dirmeier / ssnl

Simulation-based inference using SSNL

SSNL

This repository contains code to run the experiments from Simulation-based inference using surjective sequential neural likelihood estimation.

Installation

Install miniconda and create an environment using the environment file environment.yaml. This should install all required dependencies.

Usage

To run an experiment load the environment and then executre>

python main.py  \
  --outdir=results/ \
  --mode=fit \
  --config=configs/slcp/surjection.py \
  --config.training.n_rounds=${n_rounds} \
  --config.rng_seq_key=${key}

This generates a file that contains trained neural network parameters. To get posterior samples, run

python main.py  \
  --outdir=results/ \
  --mode=eval \
  --checkpoint=results/slcp-params.pkl \
  --config=configs/slcp/surjection.py \
  --round=${rounds} \
  --config.rng_seq_key=${key}

To compute the MMD between the true and the approximate posterior samples execute:

python compute_mmd.py \
  results/slcp \
  results/slcp/slcp-nuts-exact-posteriors.pkl \
  results/slcp/slcp-df.pkl