Public tests pass with relative error ~2–3 %. Private tests fail. Code already handles non-contiguous labels, duplicate edges, isolated vertices and uses the warm-up + Adam settings that match the public numbers. Is there a timeout or a specific edge-case on the private graphs?”-every obvious edge case and parameter combination has been tested.
I have done everything that can be done with the information that is publicly available
No, it is not a runtime error, because if it is, you get this error clearly as ‘runtime error’ I think . The issue may be that: some optimisers are very sensitive to the start intial vector of parameters so make sure to test random points (it take long runtime of course) but to make sure that optimisation not stuck to local minia. make sure public tests pass smoothly with optimiser that converges finally with your set up to its learning rate and number of itertaions.
yes that’s because the energies for the graphs of private tests are out from the accepted range by the grader while the public ones ok. Try random intial warmups of the Optimizer and for time sake, make the range narrow around the definite start point that allows the public tests Pass .