Truncated proposals for scalable and hassle-free simulation-based inference

Code to reproduce results from Truncated proposals for scalable and hassle-free simulation-based inference by Michael Deistler, Pedro J. Goncalves*, Jakob H. Macke*, accepted at NeurIPS 2022.

Structure of this repository

This repo contains four main folders. Each folder contains an individual README.md file which describes how the particular results can be reproduced:

• The folder benchmark contains code to run the benchmark tasks (Fig. 4)
• The folder l5pc contains code to produce the results for the pyloric network (Fig. 5) and for the L5PC (Fig. 6).
• The folder sbi is a modified version of the sbi repository which implements the truncation of the proposal.
• The folder paper contains all code to assemble the figures shown in the paper from the produced results. It also contains the code to reproduce toy examples and illustrations (Fig. 1, Fig. 3)

Completeness

This repository contains code that can be run to generate all results. In addition, it contains some files that are loaded in the notebooks in order to generate the figures (these files are also generated along the way when running the code). However, in order to maintain a reasonable file size of this repo, many intermediate results (which are not shown in the paper, e.g. trained neural networks after every round) are not stored in this repository. In addition, this repository does not include the full git history of the project. In order to recover these things, we point the reader to the development repository of this project.

Installation

First, create the conda environment and activate it:

conda env create --file environment_vm.yml
conda activate tsnpe_neurips

Then, each subfolder is to be installed separately. Note that, in the second line, you will get an error saying that sbibm 1.0.7 requires sbi==0.17.2, but you have sbi 0.18.0 which is incompatible. You can savely ignore this error.

cd benchmark; pip install -e .; cd ..
cd sbi; pip install -e .; cd ..
cd l5pc; pip install -e .; cd ..
cd paper; pip install -e .; cd ..

Installation of pyloric network simulator

In order to reproduce the results in Fig. 5 (pyloric network), you have to install the pyloric repo. Note: original results were generated with commit hash 1768cc4365d2ca24196562a4cf41a2b28cc6e647.

git clone git@github.com:mackelab/pyloric.git
cd pyloric
pip install -e .

Installation of L5PC simulator

Finally, in order to reproduce the results shown in Fig. 6, you have to set up the simulator for the layer 5 pyramidal cell (L5PC, Fig. 6). This simulator is written in Neuron. In order to compile the model, you have to:

cd l5pc/l5pc/model/x86_64
rm *.o; rm *.c; cd ..
nrnivmodl mechanisms

Large file storage

This repo is significantly smaller than the development repo. It still requires neural networks that are used to visualize results and a few datasets that are plotted, but these are no longer stored in this git repository. Instead, they are hosted on the Hugging Face dataset mackelab/tsnpe_neurips.

To reproduce the contents of l5pc/results/ locally, install the hf CLI (pip install -U "huggingface_hub[cli]") and run:

hf download mackelab/tsnpe_neurips --repo-type dataset --local-dir l5pc/results

Citation

@inproceedings{
  deistler2022truncated,
  title={Truncated proposals for scalable and hassle-free simulation-based inference},
  author={Michael Deistler and Pedro J. Goncalves and Jakob H. Macke},
  booktitle={Thirty-Sixth Conference on Neural Information Processing Systems},
  year={2022},
  url={https://openreview.net/forum?id=QW98XBAqNRa}
}

Contact

If you have questions, please reach out to michael.deistler@uni-tuebingen.de.