WIP: Yandex PID speed-up
This branch is for the current Yandex PID algorithms speeding up.
| Method | Time for 1 event (74 tracks), s | Time for 1000 events (40875 tracks), s | Time, ms/track | Comments |
|---|---|---|---|---|
| MC15TuneV1 | 2.365 (0.317) | 46.731 (33.328) | 1.090 (0.809) | Initial |
| MC15TuneCatBoostV1 | 13.233 (11.274) | 822.948 (808.976) | 19.845 (19.551) | Initial |
| MC15TuneFLAT4dV1 | 17.817 (15.644) | 478.116 (463.833) | 11.282 (10.985) | Initial |
| MC15TuneDNNV1 | 4.631 (2.495) | 425.672 (411.764) | 10.319 (10.031) | Initial |
| Preprocessing | 4.689 (2.434) | 380.135 (367.100) | 9.202 (8.938) | Initial |
| Empty PID classes | 1.882 (0.304) | 44.381 (31.660) | 1.042 (0.769) | Without PID predictions. |
| 1/6 prediction (empty) | 3.059 (0.502) | 48.918 (34.540) | 1.124 (0.834) | link |
| 0/6 prediction (empty) | 2.312 (0.336) | 48.729 (34.277) | 1.138 (0.832) | link |
Time measurement: EVENT LOOP (ANNPID)
Machine: lxplus058.cern.ch
Test script: /afs/cern.ch/work/m/mhushchy/public/davinci_test.py
DNN on python takes 0.1 ms/track in batch mode.
Edited by Mikhail Hushchyn