Regression of log-transformed energy and pt, training checkpoints - Githubissues

jpata / particleflow

Machine-learned, GPU-accelerated particle flow reconstruction

Apache License 2.0

24 stars 30 forks source link

Regression of log-transformed energy and pt, training checkpoints #343

Closed jpata closed 2 months ago

jpata commented 2 months ago

Primary changes:

use log-transformed energy and pt for regression
split classification to binary and PID multiclass losses
last attention layer queries configurable, now set to input embeddings
additional CMS stats in v2.3
generate CLIC WW->fullhad, ZH->tautau, Z->tautau in v2.1
ensure training works on AMD MI250X (LUMI)
split and monitor regression losses separately
add CLD training
option to save attention matrices for visualization

Result checkpoints:

CMS: https://indico.cern.ch/event/1458559/contributions/6140587/attachments/2931158/5147426/2024_09_19.pdf, pyg-cms_20240917_155112_841033
CLIC: 2024_09_18.pdf, pyg-clic_20240917_104006_818049