Source code for "Neural Methods for Amortised Statistical Inference"

This repository contains code for reproducing the results in the review paper "Neural Methods for Amortised Statistical Inference" by Andrew Zammit-Mangion, Matthew Sainsbury-Dale, and Raphaël Huser.

The code in this repository is made available primarily for reproducibility purposes. Readers seeking to implement neural methods for amortised statistical inference should also explore the cited software packages and their documentation. In particular, the software packages:

1. BayesFlow (Python)
2. LAMPE (Python)
3. NeuralEstimators (Julia and R)
4. sbi (Python)
5. swyft (Python)

Note that each of these packages can be interfaced from R using reticulate.

Software dependencies

We suggest that users set up a conda environment, so that the dependencies of this repository do not affect the user's current installation. In your environment, install Python, R, and Julia. An environment instruction that worked in our case was the following:

conda create -n ARSIA -c conda-forge julia=1.9.4 r-base nlopt python=3.9 tensorflow

Please ensure the R version installed is >=4.4.0 and the python version installed is >=3.9. If you do not wish to use a conda environment, then install the software directly from the following websites:

• Install Julia 1.9.4.
• Install R >= 4.4.0.
• Install Python >= 3.10.0.

Once Julia, Python and R are setup, install the Julia and R package dependencies by running the following commands from the top-level of the repository:

julia --project=. -e 'using Pkg; Pkg.instantiate(); Pkg.precompile()'

Rscript dependencies_install.R

pip install bayesflow

After installing the Julia packages you might need to load Julia and run the commands

import Pkg; Pkg.add("cuDNN")
import Pkg; Pkg.add("Flux")

Many of the R scripts call Python. Currently many R scripts use reticulate and point to the correct Python environment through the function use_condaenv(Sys.getenv("CONDA_PREFIX")). This function loads the currently-activated conda environment. If you need to modify this code please do this in src/1_Generate_GP_Data.R, src/3_fKL.R, src/4_rKL.R, src/5_rKL_Synthetic_Naive.R, and src/6_rKL_Synthetic_MutualInf.R.

Hardware requirements

In general, the fast training of neural networks requires GPUs. However, the code in this repository also runs on CPUs in a moderate amount of time. Therefore, there are no major hardware requirements for running this code.

Reproducing the results

First, download this repository and navigate to its top-level directory within terminal.

The repository is organised into folders containing source code (src), intermediate objects generated from the source code (output), and figures (fig). Checkpoints that contain pre-trained networks are available (ckpts).

The replication script is run.sh, invoked using bash run.sh from the top level of this repository. The replication script will automatically train the neural networks, generate estimates/samples from both the neural and likelihood-based estimators/samplers, and populate the fig folder with the figures and results of the manuscript.

Note that the nature of our experiments means that the run time for reproducing the results of the manuscript can be moderate (on the order of several hours).

Minor reproducibility difficulties

When training neural networks, there is often unavoidable non-determinism: see, for example, here. In our reproducible code, this does not significantly affect the "story" of the final results in the sense that each method performs similarly well in each run, but there may be some slight numerical differences each time the code is executed.