Normalizing flows

Beta-VAE experiments with normalizing flows for variational inference and generative modeling.

Project layout

• src/normalizing_flows contains reusable statistical utilities
• scripts contains the canonical Python experiment scripts
• notebooks contains archived notebook versions of the experiments
• tests contains fast numerical and repository structure checks
• requirements-legacy.txt preserves the original TensorFlow 2.5 environment

Archived notebooks are retained for provenance. Run the scripts when a repeatable Python entry point is preferred.

Setup

The supported package manager is uv.

uv sync

The default environment installs the reusable metrics package and development tools. The legacy TensorFlow experiments require the pinned environment in requirements-legacy.txt and may require an x86 Python environment because TensorFlow 2.5 does not provide Apple Silicon wheels.

Experiments

The scripts cover noisy moons, thoracic surgery preprocessing, and Beta-VAE experiments using NICE, RealNVP, MAF, and IAF flows. Dataset paths are kept in the experiment scripts and must be updated for a local dataset checkout.

uv run python scripts/beta-vae-iaf-noisy-moons.py

Experiments can be slow and may require TensorFlow, TensorFlow Probability, datasets, and a graphical backend. They are not part of the default test run.

Quality checks

uv run ruff format --check .
uv run ruff check .
uv run pytest

Tests cover the reusable two-dimensional KS and energy statistics plus script syntax and notebook-to-script parity. Full model training and notebook execution remain manual validation steps.

References

• Variational inference with normalizing flows
• Normalizing flows an introduction and review
• TensorFlow Probability
• Noisy moons
• Thoracic surgery dataset