encountered a bad mini-batch that made the loss value suddenly jump up by a
larger than normal value it could make the trainer think we converged.  Now the
test is robust to recent spikes in loss value.

This was in some of the svd routines.  However, we already had a svd routine
that used a separate svd code that is better than the NRIC derived version.  So
that's what we use everywhere now.

Now PYTHON_INCLUDE_DIR and PYTHON_LIBRARY setting up correctly for both 32/64 bit architectures

imagenet examples compile.

it simply partitions the data and is an invertible operation.

avoid compiler warnings.

if the input xml file contained unannotated images.

Fixed cudnn_hint_path detection for OSX where CUDA_CUBLAS_LIBRARIES is a list, not a single string. (#230)

let an exception sit in a thread pool and don't handle it before the
thread pool is destroyed.

training thread) out of the object into the calling code rather than
terminating the application.

Detecting C++11 capabilities did not work on OSX 10.11.6, XCode 7.3.1, Apple LLVM 7.3.0, cmake 3.6.1. First check revealed that CMAKE_CXX_KNOWN_FEATURES was empty, so subsequent matches failed. After reading http://public.kitware.com/pipermail/cmake/2014-November/059182.html changed handling to fetch property instead of relying on variable. Works on my machine.