First version of an MCMC sampler #829
Conversation
|
When the used metric is chi2, it should be halved so that it can be interpreted as a makeshift log-likelihood. I don't see that happening in the code. Can you verify you get compatible results using chi2 and llh? And does the result also give compatible results with our standard analysis? |
|
Ah right, good point. I will add the check for chi2 metrics. |
|
BTW, I have a little toy example comparing the Frequentist and Bayesian results. I won't put it into pisa, but maybe can go into the fridge? |
| Same as scaled_chain, but for the burn in phase. | ||
|
|
||
| """ | ||
| import emcee |
There was a problem hiding this comment.
May want to add a comment specifying which sampling algorithm the emcee package uses, for future reference.
There was a problem hiding this comment.
I added the option to define the sampling algorithm and pointed to the emcee docs where one can find the different options and the default
pisa/analysis/analysis.py
Outdated
| "The sampler will assume that llh=0.5*chi2.") | ||
|
|
||
| params_truth = hypo_maker.params.free | ||
| x0 = np.array([p._rescaled_value for p in params_truth]) |
There was a problem hiding this comment.
As far as I can tell, x0 is only used in the next line, and we only seem to use the length of the array. May be able to save some time and memory by extracting the length directly from params_truth.
There was a problem hiding this comment.
Good point, I simplified the code a bit
|
If you are going to put a toy example in the fridge please put it somewhere where it can easily be found, like maybe the bootcamp directory? I had to dig pretty deep to find the previous bayesian sampling work and I don't want that to happen to future analyzers. |
|
Could also have its own MCMC sampling or Bayesian analysis example folder |
First version of an simple MCMC sampler as part of the analysis class. You need to manually install emcee at the moment.
This PR is part of #804.
I also sneaked in a fix for the search for shared parameters in the detectors class.
(Again ignore the first 7 commits)