failing JuMP master tests #6
Comments
|
I think the |
|
CC @joehuchette |
|
That's weird, JuMP shouldn't introduce any auxiliary variables. Is this going through the MPB-level conic->quadratic wrapper? |
|
Probably, since I don't have the conic MPB API implemented here. Something is doubling the number of variables. |
|
Oh, the conic->quadratic wrapper will introduce extra variables, which JuMP isn't aware of. @tkelman, could you tweak the implementation of |
|
Providing a different number of variable types than variables seems like an iffy thing to do to a solver, but I'll make it a warning. I could have old variable types sitting around from a previous version of the problem that has been modified so I always need to set a type, but looks like if you introduce a new variable then the auxiliaries move to the end of the new problem. |
|
The conic->quadratic wrapper always calls @joehuchette, what happens if you go through the conic->quadratic interface with Gurobi, solve, and then add a new variable inside JuMP? |
|
It probably makes sense to do the conic->quad translation with a Wrapper type so that we can retain a variable mapping. This will also help with duals. |
|
I think everything passes now with JuMP and CoinOptServices both on master, let me know if you see otherwise. There are some failures with OS 2.10.0 that I still need to report, but that's still in a PR. |
|
If we figure out JuliaOpt/MathProgBase.jl#73, then this wrapper business will be easier. |
|
@tkelman, I'm still getting this error with CoinOptServices master: |
|
What versions of OS, Ipopt, and Cbc are in the respective |
|
Cbc 2.9.1, OS 2.9.2, and Ipopt 3.12.1. |
|
What distro? I can't seem to reproduce that locally on Ubuntu 14.04. Signal 11 is a segfault, so if you run just one of the tests that's failing to populate the osil file you should be able to reproduce by running the quoted |
|
This is Ubuntu 15.04. I'll try to dig into the segfault. |
Not sure, but it should be tested and should Just Work. I'll test it sometime over the next few days. |
|
Recompiled with debug symbols, here's the backtrace: The lines in question are: Will try running valgrind next |
|
Oh whoops, I actually just reproduced this in a docker container of 15.04 and sent a message to the OS mailing list (cc'ed you), but I wasn't getting a backtrace. |
|
I'm also getting an issue on win64, julia 0.3.8: |
|
On Windows the backtrace shows 14000+ levels of |
|
These problem seem pretty small, how can the depth get to 14000+? |
|
There are 16000 coefficients since JuMP does no presolve aggregation. |
|
We could (and probably should) combine coefficients inside of JuMP, but all of the other solvers seem to handle duplicate coefficients in a reasonable way. |
|
Yeah, was going to see how big the .nl file would get for this problem but the quadratic API isn't implemented in AmplNLWriter. May want to just take CoinOptServices off the quadratic solvers test list for now. Maybe only continuous if Bonmin and Couenne are working on this problem? |
|
In the docker container of 15.04 where I was able to reproduce these failures, it looks like they're fixed with OS 2.10.0 (#5). Still need to re-test separately on Windows. There are a separate set of new failures all in nonlinear tests, but at least they aren't segfaults AFAICT. |
On JuMP master, Julia 0.3.11, linux, I'm getting the following failures with OsilSolver.
The text was updated successfully, but these errors were encountered: