This is an implementation of a relational GIN variant, see gnns/rgin.py for a more precise description. I've verified that the collapse onto a single edge type works and has comparable results as the original GIN paper, even though I stripped out all of their batchnorms (including inside the MLPs) and replaced them by a layernorm at the end of propagation. I'll do the remaining hyperparameter search for QM9/VarMisuse over the weekend, I think, but early feedback would be helpful.
[PR under construction]
This is an implementation of a relational GIN variant, see
gnns/rgin.py
for a more precise description. I've verified that the collapse onto a single edge type works and has comparable results as the original GIN paper, even though I stripped out all of their batchnorms (including inside the MLPs) and replaced them by a layernorm at the end of propagation. I'll do the remaining hyperparameter search for QM9/VarMisuse over the weekend, I think, but early feedback would be helpful.