Scikit-Learner 📈

A web-based machine learning application for training and comparing regression and classification models. This runs scikit-learn directly in the user's browser via Pyodide, so the whole app deploys as a static website.

Features

• 27 Regression Models across 6 categories
• 22 Classification Models across 6 categories
• Interactive Plotly visualizations — scatter, residuals, predicted vs actual, ROC, confusion matrix, comparison bar chart
• Cross-Validation (3 / 5 / 10 folds)
• Sample Datasets — Iris, Wine, Breast Cancer, Digits (classification); Diabetes, Boston-synthetic, Airfoil, Synthetic (regression)
• Model Export — joblib bytes, single-file or zipped bundle

How it works

┌───────────────────────────────────────────────────────────┐
│ Browser │
│ ┌─────────────────────────────────────────────────────┐ │
│ │ index.html + Bootstrap + Plotly │ │
│ │ ↓ pyCall('train', [...]) │ │
│ │ ┌────────────────────────────────────────────────┐ │ │
│ │ │ pyodide-bridge.js │ │ │
│ │ │ • loads Pyodide from JSDelivr CDN │ │ │
│ │ │ • installs scikit-learn / pandas / numpy / │ │ │
│ │ │ scipy / joblib │ │ │
│ │ │ • runs frontend/py/learner.py inside Pyodide │ │ │
│ │ │ • thin pyCall / pyCallBinary wrappers │ │ │
│ │ └────────────────────────────────────────────────┘ │ │
│ └─────────────────────────────────────────────────────┘ │
└───────────────────────────────────────────────────────────┘
(no network calls after first load)

First load: ~10 s (downloads Pyodide runtime + sklearn wheel, ~15 MB total). Subsequent loads: ~1 s thanks to browser cache.

Deploy

Any static host works. Example with Python's stdlib server:

python3 -m http.server -d frontend 8080
open http://localhost:8080/

Local development

No backend to start. Edit any file under frontend/, reload the browser.

If you change frontend/py/learner.py, the browser fetches it fresh on reload — but Pyodide doesn't pick up the change until the module is re-imported. Hard-reload (Cmd-Shift-R / Ctrl-F5) or open a new tab.

Testing

A Playwright end-to-end spec covers Pyodide bootstrap, sample loading, training, predictions, export, and the UI scatter-plot render — 8 assertions, runs against either a local python -m http.server -d frontend or the public URL.

Caveats (WASM)

• Pyodide initial load adds ~10 s and ~15 MB of one-time download. Loading overlay covers it.
• CSV upload capped at 20 MB (Pyodide's WASM heap).
• The airfoil dataset is bundled as frontend/data/airfoil.csv because Pyodide can't reach fetch_openml from inside the browser.
• Boston-housing uses the synthetic generator (real Boston was removed from sklearn ≥1.2).

License

BSD