avoid seed to be zero
To be able to use --X-rtd TMCSO_AdaptivePseudoAsimov=0 --X-rtd TMCSO_PseudoAsimov=0 in bbgg, also added the option as part of the special combine options to bbgg
@fmonti could you please triple check?
Merge request reports
Activity
assigned to @fmonti
What's the effect of the TMCSO_AdaptivePseudoAsimov=0 and TMCSO_PseudoAsimov=0 options,
I let @fmonti comment here, that is for the bbgg needs
and are we sure that this does not negatively influence all other combine commands?
Yes, we are sure it does not affect other channels or combo. I redid all the other channels and combos with it in the context of the H+HH combination and all went well
Also, wouldn't a simple
+ 1
be enough, instead of thissum_seed
?hummm, probably, but I remember I had problems with the fact that the seed appear on the result file name from combine, therefore it needs to be well defined for the mv part of the command.
added medium-priority label
added fit-optimization label
hummm, probably, but I remember I had problems with the fact that the seed appear on the result file name from combine, therefore it needs to be well defined for the mv part of the command.
I did a check again and the
{self.branch}
part, e.g. inhiggsCombineTest.AsymptoticLimits.mH{self.mass_int}.{self.branch}.root
should cover this :) We would just have to consistently add the same number / +1. If you agree I would add this to all commands (and resolve conflicts), and then the only point left is the additional options that I would leave to @fmonti .
Edited by Marcel RiegerHi, the options TMCSO_AdaptivePseudoAsimov=0 and TMCSO_PseudoAsimov=0 avoid that in the channels with an unbinned dataset, combine decides to generate a pseudoasimov dataset (weighted unbinned events) instead of the histograms. In the HH combination this option will affect at most the bbgg channel.
An unfortunate combination of settings resulted in a very wiggly likelihood scan from the bbgg+tthgg channels in the h+hh combination:
- without TMCSO_AdaptivePseudoAsimov=0 and TMCSO_PseudoAsimov=0, combine decided to generate a pseudoasimov dataset
- the pseudoasimov dataset used to be re-generated at each point of the likelihood scan (this is not the case anymore, right @acarvalh ?)
- with seed=0 the seed was randomized at each point of the scan
Edited by Fabio MontiAlso I would avoid to include
--X-rtd MINIMIZER_skipDiscreteIterations
unless we have a strong need of further reducing the resource usage (but bbgg is not the main offender)In order to use that trick properly, we should first perform a fit floating the pdfindices and saving the fit results. Then at each generation of the asimov we should manually set all the pdf indices to their best fit values.
Edited by Fabio Montiadded 169 commits
-
4ba4d9a9...d01918e8 - 168 commits from branch
master
- 56122b9b - Merge branch 'master' into avoid_zero_seed.
-
4ba4d9a9...d01918e8 - 168 commits from branch
added 1 commit
- 1e8280f6 - Consistently use ammended seed, use python 3 formats in some places.
I've added the seed adjustment in all tasks and started a full test pipeline in https://gitlab.cern.ch/hh/tools/inference/-/pipelines/5372327.
When this passes, we could merge.
Edited by Marcel Rieger