Joosep Pata

implement a per-particle weight to predict a MET correction

1

similar to DeepMET.

investigate HEPT

1

https://github.com/Graph-COM/HEPT/tree/main/src/models/attention https://arxiv.org/abs/2402.12535

It seems the cluster shape variables for ECAL are not yet filled. It looks like the PFRecHit->depth() is always 1 for ECAL hits, so the features with depth 2, 3,...

Provide an efficient inference implementation using sparsification/quantization

1

Goal: reduce inference time of the model using quantization We made some CPU inference performance results public for 2021 in CMS, https://cds.cern.ch/record/2792320/files/DP2021_030.pdf slide 16, “For context, on a single CPU...

[WIP] Training on AMD / ROCm

5

Currently, there are the following issues on LUMI / AMD MI250x for pytorch: - ROCm flash attention is not supported natively in pytorch, and must be used via the external...

Evaluating the dataset number of steps in tensorflow is currently slow when the loop is IO-bound in a single process, because we use `tf.data.Dataset.from_generator`, which uses Python underneath and doesn't...

https://arxiv.org/pdf/2310.16121.pdf

Since the port is hardcoded, starting several different multi-GPU runs on the same host doesn't work currently: https://github.com/jpata/particleflow/blob/main/mlpf/pyg_pipeline.py#L76 A fix would be to auto-select the port, or configure it via...

https://github.com/pbelcak/fastfeedforward. - [ ] Read the paper - [ ] Does it speed up wrt. layer width or wrt. number of elements per event? - [ ] Quick test for...

a fast/efficient graph constructor was recently proposed here: https://github.com/mieskolainen/hypertrack https://indico.jlab.org/event/459/contributions/11748/attachments/9580/14256/HyperTrack_Mieskolainen_CHEP2023_v1.pdf we should give a try if this works for graph construction for us.