gwflow

Normalizing flows for gravitational-wave analyses.

gwflow builds on zuko and implements normalizing flows with problem-specific modifications.

Installation

gwflow is currently only available from source:

pip install git+git@github.com:mj-will/gwflow.git

Flows

GWCalFlow

This flow is designed for use in analyses where there are large number of (simple) calibration parameters with approximately Gaussian distributions.

It approximates the distribution over the calibration parameters separately to the main GW parameters. It supports two options for handling the calibration parameters:

• calibration_model='gaussian': fit a Gaussian with a diagonal covariance matrix
• calibration_model='nn': fit a neural network that estimates the per-parameter mean and variance for of a diagonal Gaussian

Usage

GWCalFlow supports three different ways of specifying which parameters are 'GW' parameters and which are calibration parameters.

We provided a factory class than can be used to instantiate the flow using various different methods:

from gwflow import GWClaFlow

Then, assuming data lies in an n-dimensional space with parameters that described the signal (GW parameters) and calibration parameters. The objects returned in each support the same functionality as standard flows from zuko.

In all cases, additional keyword arguments can be passed when instantiating the class and these will be used to configure e.g. the normalizing flow.

Parameters names

Using a list of parameter names and regex:

flow = GWCalFlow(
    parameters=["chirp_mass", "mass_ratio", ..., "recalib_H1_amplitude_0", ...],
    calib_regex=".*calib.*",
)

Indices

Using the indices for GW and calibration parameters

flow = GWCalFlow(
    gw_idx=[0, 1, 3],
    cal_idx=[2, 5]
)

Slices

Using slices, assuming the GW parameters are first

flow = GWCalFlow(
    gw_dim=15,
    cal_dim=20,
)

Basic usage

Flows are used in the same way as zuko flows:

flow = GWFlowClas(gw_dim=15, cal_dim=20, context=4, transforms=3)

x = flow().sample((10,))

log_prob = flow().log_prob(x)

# If the flow is conditional
x = flow(c).sample((10,))

For more details, see the zuko documentation

Note: since the transform only applies to a subset of the parameter space, the flow does not expose this attribute.

Citation

If you use gwflow in your work, please cite our DOI (to be added) and zuko.