GMMconv only learn to aggregate features from the node itself (self-loop)

[Icecream-blue-sky]

When I was using GMMconv to learn spatial relation of objects in object detection task, I found that the GMMconv only learn to aggregate features from the node itself. The edge between object i and object i is much larger than that between object i and other object. I use polar pseduo coordinates based on objects' centers, which is like below:

I send them to GMMconv after normalizing the pseduo coordinates. In the graph, the node feature is the object feature. The graph have self-loop before sending to GMMconv.

edge_spatial = edge_spatial/torch.tensor([1333, 2*math.pi]).unsqueeze(0).cuda() # 1333 is the max width or height of the image.

I set the aggr = 'mean', root-weight = True (which may be redundant? Is it the cause?), and kernel_size=16.
What is the cause of this problem? Should I do special handling for self-loop edge? Since its polar pseduo coordinate is (0,0).

[rusty1s]

How do you determine that GMMConv ignores neighboring node information? As you correctly pointed out, there exists two ways to incorporate "self-loops":

1. root_weight=True will learn a separate transformation/skip-connection for central node features
2. As part of a self-loop: The layer can only separate neighbor and central information based on pseudo coordinates.

IMO, dropping the self-loop and apply root_weight=True is probably the better option.

[Icecream-blue-sky]

How do you determine that GMMConv ignores neighboring node information? As you correctly pointed out, there exists two ways to incorporate "self-loops":

1. root_weight=True will learn a separate transformation/skip-connection for central node features
2. As part of a self-loop: The layer can only separate neighbor and central information based on pseudo coordinates.

IMO, dropping the self-loop and apply root_weight=True is probably the better option.

Maybe you are right, I'll try. Thanks.