Adaptive line search for Frank-Wolfe breaks tests

JuliaDecisionFocusedLearning / InferOpt.jl

Combinatorial optimization layers for machine learning pipelines

https://juliadecisionfocusedlearning.github.io/InferOpt.jl/

MIT License

113 stars 4 forks source link

Adaptive line search for Frank-Wolfe breaks tests #40

Open gdalle opened 1 year ago

gdalle commented 1 year ago

Frank-Wolfe related tests fail due to the linear solver for implicit differentiation not converging

Main suspect: https://github.com/ZIB-IOL/FrankWolfe.jl/issues/387

gdalle commented 1 year ago

Possible solution: #39

gdalle commented 1 year ago

We still need to diagnose why this happened in the first place

gdalle commented 1 year ago

Diagnosis will be easier once #65 is merged. If we switch the line search strategy in the new tests I think the bugs pop up again

BatyLeo commented 1 year ago

@gdalle the latest FrankWolfe.jl release seems to have broken our RegularizedFrankWolfe tests. I'm facing the following error locally on main (on paths and ranking tests, argmax seems fine):

 Warning: IterativeLinearSolver failed, result contains NaNs
└ @ ImplicitDifferentiation ~/.julia/packages/ImplicitDifferentiation/HJshL/src/linear_solver.jl:34

gdalle commented 1 year ago

It may also be due to the latest release of ImplicitDifferentiation. I will take a look