Deep Learning exercises

Student material for the HPI Deep Learning course, taught by Christoph Lippert with Noel Danz, Pedro De Macedo Gomes, and Anees Hashmi.

Start with Week 1: fitting a line and its worksheet. The repository includes the runnable experiment, its synthetic data generator, and the lecture figures. It works independently of the book repository. Instructor notes and worksheet answer keys are maintained separately.

Week 2: logistic regression includes a runnable student notebook and classroom worksheet for two 90-minute sessions. It follows Sections 8–10 of the book using the same bundled two-feature WDBC dataset, split, and numerical conventions.

Read the rendered classroom worksheets online. Updates are published automatically after merging into main. For the one-time Pages setup and the permanent link to add in Moodle, see publishing instructions.

Get the code

git clone https://github.com/HealthML/deep-learning-exercises.git
cd deep-learning-exercises

For an existing checkout, commit or save your own changes before updating with git pull --ff-only. You can change the starter locally or keep experiment copies under exercises/week01/output/.

Install Python and create an environment

Use uv to manage Python and the course packages. Install it once using the official installation instructions:

macOS and Linux

curl -LsSf https://astral.sh/uv/install.sh | sh

Windows PowerShell

winget install --id=astral-sh.uv -e

Open a new terminal after installation so uv is available on your PATH. Then run these commands from the repository root on all platforms:

uv sync --locked
uv run --locked python tools/check_setup.py
uv run --locked python exercises/week01/starter.py

uv sync --locked downloads Python 3.11 if needed, creates .venv, and installs the student package in editable mode together with the exact dependency versions in uv.lock. Python 3.11 is the default in .python-version; CI also checks Python 3.12 and 3.13. No separate Python installation or environment activation is required. Use uv run --locked python ... for worksheet commands on every platform, including Windows PowerShell.

The project selects CPU-only PyTorch wheels on Linux and Windows; macOS uses PyPI. On macOS, the PyTorch dependency supports Apple Silicon with macOS 11 or newer. All tensors in Week 1 stay on CPU. The dependency and index settings live in pyproject.toml, so setup and CI use the same choices.

If you already have a .venv from the previous setup, uv sync --locked updates it. If it uses a different Python version, uv recreates it for Python 3.11.

In a new terminal

From the repository root, continue with uv run --locked python .... After updating the repository with git pull --ff-only, run uv sync --locked to apply any dependency changes.

Week 2 notebooks

Install the optional notebook tools and open the Week 2 experiments:

uv sync --locked --group notebook
uv run --locked --group notebook jupyter lab exercises/week02/logistic_regression.ipynb

Choose this repository's .venv as your notebook kernel if prompted. The experiments use NumPy and CPU PyTorch; no book package or data download is required. All cells run as delivered. The test checkpoint is disabled by default until the model, threshold, and written validation work are frozen. See the Week 2 instructions.

Troubleshooting

• uv is not found: reopen the terminal after installing uv and follow the installer's PATH instructions.
• No module named dlexercises: run uv sync --locked from the repository root and use uv run --locked python tools/check_setup.py. The script reports the interpreter and package location so you can identify an environment mismatch.
• The lockfile needs updating: after changing dependencies, use uv add (or uv remove) and commit both pyproject.toml and uv.lock. For an intentional dependency refresh, run uv lock --upgrade, uv sync --locked, and the checks below before committing the updated lockfile.
• No plot window: use --save fit.png instead of --plot; on Linux servers, set MPLBACKEND=Agg if needed. A graphical desktop is needed for a plot window.
• No compatible PyTorch wheel: use the pinned Python 3.11 environment and check your platform against the official PyTorch installer.
• Quarto or Jupyter is missing: both are optional for this initial exercise. Read the Markdown worksheet and run the Python script directly.

Files and checks

Path	Purpose
exercises/week01/starter.py	Manual parameter trials and guided fitting experiments
exercises/week01/worksheet.md	Questions and spaces for your explanations
exercises/week01/worksheet.qmd	Optional Quarto source of the same worksheet
exercises/week01/figures/	Lecture visuals accompanying the worksheet
src/dlexercises/data.py	Reproducible synthetic sensor data
pyproject.toml / uv.lock	Dependencies, platform-specific PyTorch index, and exact versions
.python-version	Default Python version managed by uv
tools/check_setup.py	Interpreter, dependency, data, and image-export check
tools/smoke_check.py	Run each experiment mode and check CLI/image behavior
exercises/week02/logistic_regression.ipynb	Prediction-first experiments for two logistic-regression sessions
exercises/week02/worksheet.md	Timed pair tasks, individual exit tickets, and optional depth
src/dlexercises/week02.py	Transparent logistic loss, derivatives, fitting, and evaluation helpers
src/dlexercises/datasets/wdbc/	Bundled two-feature data with attribution and checksums
tools/check_week02.py	Execute all notebook cells and check data, derivatives, solvers, and test protection

You can run the repository checks with:

uv run --locked python -m compileall -q src exercises tools
uv run --locked python tools/smoke_check.py
uv run --locked python tools/check_week02.py

Attribution and license

Week 1 material is adapted from HealthML/deep-learning-book at commit 7919dce810877339178c33725e17959c83c02d47 by Christoph Lippert and contributors. The dataset and equations match the course book and slides. See LICENSE.md: original source code is MIT; prose, exercises, and figures are CC BY-SA 4.0.

Week 2 material and helpers are adapted from Chapter 02 of the same book at commit 7803e0bcbbe1b20d7563db09cee97f4976242e6b. WDBC data is CC BY 4.0; the bundled data notes credit Wolberg, Mangasarian, Street, and Street and record the UCI DOI, exact extraction, checksums, and book split.