the official ‘deb’ from NVIDIA it not well tested on latest version of ubuntu, so I would avoid packages from NVIDIA as much as I can. Also for cuDNN, it’s just a huge library file and a single header, only extra change is
LD_LIBRARY (Ubuntu didn’t ship it in official repository, I guess it’s because license)
when I learned consistent hashing in class, there is a paper called ‘Chord: A Scalable Peer-to-peer Lookup Service for Internet Applications’ in 2001, I didn’t realize there is one in 1997, though they do have a common author ‘ David Karge’
Actually you can copy the file anywhere, just make sure to update your LD_LIBRARY_PATH so
libcudnn.so can be found, i.e. I extracted the downloaded cuDNN to
/home/at15/app/cuda(it has just two folder
lib64 ) and in my
.bashrc I added `export LD_LIBRARY_PATH=/home/at15/app/cuda/lib64:$LD_LIBRARY_PATH`.
Then open a new shell,
import tensorflow as tf should work.
I run dpkg -L but it seems the only thing related with sample is man page and pdf, so I guess the deb package from Ubuntu didn’t include it.
As for the compiler, I end up with gcc 5.4 because gcc 6 has breaking changes in header files, and clang-3.8 is using those header files as well (end up having tons of error from math.h). Many library requires C+11, and I like the auto keyword.