Closed sathvikbhagavan closed 9 months ago
@ArnoStrouwen, is there any reason we use old version of Tracker and Flux? I bumped the compat of Flux but it causes downgrade CI to fail - https://github.com/SciML/Surrogates.jl/actions/runs/7355008364/job/20022994040?pr=460
All modified and coverable lines are covered by tests :white_check_mark:
Comparison is base (
1cff140
) 77.93% compared to head (4320779
) 78.12%.
:umbrella: View full report in Codecov by Sentry.
:loudspeaker: Have feedback on the report? Share it here.
In my PR, the CI fails for both regular and downgrade CI: https://github.com/SciML/Surrogates.jl/pull/459 The failure is thus not related to an old version of something, but just a test that is random?
I would put the flux compat to both 0.13 and 0.14. But if that turns out not to be possible, we will have to bump other packages also, which will then cause even more packages to need a bump and so on.
The lower bound versions for downgrade CI are tedious to set up, but I'm quite experienced at it now, so I can do that part, as long as the random test result is fixed here.
Yes, random test failures should be fixed
Context
RandomSampling
like in https://github.com/SciML/Surrogates.jl/pull/459. This happens when one of the clusters in the Gaussian Mixture Model is empty. We should handle it more elegantly. I will look into it in a subsequent PR. To prevent spurious failures, I have changed it to use a deterministic method.