tensorflow gpu support

scomme · 2017-09-16 10:37:23

I have problem with tensorflow gpu support: I installed cuda toolkit 8.0 and cudnn6 through package manager in /opt/cuda/…
I need tensorflow for keras framework…when I build the project the output is: ImportError: libcuda.so.1: cannot open shared object file: No such file or directory
In my bashrc I have:
export PATH=/opt/cuda/bin:$PATH export LD_LIBRARY_PATH=/opt/cuda/lib64:$LD_LIBRARY_PATH
export CUDA_HOME=/opt/cuda
how can i resolve?

Trilby · 2017-09-16 11:35:04

How are you builing this software? What command(s) are you using? Do you know if bashrc is even sourced in the process that is doing the build? Are those variables being over-written by something in the build? Can you echo those variables right before the actual compiler command to see what they are in the build process? Are you using nvcc or gcc? If gcc, are you using gcc5 or trying to use gcc6?

Last edited by Trilby (2017-09-16 11:37:41)

scomme · 2017-09-16 12:29:21

I use tensorflow for keras framework, i'm using Pycharm IDE and I build my project it with that, for installing tensorflow I used anaconda following the guide on tensorflow and I installed tensorflow in /usr/bin/anaconda3/..
I don't know if the variable are overwritten...how can i see it?

Slithery · 2017-09-16 13:24:36

scomme wrote:

for installing tensorflow I used anaconda following the guide on tensorflow and I installed tensorflow in /usr/bin/anaconda3/..

Why?

The tensorflow package is in the official repositories, by using anaconda your likely to run into all sorts of problems.
Did you also install cuda and cudann6 without using pacman?

Last edited by Slithery (2017-09-16 13:26:30)

Trilby · 2017-09-16 14:07:25

scomme wrote:

I don't know if the variable are overwritten...how can i see it?

I have no idea in that setup.

In addition to Slithery's point, in general if you want help troubleshooting you should make an effort to reduce the number of relevant variables yourself first. This might be a problem with the anaconda installed tensorflow, it might be a problem with your cuda install, it might be a problem with the compiler you are using, it might be a problem with what pycharm uses for compiler settings. It might be a dozen other things.

Rather than having us try to ask about every single one of these, see what you can rule out. Build the project directly without all these middle-man tools (IDE, etc) and report what build commands were used and what the output was.

According to the wiki, it looks like the default gcc would not be expected to build such software properly. You either need to use gcc5, or nvcc. Have you configured pycharm to use one of these, or is it using the default gcc?

Last edited by Trilby (2017-09-16 14:08:45)

Arch Linux

#1 2017-09-16 10:37:23

tensorflow gpu support

#2 2017-09-16 11:35:04

Re: tensorflow gpu support

#3 2017-09-16 12:29:21

Re: tensorflow gpu support

#4 2017-09-16 13:24:36

Re: tensorflow gpu support

#5 2017-09-16 14:07:25

Re: tensorflow gpu support

Board footer