WIP: SDP and Block structure #156

joehuchette · 2014-04-03T21:27:14Z

This pull request adds SDP functionality to JuMP. Primal, dual, and matrix constraints can be used in modeling. Currently only Mosek support is available, although that will change soon. Left remaining before a merge:

Extensive tests on types, operators, etc.
Construct useful example suite
Documentation
~~Macros for setObjective, addConstraint~~
Absolute values (Absolute values in objective/constraints #48)
Handle MatrixExpr/NormExpr in setObjective
matrix operations on MatrixVar/MatrixExpr's

I'm still not quite sure what functionality will be most useful in practice (for instance, converting Lyapanov equations seems like it would be quite handy), so this should probably be considered an experimental feature for a release cycle or two, at the very least.

mlubin · 2014-04-03T21:29:02Z

We're going to run out of features to add to JuMP soon.

joehuchette · 2014-04-04T03:23:45Z

Not a bad problem to have.

IainNZ · 2014-04-04T13:55:32Z

src/sdp.jl

+
+SDPData() = SDPData(MatrixVar[], {}, {}, String[], MatrixConstraint[],PrimalConstraint[],DualConstraint[],MatrixFuncExpr(),{})
+
+function SDPModel(;solver=MosekSolver())


Is this permanent (having a separate constructor)? Or will it eventually just be like nonlinear models?

No reason not to make it the same, I guess.

IainNZ · 2014-04-04T14:04:57Z

Very cool Joey, very very cool. Once you have an example in there I'll try and code up some of my own and can help with other aspects of it.

IainNZ · 2014-04-07T19:51:32Z

Didn't get very far :(

julia> using JuMP

julia> m=Model()
Feasibility problem with:
 * 0 linear constraints
 * 0 variables
Solver set to GLPK

julia> n = 4
4

julia> @defSDPVar(m, X[n])
ERROR: n not defined

IainNZ · 2014-04-07T19:52:10Z

then

julia> @defSDPVar(m, X[4])

julia> X
Evaluation succeeded, but an error occurred while showing value of type MatrixVar:
ERROR: m not defined
 in show at /home/idunning/.julia/v0.3/JuMP/src/sdp.jl:109
 in anonymous at show.jl:1042
 in showlimited at show.jl:1041
 in display at multimedia.jl:118
 in display at multimedia.jl:150

IainNZ · 2014-04-07T19:54:35Z

OK fixed second one

joehuchette · 2014-04-07T19:55:06Z

Thanks, I think I just broke the first by accident a few commits ago.

joehuchette · 2014-04-07T19:55:27Z

*Putting macro hat on"

IainNZ · 2014-04-07T19:55:48Z

OK, I was about to fix it, but I'll let you. Just need to escape I think

mlubin · 2014-04-07T19:57:39Z

Relatedly, some tests should be added to hygiene.jl for the new macros.

IainNZ · 2014-04-07T19:58:47Z

We should bump Mosek at some point too, those deprecation warnings on released version are a pain.
New error:

idunning@IAIN-DESKTOP:~/.../JuMP/examples$ julia maxcut_sdp.jl 
ERROR: setupSDPVar! not defined
 in solveSDP at /home/idunning/.julia/v0.3/JuMP/src/sdp_solve.jl:184

IainNZ · 2014-04-07T19:59:36Z

Ah I see the "real" function name doesn't have a !

joehuchette · 2014-04-07T20:00:25Z

Haha yeah, you caught me in the midst of some refactoring.

IainNZ · 2014-04-07T20:02:50Z

No problem, will just change locally

IainNZ · 2014-04-07T20:05:11Z

Hah maybe I should try later

ERROR: dim not defined
 in setupSDPVar at /home/idunning/.julia/v0.3/JuMP/src/sdp_solve.jl:138
 in solveSDP at /home/idunning/.julia/v0.3/JuMP/src/sdp_solve.jl:184

EDIT:
Making it

var.dim

seemed to work, actually solved it!

joehuchette · 2014-04-07T20:07:52Z

Hahaha yeah, sorry, I've been pushing broken code today trying to get SDP bounds working properly. Should've branched on my branch...

joehuchette · 2014-04-07T20:08:08Z

I'll let you know when things are fixed up later today.

IainNZ · 2014-04-07T21:39:08Z

Alright, my contribution: the Goemans-Williamson algorithm for approximating MAXCUT.
Unfortunately doesn't have any interesting contraints...

joehuchette · 2014-04-07T21:45:51Z

Very cool. I'm hoping to find a practical example that mixes in different forms of constraints (primal, dual, etc), but nothing has come to me yet...

IainNZ · 2014-04-07T21:51:43Z

Would be nice to define

getValue(X[1,3])
ERROR: no method getValue(MatrixFuncVar)

joehuchette · 2014-04-07T21:55:10Z

Yes, definitely need that

IainNZ · 2014-04-07T22:01:55Z

Changing objective sense/function after solving once doesn't seem to work (although it looks like it should at glance, so not sure whats going on). I've committed another example anyway - it fails because of this though.

joehuchette · 2014-04-07T22:09:50Z

That's very strange, the sense is definitely getting communicated through to Mosek. Any case, thanks for the other nice example!

mlubin · 2014-04-30T21:18:11Z

Just added an example, worked on the first try!

joehuchette · 2014-04-30T21:41:14Z

Haha always a good sign, thanks!

add missing field in Model constructor change behavior for x'*Q*x minor fixes work towards fixing tests

coveralls · 2014-05-25T07:13:30Z

Coverage decreased (-4.15%) when pulling b179e0e on sdp into 7197c7d on master.

coveralls · 2014-05-29T15:53:49Z

Coverage decreased (-3.93%) when pulling 4a9a50c on sdp into 7197c7d on master.

joehuchette · 2014-07-07T18:21:10Z

@mlubin @IainNZ any idea how to debug module loading speed? This branch is painfully slow compared to master.

mlubin · 2014-07-07T18:24:42Z

Given that this more than doubles the size of the JuMP source, would it make sense/be feasible to perhaps keep this in a separate package until Julia gets fast module loading?

joehuchette · 2014-07-07T18:34:03Z

It would take a bit of work, but might be feasible. For instance, things like [x] are no longer valid in the operator code, since Base.vcat(::Variable) gets overwritten. There's some stuff in here that I would really like in JuMP proper, as well.

mlubin · 2014-07-07T18:36:08Z

What stuff, specifically? If it's a relatively small chunk we can definitely merge it and put the bulk of the SDP code in a separate package.

joehuchette · 2014-07-07T18:40:05Z

The norm stuff and x'*Q*x support would be great to have in master; everything that is not really tied to SDP, essentially. The problem is that as it's currently written, it's not possible to quarantine this support (which is probably a design flaw).

IainNZ · 2014-07-07T19:40:40Z

Is it more than doubling startup time?

joehuchette · 2014-07-07T19:47:16Z

Yeah, roughly.

IainNZ · 2014-07-07T23:27:26Z

Interesting. If it was like, 4x startup time for 2x code I would have been hopeful that we could pull something off.

Can we do a conditional include when someone wants to use SDP? It'd be ugly... but its an ugly situation.

joehuchette · 2014-07-08T00:55:27Z

Without a mechanism to pass arguments to using JuMP, I'm not quite sure how to do that. You're probably right, though; it makes sense that doubling the code base will double the load time.

IainNZ · 2014-07-08T01:03:38Z

I was thinkng like

@load_JuMP_with_SDP()

or something wacky like that that would maybe do

using JuMP
include(Pkg.dir("JuMP") * "sdp.jl")

or whatever

joehuchette · 2014-08-15T18:36:41Z

I rebased this locally and I have JuMP load time going from ~4 sec on master to ~5 sec on SDP. I think that seems reasonable to me, so maybe 0.6 (or immediately after 0.6 lands) is an acceptable target for this PR.

mlubin · 2014-08-15T23:20:22Z

Not bad at all. I'd prefer to merge this after 0.6.

joehuchette · 2014-08-15T23:27:45Z

I agree. I think right after 0.6 would be ideal to try to work out as many kinks as possible before the next major release.

mlubin · 2014-12-11T18:39:17Z

I guess we can untag this for 0.7?

joehuchette · 2014-12-11T18:41:04Z

Yep.

IainNZ · 2015-06-09T01:26:47Z

🌟 sdp branch 1️⃣ gone but never forgotten 🌟

IainNZ reviewed Apr 4, 2014
View reviewed changes

joehuchette added 3 commits May 25, 2014 01:40

initial port of SDP

e7ecccb

add missing field in Model constructor change behavior for x'*Q*x minor fixes work towards fixing tests

rebase and add explicit annotations to imported functions

e435b73

tests pass (but still incomplete)

b179e0e

slight fix

4a9a50c

joehuchette removed this from the 0.6 milestone Jul 24, 2014

joehuchette added this to the 0.7 milestone Aug 15, 2014

joehuchette removed this from the 0.7 milestone Dec 11, 2014

joehuchette closed this Jun 9, 2015

joehuchette deleted the sdp branch September 5, 2015 15:55

blegat pushed a commit that referenced this pull request Oct 16, 2018

Grep for parameters starting with 'CPXPARAM' (#156)

8d84c86


		SDPData() = SDPData(MatrixVar[], {}, {}, String[], MatrixConstraint[],PrimalConstraint[],DualConstraint[],MatrixFuncExpr(),{})

		function SDPModel(;solver=MosekSolver())

WIP: SDP and Block structure #156

WIP: SDP and Block structure #156

Conversation

joehuchette commented Apr 3, 2014

mlubin commented Apr 3, 2014

joehuchette commented Apr 4, 2014

IainNZ Apr 4, 2014

Choose a reason for hiding this comment

joehuchette Apr 4, 2014

Choose a reason for hiding this comment

IainNZ commented Apr 4, 2014

IainNZ commented Apr 7, 2014

IainNZ commented Apr 7, 2014

IainNZ commented Apr 7, 2014

joehuchette commented Apr 7, 2014

joehuchette commented Apr 7, 2014

IainNZ commented Apr 7, 2014

mlubin commented Apr 7, 2014

IainNZ commented Apr 7, 2014

IainNZ commented Apr 7, 2014

joehuchette commented Apr 7, 2014

IainNZ commented Apr 7, 2014

IainNZ commented Apr 7, 2014

joehuchette commented Apr 7, 2014

joehuchette commented Apr 7, 2014

IainNZ commented Apr 7, 2014

joehuchette commented Apr 7, 2014

IainNZ commented Apr 7, 2014

joehuchette commented Apr 7, 2014

IainNZ commented Apr 7, 2014

joehuchette commented Apr 7, 2014

mlubin commented Apr 30, 2014

joehuchette commented Apr 30, 2014

coveralls commented May 25, 2014

coveralls commented May 29, 2014

joehuchette commented Jul 7, 2014

mlubin commented Jul 7, 2014

joehuchette commented Jul 7, 2014

mlubin commented Jul 7, 2014

joehuchette commented Jul 7, 2014

IainNZ commented Jul 7, 2014

joehuchette commented Jul 7, 2014

IainNZ commented Jul 7, 2014

joehuchette commented Jul 8, 2014

IainNZ commented Jul 8, 2014

joehuchette commented Aug 15, 2014

mlubin commented Aug 15, 2014

joehuchette commented Aug 15, 2014

mlubin commented Dec 11, 2014

joehuchette commented Dec 11, 2014

IainNZ commented Jun 9, 2015