Flow-Matching-Posterior-Estimation

This is the companion repository for the NeurIPS-2023 paper Flow Matching for Scalable Simulation Based Inference, which introduces a simulation-based inference (SBI) method based on flow matching, called FMPE. It contains code to run the experiments for an SBI benchmark and gravitational wave inference as a challenging real world example.

This repository builds on the dingo package for the implementation of the base methods for flow matching and the pipeline for the GW experiment.

Set up

First, clone the FMPE branch of the dingo repository.

git clone --branch FMPE https://github.com/dingo-gw/dingo.git

Next, create a new virtual environment.

python3 -m venv fmpe-venv
source fmpe-venv/bin/activate

Install dingo in this environment.

cd dingo
pip install -e ."[dev]"

Note that while dingo can in general be installed from PyPI with pip, this manual installation is required as the FMPE code is not yet contained in the main branch.

Finally, install the sbibm package for the benchmark experiments.

pip install sbibm

Experiments

SBI Benchmark

Training and evaluation scripts available in ./sbi-benchmark. To train an FMPE model, run

python run_sbibm.py --train_dir </path/to/train_dir>

where the training directory contains a settings.yaml file. Example settings can be found in ./sbi-benchmark/settings.yaml.

Gravitational Wave inference

We here provide a full inference example for the real GW event GW150914. This builds on the full dingo pipeline, although it applies some simplifications (e.g., no full noise conditioning).

This example follows the steps from the dingo NPE tutorial. To run this, navigate to the ./gravitational_waves directory and follow the instructions below.

1. Generate a waveform dataset

We first generate a dataset of GW polarizations, which is later used to train the model.

dingo_generate_dataset --settings datasets/waveform_dataset_settings.yaml --out_file 
datasets/waveform_dataset.hdf5 --num_processes 8

The number of processes should be adjusted to the number of available CPU cores.

2. Generate an ASD dataset

To train Dingo, one needs to provide an estimate for the detector noise amplitude spectral density (ASD). For simplicity, we provide a dataset in ./datasets/asd_dataset_GW150914.hdf5, so this step can be skipped.

3. Train the Dingo network

Once the datasets are available, we train the FMPE network.

dingo_train --settings_file training/train_settings.yaml --train_dir training

Depending on the available resources some settings may need to be adjusted. This includes

• model/... for the size of the model,
• local/num_workers for number of workers (set this to number of available CPU cores -1; dingo requires some heavy preprocessing, so if this number is too low, it may become the bottleneck),
• local/device for the accelerator (typically cuda, cpu is also possible, but extremely slow) and
• training/stage_0/batch_size for the batch size (reduce this if you run out of memory).

Please also refer to the documentation: https://dingo-gw.readthedocs.io.

References

@inproceedings{wildberger2023flow,
    title={Flow Matching for Scalable Simulation-Based Inference},
    author={Jonas Bernhard Wildberger and Maximilian Dax and Simon Buchholz and Stephen R Green and Jakob H. Macke and Bernhard Sch{\"o}lkopf},
    booktitle={Thirty-seventh Conference on Neural Information Processing Systems},
    year={2023},
    url={https://openreview.net/forum?id=D2cS6SoYlP},
    eprint={2305.17161},
    archivePrefix={arXiv},
    primaryClass={cs.LG},
}

If you use the dingo code for the GW example, please also cite the following (and refer to the Reference section of the dingo README for other optional citations).

@article{Dax:2021tsq,
    author = {Dax, Maximilian and Green, Stephen R. and Gair, Jonathan and Macke, Jakob H. and Buonanno, Alessandra and Sch\"olkopf, Bernhard},
    title = "{Real-Time Gravitational Wave Science with Neural Posterior Estimation}",
    eprint = "2106.12594",
    archivePrefix = "arXiv",
    primaryClass = "gr-qc",
    reportNumber = "LIGO-P2100223",
    doi = "10.1103/PhysRevLett.127.241103",
    journal = "Phys. Rev. Lett.",
    volume = "127",
    number = "24",
    pages = "241103",
    year = "2021"
}