BLAS := open in Makefile.config doesn't work properly for the Imagenet tutorial/example

[Leslie N. Smith]

In trying the Imagenet tutorial/example, it worked fine for me when the Makefile.config contained BLAS := atlas.  However, when BLAS := open, it didn't work the same.  More specifically, the best accuracy results I've obtained so far (for testing the validation data every 1000 iterations) goes like this:

I1119 08:30:35.746019 22547 solver.cpp:298]     Test net output #0: accuracy = 0.001

I1119 08:51:27.629741 22547 solver.cpp:298]     Test net output #0: accuracy = 0.0006

I1119 09:12:23.660038 22547 solver.cpp:298]     Test net output #0: accuracy = 0.001

I1119 09:33:21.002056 22547 solver.cpp:298]     Test net output #0: accuracy = 0.001

I1119 09:54:17.218142 22547 solver.cpp:298]     Test net output #0: accuracy = 0.00092

I1119 10:15:14.281661 22547 solver.cpp:298]     Test net output #0: accuracy = 0.001

I1119 10:40:32.022413 22547 solver.cpp:298]     Test net output #0: accuracy = 0.00164

I1119 11:21:45.063426 22547 solver.cpp:298]     Test net output #0: accuracy = 0.00314

I1119 12:04:23.061851 22547 solver.cpp:298]     Test net output #0: accuracy = 0.01006

On the other hand, when the Makefile.config contained BLAS := atlas, here is an equivalent accuracy tests:

I1119 05:34:19.386816  6530 solver.cpp:298]     Test net output #0: accuracy = 0.001

I1119 05:56:12.723809  6530 solver.cpp:298]     Test net output #0: accuracy = 0.00138

I1119 06:18:02.217006  6530 solver.cpp:298]     Test net output #0: accuracy = 0.004

I1119 06:38:35.590013  6530 solver.cpp:298]     Test net output #0: accuracy = 0.0144001

I1119 06:59:07.464596  6530 solver.cpp:298]     Test net output #0: accuracy = 0.0539804

I1119 07:19:32.936643  6530 solver.cpp:298]     Test net output #0: accuracy = 0.0974601

I1119 07:40:05.831581  6530 solver.cpp:298]     Test net output #0: accuracy = 0.13342

I1119 08:00:24.890449  6530 solver.cpp:298]     Test net output #0: accuracy = 0.16336

I1119 08:20:38.903945  6530 solver.cpp:298]     Test net output #0: accuracy = 0.1846

On the other hand, I also compared open vs atlas libraries with CIFAR-10 tutorial/example and both libraries give similar results (plus openBlas is much faster because of multithreading).

Has anyone seen this before?  If so, is there a problem with my openBlas?  Any insights would be appreciated.

Thanks (in advance),

Leslie