Support higher-dimensional arrays in dot
#85
Conversation
Codecov Report
@@ Coverage Diff @@
## master #85 +/- ##
=======================================
Coverage 97.86% 97.86%
=======================================
Files 9 9
Lines 515 515
=======================================
Hits 504 504
Misses 11 11
Continue to review full report at Codecov.
|
test/functions.jl
Outdated
| show(buf, x ⋅ w) | ||
| @test String(take!(buf)) == "0.1 * x1 + 0.3 * x2 + 0.2 * x3 + 0.4 * x4 + 0.0" | ||
| show(buf, w ⋅ x) | ||
| @test String(take!(buf)) == "0.1 * x1 + 0.3 * x2 + 0.2 * x3 + 0.4 * x4 + 0.0" |
There was a problem hiding this comment.
Not sure I like involving show in this test. So far I've usually just evaluated the expressions with random variable values and ensured that the results are the same. But I think we should just overload == (and hash) so that 0.1 * x1 + 0.2 * x2 + 0.0 == 0.2 * x2 + 0.1 * x1 + 0.0 == 0.1 * x1 + 0.1 * x2 + 0.1 * x2 + 0.0. What do you think?
There was a problem hiding this comment.
Once #86 is merged, could you use that here instead?
|
The MathOptInterface wrappers for Gurobi (and other LQOI-based solvers) currently are just quite inefficient. Again, if your problem doesn't involve integer variables, try OSQP. If not (or maybe, regardless), you'll have to open issues with the MOI wrapper packages. The zero allocation results in the readme are of course predicated on having a solver interface that doesn't allocate when you do problem modification. |
|
Done. I keep getting fooled by the Q in OSQP, forgetting that of course it can solve linear problems too. It's definitely better than Gurobi or GLPK. Though I've now benchmarked it against https://github.com/robertfeldt/EarthMoversDistance.jl.git and getting results like these:
So I'll probably be using the dedicated solver. |
|
Great, thanks. Yeah, OSQP is definitely best when the cost function Hessian is positive definite, and a proper dedicated LP algorithm is always going to win. |
|
Interestingly, at least for the small number of bins I think OSQP beat the linear solvers. But everyone got beat by the specialized implementation that takes advantage of particular features of "transportation problems." |
I'm implementing the earth mover distance and the most natural representation uses a matrix of
Variables and requires taking its dot-product with some weights.Incidentally, in profiling I'm finding there's still quite a lot of allocation, which I'm guessing is a consequence of this line. Any thoughts on how to get around it? Also this line is insanely slow. Any tips for overcoming such problems?