-
Notifications
You must be signed in to change notification settings - Fork 2
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
DNN Indexing Bug Fix + Normalization Improvements #100
base: CMSSW_14_1_0_pre3_LST_X_LSTCore_realfiles_batch1_devel
Are you sure you want to change the base?
DNN Indexing Bug Fix + Normalization Improvements #100
Conversation
/run all |
The PR was built and ran successfully in standalone mode. Here are some of the comparison plots. The full set of validation and comparison plots can be found here. Here is a timing comparison:
|
Efficiency increases slightly for pT5's but no huge changes. Fake rate decreases in transition region, increases in barrel. Note: Plots below exaggerate the fake rate increase in the barrel because the y axis changes. The AUC of the model increases significantly (from .88 -> .91) from fixing the bug but this doesn't translate to significantly better performance. Tracks in the transition region become easier to reject than tracks in the barrel. It seems like the eta value of the hit shared between the two T3's offers a lot of rejection power in that region using radius matching (inner vs. bridge, outer vs. bridge) Overall, I don't think there is a large difference in physics performance or timing from fixing the bug. If the fake rate increase in the barrel is a concern, the easiest solution would be to choose a tighter cut in the barrel to match the performance of the current DNN in that region. |
The PR was built and ran successfully with CMSSW. Here are some plots. OOTB All TracksThe full set of validation and comparison plots can be found here. |
Fixes a bug related to the hit indexing in the DNN train script. Also adds other hits to write_lst_ntuple. Will add more to this PR soon.