[mlpack-svn] [MLPACK] #300: allknn fails for mnist8m dataset

[MLPACK Trac]

#300: allknn fails for mnist8m dataset
----------------------+-----------------------------------------------------
  Reporter:  rozyang  |        Owner:  rcurtin 
      Type:  defect   |       Status:  accepted
  Priority:  major    |    Milestone:          
 Component:  mlpack   |   Resolution:          
  Keywords:           |     Blocking:          
Blocked By:  302      |  
----------------------+-----------------------------------------------------

Comment (by rcurtin):

 Okay, digging into this I think I may have isolated the issue.  Here is
 what I think happened:

 I suggested that you add `#define ARMA_64BIT_WORD` to your program.  This
 was successful and you were able to rebuild the program just fine and now
 the matrix loaded.  But `libmlpack.so.1` is compiled without
 `ARMA_64BIT_WORD`, meaning that the structs `arma::vec` (for instance) in
 `libmlpack.so.1` and `allknn` are different.  So when
 `NearestNeighborSort::InsertDistance()` (which is in `libmlpack.so.1` and
 not `allknn`) is called, everything is misaligned and a segfault happens.

 So, I should not have recommended that you add `#define ARMA_64BIT_WORD`
 to the top of your program, and I'll remember not to make that suggestion
 in the future.  Sorry about that.  (That *can* work in some situations,
 but those situations are specifically when your program happens to only
 use templated functions and classes and nothing at all that has been
 compiled into `libmlpack.so`.)

 Can you try recompiling mlpack against a version of Armadillo with
 ARMA_64BIT_WORD enabled?  You can do this by changing
 `include/armadillo_bits/config.hpp` and uncommenting the line `// #define
 ARMA_64BIT_WORD`.  I think this will solve your problem; can you let me
 know if it does?  If not, I'll keep digging... :)

-- 
Ticket URL: <http://trac.research.cc.gatech.edu/fastlab/ticket/300#comment:14>
MLPACK <www.fast-lab.org>
MLPACK is an intuitive, fast, and scalable C++ machine learning library developed at Georgia Tech.