[KathrineOls1]

This change removes three dense identity matrices and seven identity multiplications from SCCNNLayer.forward. Multiplying a signal by an identity matrix is unnecessary, and allocating these matrices adds memory and computation costs that grow with the number of cells in a batch.

The implementation uses the signals directly while preserving the filter layout, learned parameters, and surrounding operations. It also removes the unused node, edge, and triangle counts.

A small regression test isolates the seven identity terms and checks outputs and input/parameter gradients for sc_order=2 and sc_order=3. This is made by ChatGPT (GPT-6 Astra).

Validation:

• All 5 SCCNN tests passed.
• Ruff lint and formatting checks passed.
• Broader library suite: 145 passed; 3 DHGCN tests failed because pyg-lib was unavailable locally.
• A separate local comparison against the original implementation passed 48 configurations over three Adam updates each, with exact agreement in outputs, losses, gradients, weights, and optimizer state.

This optimization comes from a joint INF319 project by Christina Holthe and Kathrine Nordnes Olsen (at University of Bergen)

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 96.79%. Comparing base (983d5b2) to head (bc69a38).

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #396      +/-   ##
==========================================
+ Coverage   96.36%   96.79%   +0.42%     
==========================================
  Files          58       58              
  Lines        2092     2090       -2     
==========================================
+ Hits         2016     2023       +7     
+ Misses         76       67       -9     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.