Pruning dependencies

[kaandocal]

I've noticed that GalacticOptim takes ages to load after precompilation for me (more than a minute), which is not good news for an interface package. Part of this is the dependency on many packages which do not appear essential on a first glance:

• DiffEqBase: The relevant functionality seems to be NoAD and NullParameters which are now in SciML
• ModelingToolkit: I didn't find any code that uses this (maybe a test dependency?)
• Optim & Flux: Why not use @require as with the other packages? Flux in particular is a heavy package and introduces a lot of dependencies.

This PR removes DiffEqBase, marks ModelingToolkit as a test dependency and downgrades Optim and Flux to @require status. Loading GalacticOptim now takes half a minute, and most of that is due to the various differentiation packages. I propose demoting all of these to @require status (except DiffResults) so that users wishing to use GalacticOptim with eg. ForwardDiff only do not need to load 4 other packages.

Note: This version of the package does not @reexport Optim or Flux. Since this package is still in active development I hope this API breakage is acceptable.

Note 2: Since Flux does not yet have an abstract optimiser type (see here) I used a Union for now, but this is a compatibility issue.

[ChrisRackauckas]

Suggested change

	function __solve(prob::OptimizationProblem, opt, data = DEFAULT_DATA;
	function __solve(prob::OptimizationProblem, opt::AbstractFluxOptimiser , data = DEFAULT_DATA;

[ChrisRackauckas]

Looks good, can you version bump to 2.0?

[Vaibhavdixit02]

I propose demoting all of these to @require status (except DiffResults) so that users wishing to use GalacticOptim with eg. ForwardDiff only do not need to load 4 other packages.

Yeah change that as well.

[ChrisRackauckas]

Looks like this needs more test usings

[codecov]

Codecov Report

Merging #135 (63d2a90) into master (65af75c) will decrease coverage by 25.16%.
The diff coverage is 62.07%.

@@             Coverage Diff             @@
##           master     #135       +/-   ##
===========================================
- Coverage   87.30%   62.14%   -25.17%     
===========================================
  Files           3       18       +15     
  Lines         386      560      +174     
===========================================
+ Hits          337      348       +11     
- Misses         49      212      +163     

Impacted Files	Coverage Δ
src/solve/multistartoptimization.jl	0.00% <0.00%> (ø)
src/solve/quaddirect.jl	0.00% <0.00%> (ø)
src/solve_flux.jl	0.00% <0.00%> (ø)
src/solve_optim.jl	0.00% <0.00%> (ø)
src/solve/solve.jl	33.33% <33.33%> (ø)
src/solve/evolutionary.jl	79.16% <79.16%> (ø)
src/solve/blackboxoptim.jl	80.76% <80.76%> (ø)
src/solve/cmaevolutionstrategy.jl	81.81% <81.81%> (ø)
src/solve/optim.jl	87.75% <87.75%> (ø)
src/solve/flux.jl	89.18% <89.18%> (ø)
... and 25 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 65af75c...63d2a90. Read the comment docs.

[kaandocal]

Looks good, can you version bump to 2.0?

I tried, but the minibatch/diffeqflux tests print a warning and ultimately fail:

┌ Warning: Error requiring `ForwardDiff` from `LoopVectorization`                                                                                                                                                  
│   exception =                                                                                                                                                                                                    
│    LoadError: UndefVarError: AbstractSIMD not defined                                                                                                                                                            
│    Stacktrace:                                                                                                                                                                                                   
│      [1] top-level scope                                                                                                                                                                                         
│        @ ~/.julia/packages/LoopVectorization/MVMVN/src/simdfunctionals/vmap_grad_forwarddiff.jl:32
(...)

Seems like this is a versioning issue (haven't been able to track this down precisely)

[ChrisRackauckas]

@chriselrod ?

[kaandocal]

I propose demoting all of these to @require status (except DiffResults) so that users wishing to use GalacticOptim with eg. ForwardDiff only do not need to load 4 other packages.

Yeah change that as well.

I uploaded a commit for this as well. This required reorganising the code (with significantly reduced coverage although I didn't add or remove anything). Feel free to rearrange things as you please! Although many AD/optimisation backend are really not tested at all...

[chriselrod]

@chriselrod ?

Should be fixed on master. I'll issue a release in the next few hours.
The @require for adding ChainRules support for LoopVectorization.vmap causes a number of issues for a feature that I'm not sure if anyone actually uses.

[ChrisRackauckas]

The @require for adding ChainRules support for LoopVectorization.vmap causes a number of issues for a feature that I'm not sure if anyone actually uses.

Can you pick that out to a package or something?

[ChrisRackauckas]

This was rebased and merged in #140 . Next time you might want to work on a branch other than master.

Thanks!