Narrow-Parameter Problem in NPE + Importance Sampling

Reproduction code for the experiments from our study of Dingo-IS (Dax et al. 2023).
Compares forward KL, reverse KL, alpha-divergence, annealed training, and synthetic marginalization on 2D and 3D toy posteriors.

Setup

git clone <repo>
cd narrow_posterior_IS
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

The code automatically uses CUDA if available.

Run

# 2D experiments (less than 10min on T4 GPU)
python run_experiments.py 2d

# 3D experiments (slightly more)
python run_experiments.py 3d

# Options
python run_experiments.py 2d --out_dir figures --n_epochs 400 --n_is 20000 --seed 42

Figures are saved as PDF/PNG to figures/ (or --out_dir).

Structure

├── models/              # Real-NVP normalizing flow
├── posteriors/          # 2D and 3D toy posteriors
├── training/            # fKL, rKL, alpha-div, annealed objectives
├── inference/           # IS utilities + synthetic marginalization
├── plotting/            # all figure generation
├── report/              # The 5-page PDF report
├── run_experiments.py   # main script
└── everything.ipynb     # detailed and commented notebook