HftKnn problem
Oct 1 2015
#########
1. Several iterations for the same 1st hit does not work. If 1st iter failed the next is also failed
2. StvKNSelector::Relink improved. Founded shortest diostance to the 4th neigbour is used as a cut for the next search
StvKNSeedFinder::NextSeed 56% ==>43%
More precise StvKNSelector::Relink 32% ==> 8%
Maping of the search became effective, it was not before.
Total time 0.030/0.017 ==> 0.024/0.18 per 20 % became faster but still 30% slower then default
3. The angle of selection cone is 88d now and direction from the 1st hit to cnter of TPC.
Probably it will be better to use angle which which contains +- meter in Z and direction between. For this case
angle will be less and hence nomber of hits in cone will be less too.
4. Right now all the hits are sorted by X,Y,Z. But for this case surrounding box could be too big when track is between Y & Z axis
The idea is to add 2 dimensions, X+Y and X-Y. Then surrounding box for track between for additional dimensions is better.
=====================================================================================================
Good order for StiKnn, xx+yy-0.01*zz
==========================
Bad order for default Sti, xx+yy+zz
==========================
Now different Order for KNN xx+yy-0.1*zz
===============
IAfter long time problem was found and fixed.
Now BAD info:
Short test for one event.
StiDef: <nGlob> = 2097 <nPrim> = 983 -17%
StiKnn:<nGlob> = 1747 <nPrim> = 792 -19%
To fix efficiency tuning is needed
StiDef: CPU/track = 0.02
StiKnn: CPU/track =0.05
For timing also tuning is needed
GOOD info
StiDef From one seed 10% of accepted tracks
StiKnn From one seed 70% of accepted tracks
That means KNN one seed is much more expencive, but quality is much better
- perev's blog
- Login or register to post comments