Materials

Slides and notebooks used during the school. This page grows as the week progresses — check back for updates.

TipNotebooks with exercises

Notebooks live in their own repository
The notebooks and case-study exercises are in atrium-school-ml-lessons, so you can clone them, open them in Colab, and keep working after the school without dragging this website along. The slides stay here.

Most datasets are not in the repository
Where each dataset actually lives is recorded in datasets.yml.

NoteThree ways to run the notebooks

Every notebook starts with the same setup cell. It works out where it is running, fetches the repository if it has to, installs only what is missing, and puts the caches somewhere sensible. You do not have to edit anything to switch platforms.

Google Colab

Nothing to install. Click the Open in Colab badge on a notebook below, then run the first cell — it clones this repository into the session for you.

WarningSave a copy before you edit

A notebook opened from GitHub does not keep your changes. Use File → Save a copy in Drive first, and work in that copy.

Colab wipes its disk when the runtime is recycled, so model weights are re-downloaded in a new session. To keep them, change the last line of the setup cell, before you run anything else, to:

setup("torch", "transformers", drive=True)

and approve the Google Drive prompt. The caches then live in MyDrive/atrium-school.

School JupyterHub

Nothing to install, and nothing to spawn. We will give you a URL on the first morning; open it and you are in JupyterLab with the server already running.

Everyone’s server shares the same home directory. In the file browser you will find:

_atrium-school-ml-lessons/   the notebooks — we keep this up to date, do not edit here
_atrium-data/                the datasets, shared by everyone
<your name>/                 your own folder: work here

Copy the notebook you need from the day’s folder in _atrium-school-ml-lessons/ into your own folder (right-click → Copy, then Paste in your folder), open the copy, and run the first cell as usual. It finds the shared repository and installs nothing — the packages, model weights and data are already in place.

WarningCaution

Your work is saved, the machine is not
Your folder persists, so notebooks you edit and files you save are still there tomorrow. The server itself may be restarted between sessions; if a notebook stops responding, Kernel → Restart Kernel and re-run from the top.

Shared means shared
Everyone can see — and delete — everything in the home directory. Stay in your own folder.

Something does not look right?
If the URL asks you to choose a server configuration, or you land somewhere that looks nothing like the above, tell an instructor — do not guess.

Your own machine

Follow the setup guide, then start Jupyter (or open the notebook in your IDE) anywhere inside the clone:

git clone https://github.com/arubrno/atrium-school-ml-lessons.git
cd atrium-school-ml-lessons
python -m venv .venv && source .venv/bin/activate   # .venv\Scripts\activate on Windows
pip install -r requirements.txt \
    --index-url https://download.pytorch.org/whl/cpu \
    --extra-index-url https://pypi.org/simple

The setup cell finds the repository around it and installs nothing further.

ImportantWork on a copy

Copy a notebook before you edit it (cp clip_zero_shot.ipynb my-clip.ipynb). We will push updates during the week, and git pull on a notebook you have changed produces a merge conflict in raw JSON that nobody enjoys resolving. On the hub, copy it into your own folder; on Colab, Save a copy in Drive.

Monday

1. Introduction to computer vision — What computer vision is, the family of tasks, discriminative vs. generative models, a live zero-shot demo with CLIP, and three rounds of group discussion.
2. CLIP zero-shot classification — the notebook behind the demo. Classifying artefact photographs with no training and no annotation, and three ways it misleads you.
   • New to Python? Use the beginner version — the same notebook with the code spelled out step by step and commented line by line, plus an optional look inside the model.
3. Coding with agents
4. Python for computer vision — enough Python to read, change and debug the code an AI tool writes for you, then NumPy, Matplotlib and OpenCV. It ends with a first computer-vision workflow that separates the Venus of Dolní Věstonice from its background and measures it.
   • Introduction to Python for computer vision — why learn Python when AI writes code, and how to turn a task into an algorithm before writing any.
   • Python fundamentals for computer vision — the hands-on part. The code cells are empty and you fill them in with the instructor: notebooks, Python basics, NumPy, Matplotlib, and OpenCV on the Venus photograph.
   • Fallen behind? Use the filled-in version — the same notebook with every cell filled in, for catching up or checking your version.

Ice-breaker event in the evening.

Tuesday

1. Artefacts: datasets and annotation — Where the AMČR-PAS artefact photographs come from, what “annotated” actually means, the COCO, YOLO and Pascal VOC formats, a live demo of CVAT and SAM 2, and the three problems the dataset has: class imbalance, vocabularies that do not line up, and the metadata nobody has time to record.
   • Hands-on annotation in CVAT — we will give you its address in the session. Everything after the hands-on block is a reading of what you annotate there.
   • ArchaeoTag — the guessing game. Bring a phone.

Participants’ project presentations in the afternoon; walking meeting in Brno in the evening.

Name	Topic	Time
Zabya Abo Aljadayel	Reconstructing Palaeoenvironmental Change and Climatic Variability in Southwestern Portugal over the Last 2.500 years: A Multi-Proxy Geoarchaeological Study of Miróbriga and Barbaroxa de Baixo Wetland	10 min
Gianmarco Borgi	CNN for hyperspectral reconstruction	10 min
Dionysios Danelatos	Approaching Stylistic Variation in Upper Palaeolithic Zoomorphic Figures Using Artificial Intelligence & Geometric Morphometrics	15 min
Cherene De Bruyn	The Fast and the Shallow: Drone images for clandestine grave location	15 min
Ákos Ekrik	CNN for Stone Carvings	10 min
Konstantina Eleftheriadi	Leveraging Monocular Depth Estimation for the 3D Digitization of Cultural Heritage Artifacts (GitHub, Colab)	10 min
Hüseyin Erdoğan	Rupe Magna petroglyphs (Zenodo)	10 min
Agata Maria Gaszka	Exploring Bipolar Technology through Deep Learning Methods: Biomechanical Analysis and Artefact Classification	20 min
Cheng Liu	Human Stone Toolmaking Action Grammar (HSTAG): A Challenging Benchmark for Fine-grained Motor Behavior Recognition	15 min

Thursday

1. Satellite imagery: cropmarks in Sentinel-2 — buried ditches and walls change how crops grow above them, and that shows from space. Using Sentinel-2 imagery of the Bronze Age fortification at Cornești-Iarcuri (Romania), you learn what spectral bands and NDVI are, make true- and false-colour composites, bring out the cropmark with background removal and a ridge filter, train a Random Forest on hand-placed points and retrain it with hard negatives, then export the result as polygons for GIS.
   • Cropmarks in Sentinel-2 — the hands-on notebook. The code cells are empty and you fill them in with the instructor.
   • Fallen behind? Use the filled-in version — the same notebook with every cell filled in, for catching up or checking your version.
2. Rock art case study (PDF)

NoteReading

The MAIA COST Action Zotero library is an open, shared bibliography on AI in archaeology. Most examples used in the slides come from there.