glmlp

Extension of gltknd making the Update procedure more complicated and in line with particleNet. Here each update consists of 3 learnable Dense Layers with included Biases (thus breaking Graph Permutation Symmetry). In between each of these Layers is a Batch Normalisazion Layer and an activation (mlpact)

Arguments

  • gs: The gs Number of Nodes of the Graph (Graph Size)
  • param: The Number of Features for each node in the Feature Vector
  • keepconst: The first keepconst Features are keept unchanced
  • iterations=1: repeat the Actions of this Layer iterations time
  • alinearity=[-1.0,1.0]: activation of this Layer, explained better in glm
  • initializer=”glorot_uniform”: Initializer of this Layer
  • i1: Size after the first Dense Layer
  • i2: Size after the second Dense Layer
  • mlpact=K.relu: Activation after each Dense Update Step. Requires to be a function
  • momentum=0.99: Momentum of the BatchNormalisation
  • k=16: Number of Average Connections in the Graph. Can be ignored, and is ignored in glm, but migth help the Network converge