status == CUDNN_STATUS_SUCCESS (3 vs. 0) CUDNN_STATUS_BAD_PARAM about openpose

[RainBoltz]

since this error message often appears in other works(experienced):
1. if gpu compute capability is under 3.0, maybe try not to compile with cudnn (comment USE_CUDNN in Makefile.config, both openpose and caffe) (updated)
2. maybe the GPU is out of use, just close the processes that are using GPU

[gineshidalgo99]

For @RainBoltz comment, thank you for point 2! About point 1, OpenPose must be compiled with cuDNN support (by default it is enabled in both OpenPose and the Caffe we provide in 3rdparty/caffe). If you disable it, Caffe would ask for more than 12 GB of GPU memory for our model.

@felipheggaliza Which Ubuntu, CUDA and cuDNN versions are you using? The script only works for CUDA 8, otherwise you must use the manual steps to install it: doc/installation.md. Thanks!

[vermicelli]

I got this error too. I'm using cuda8.0 and cudnn5.1.10.

[vermicelli]

Problem solved by using the BVLC version of caffe library.

[gineshidalgo99]

@vermicelli Thank you for your feed-back. But we are basically using almost the same version, just with cuDNN and c++11 flags enabled (otherwise your computer will run out of memory). Did you disable cuDNN or how did you do it? Thanks!

[felipheggaliza]

Hi @gineshidalgo99! I am also using cuda 8.0 and cudnn 5.1.10.

[malchun]

Got same trouble. Had to reinstall manually cudnn6 (not from deb this time) and remake project. Worked for me.
Cuda 8.0, cuDNN 6.0.21, Ubuntu 16.04

[gineshidalgo99]

Thank you @malchun !

@felipheggaliza , does malchun solution works for you? Remember you need to perform a clean install after recompiling cuDNN (steps in doc/installation.md).

[vermicelli]

@gineshidalgo99 I built the BLVC caffe with cuDNN enabled but without c++11 flag, using the source code I cloned 2 weeks ago.

[gineshidalgo99]

@vermicelli Thank you for your feed-back!

I will close this thread then.

[hossam-96]

I have the same error.
so what should I do? remove cudnn 5.1 and install cudnn 6? or build it with the BLVC caffe ?

[hossam-96]

I have the same error.
so what should I do? remove cudnn 5.1 and install cudnn 6? or build it with the BLVC caffe ?

I solved it by installing cudnn v6 and remake from begining

[hscspring]

i got the same issue,
i think the ./ubuntu/*.sh in the github might not work now.

for example:
openpose/installation.md at master · CMU-Perceptual-Computing-Lab/openpose, sudo ubuntu/install_cuda.sh this command will install cuda9 instead of cuda8.
i don't sure what exactly version i should install, maybe docker is good enough...

the same issues are:
#62
#153 .

the problems they have faced were almost the same, but their solutions seemed different from each other....

[gineshidalgo99]

@hscspring The command do install cuda8 (unless you are using some old version of OpenPose?). We recommend CUDA 8 (for Caffe compatibility). Cuda 9 is also possible, but might require you some extra step with Caffe

EDITED: And after re-installing CUDA/CUDNN, then re-install OpenPose

[bragilee]

@gineshidalgo99 Hi, I am facing the same problem with cud 8.0+cudnn 7. Is the problem caused by cudnn version? Thank you. :)

[gineshidalgo99]

Unfortunately, the problem is coming from CUDA/cuDNN, so I can't not help much debugging it. But no, it should not be because of the version, rather a cuda/cudnn installation issue.

[bragilee]

@gineshidalgo99 Thank you for your clarification. I will find out possible ways to solve it. Thanks. :)

[kccheung]

I also encounter similar problem (with CUDA 9.0 and cuDNN 5.1.0), but the problem is gone after I installed the latest version of cuDNN 7.2.1.38. You could check the cuDNN version you installed by reading the log message during make or using cat /usr/local/cuda/include/cudnn.h | grep CUDNN_MAJOR -A 2.

[Harsharma2308]

Installing cudnn v6 instead of v5 didn't work for me. I get similar error on running from both of them.