This implements a new feature to clip values of the (transformed) node representations before summing them up. This is particularly important for using the layer using the (weighted) summation modes ("sigmoid", "none"), as those otherwise can lead to values of extremely large magnitudes. [I've noticed by trying them out in the MoLeRVAE setting.] Clipping before summation allows the layer to still perform tasks such as counting the number of nodes, which would be limited by clipping after summation.
This implements a new feature to clip values of the (transformed) node representations before summing them up. This is particularly important for using the layer using the (weighted) summation modes ("sigmoid", "none"), as those otherwise can lead to values of extremely large magnitudes. [I've noticed by trying them out in the MoLeRVAE setting.] Clipping before summation allows the layer to still perform tasks such as counting the number of nodes, which would be limited by clipping after summation.