Removing receiver conditioning in graph network

[ttdelgadott]

I'm trying to remove conditioning on receiver mesh modes in the edge model of the grid2mesh encoder (as mentioned in the FGN paper). However, I'm noticing quite a few of artifacts in longer rollouts.

In typed_graph_net.py, I removed the conditioning in both the _edge_update function at the InteractionNetwork instantiation. Is there anything else I'm missing? Any code examples would be extremely helpful!

[DominicMasters]

Hi!

So just to start the motivation behind this change is purely about efficiency, the gather from the mesh input features is quite expensive (particularly in the backwards pass!) so removing it can have a moderate boost to throughput. While removing any of the input features to the GNN layers would usually also have a negative impact on accuracy, we can see that these particular ones only hold redundant information so should be safe to remove. Specifically they contain the absolute positions of the mesh points, however these are already computable from the grid and edge features which contain the absolute grid positions and relative position between grid and mesh nodes, respectively.

So my first piece of advice would be to not get too hung up on this particular feature as it is really a small performance optimisation rather than a critical feature.

Now if you are still keen to implement this i would first confirm that you really are only removing the receiver features from just the grid2mesh. The equivalent receiver features in the mesh2grid are definitely important as they hold the residual connections which contain lots the high frequency information from the grid. Removing these could conceivable cause artifacts in rollout so this would be my initial guess at the bug.

If this does not resolve you issue please feel free to share the changes youve made so we can better understand if your code matches what we have internally.

Thanks!
Dom