Add support for starting values #111

blegat · 2018-09-27T15:08:00Z

@ccoffrin Can you test it and report whether it works as expected for your use case ?

Closes #109

ccoffrin · 2018-09-27T19:29:03Z

Using JuMP master, MOI v0.6.1 and this branch I am getting,

ERROR: MethodError: no method matching allocate(::SCS.Optimizer, ::MathOptInterface.ObjectiveFunction{MathOptInterface.ScalarQuadraticFunction{Float64}}, ::MathOptInterface.ScalarQuadraticFunction{Float64})
Closest candidates are:
  allocate(::MathOptInterface.Utilities.AbstractModel, ::Any...) at /home/.julia/v0.6/MathOptInterface/src/Utilities/model.jl:339
  allocate(::MathOptInterface.Utilities.MockOptimizer, ::MathOptInterface.ObjectiveFunction, ::Any) at /home/.julia/v0.6/MathOptInterface/src/Utilities/mockoptimizer.jl:285
  allocate(::MathOptInterface.Utilities.MockOptimizer, ::Union{MathOptInterface.AbstractConstraintAttribute, MathOptInterface.AbstractModelAttribute, MathOptInterface.AbstractOptimizerAttribute, MathOptInterface.AbstractVariableAttribute}, ::Any) at /home/.julia/v0.6/MathOptInterface/src/Utilities/mockoptimizer.jl:284
  ...
Stacktrace:
 [1] _pass_attributes(::SCS.Optimizer, ::MathOptInterface.Utilities.UniversalFallback{JuMP.JuMPMOIModel{Float64}}, ::Bool, ::MathOptInterface.Utilities.IndexMap, ::Array{MathOptInterface.AbstractModelAttribute,1}, ::Tuple{}, ::Tuple{}, ::Tuple{}, ::MathOptInterface.Utilities.#allocate) at /home/.julia/v0.6/MathOptInterface/src/Utilities/copy.jl:59
 [2] allocate_load(::SCS.Optimizer, ::MathOptInterface.Utilities.UniversalFallback{JuMP.JuMPMOIModel{Float64}}, ::Bool) at /home/.julia/v0.6/MathOptInterface/src/Utilities/copy.jl:248
 [3] (::MathOptInterface.#kw##copy_to)(::Array{Any,1}, ::MathOptInterface.#copy_to, ::SCS.Optimizer, ::MathOptInterface.Utilities.UniversalFallback{JuMP.JuMPMOIModel{Float64}}) at ./<missing>:0
 [4] attachoptimizer!(::MathOptInterface.Utilities.CachingOptimizer{MathOptInterface.AbstractOptimizer,MathOptInterface.Utilities.UniversalFallback{JuMP.JuMPMOIModel{Float64}}}) at /home/.julia/v0.6/MathOptInterface/src/Utilities/cachingoptimizer.jl:125
 [5] optimize!(::MathOptInterface.Utilities.CachingOptimizer{MathOptInterface.AbstractOptimizer,MathOptInterface.Utilities.UniversalFallback{JuMP.JuMPMOIModel{Float64}}}) at /home/.julia/v0.6/MathOptInterface/src/Utilities/cachingoptimizer.jl:158
 [6] optimize!(::MathOptInterface.Bridges.LazyBridgeOptimizer{MathOptInterface.Utilities.CachingOptimizer{MathOptInterface.AbstractOptimizer,MathOptInterface.Utilities.UniversalFallback{JuMP.JuMPMOIModel{Float64}}},MathOptInterface.Bridges.AllBridgedConstraints{Float64}}) at /home/.julia/v0.6/MathOptInterface/src/Bridges/bridgeoptimizer.jl:73
 [7] #optimize!#92(::Bool, ::Function, ::JuMP.Model, ::Void) at /home/.julia/v0.6/JuMP/src/optimizer_interface.jl:65
 [8] macro expansion at ./util.jl:378 [inlined]
 [9] optimize!(::PowerModels.GenericPowerModel{PowerModels.SDPWRMForm}) at /home/.julia/v0.6/PowerModels/src/core/base.jl:126
 [10] #solve_generic_model#83(::Function, ::Function, ::PowerModels.GenericPowerModel{PowerModels.SDPWRMForm}, ::SCS.Optimizer) at /home/.julia/v0.6/PowerModels/src/core/base.jl:201
 [11] (::PowerModels.#kw##solve_generic_model)(::Array{Any,1}, ::PowerModels.#solve_generic_model, ::PowerModels.GenericPowerModel{PowerModels.SDPWRMForm}, ::SCS.Optimizer) at ./<missing>:0
 [12] #run_generic_model#80(::Function, ::Array{Any,1}, ::Function, ::Dict{String,Any}, ::Type{T} where T, ::SCS.Optimizer, ::Function) at /home/.julia/v0.6/PowerModels/src/core/base.jl:171
 [13] run_generic_model(::Dict{String,Any}, ::Type{T} where T, ::SCS.Optimizer, ::Function) at /home/.julia/v0.6/PowerModels/src/core/base.jl:169
 [14] #run_generic_model#79(::Array{Any,1}, ::Function, ::String, ::Type{T} where T, ::SCS.Optimizer, ::Function) at /home/.julia/v0.6/PowerModels/src/core/base.jl:164
 [15] #run_opf#325 at /home/.julia/v0.6/PowerModels/src/prob/opf.jl:15 [inlined]
 [16] run_opf(::String, ::Type{T} where T, ::SCS.Optimizer) at /home/.julia/v0.6/PowerModels/src/prob/opf.jl:15

blegat · 2018-09-27T20:22:10Z

SCS does not support quadratic objective function. We might create a bridge that transforms into SOC constraint in MOI but this is not done yet. No change should be done in SCS for that, this is a missing feature in MOI.

blegat · 2018-09-27T20:24:50Z

Once we have a bridge to move the objective to a constraint as described in jump-dev/MathOptInterface.jl#529, it remains to transform to quadratic constraint to a SOC constraint but there is an open PR for that:
jump-dev/MathOptInterface.jl#483

mlubin · 2018-09-27T20:24:58Z

The MethodError is still a bug via http://www.juliaopt.org/JuMP.jl/latest/style.html#User-facing-MethodError-1 though. (Not related to this PR though.)

mlubin · 2018-09-27T20:26:29Z

src/MOIWrapper.jl

@@ -259,13 +260,46 @@ function MOIU.load_variables(optimizer::Optimizer, nvars::Integer)
    b = zeros(m)
    c = zeros(nvars)
    optimizer.data = ModelData(m, nvars, I, J, V, b, 0., c)
+    if length(optimizer.sol.primal) != nvars


What could the length be if not nvars? Would isempty also work as a test? Same below.

sol contains the solution of the previous optimization. If the user has added new variables, then it is not nvars

Does this interface allow adding variables and/or constraints?

You can copy a model with a different number of variables/constraints. Normally, they will arrive in the same order.

Could this if be replaced with resize! then?

ccoffrin · 2018-09-27T21:24:32Z

I think there may be a deeper issue here. As far as I can tell PowerModels is posting a linear objective in the JuMP model. See this code, https://github.com/lanl-ansi/PowerModels.jl/blob/moi/src/core/objective.jl#L104

Is it possible that JuMP is interpreting the objective as quadratic when it is linear?

mlubin · 2018-09-28T00:38:38Z

Do we have a test for this functionality anywhere?

blegat · 2018-09-28T07:02:48Z

Yes, it is tested:
https://github.com/JuliaOpt/JuMP.jl/blob/500945cac4949dfd9e30cb0ca68450f19f4ff34d/test/objective.jl#L24
and
https://github.com/JuliaOpt/JuMP.jl/blob/500945cac4949dfd9e30cb0ca68450f19f4ff34d/src/aff_expr.jl#L317

ccoffrin · 2018-09-28T14:21:31Z

Do you have a test were the model has conic constraints and a linear objective?

ccoffrin · 2018-09-28T16:14:29Z

In any case, this PR seems to have resolved #109, should we move this later point to a separate thread?

blegat · 2018-09-28T17:00:24Z

Yes we can open an issue on JuMP, if you could produce a MWE producing a quadratic objective when it should be linear, it would help. It is probably because of a promotion that shouldn't happen.
For instance if you do

@objective(model, [1, 2]' * [x, y])

if the product is handled by a function that start by promoting the two vectors to the same type, it will promote them to affine expressions and hence the result would be quadratic.
I would guess that you have somthing similar.

mlubin · 2018-09-28T17:06:16Z

src/MOIWrapper.jl

@@ -259,13 +260,46 @@ function MOIU.load_variables(optimizer::Optimizer, nvars::Integer)
    b = zeros(m)
    c = zeros(nvars)
    optimizer.data = ModelData(m, nvars, I, J, V, b, 0., c)
+    if length(optimizer.sol.primal) != nvars


Does this interface allow adding variables and/or constraints?

blegat · 2018-09-28T17:22:52Z

@ccoffrin Can you retest now that a7591c6 has been added.
I forgot to set warm_start to SCS_solve so it was using zero as warm start instead of using the provided value.
Now you should get a better result

ccoffrin · 2018-09-28T23:26:13Z

@blegat, I confirmed on the latest commit on this branch I still have the issue with the objective function being quadratic. Probably the starting value is improved; but I don't know yet.

On a side note, I never used to provide starting values to SCS in PowerModels. Now I am because of the all or nothing start value semantics.

ccoffrin · 2018-09-29T00:11:32Z

@blegat good news, I found a bug in my code, which was the cause of the quadratic objective function. That is fixed now.

Now I am getting this error,

ERROR: MethodError: no method matching _allocate_constraint(::SCS.ConeData, ::MathOptInterface.VectorAffineFunction{Float64}, ::MathOptInterface.PositiveSemidefiniteConeSquare)
Closest candidates are:
  _allocate_constraint(::SCS.ConeData, ::Any, ::Union{MathOptInterface.EqualTo, MathOptInterface.Zeros}) at home/.julia/v0.6/SCS/src/MOIWrapper.jl:100
  _allocate_constraint(::SCS.ConeData, ::Any, ::Union{MathOptInterface.GreaterThan, MathOptInterface.LessThan, MathOptInterface.Nonnegatives, MathOptInterface.Nonpositives}) at home/.julia/v0.6/SCS/src/MOIWrapper.jl:106
  _allocate_constraint(::SCS.ConeData, ::Any, ::MathOptInterface.SecondOrderCone) at home/.julia/v0.6/SCS/src/MOIWrapper.jl:112
  ...
Stacktrace:
 [1] allocate_constraint at home/.julia/v0.6/SCS/src/MOIWrapper.jl:132 [inlined]
 [2] allocate_constraints(::SCS.Optimizer, ::MathOptInterface.Utilities.UniversalFallback{JuMP.JuMPMOIModel{Float64}}, ::Bool, ::MathOptInterface.Utilities.IndexMap, ::Type{MathOptInterface.VectorAffineFunction{Float64}}, ::Type{MathOptInterface.PositiveSemidefiniteConeSquare}) at home/.julia/v0.6/MathOptInterface/src/Utilities/copy.jl:205
 [3] allocate_load(::SCS.Optimizer, ::MathOptInterface.Utilities.UniversalFallback{JuMP.JuMPMOIModel{Float64}}, ::Bool) at home/.julia/v0.6/MathOptInterface/src/Utilities/copy.jl:253

Recommendations?

blegat · 2018-09-29T00:23:29Z

This cone is not supported by SCS and should be bridged. It should be bridged automatically if you use JuMP master with MOI v0.6.1.
You should also normally receive the error message of jump-dev/JuMP.jl#1476 instead of this cryptic method error.

ccoffrin · 2018-09-29T05:36:38Z

This works with SCS in JuMP v0.18 / MPB, ```julia WR = @variable(m, [1:n, 1:n], Symmetric) WI = @variable(m, [1:n, 1:n]) SDconstraint(pm.model, [WR WI; -WI WR] >= 0) ``` What would be the translation of this into JuMP v0.19 / MOI?

blegat · 2018-09-29T08:41:10Z

What would be the translation of this into JuMP v0.19 / MOI?

Same thing.

I just realized that you might get the error you give in #111 (comment) if you do not provide the solver in the Model constructor but rather in the optimize! call. Is that the case ?

ccoffrin · 2018-09-29T22:06:27Z

To finish up this discussion. You are correct, specifying the optimizer in the JuMP model constructor resolved the most recent issue. The only remaining issue I have at the moment is with setting the solver parameters (i.e. #113). In any case, this starting point fix is working great for me.

Add support for starting values

d822246

blegat mentioned this pull request Sep 27, 2018

Objective function bridge jump-dev/MathOptInterface.jl#529

Closed

mlubin reviewed Sep 27, 2018

View reviewed changes

blegat mentioned this pull request Sep 27, 2018

Nice error for unsupported objective jump-dev/JuMP.jl#1500

Closed

Add comment

a7591c6

mlubin approved these changes Sep 28, 2018

View reviewed changes

blegat merged commit fa37cd9 into master Sep 30, 2018

blegat mentioned this pull request Jan 14, 2019

Add support for VariablePrimalStart and ConstraintDualStart jump-dev/ECOS.jl#79

Closed

odow deleted the bl/start branch October 8, 2020 01:09

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add support for starting values #111

Add support for starting values #111

blegat commented Sep 27, 2018 •

edited

Loading

ccoffrin commented Sep 27, 2018

blegat commented Sep 27, 2018

blegat commented Sep 27, 2018

mlubin commented Sep 27, 2018

mlubin Sep 27, 2018

blegat Sep 28, 2018

mlubin Sep 28, 2018

blegat Sep 28, 2018

mlubin Sep 28, 2018

ccoffrin commented Sep 27, 2018

mlubin commented Sep 28, 2018

blegat commented Sep 28, 2018

ccoffrin commented Sep 28, 2018

ccoffrin commented Sep 28, 2018

blegat commented Sep 28, 2018

mlubin Sep 28, 2018

blegat commented Sep 28, 2018

ccoffrin commented Sep 28, 2018

ccoffrin commented Sep 29, 2018

blegat commented Sep 29, 2018

ccoffrin commented Sep 29, 2018 via email •

edited by blegat

Loading

blegat commented Sep 29, 2018

ccoffrin commented Sep 29, 2018

Add support for starting values #111

Add support for starting values #111

Conversation

blegat commented Sep 27, 2018 • edited Loading

ccoffrin commented Sep 27, 2018

blegat commented Sep 27, 2018

blegat commented Sep 27, 2018

mlubin commented Sep 27, 2018

mlubin Sep 27, 2018

Choose a reason for hiding this comment

blegat Sep 28, 2018

Choose a reason for hiding this comment

mlubin Sep 28, 2018

Choose a reason for hiding this comment

blegat Sep 28, 2018

Choose a reason for hiding this comment

mlubin Sep 28, 2018

Choose a reason for hiding this comment

ccoffrin commented Sep 27, 2018

mlubin commented Sep 28, 2018

blegat commented Sep 28, 2018

ccoffrin commented Sep 28, 2018

ccoffrin commented Sep 28, 2018

blegat commented Sep 28, 2018

mlubin Sep 28, 2018

Choose a reason for hiding this comment

blegat commented Sep 28, 2018

ccoffrin commented Sep 28, 2018

ccoffrin commented Sep 29, 2018

blegat commented Sep 29, 2018

ccoffrin commented Sep 29, 2018 via email • edited by blegat Loading

blegat commented Sep 29, 2018

ccoffrin commented Sep 29, 2018

blegat commented Sep 27, 2018 •

edited

Loading

ccoffrin commented Sep 29, 2018 via email •

edited by blegat

Loading