Joosep Pata comments

Results 115 comments of


                                            Joosep Pata

[WIP] PyTorch-based training on Gaudi HPU

https://github.com/jpata/particleflow/pull/310 is merged, you can try updating.

Provide an efficient inference implementation using sparsification/quantization

Basically, to summarize: - with @raj2022 we saw that it's possible to quantize the model to int8 in pytorch using post-training stating quantization, following the recipe in https://github.com/jpata/particleflow/blob/main/notebooks/clic/mlpf-pytorch-transformer-standalone.ipynb - the...

implement a per-particle weight to predict a MET correction

First implementation in https://github.com/jpata/particleflow/pull/320.

phase2 PV sorting is not using HGCAL inputs (only the barrel and HF contribute to ak4CaloJetsForTrk) will be needed for jetCore as well

type tracking

phase2 PV sorting is not using HGCAL inputs (only the barrel and HF contribute to ak4CaloJetsForTrk) will be needed for jetCore as well

@cms-sw/upgrade-l2 is this still relevant?

Infinite(?) recursion in Basic2DGenericTopoClusterizer::buildTopoCluster() in WF 138.3

Thanks, I'm able to reproduce this with the latest ASAN build CMSSW_12_4_ASAN_X_2022-03-25-1100. I'm going to check CMSSW_12_4_ASAN_X_2022-03-21-1100 as well. EDIT: also reproducible in this release. Are earlier ASAN builds available...

Infinite(?) recursion in Basic2DGenericTopoClusterizer::buildTopoCluster() in WF 138.3

I looked into this a bit, and it seems `buildTopoCluster` recurses over all the rechits, which can easily be 50k or more. https://github.com/cms-sw/cmssw/blob/5f75b2d25170c527f84074e9e817231cedd8ba8b/RecoParticleFlow/PFClusterProducer/plugins/Basic2DGenericTopoClusterizer.cc#L121 This does not crash the normal release,...