EMPSN seemingly forgets to increment some tensors that hold node indices during batching, which results in the nodes of the first few graphs being used for all graphs in a batch (default batch size 96). Specifically, the __inc__ function defined in CombinatorialComplexData does not increment attributes that contain x_. These attributes hold node indices and thus should be incremented according to my understanding. This implies that for most graphs in the batch, the node and cell features are completely random (but not the geometric features). I reached out to the authors via email but did not hear back. We could open an issue in their repo. In the meantime, I want to fix the bug and make it a command line option to use the old or new version such that we can still reproduce EMPSN.
:bug: Potential bug discovery
EMPSN seemingly forgets to increment some tensors that hold node indices during batching, which results in the nodes of the first few graphs being used for all graphs in a batch (default batch size 96). Specifically, the
__inc__
function defined inCombinatorialComplexData
does not increment attributes that containx_
. These attributes hold node indices and thus should be incremented according to my understanding. This implies that for most graphs in the batch, the node and cell features are completely random (but not the geometric features). I reached out to the authors via email but did not hear back. We could open an issue in their repo. In the meantime, I want to fix the bug and make it a command line option to use the old or new version such that we can still reproduce EMPSN.