Fitting

After downloading/simulating data, creating a transient object, specifying a model, and creating a prior we now come to the exciting part; Fitting the model to data!

To fit our model to data we have to specify a sampler and sampler settings. The likelihood is set by default depending on the transient/data but one can use a different one or write their own as explained in the likelihood documentation.

Installing redback with minimal requirements will install the default sampler dynesty. Installing optional requirements will also install nestle. We generally find dynesty to be more reliable/robust but nestle is much faster. We note that dynesty has checkpointing, as do many other samplers.

Samplers

As we use bilby under the hood, we have access to several different samplers. Cross-checking results with different samplers is often a great way to ensure results are robust so we encourage users to install multiple samplers and fit with different samplers.

Nested samplers

  • Dynesty:

  • Nestle

  • CPNest

  • PyMultiNest

  • PyPolyChord

  • UltraNest

  • DNest4

  • Nessai

MCMC samplers

  • bilby-mcmc

  • emcee

  • ptemcee

  • pymc3

  • zeus

A full up to date list of samplers can be found in the bilby documentation. This page also provides guidance on how to install these samplers, while the bilby API provides information on the sampler settings for each sampler.

In redback, having created a transient object, specified a model, priors, fitting is done in a single line.

result = redback.fit_model(name='GRB', model=model, sampler='dynesty',
                           nlive=200, transient=afterglow, prior=priors,
                           data_mode='luminosity', **kwargs)

Here

  • name: GRB is a string/name of transient fitting, this is used to .

  • model: is a string referring to a function implemented in redback. Or a function the user has implemented.

  • sampler: is a string referring to the sampler. It could be a string referring to any name of a sampler implemented in bilby.

  • nlive: is the number of live points to sample with. Higher = better. Typically we would use nlive=1000/2000 but this depends on the sampler.

  • transient: the transient object

  • prior: the prior object

  • data_mode: type of data to fit.

  • kwargs: Additional keyword arguments to pass to fit_grb, such as the likelihood, or things required by the sampler, label of the result file, directory where results are saved to etc.

We note that some samplers have multiprocessing, which you can see how to use here. We will soon implement some GPU models and parallel bilby functionality for more rapid inference workflows.