Normalizing Flows

Code accompanying mariogemoll.com/normalizing-flows.

A normalizing flow trained on the two-moons dataset, with interactive browser visualizations that show the change-of-variables principle, the individual flow layers, training, and sample generation.

Layout

• py/ — PyTorch reference implementation and training notebook (normalizing-flows.ipynb).
• ts/ — TypeScript / TensorFlow.js port that powers the interactive widgets on the page. Entry point: ts/src/page.ts, bundled to ts/dist/page.js via esbuild. Trained weights live in ts/model.json / ts/model.weights.bin.
• sh/checkpy.sh — lint the Python code with ruff and verify the notebook.

Python

cd py
uv sync
uv run jupyter lab normalizing-flows.ipynb

TypeScript

cd ts
npm install
npm run build

Then open ts/page.html in a browser (WebGPU-capable recommended).