[mlpack-svn] [mlpack] add a test case for matrix sensing SDP (#373)

[Ryan Curtin]

Looks good to me.  I should buy you lunch for writing more comprehensive tests for the code I originally wrote (and also making it work for something useful!).  Two comments, neither of which need to be entirely addressed before I merge it:

Can you add the description of the test to the pull request as a comment before the test?  To me, at least, it's non-obvious what the test is doing based on the name; I'm not entirely an expert in the matrix recovery area.

The second thought is that on my i7 machine in debug mode, the test takes 90 seconds (so, probably a few minutes on older machines).  This isn't too terribly long (especially since CFTest can take up to an hour (#354)...), but one of the things I notice while watching it is that the LBFGS iterations converge most of the way to the optimum in a handful of iterations, but will spend hundreds of iterations before the convergence criterion is met.  I wonder, if the convergence tolerance was loosened, would LRSDP run much faster and still provide adequate solutions?  I doubt that I'll have the time to dig too deep into that, but if you are thinking that the solver is taking too long, I bet that's a good place to start looking to accelerate it.

---
Reply to this email directly or view it on GitHub:
https://github.com/mlpack/mlpack/pull/373#issuecomment-68667940
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mailman.cc.gatech.edu/pipermail/mlpack-git/attachments/20150104/7d5c9a70/attachment.html>