Optimize twice

[laradicp]

No description provided.

[guimarqu]

Let's focus on reformulation with direct_model = true, it looks like there is a bug because we must reinitialize parts of the Annotations data structure. This structure is used to store to which subproblem the constraints and variables belong. It is initialized when we call Coluna.Optimizer. Then, it is populated when BlockDecomposition assigns variables and constraints to subproblems, for instance :

Coluna.jl/src/MOIwrapper.jl

Lines 378 to 385 in 4d32761

	function MOI.set(
	model::Coluna.Optimizer, ::BD.VariableDecomposition, varid::MOI.VariableIndex,
	annotation::BD.Annotation
	)
	store!(model.annotations, annotation, model.vars[varid])
	return
	end

and it is populated a last time when Coluna does the reformulation (calls to store! in buildformulation).

To fix the bug, you must keep what is stored in MOI wrapper (because the model passes through MOI only once when direct_model = true) and erase what is stored in reformulate!. I think only ann_per_form needs to be reinitialized.

[guimarqu]

I think for the first (with no reformulation and direct_model=true), when you write the model, the original formulation is stored in the formulation and the buffer of the subsolver. When you call optimize!, the content of the buffer is pushed in the subsolver and then erased. When you call optimize! a second time, the buffer is empty and thus the subsolver solves an empty formulation.

[codecov]

Codecov Report

Merging #462 (4de9806) into master (8d13bfd) will decrease coverage by 0.05%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master     #462      +/-   ##
==========================================
- Coverage   80.90%   80.84%   -0.06%     
==========================================
  Files          51       51              
  Lines        4351     4354       +3     
==========================================
  Hits         3520     3520              
- Misses        831      834       +3     

Impacted Files	Coverage Δ
src/decomposition.jl	98.50% <100.00%> (ø)
src/optimize.jl	93.84% <100.00%> (+0.40%)	⬆️
src/Algorithm/branching/branchinggroup.jl	86.51% <0.00%> (-1.13%)	⬇️
src/Algorithm/branching/branchingalgo.jl	90.75% <0.00%> (-0.85%)	⬇️
src/Algorithm/storage.jl	73.77% <0.00%> (-0.82%)	⬇️
src/MOIcallbacks.jl	94.23% <0.00%> (-0.11%)	⬇️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 8d13bfd...4de9806. Read the comment docs.

[guimarqu]

Since optimize!(model) returns nothing, you don't test if it's really working. You should retrieve the objective value & the solution of the two runs and compare them.

For the reformulation, you can use the generalized assignment (using the call to ColunaDemos such as in full instance tests).