Usage of L1 hits as "subsampling"
We currently use random sampling in order to subsample the hits when they exceed some boundary. I would like to see how using L1 hits instead as a way of reducing the size of the event files with minimal loss of information.
It might not be enough to get under the current sample size (5000) for ARCA, but it could be complemented by random sampling if needed still.
Activity
It may indeed be useful @gvermarien. I understood from @dguderian that the h5 file obtained with h5extractor maintains the hits information, therefore also L1 hits. So we would need a function in OrcaSong to keep such hits I imagine. Right?
Funny you mention it today; I was just thinking about that. I agree that the selection should be:
- L1
- L0 with the smallest |t|
instead of randomly drawing. The argument was that this way the selection is less biased. But I think we can be save to say that a proper selection should contain more info on average.
Edited by Daniel Guderian
Earlier I talked to @adomi and she told me we can use the https://km3py.pages.km3net.de/km3pipe/_modules/km3modules/hits.html#count_multiplicities function in order to process this during the orcasong conversion.
Maybe you can also just use the "triggered" column like here https://git.km3net.de/ml/OrcaSong/-/blob/master/orcasong/modules.py#L86 ?
-
I have created a quick and dirty implementation to subsample based on only triggered hits in this branch: https://git.km3net.de/ml/OrcaSong/-/tree/triggered_hits it is by no means ready to be integrated, but I can hopefully demonstrate its usefulness soon.
-
after our discussion I quickly added your changes to the master here: 770d8a8d
