Package design

[gdalle]

• Simplify code
• Keep terminology experience/imitation? or losses/layers?

Layers and losses

• Callable structs
• Passing kwargs possible
• Fields:
  • solver
  • solution_cache (instance-specific?)
  • hyperparameters for black box / perturbation / regularization
  • parallelization (within an instance?)
  • identify objective and constraint type?
• Modules InferOpt.Layers and InferOpt.Losses
• Probabilistic unification might be debatable, especially for regularization. Useful for pushforward when cost $c$ is not differentiable or defined inside the polytope.
• Rebase all Perturbed on the same infrastructure? Some fit inside FYL and some not (generic computation?)
• Rebase everything on GeneralizedMaximizer to be linear in the predicted parameter only?
• PerturbedOracle is the only non-ILP tool we have: put it elsewhere with reparametrization trick, DICE (mini Storchastic) and focus InferOpt on ILP?

Tests

• With neural networks: only on one problem
• Black box: check piecewise linear interpolation
• Perturbed: on a known differentiable function, check that the behavior as $\epsilon \to 0$ and $M \to \infty$ is expected
• Regularized: check against DiffOpt
• Check rrule with finite differences (assuming the forward pass is coherent)

Benchmarks

• Reimplement the instance generators, solvers and ML algorithms from the "Decision-Focused Learning" review