divcla

classic graph abstractor, also does not chance the paramsize, just goes from one param to c params, and has one learnable matrix (which is const between the elements). Works by usual parameter divergence, and then by abstracting the graphs, with the constant learnable one Best handled by decompress

Arguments:

• g: a grap object
• c: the abstraction factor
• m: a constant, defined for example by getm()
• repeat=1: repeat this layer repeat time (multiple divergences)

Returns:

• g: the updated grap object