Skip to content

Model compression and resolution calculation

Simon Poon requested to merge github/fork/hanhiller/hyperparams into main

Created by: hanhiller

  • removed redundant inputs to model (px and py— they are used in the puppi calculation still)

    • NOTE: Currently these are still fed into the model

  • removed redundant/useless pdgID information

    • now neg and pos. pdgIDs map to the same particle type (charge already specified)
    • removed pdgIDs which do not appear in the data set [0,1,2]

  • embedding output dimension=2 (was 8)

  • fixed bug with puppi resolution calculation (was being scaled by ML response correction)

    • Now puppi resolutions will remain fixed between trainings

  • added calculation of the average difference of ML and puppi resolutions (over all bins weighted by the number of events in each bin) (greater average difference = better model)

    • This metric gives us a way to qualitatively compare the resolutions between different models

  • added error bars to the METx and METy plots ( these upper and lower errors are the widths of the distribution at points slightly wider and narrower than 1 SD )

  • For comparisons sake, here are the average resolution differences for the default (current main) model— 3 layers (64,32,160

    • Root workflow: average xRes difference=5.21, average yRes difference=5.30
    • h5 workflow: average xRes difference=4.98, average yRes difference=4.91

  • quantized model trains properly

    • 16,6 bit model out preforms the default model (average xRes dif=5.59, average yRes dif=5.54)

  • 2 layers (16-32) seems to train well

    • h5 workflow: ~45min, average xRes dif=4.92, yRes difference=5.20
    • root work flow: 5.5 hrs, average xRes dif=3.92, average yRes dif =4.04

Merge request reports