Fitting with nested sampling (UltraNest or MultiNest)

[gabrielastro]

Fitting with nested sampling (UltraNest or MultiNest)

[tomasstolker]

I think that all the species-related points may have been addressed in the meanwhile.

For setting some of the MultiNest/UltraNest options, you can use the kwargs_multinest and kwargs_ultranest parameters in FitModel. You can have a look at the documentation pages of those packages for more details.

To look at the intermediate sampling results, you can have a look at the output folders from MultiNest/UltraNest.

In general, I would suggest to not add all data at once in case of a high SNR spectrum with broad wavelength range. Try to increase stepwise the complexity of the fit. Typically a fit shouldn't take longer than several minutes.

Point 8: if you leave out a parameter from bounds, then it will automatically set the available range from the grid in case the parameter is mandatory. So in this case I would just not include 'feh'. If you do want to check if the parameter is needed, then you can use get_parameters() or get_points() of ReadModel.

[gabrielastro]

Thanks for your helpful tips!

1. For the documentation of normal_prior of FitModel:
   "The parameter is not used if the argument is set to None."
   →
   "A linear flat prior betwen the natural bounds is used for
   the parameters for which a prior was not set explicitly through bounds"
   or something like this would be more accurate.

2. Indeed, Ctrl+C will not work for PyMultiNest, as Johannes Buchner explained. Avoiding mistakes is better 😉.

3. Thanks!

4. Indeed, the kwargs_* you added answer this!

5. Ok, starting with small complexity and increasing sounds good.

6. I do not get the warning anymore, so, thank you.

7. If I am not mistaken, the priors (including parameter ranges) used for the fits are not stored in the database. Might this be a good idea, since they do make a difference? While at it, you could maybe add the number of likelihood evaluations or CPU time as a reminder of how expensive a good run was :).

8. Thanks for pointing to those functions! Sorry for not being clear but I meant that I accidentally set a parameter that is not one of the grid parameters. This could either get ignored or, better, give a single error and stop, instead of keeping on printing the error message. But here again, the best solution is not to pass accidentally unneeded parameters 😉.

[tomasstolker]

Consider it done! The priors are stored in the database as the bounds and normal_prior groups of the sampling results.

[gabrielastro]

Excellent! Thank you. I will not wait for the next run to confirm it but thank you already 😉.