Multi-β VAE Benchmark Demo

Status: Demo-only benchmark walkthrough Device: CPU-compatible

This walkthrough keeps the retained multi-β VAE benchmark example, but the shipped contract is explicit: it runs a mock-model demo through the retained suite with demo-mode image metrics. It is useful for understanding the API shape, not for claiming a benchmark-grade VAE evaluation stack out of the box.

Files

• Python script: multi_beta_vae_benchmark_demo.py
• Jupyter notebook: multi_beta_vae_benchmark_demo.ipynb

Run It

python examples/generative_models/vae/multi_beta_vae_benchmark_demo.py
jupyter lab examples/generative_models/vae/multi_beta_vae_benchmark_demo.ipynb

What This Demo Actually Uses

• MultiBetaVAEBenchmarkSuite(..., demo_mode=True)
• retained disentanglement metrics plus demo-mode FIDMetric and LPIPSMetric
• a checked-in mock MultiBetaVAE implementation that exposes the expected benchmark interface
• small CelebA-style batches for a quick local walkthrough

Why It Is Demo-Only

• the shipped suite relies on retained mock perceptual/image-quality backends in demo mode
• the example model is a pedagogical mock, not a benchmark-ready trained VAE
• the public benchmark surface is Python-first and intentionally narrower than the historical docs implied

Use This When

Use this pair when you want to understand the retained suite inputs and outputs, compare mock β-VAE quality tiers, or swap in your own model implementation while keeping the demo-only expectations explicit.