gcomgpool

Takes an Adjacency Matrix and a Feature Vector, orders the Feature Vector by its last Feature and the Adjacency Matrix in the same way. Then uses a learnable Transformation to go from each neighbouring c vectors of size param into one vector of size paramo. and uses mode to go from a big adjacency Matrix of size gs to one of size gs, after which it roundes the resulting Adjacency Matrix to either 0 or 1 and returns both the new Adjacency Matrix and the new Feature Vector. Basically a simpler (and not used by me) version of the function abstr

Arguments

• gs: The initial Number of Nodes of the Graph (Graph Size).
• param: The initial Number of Features in each Node.
• paramo: The resulting Number of Features in each Node.
• metrik_init=”glorot_uniform”: The initializer of the Feautre Vector tranformation
• learnable=True: weather the Transformation is learnable
• mode=”mean”: Either “mean”, “max” or “min”, The Transformation to go combine each subgraph into a single node
• cut=0.5: Where to round the graph after the mode transformation, Each Value above cut will be set to 1 and each value below to 0, each Value that is equal to cut will be also set to 1
• c_const=1000: Since the Rounding is implementet using relus (relu(1+c_const*(X-cut))-relu(c_const*(X-cut))) , you require a numerical constant to decide how fast the rounding function goes from 0 to 1. This could be set higher, but this could result in diverging gradients.