Stacking for Simulation-Based Inference

Code to reproduce the stacking experiments of the paper Simulation based stacking (Y. Yao, B. Régaldo-Saint Blancard & J. Domke).

Install

For this work, we worked with a Python 3.10.10 environment and PyTorch v2.0.1. First follow PyTorch installation instructions here.

Then, within the relevant environment, run:

pip install wandb
pip install https://github.com/eventlet/eventlet/archive/master.zip
pip install sbi
pip install sbibm
pip install altair==v4.2.2
pip install -e .

To use the SIR model, also follow this installation procedure.

General overview

We do not provide here the neural posterior inferences used in the paper or the different quantities we can extract from them (log probabilities, intervals, ranks, etc). However, we provide the necessary material to recompute this on the sbibm examples.

For neural posterior estimation, create a WandB sweep configuration file as in scripts/sweeps_yaml/ and run the sweep using scripts/sweep_scheduler.ipynb.

Then, for each neural posterior estimation, estimate statistics using compute_stats function from sbi_stacking/stats.py file.

Once these statistics are computed, stacking experiments can be conducted using the functions of the sbi_stacking/stacking.py file. Notebooks from nb/ directory allow to reproduce the stacking experiments of the paper.