Joosep Pata

Currently we are using a simple matching-based loss based on individual particles/elements. It might be interesting to try a contrastive event-based loss, such as the one proposed in this paper....

- https://openaccess.thecvf.com/content/CVPR2023/papers/Liu_FlatFormer_Flattened_Window_Attention_for_Efficient_Point_Cloud_Transformer_CVPR_2023_paper.pdf - https://github.com/mit-han-lab/flatformer - https://www.youtube.com/watch?v=BRVxeRRiPUY

https://inspirehep.net/literature/2609884

https://arxiv.org/abs/2310.05869

Kenneth suggested uploading the singularity image to cvmfs: - https://indico.cern.ch/event/764570/contributions/3173502/attachments/1735975/2807816/CVMFS-unpacked.pdf - https://gitlab.cern.ch/unpacked/sync Currently the latest TF+nvidia image is here: https://hep.kbfi.ee/~joosep/tf-2.13.0.simg The pytorch image is here: https://hep.kbfi.ee/~joosep/pytorch.simg

PR testing, memory info

1

We are trying to profile the PR https://github.com/cms-data/RecoHGCal-TICL/pull/3 and noticed that with "enable profiling", the memory profiles are not part of the results. Recalling past discussion, I understand `igprof mp`...

MLPF datasets v2.0: track pythia-level genjets, genmet in datasets; add per-particle ispu flag; generate CMS ttbar+noPU [WIP]

1

Regenerated all relevant datasets as v2.0: - [x] CMS: new CMSSW generation with CMSSW_14, new postprocessing, new tfds - [x] add CMS ttbar no-pileup sample as a cross-check - [...

Based on the discussion with Jennifer, it might be good to have CHS available as well in addition to PUPPI, as PUPPI might require retuning given the updated reconstruction. Need...

Unify these scripts to reduce duplication: - https://github.com/jpata/particleflow/blob/main/scripts/clic/postprocessing.py - https://github.com/jpata/particleflow/blob/main/scripts/clic/postprocessing_hits.py - https://github.com/jpata/particleflow/blob/main/scripts/fccee_cld/postprocessing.py

Check which elements attend to which other elements throughout the layers.