Add a zero-likelihood mode
Adds a test-mode for running with zero likelihood.
- Adds a ZeroLikelihood which can be instantiated from any other likelihood as a drop-in replacement
- To use on any example, run with
--bilby-zero-likelihood-modein the command line - A test has been added which runs a 15d CBC problem and checks the 1D posteriors agree with the priors.
- The next is added to the C.I nightly test
Merge request reports
Activity
@vivien and @cjhaster I recall this being raised by one of you on the review call this morning.
What exactly should we do here. Is it sufficient just to run the standard 15 d CBC example and show that the posteriors are like the prior (ish) or do we need some quantification?
I've attached the 1D posteriors and trace plots from a quick run I did on the 15d example. It takes about ~10s to run, then 1 minute to reconstruct the marginalized parameters. The full corner plot is too large to upload here.
@gregory.ashton would it be possible to add the KS test p-values between the prior samples and the known analytical prior distributions?
-
It is certainly possible, at the moment, all this really does it provide a simple functionality to run with zero likelihood.
It sounds like what people really want is a "test" of this. I suggest this become part the nightly C.I test. Does it need to do the full 15d parameter space, or is a subset enough?
added 1 commit
- 998471a7 - Adds a test to the nightly tests which samples from the prior
changed milestone to %0.5.3
mentioned in merge request !369 (closed)
mentioned in commit c04a7e97