[mlpack] make test failure

[Ryan Curtin]

On Sun, Mar 09, 2014 at 12:13:18AM -0500, Ryan Curtin wrote:
> On Thu, Mar 06, 2014 at 11:57:52AM +0800, Li Dong wrote:
> > rmbp:mlpack-1.0.8 dongli$ bin/mlpack_test -t NMFTest/SparseNMFRandomDivTest
> > Running 1 test case...
> > /Library/Caches/Homebrew/mlpack-1.0.8/src/mlpack/tests/nmf_test.cpp:239: fatal error in "SparseNMFRandomDivTest": difference{nan%} between h(i){nan} and dh(i){nan} exceeds 1.0000000000000001e-05%
> > 
> > *** 1 failure detected in test suite “MLPACKTest"
> > 
> > Sleep is more important! :)
> 
> Hello Li,
> 
> I'm sorry for the slow response.  In this case the test error is an edge
> case where the matrices approximated by NMF have elements that converge
> to zero, then there is a line where there is a division by those
> elements, so then suddenly there are NaNs everywhere.
> 
> In trunk, I've committed a fix with r16349, but for mlpack-1.0.8 you can
> either choose to ignore the error, or apply the backported patch I've
> attached.
> 
> But with either thing you choose, mlpack is still functioning the same
> on OS X as it would on Linux (note that the test would pass if it
> considered, more carefully, what to do when h(i) and dh(i) were both
> NaNs).
> 
> Let me know if I can clarify any further.

Ah, it's helpful when I base the patch in mlpack-1.0.8 not
mlpack-1.0.8/src/mlpack/methods/nmf/... try this one instead.  Sorry
about that.

-- 
Ryan Curtin    | "Get out of the way!"
ryan at ratml.org |   - Train Engineer Assistant
-------------- next part --------------
A non-text attachment was scrubbed...
Name: nan-handling.diff
Type: text/x-diff
Size: 1791 bytes
Desc: not available
URL: <http://mailman.cc.gatech.edu/pipermail/mlpack/attachments/20140309/23ef75b1/attachment-0001.diff>