jimgw 0.2.0

Gravitatioanl wave data analysis tool in Jax

Jim - A JAX-based gravitational-wave inference toolkit

Jim comprises a set of tools for estimating parameters of gravitational-wave sources thorugh Bayesian inference. At its core, Jim relies on the JAX-based sampler flowMC, which leverages normalizing flows to enhance the convergence of a gradient-based MCMC sampler.

Since its based on JAX, Jim can also leverage hardware acceleration to achieve significant speedups on GPUs. Jim also takes advantage of likelihood-heterodyining, (Cornish 2010, Cornish 2021) to compute the gravitational-wave likelihood more efficiently.

See the accompanying paper, Wong, Isi, Edwards (2023) for details.

[!WARNING]
Jim is under heavy development, so API is constantly changing. Use at your own risk! One way to mitigate this inconvience is to make your own fork over a version for now. We expect to hit a stable version this year. Stay tuned.

[Documentatation and examples are a work in progress]

Installation

You may install the latest released version of Jim through pip by doing

pip install jimGW

You may install the bleeding edge version by cloning this repo, or doing

pip install git+https://github.com/kazewong/jim

If you would like to take advantage of CUDA, you will additionally need to install a specific version of JAX by doing

pip install --upgrade "jax[cuda12_pip]" -f https://storage.googleapis.com/jax-releases/jax_cuda_releases.html

NOTE: Jim is only currently compatible with Python 3.10.

Performance

The performance of Jim will vary depending on the hardware available. Under optimal conditions, the CUDA installation can achieve parameter estimation in ~1 min on an Nvidia A100 GPU for a binary neutron star (see paper for details). If a GPU is not available, JAX will fall back on CPUs, and you will see a message like this on execution:

No GPU/TPU found, falling back to CPU.

Directory

Parameter estimation examples are in example/ParameterEstimation.

Attribution

Please cite the accompanying paper, Wong, Isi, Edwards (2023).