This small PR addresses the recently discovered bug, detailed in #17. The PR fixes the bug by modifying the __inc__ method of CombinatorialComplexData such that the x_i tensors are also incremented. Furthermore, the PR introduces a script argument (--enable_indexing_bug) such that we can choose to use the buggy version, in order to be able to reproduce EMPSN.
I trained a few models on subsets of data with the exact same specs twice, once with the bug enabled and once with it fixed. In all specs, the fixed version achieved better performance. This is further evidence that what we observed was indeed a bug.
This small PR addresses the recently discovered bug, detailed in #17. The PR fixes the bug by modifying the
__inc__
method ofCombinatorialComplexData
such that thex_i
tensors are also incremented. Furthermore, the PR introduces a script argument (--enable_indexing_bug
) such that we can choose to use the buggy version, in order to be able to reproduce EMPSN.I trained a few models on subsets of data with the exact same specs twice, once with the bug enabled and once with it fixed. In all specs, the fixed version achieved better performance. This is further evidence that what we observed was indeed a bug.