EventProcessor: Optimize Laplacian & Boundary Computation

[Ankitaghavate]

The current EventProcessor code for computing Laplacians and boundaries is long and repetitive. It should be simplified to handle dynamic simplices, reduce repeated code, and improve memory usage when saving chunks. Running compute_lapl_and_bounds() on large datasets can be inefficient, and feature vectors for nodes, edges, and faces should be dynamically handled for all events.

Environment: Python 3.10+, PyTorch, gudhi, awkward, numpy, scnn.

[Abhi1366-create]

Hi, I’d like to work on this issue.

After reviewing compute_lapl_and_bounds() in EventProcessor, I noticed a few areas that could be improved:
The current implementation handles simplices using hardcoded checks (len(boundaries) == 1 or 2), which makes it non-scalable if max_dimension changes.
Laplacian and boundary tensor construction logic is repeated across branches.
Feature initialization (xs_temp) is tightly coupled to fixed dimensions (nodes/edges/faces), which prevents handling dynamic simplicial complexes.
Results are accumulated in memory before chunk saving, which may become inefficient for large datasets.

My proposed approach:

Refactor the Laplacian and boundary construction into a dimension-agnostic loop over simplex degrees.
Dynamically generate feature tensors based on the actual number of simplices per dimension.
Remove duplicated logic and centralize tensor conversion (coo2tensor, normalize) into reusable blocks.
Improve memory usage by reducing intermediate storage and ensuring chunk flushing is handled more efficiently.
Please let me know if this aligns with the intended direction, or if there are additional constraints I should consider before submitting a PR.

Thanks!