Model vs dumper differences
I am seeing some issues with the script umami/scripts/check_lwtnn-model.py
where for trigger models (DL1d and DIPS) the output of the script compared to what the training-dataset-dumper creates is, what I hope, incorrect. For DIPS we were thinking that the script umami/scripts/check_lwtnn-model.py
is selecting the "wrong" tracks
dataset. But then I ran this on a new trigger DL1d training (that is doing better than DL1r) and the differences are "bad" in comparison to what the docs say one should get for a good model. I get for u-jet rejection predicted/training-dataset-dumper output differences:
Differences off 1e-6 99.67 %
Differences off 2e-6 99.38 %
Differences off 3e-6 99.0 %
Differences off 4e-6 98.76 %
Differences off 5e-6 98.51 %
Differences off 1e-5 97.47 %
I am opening this issue to shed some light on what may be going on.