Application to 3D dataset

[Berumott0]

Hello,
I am interested in applying the GNS approach to my dataset. Specifically, I intend to use ceramic fragments as nodes to predict the location of the fragments after the ceramic is impacted. However, after expanding my test dataset to 3D, I am having a bit of a problem with the dimensionality of the boundary distances, and the boundaries are never dimensionally aligned with the most_recent_position.
Can you suggest how best to modify the current framework to accommodate the 3D dataset? Any guidance or advice would be appreciated.

[yjchoi1]

We have data format documentation here. If the data is 3D, the "bounds" keyword in metadata should have 3 items where each item corresponds to the range of boundary for each dimension ((e.g., "bounds": [[0.1, 0.9], [0.1, 0.9], [0.1, 0.9]])). Also, please check if the node feature length nnode_in aligns with your data.

[li-ruidong]

We have data format documentation here. If the data is 3D, the "bounds" keyword in metadata should have 3 items where each item corresponds to the range of boundary for each dimension ((e.g., "bounds": [[0.1, 0.9], [0.1, 0.9], [0.1, 0.9]])). Also, please check if the node feature length nnode_in aligns with your data.

I created a dataset myself, where the bounds are [[0, 10], [0, 10], [0, 40]]. However, during training, I encountered all loss values being NaN. Do these bounds and positions need to be normalized before use?