Remove JuMP dependency

[blegat]

This is an attempt to simplify the code and remove the JuMP dependency. The hope is to simply use a cache and copy_to to generalize Juniper to any constraint type.

[ccoffrin]

Nice, I did not realize this was even possible!

[blegat]

Another advantage for copy_to is that it will be faster than the current one:

Juniper.jl/src/model.jl

Line 43 in eca68b3

MOI.add_constraint(backend, constr[1], constr[2])

which has type unstability issues

[codecov]

Codecov Report

Merging #207 (0c3cdd0) into master (661f586) will decrease coverage by 1.20%.
The diff coverage is 79.28%.

@@            Coverage Diff             @@
##           master     #207      +/-   ##
==========================================
- Coverage   92.19%   90.98%   -1.21%     
==========================================
  Files          22       20       -2     
  Lines        2012     1842     -170     
==========================================
- Hits         1855     1676     -179     
- Misses        157      166       +9     

Impacted Files	Coverage Δ
src/Juniper.jl	100.00% <ø> (ø)
src/printing.jl	76.47% <ø> (-2.48%)	⬇️
src/types.jl	100.00% <ø> (ø)
src/util.jl	69.33% <56.60%> (-18.44%)	⬇️
src/MOI_wrapper/MOI_wrapper.jl	84.26% <78.57%> (+1.10%)	⬆️
src/BnBTree.jl	93.71% <93.33%> (+0.27%)	⬆️
src/fpump.jl	88.60% <93.33%> (-3.40%)	⬇️
src/filter.jl	96.55% <96.55%> (ø)
src/MOI_wrapper/results.jl	80.76% <100.00%> (ø)
src/bb_inits_and_defaults.jl	100.00% <100.00%> (ø)
... and 6 more

---

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 661f586...0c3cdd0. Read the comment docs.

[blegat]

Should be ready for review :)
With this PR, Juniper should now work for any constraint type (so it should for instance work for MISDP if the nlp solver supports PSD constraints) although it probably still needs some work to support other mixed-integer sets such as Indicator, SOS, Semiinteger, ... as currently, it will try to copy it to the nonlinear solver and it will result in an UnsupportedConstraint` error. We probably just need to filter out but I'll leave that for another PR.

[Wikunia]

Thanks a lot! Overall it looks much cleaner.

Just some basic questions for now beside the one comment I've added:
Will MOI use bridges automatically for the MIP part in the feasibility pump if for example the solver doesn't support >= just <= ? That's more a general MOI question as you know that better than me 😉

It seems like we don't need the extra register for user defined functions anymore, right?
If that is actually the case first of all I'm amazed 🎉
Would be great if you can remove https://lanl-ansi.github.io/Juniper.jl/stable/options/#registered_functions::Union%7BNothing,Vector%7BRegisteredFunction%7D%7D-%5Bnothing%5D from the docs then as well.

@ccoffrin before we merge it would be in general nice to run the regression tests again to see if any problem fails or if there is speedup/slowdown.

[blegat]

Will MOI use bridges automatically for the MIP part in the feasibility pump if for example the solver doesn't support >= just <= ?

Yes, we use MOI.instantiate(solver, with_bridge_type=Float64) which adds a bridge layer will all the bridges just like you have with JuMP.

It seems like we don't need the extra register for user defined functions anymore, right?

Yes :) will remove

[ccoffrin]

@Wikunia, sure thing I'll do the full regression.

[ccoffrin]

While testing this I was wondering if JuMP can now be moved from deps to extras on the Project.toml?

[blegat]

I was wondering if JuMP can now be moved from deps to extras on the Project.toml?

Yes, done :)

[Wikunia]

I have some great news @blegat & @ccoffrin

With one processor the MOI version is quite a bit better

With four cores it is better earlier on and then they solve the same number of instances

For 8 cores it looks quite similar to the four cores one ;)

[ccoffrin]

Nice! Shall we merge and release then?

Maybe we should tag the current master before merging?

[Wikunia]

You mean current master as v0.6.6 and this as v0.7.0 ?
The current one doesn't have any new changes beside github action, right? So it's not strictly necessary I assume?

[ccoffrin]

You are correct. Nothing new on master, no v0.6.6 release is required.