- Jupyter Notebook 86.4%
- Python 13.3%
- JavaScript 0.2%
Feiko
9119b9169e
Nothing in the course trained anything substantial: the heaviest notebook was the Fashion-MNIST CNN at 2m15s, and everything else finished in seconds. This one takes seven minutes on an RTX 5070 and is the only notebook that uses a GPU. It also fills a real gap. The transformers appendix has a slide on pretraining and fine-tuning; nothing demonstrated it. Every deep learning notebook so far either trained something small from scratch or used a pretrained model frozen. The task is a Yelp review's star rating, 1 to 5, from its text - deliberately harder than Exercise 10's one-star-against-five, and ordered, so mean absolute error in stars is reported alongside accuracy. Three routes on one split: baseline (majority) 0.352 acc 0.890 MAE TF-IDF + logistic regression 0.527 0.617 2.4s frozen embeddings + logreg 0.510 0.629 15.7s DistilBERT, fine-tuned 0.603 0.447 386.3s The middle rung is worse than the first, which is the notebook's real subject. A frozen general-purpose sentence encoder is optimised for semantic similarity, not rating intensity, and 384 dimensions throws away the specific words that carry the signal. Using a big model is not the same as adapting one. The negation probes end on a limit rather than a win. TF-IDF rates "I would not say this place is good" at five stars and the fine-tuned model rates it one - but all three models still misread "Not the worst meal I have ever had". The notebook says so instead of stopping at the good example. Runs on CPU too: it detects the absence of CUDA, drops to 2,000 reviews and one epoch, and reports that it did. metadata.execution.timeout is raised to 5400s because nbmake caps cells at 300 by default. The GPU build of PyTorch is deliberately NOT added to pyproject or uv.lock - students keep the CPU build that works everywhere. Installing CUDA is a documented opt-in in demos/showcases/README.md, along with the two traps: `uv run` re-syncs and reverts torch, and TensorFlow has had no native-Windows GPU support since 2.11. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> |
||
|---|---|---|
| .forgejo/workflows | ||
| articles | ||
| books | ||
| build | ||
| demos | ||
| exercises | ||
| original-deck | ||
| src/slides | ||
| tutorials | ||
| .gitattributes | ||
| .gitignore | ||
| conftest.py | ||
| pyproject.toml | ||
| README.md | ||
| uv.lock | ||
DSC700 : Machine Learning with Python
Four-day SpiralTrain course. Classical machine learning with scikit-learn, from the Estimator API through to a first neural network.
For students
Two commands and you are running:
uv sync --all-groups
uv run jupyter lab
Full instructions, including installing uv itself, are in SETUP.md.
| Where | What |
|---|---|
DSC700-Machine-Learning-with-Python.pdf |
The slides |
exercises/ |
Eleven exercises with reference solutions, starter notebooks and data. Start at the index |
demos/ |
Worked examples per topic, shown in class |
demos/showcases/ |
Sixteen longer worked examples on real external data. Index |
demos/apps/ |
Two interactive apps for live parameter fiddling. Index |
books/ |
The Python Data Science Handbook, in full |
articles/, tutorials/ |
Background reading |
For the trainer
Slides
Built from Marp Markdown in src/slides/ through the shared
Presentation-Template toolchain.
bun install --cwd ../Presentation-Template/build # once per machine
bun run --cwd build build # produce the PDF
bun run --cwd build serve # live preview, press P for presenter mode
bun run --cwd build lint # check slides fit 16:10
One module per directory, NN-topic/NN-topic.md, plus 00-preamble. Shared
images live in src/slides/images/ and are referenced as ../images/x.png;
the build inlines them as data URIs. The course schedule slides are generated,
so do not write them by hand.
Modules 01–10 are the four scheduled days; 11-more-models and
12-transformers carry appendix: true and are not scheduled.
The figures for the newer chapters are generated from the same code the demos run, rather than exported from the old PowerPoint:
uv run python src/slides/make_figures.py # all of them
uv run python src/slides/make_figures.py clustering # one group
They land in src/slides/images/gen_*.png. The gen_ prefix keeps them
separate from the inherited exports, which must not be overwritten. Rerun after
changing a demo whose result appears on a slide.
bun run --cwd build lint enforces the house style: no bold, no em dashes, no
colon left dangling at the end of a wrapped line, and nothing taller than
800px. Run it before building.
Notebooks
uv run pytest --nbmake -n 8 demos/ exercises/ # all of them
uv run pytest --nbmake -n 8 --offline demos/ exercises/ # skip downloads
Three showcase notebooks download on first run and are worth warming up before you teach: the CNN pulls Fashion-MNIST (~30 MB), the embeddings demo pulls a sentence model (~90 MB), and the fine-tuning demo pulls DistilBERT (~250 MB). All cache afterwards.
--skip-pretrained deselects the three transformer notebooks, which is the
fast path: the suite drops from about thirteen minutes to six. The fine-tuning
notebook alone is seven minutes on a GPU, and it is the only notebook that
wants one — see the showcase index for the CUDA
install.
uv run python demos/data/fetch_datasets.py # rebuild the showcase CSVs
That one is trainer-only and needs KAGGLE_USERNAME and KAGGLE_KEY in the
environment for five of its nine sources; the CSVs are committed, so students
never run it. quakes.csv is a live USGS feed, so rerunning it changes which
earthquakes the DBSCAN demo clusters.
CI runs the full execution weekly and on demand, so library releases surface
there rather than in a classroom; a push only gets the fast check that no
notebook carries committed outputs. Before teaching, trigger the Notebooks
workflow by hand rather than trusting the last push. Notebooks are committed
without outputs. uv run nbstripout --install sets up the filter.
The student bundle
bun run --cwd build build # slides PDF
bun run --cwd build exercises # exercises PDF + HTML
bash build/make-student-zip.sh # -> student.zip
Compiled decks, notebooks and reading material; no slide sources, no build
wrapper, no trainer-only files. The Build and Release PDFs workflow rebuilds
it on every push to main and republishes it, alongside both PDFs, as the
latest release, the same shape the Go and C# courses use.
The original deck
original-deck/ holds the inherited PowerPoint and its PDF export, untouched.
The Marp deck replaces it, but if a rebuild ever looks wrong in front of a
class, that is the version this course was taught from before the conversion.
Layout
src/slides/ Marp source, one directory per module
make_figures.py generates the gen_*.png figures from the demo code
exercises/ student workbook, solutions and case studies
demos/ per-topic notebooks used in class
plotstyle.py shared palette and cluster-plot helpers
data/ the showcase datasets, plus the script that rebuilds them
showcases/ longer worked examples on real external data
apps/ marimo and Streamlit apps for live demonstration
books/ articles/ tutorials/ reference material
original-deck/ the inherited PowerPoint, kept as a fallback
build/ thin wrapper over ../Presentation-Template