Following how invariants are computed in the spatial task using scatter add indexing, which is required for large graphs such as the spatial task, this issue should introduce an option to compute the invariants without the need for the padded distance matrix used for qm9. Additionally, it would be useful to benchmark the difference in speed per training epoch using each method for invariant computation.
mauriciogtec
commented
2 months ago
Following how invariants are computed in the spatial task using scatter add indexing, which is required for large graphs such as the spatial task, this issue should introduce an option to compute the invariants without the need for the padded distance matrix used for qm9. Additionally, it would be useful to benchmark the difference in speed per training epoch using each method for invariant computation.