Closed sathvikbhagavan closed 1 year ago
@ChrisRackauckas
I investigated a bit about the Surrogates test failures and wanted to summarize them here:
Tests in SurrogatesMOE fail spuriously - i. Successful run - https://github.com/SciML/Surrogates.jl/actions/runs/6248364164?pr=441 ii. Unsuccessful run - https://github.com/SciML/Surrogates.jl/actions/runs/6248569007?pr=441 This happens when one of the clusters do not have any test points which is not handled (not 100% sure)
using QuasiMonteCarlo 0.2.19 (latest of 0.2) fails tests in GEKPLS
tests. But using 0.2.16 works and passes all tests. The reason explained in https://github.com/SciML/Surrogates.jl/pull/420
Cannot bump QuasiMonteCarlo to 0.3 as for some reason https://github.com/SciML/QuasiMonteCarlo.jl/blob/master/src/Section.jl is commented and it complains about SectionSample
not defined.
I had to bump the compat of ExtendableSparse
to 1 as I get this error
ERROR: MethodError: \(::Symmetric{Float64, ExtendableSparseMatrix{Float64, Int64}}, ::Vector{Float64}) is ambiguous
with 0.6 in julia 1.9.
After bumping, it works.
So, I think (1) can be fixed but I am not sure what's happening with QuasiMonteCarlo currently.
Okay, we should just bound QMC to 0.2.16, set the RNG seed for 1. Let's see how far that takes us.
I think the reason MOE tests were failing spuriously as division between train and test data was not same each run because of using BitArray(undef, n)
in https://github.com/SciML/Surrogates.jl/blob/master/lib/SurrogatesMOE/src/SurrogatesMOE.jl#L144
as all elements are not false all the time (we want all of them to be false)
This is fixed in 6848450
julia> for _ in 1:10
@info sum(BitArray(undef, 150))
end
[ Info: 26
[ Info: 49
[ Info: 4
[ Info: 3
[ Info: 25
[ Info: 49
[ Info: 7
[ Info: 3
[ Info: 51
[ Info: 48
Codecov Report
0.00% <0.00%> (ø)
84.48% <ø> (+1.72%)
... and 13 files with indirect coverage changes
:mega: We’re building smart automated test selection to slash your CI/CD build times. Learn more