Caffe on GPU1 and GPU2

[Predrag Punosevac]

Dear Autonians,

In-spite of beautiful how-tos posted by Arne Suppe, Dougal Sutherland,
and private notes shared by Benedikt Boecking few people still complain
that Caffe deep learning framework is "not available" on GPU machines. 

Well I just built it myself and tested it including pycaffe and matcaffe
(MATLAB interface). I have not built cudnn. You will have to register
and build it yourself.

The binaries are located in 

/opt/caffe/build/tools

If you get an error that CUDA Libraries are not available you will need
to set explicitly LD_LIBRARY_PATH variable in your shell profile. The
same goes for binaries itself. Please adjust your shell profiles
accordingly (no hand-holding this time).

My "official" version of Caffe is built using ATLAS library. Already
mentioned tutorials explain why we used ATLAS. It is also possible to
build it using Open-BLAS. Please no complains about this choice. You can
build in your directory whatever you fancy.

I also played with Caffe myself little bit. That thing requires serious
reading of documentation before you can do anything with it even if you
are an expert in machine learning. If you come to my office for help you
will be quizzed on existing documentation before getting any help :)

Cheers,
Predrag