MiniAOD b-tag tweaks #8370
MiniAOD b-tag tweaks #8370
Conversation
… removed them from CMSTopTag fat jets
To correctly store daughters of AK8 jets it is important to use subjets actually derived from AK8 jets. In this case these are the SoftDrop subjets while the CMSTopTag subjets are derived from CA8 jets and could in principle contains extra constituents not clustered inside AK8 jets.
|
@ferencek is there a twin PR on 75X? |
|
A new Pull Request was created by @ferencek (Dinko Ferencek) for CMSSW_7_4_X. MiniAOD b-tag tweaks It involves the following packages: PhysicsTools/PatAlgos @cmsbuild, @vadler, @nclopezo, @monttj can you please review it and eventually sign? Thanks. |
|
@arizzi, I first wanted this to be reviewed and tested for 74X. Also, the policy on automatic vs manual forward porting is a bit confusing so I did not want to rush with a 75X PR. If and when needed, a 75X PR will be created. |
|
Just a general comment. The way we handle AK8 daughters at the moment is a but fragile. If somebody in the future reverses the order of subjet collections or adds a new one in front, things might not work as intended. Should there be a comment left in the cfg file? |
|
@ferencek yes, that would definitely help. |
|
Ah, ok, now I understand... indeed we assume that the subjets only contain A protection could be to change these lines https://github.com/cms-sw/cmssw/blob/CMSSW_7_4_X/PhysicsTools/PatUtils/plugins/JetSubstructurePacker.cc#L77-L78 On Wed, Mar 18, 2015 at 7:00 PM, rappoccio notifications@github.com wrote:
|
|
@gpetruc, this is in principle OK from the point of view of ensuring that AK8 daughters are all stored correctly but this is not necessarily optimal from the event size point of view. CA8 and AK8 jets generally tend to have quite different shapes so I wouldn't be surprised that subjets derived from CA8 jets quite often contain some extra constituents wrt to AK8 jets. So a possible (and maybe not so unlikely) scenario is that because of a single extra constituent in one or more of the subjets, we would duplicate a lot of daughter pointers. |
|
@ferencek Even the CA8 jets are preclustered with AK8, though, so I'm not sure this happens in practice. You may be right but I think these corner cases are pretty rare. |
|
@rappoccio, aah, indeed, preclustering would reduce the likelihood of this happening (I somehow keep forgetting about preclustering ;) ). However, in busy final states, where it maybe matters even more not to duplicate stuff, this would probably not be so unlikely. |
ferencek
force-pushed
the
MiniAODBTagTweaks_from-CMSSW_7_4_X_2015-03-17-1400
branch
from
abfb5c7
to
929961c
Compare
|
Yes, i understand.
However i suppose there is no easy way to be able to tell for sure whether
the sub-jets are made from ak8 or not at configuration or creation time.
If we were to put instead an asser that daughters are all in the AK8, it
could happen that all test events pass but the thing eventually crashes in
production (after all, in the tests i had made on some events in the qcd
high pt relval i had not found any spurious daughter even if i was using
the cmsTopTag and not the softDrop...).
Other bright ideas?
Saving separately the the spurious daughters and asking people to veto them
seems a bit a complication.
Giovanni
|
|
OK, I added a comment but it would probably still be good to implement Giovanni's suggestion. |
…ontaining extra constituents
ferencek
force-pushed
the
MiniAODBTagTweaks_from-CMSSW_7_4_X_2015-03-17-1400
branch
from
8d6ce63
to
2a05498
Compare
|
@gpetruc, please have a look at the latest commit and let me know if there is anything missing or can be improved. |
|
@ferencek the modification looks ok to me |
…W_7_4_X_2015-03-17-1400 MiniAOD b-tag tweaks
This PR introduces the following changes:
Observed increase in the event size is at sub-per-mille level (tested on 9k events from /RelValProdTTbar_13/CMSSW_7_4_0_pre8-MCRUN2_74_V7-v1/AODSIM).
@gpetruc @arizzi @rappoccio @dmajumder