Train Lenet on MNIST: Check failed: error == cudaSuccess (11 vs. 0) invalid argument

[Majid Azimi]

Hi,

I am trying to install caffe on a 32bit  linux(Debian). version of CUDA is 5.5 (I don't have root priviledge to update it)

after successfully "make all" caffe in training Lenet on MNIST, it gives the following error:


I0121 23:01:30.986553 27837 solver.cpp:294] Iteration 0, Testing net (#0)
F0121 23:01:30.986930 27837 math_functions.cpp:91] Check failed: error == cudaSuccess (11 vs. 0)  invalid argument

by setting CPU rather GPU this problem is solved and  lenet is successfully trained on MNIST.

anaconda was used to install caffe.


is the problem the version of CUDA?
if so, is there anyway to to install recent version of CUDA without root permission?(based on my searhes it's not possible)


not using anaconda in making caffe, encounter the following error:

struct cudaDeviceProp' has no member named 'isMultiGpuBoard error on the recent version.


Thanks already for your constructive help

[Jan C Peters]

Yes it is possible to install the CUDA SDK to a local directory. If you download the binary installer from NVidia it will ask you for a path where to install it. If you choose a directory you have full access to (probably somewhere in your home folder) installation should work fine. Of course you need to make the headers, libraries and binaries accessible to your system (e.g. give a corresponding -I to gcc for compilation, or set LD_LIBRARY_PATH to the cuda libs dir when running a CUDA application). The only thing you cannot change without root permissions is the installed nvidia driver. You can do a CUDA device query to see that version. If that version is too low (driver too old) it could be a problem. But then you can talk to your administrator to get a recent version installed.

Jan

[S.Majid Azimi]

Hi Jan,

I gave it a try. when I enable --silent --driver flags. I get the message saying that it is not compatible with the architechture of system, but I don't get it. I'm trying to install the nvidia's driver as well plus cuda 6.5 32 bit toolkit in a  local path.

best

Majid

--
You received this message because you are subscribed to the Google Groups "Caffe Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to caffe-users...@googlegroups.com.
To post to this group, send email to caffe...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/caffe-users/e50ed5c1-025c-4066-8923-52784e7a8af7%40googlegroups.com.

For more options, visit https://groups.google.com/d/optout.

[Jan C Peters]

As I said, you can install the toolkit into an arbitrary location, but NOT the driver! The driver essentially consists of a kernel module which of course can not be put anywhere. To separately install driver, toolkit and samples, run the binary installer with the --extract flag, to break it into three separate installers for these respective components. And you cannot install the driver without root permissions.

Jan

[S.Majid Azimi]

thanks Jan, got it. I installed it successfully.

To view this discussion on the web visit https://groups.google.com/d/msgid/caffe-users/7baf9e95-38d2-4deb-803e-aee77347be27%40googlegroups.com.