config.yaml

net: deepset # deepset or particlenet

epochs: 30
shuffle_buffer: 64

loss: mean_absolute_error
optimizer: adam
lr: 1.e-3

callbacks:
  reduce_lr_on_plateau:
    factor: 0.5
    min_lr: 1.e-8
    min_delta: 1.e-4
    patience: 5
  early_stopping:
    min_delta: 1.e-4
    patience: 7

model:
  deepset:
    type: mlp # mlp or resnet
    activation: relu
    initializer: he_normal
    batch_norm: False
    dropout: 0
    units: [512, 512, 512, 512, 512]

  particlenet:
    activation: relu
    initializer: he_normal
    batch_norm: False
    shortcut: False
    pooling: average # average or max
    num_points: 100
    K: 16
    channels: [
      [64, 64, 64],
      [128, 128, 128],
      [256, 256, 256]
    ]
    dropout: 0
    units: [256, 128, 64]

data:
  batch_size: 256
  train_size: 0.6
  validation_size: 0.2
  test_size: 0.2

  num_points: 100 # ParticleNet point cutoff
  
  features:
    jet: 
      numerical: [
        pt, eta, mass, phi, area, qgl_axis2, qgl_ptD, qgl_mult
      ]
      categorical: [
        puId, partonFlavour
      ]
    pf: 
      numerical: [
        pt, eta, phi, d0, dz, d0Err, dzErr, trkChi2, vtxChi2, puppiWeight, puppiWeightNoLep
      ]
      categorical: [
        charge, lostInnerHits, pdgId, pvAssocQuality, trkQuality
      ]
      synthetic: [
        rel_pt, rel_eta, rel_phi
      ]
  
  transforms:
    numerical:
      jet_pt: log
      jet_mass: log
      jet_qgl_axis2: sqrt
      jet_qgl_mult: sqrt
      pf_d0: abs
      pf_dz: abs
      pf_rel_pt: sqrt
    categorical:
      jet_puId: [0, 4, 6, 7]
      jet_partonFlavour: [-5, -4, -3, -2, -1, 0, 1, 2, 3, 4, 5, 21]
      pf_charge: [-1, 0, 1]
      pf_lostInnerHits: [-1, 0, 1, 2]
      pf_pdgId: [-211, -13, -11, 1, 2, 11, 13, 22, 130, 211]
      pf_pvAssocQuality: [0, 1, 4, 5, 6, 7]
      pf_trkQuality: [0, 1, 5]