Custom loss function using nodes predicted output

[RaphaelChristienECTL]

Hello everybody !
I am really happy with spektral, thanks for this !

I currently have a model predicting a sequence of real numbers (representing a trajectory, x1, y1, x2, y2 etc.)
for each graph node (a moving object).

These moving objects must not get too close from each other.
To push the model to respect this, I would like to build a custom loss function including, for each node :
1- the difference between its predicted and actual trajectory (I have this),
2- a great penalty if one neighbor is predicted to get too close (e.g. the minimum distance between node1 and node2 is lower than a prescribed threshold).

Any suggestion on how to do this second step ? I guess this would involve looping over y_preds and computing
y_prednode1, y_prednode2 differences etc. but I am not clear how to ensure I am checking connected nodes ?
Any other suggestion are welcome :-) !

Cheers !

[danielegrattarola]

Hello,

to compute sparse pairwise differences between neighbours, you can do something similar to the EdgeConv layer.
Have a look at the source code here: https://github.com/danielegrattarola/spektral/blob/master/spektral/layers/convolutional/edge_conv.py#L9

cheers

[RaphaelChristienECTL]

Thanks very much for the code pointer, I will check this :-)