GraphSAGE on 3d data

Questions
#1

Is there anyway to run the GSAGE on 3d data? How can we modify the dgl graph sage model to run on 3d data?

#2

I assume you want to run GraphSAGE on point cloud data? We have a few examples in dgl/examples/pytorch/pointcloud at master · dmlc/dgl · GitHub and you could start there.

#3

I want to run it on 3d images.

#4

Not sure what you meant by 3D images. But if an image is expressed as a 4D tensor, then perhaps you could try PyTorch’s Conv3d module?

#5

Yes, my images can be expressed as 4d tensor. Do you think, can I apply the GSAGE from deep graph library on this?

#6

in order to apply graph on images you should:

  1. understand what features of the image you want that your GNN will process (ex. intensity values)
  2. assign data extracted from the image to nodes
  3. build relations between nodes based on what you want (ex. adjacency)
  4. once you have the graphs you can feed them to the GNN.

Since your question is too generic, I assume that you are a newbie in this field. So, go deeper on what i said above.

1 Like
#7

If your images is a 4D dense tensor, my general impression is that you don’t have to use GraphSAGE. It is best suited for the dense convolution module torch.nn.Conv3d.

1 Like
#8

Thanks for suggestion.

#9

This topic was automatically closed 30 days after the last reply. New replies are no longer allowed.