search

microsoft / gated-graph-neural-network-samples

Sample Code for Gated Graph Neural Networks
MIT License
1.03k stars 264 forks source link

where are the gates during propagation #20

Closed chengmengli06 closed 5 years ago

chengmengli06 commented 5 years ago

gates seems not used during propagation process

mmjb commented 5 years ago

The propagation (apart from the GCN implementation) uses gated recurrent units to combine the incoming information and the old node state, see e.g. here: https://github.com/Microsoft/gated-graph-neural-network-samples/blob/ae106c3e3e44e1f208ea606516ecc23e1a3ed041/chem_tensorflow_sparse.py#L213

chengmengli06 commented 5 years ago

why chem_tensorflow_dense.py has no gates, the dense one and the sparse one should be the same, but they does not looks to be the same? @mmjb

mmjb commented 5 years ago

It does, see here: https://github.com/Microsoft/gated-graph-neural-network-samples/blob/ae106c3e3e44e1f208ea606516ecc23e1a3ed041/chem_tensorflow_dense.py#L115

chengmengli06 commented 5 years ago

thks, if not using gates, is it able to train 4 layer gnn to similar precision? the paper does not compare the cases using gates and not using gates.

mmjb commented 5 years ago

We haven't explored this in detail, so I can't give definite answers, but in brief experiments, plain RNN cells performed less well. (This difference may go away with hyperparameter tuning, though)