Output problems for summary and stats files

[jtlz2]

Hi - some of the output files are not produced, but I haven't tracked down why yet (sorry!), e.g. 1-summary.txt. If the file does not already exist, it is produced but is empty.

1-post_separate.dat is empty when multimodal=True.

1-stats.txt is incomplete:

Global Evidence: -0.690049101357911621E+01 +/- 0.602774045152852864E-01

Local Mode Properties

Total Modes Found: 2

This is for all combinations of the verbose, write_output and multimodal keywords.

pymultinest.Analyzer(n_params = n_params).get_stats()['modes']
analysing data from chains/1-.txt
[]

is empty (whether multimodal is True or False), so I don't know where to get e.g. the MAP parameters from except by calculating those and other stats myself.

I guess you'll want a minimal example..?

[JohannesBuchner]

Please take a look at issue 3 (#3), it might be the same problem. Ultimately, it's a MultiNest problem I think, MultiNest shouldn't suddenly decide not to write out stuff. I am unsure what kind of likelihood functions trigger this problem.

You should however not rely on MultiNest writing out modes. If only one peak exists, then there will be no mode statistic written. You can get your statistics from looking at the distribution (equal weights, loaded for you in Analyser.get_equal_weighted_posterior) -- if you want to assume a Gaussian approximation there (which may or may not be justified). The MAP should be the last line in the prefix.txt file (loaded for you in Analyser.get_data). If you would like this functionality to be in PyMultiNest perhaps we can think of a generic way to do it.

[jtlz2]

Hey - yes I think that's the same problem. I'm very keen to help track down the problem as I had high hopes for PyMultiNest, and if I can't get it working soon I'll have to revert to MultiNest itself..

It would be really good to have the multiple mode statistics - should I rather write to Farhan about this?

I've coded up a simplified version of the object-detection example, if it's any use, where A and sig are fixed and the search is for two gaussians in x,y only. 6 modes are reported (I think some are edge artifacts), but no stats are written except what I describe above: prefix-stats.dat is:

Global Evidence: -0.142335426023810392E+02 +/- 0.597318229859429492E-01

Local Mode Properties

Total Modes Found: 6

and prefix-summary.txt is empty.

Shall I keep simplifying the example until it starts working (as for your demo)?

It would definitely be good to be able to access all of the stats from within PyMultiNest (assuming they can be recovered, and if I've understood you correctly).

So let me know how I can help!

[JohannesBuchner]

Yes, please write him with your code. However, I would like to stress again that I think this is not a Python issue and you will get the same result if you write your likelihood function in C or Fortran. We don't do anything magical in PyMultiNest. This is somehow a weird bug in MultiNest that gets only triggered sometimes. If you can find a simple likelihood function that triggers this issue, Farhan will be able to investigate writing a C/Fortran function. Alternatively, you can log the values your likelihood function returns.

[jtlz2]

Point taken - I'll get on it :)

[JohannesBuchner]

possibly resolved with issue #16. If not, please re-open.

[jtlz2]

This is fantastic news - thank you - that exactly fixes it!

On 3 July 2013 03:25, Johannes Buchner notifications@github.com wrote:

possibly resolved with issue #16#16.
If not, please re-open.

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/12#issuecomment-20386114
.