modeling choices.

---

Part 3 — Contest (dataset TBD)

You may use any model and any library (your Part 2 model, scikit-learn, …) — you're graded on predictions, not implementation. When the contest dataset is released you'll get two files in data/:

• contest_train.csv — labeled training data (a label column + feature columns).
• contest_test.csv — the test set: an example_id column + the same features, with labels hidden.

You will train a model, predict a label for every example_id, and save a CSV with columns example_id,prediction to contest/submission.csv (see contest/sample_submission.csv). Validate the format before submitting:

python hw1_tests.py contest/submission.csv

A submission that does not pass the validator will not be scored. The exact metric, deadline, and leaderboard will be announced with the dataset.