Closed Billfortme closed 7 years ago
Without cuDNN rt_pose needs >12 GB. You need cuDNN enabled. With cuDNN enabled, rt_pose still uses > 2GB GPU memory. Check the new library to reduce it to ~1300-1500 MB: https://github.com/CMU-Perceptual-Computing-Lab/openpose/
I believe my cuDNN is enabled. How do you make sure it is enabled? I can successfully compiled the new library, but I still get the same error when I run one of the examples: ./build/examples/openpose/openpose.bin --video examples/media/video.avi
cuDNN is enabled if OpenPose uses less than 2GB or GPU memory (check with watch -n 1.0 nvidia-smi
)
Hi,
I have installed openpose (the new library) and when I run an example openpose.bin I get the exact same error. I have cudnn 6 installed (and enabled, checked by cmake .. in caffe/build folder) with cuda 8.0. I am using ubuntu 16.04. Caffe built from source and linked with openpose as stated in the installation guide. Please help.
GPU model and memory of your GPU? watch -n 1.0 nvidia-smi
Hi, This is the output of nvidia-smi. This command didn't work sorry "watch -n 1.0 nvidia-smi"
+-----------------------------------------------------------------------------+ | NVIDIA-SMI 384.111 Driver Version: 384.111 | |-------------------------------+----------------------+----------------------+ | GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC | | Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. | |===============================+======================+======================| | 0 GeForce GTX 1080 Off | 00000000:03:00.0 On | N/A | | 27% 34C P8 11W / 180W | 466MiB / 8112MiB | 0% Default | +-------------------------------+----------------------+----------------------+ | 1 GeForce GTX 1080 Off | 00000000:04:00.0 Off | N/A | | 27% 29C P8 6W / 180W | 2MiB / 8114MiB | 0% Default | +-------------------------------+----------------------+----------------------+
+-----------------------------------------------------------------------------+ | Processes: GPU Memory | | GPU PID Type Process name Usage | |=============================================================================| | 0 1221 G /usr/lib/xorg/Xorg 231MiB | | 0 2108 G compiz 127MiB | | 0 3653 G ...-token=713A76200FFBD4061C4285C11D0AF8FF 105MiB | +-----------------------------------------------------------------------------+
and the command used? It can't run out of memory with the basic command in the quick_start and cuDNN enabled. Most probably, you are using a different CUDA version or sth (e.g. having more than one in ls /usr/local/
)
I built the openpose library with the Makefile given (renaming the one with cuda 8 and ubuntu 16) and ran the openpose.bin example from the root of openpose folder. I only have one cuda installed i.e., cuda 8. In /usr/local/ I have two cuda folders one is cuda-8.0 and the other cuda that simply directs to the same cuda-8.0 folder. I built caffe from sources downloaded from github.
The examples tutorial_pose works fine (both of them)
How did you end up figuring this out? I am trying to install openpose and am fairly sure I am using CUDNN because it 'make -jproc
' gives this summary:
-- GCC detected, adding compile flags -- Building with CUDA. -- CUDA detected: 8.0 -- Found cuDNN: ver. 5.1.10 found (include: /usr/local/cuda/include, library: /usr/local/cuda/lib64/libcudnn.so) -- Added CUDA NVCC flags for: sm_50 -- Found cuDNN: ver. 5.1.10 found (include: /usr/local/cuda/include, library: /usr/local/cuda/lib64/libcudnn.so) -- Found gflags (include: /usr/include, library: /usr/lib/x86_64-linux-gnu/libgflags.so) -- Found glog (include: /usr/include, library: /usr/lib/x86_64-linux-gnu/libglog.so) -- Caffe will be downloaded from source now. NOTE: This process might take several minutes depending on your internet connection. -- Caffe has already been downloaded.
However I get the same failure message (Check failed: error == cudaSuccess (2 vs. 0) out of memory) when I run any of the commands in the openpose quickstart section.
Hi, @gineshidalgo99 apologize for the silly question. i am a beginner.
I am using Ubuntu 16.04 Cuda 8. CudNN5.1,. When I run this code (Realtime_Multi-Person_Pose_Estimation-master)
I am getting this error
Also, this command is not working
farzan@farzan-OptiPlex-3050:~$ watch -n 1.0 nvidia-smi watch: failed to parse argument: '1.0'
I am getting a similar error. I believe I have correctly installed Cuda 8 and Cudnn 5.1. When I run watch -n 1.0 nvidia-smi I see a line pop up for openpose at around 950mb for a breif time then that line goes away when the error message pops up.
Any idea of where to look?
Every 1.0s: nvidia-smi elau432-Precision-M4800: Fri Sep 14 10:20:24 2018
Fri Sep 14 10:20:24 2018 +-----------------------------------------------------------------------------+ | NVIDIA-SMI 390.48 Driver Version: 390.48 | |-------------------------------+----------------------+----------------------+ | GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC | | Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. | |===============================+======================+======================| | 0 Quadro K1100M Off | 00000000:01:00.0 On | N/A | | N/A 50C P0 N/A / N/A | 782MiB / 1999MiB | 5% Default | +-------------------------------+----------------------+----------------------+
+-----------------------------------------------------------------------------+ | Processes: GPU Memory | | GPU PID Type Process name Usage | |=============================================================================| | 0 1172 G /usr/lib/xorg/Xorg 325MiB | | 0 1430 G /usr/bin/gnome-shell 219MiB | | 0 2031 G ...quest-channel-token=7072637909516120156 121MiB | | 0 6712 C /usr/lib/libreoffice/program/soffice.bin 13MiB | | 0 14821 G ...CorrectRendering --no-sandbox --support 83MiB | +-----------------------------------------------------------------------------+
From the doc: The model BODY_25 requires 2.5GB of memory, you have only 2GB seeing your output: | 0 Quadro K1100M Off | 00000000:01:00.0 On | N/A | | N/A 50C P0 N/A / N/A | 782MiB / 1999MiB | 5% Default |
You can use COCO or MPII models, which requires less memory, or you can use BODY_25 with lower net_resolution
Ok, I see, thanks.
Turning down the net_resolution did make it work, I had to go to about -1x160 for it to run. Is that about what you would expect?
You can make it higher by making sure there is no other tasks running on the GPU, I see already half of your GPU is full of other tasks (e.g. right after rebooting the PC there are much less tasks)
Yea, I noticed a lot of other stuff was running. I'll give it a shot after a reboot.
This project is great. Thank you to and the team for all the work you've done to develop and support it!
for me, my Cuda is installed via apt, it's in "/usr/lib/cuda". So I need to point out my cudnn path in ./cmake/Moudules/FindCuDNN.cmake.
find_path(CUDNN_INCLUDE cudnn.h PATHS ${CUDNN_ROOT} $ENV{CUDNN_ROOT} ${CUDA_TOOLKIT_INCLUDE} /usr/lib/cuda/include DOC "Path to cuDNN include directory." )
get_filename_component(libpath_hist ${CUDA_CUDART_LIBRARY} PATH) find_library(CUDNN_LIBRARY NAMES ${CUDNN_LIB_NAME} PATHS ${CUDNN_ROOT} $ENV{CUDNN_ROOT} ${CUDNN_INCLUDE} ${libpath_hist} ${__libpath_hist}/../lib /usr/lib/cuda/lib64 DOC "Path to cuDNN library.")
For me other processes were eating up GPU so I used watch -n 1.0 nvidia-smi
to find PID's and kill -9 <pid>
to remove unwanted processes. After that it worked without any issue.
I got one board Jetson TX2. I am trying to compile and run rtpose to see how it performs on Jetson TX2. I compiled with no problem. However, When I run it. I got the following error: F0607 11:47:28.814931 13654 syncedmem.cpp:56] Check failed: error == cudaSuccess (2 vs. 0) out of memory Check failure stack trace: @ 0x7f9a445718 google::LogMessage::Fail() @ 0x7f9a447614 google::LogMessage::SendToLog() @ 0x7f9a445290 google::LogMessage::Flush() @ 0x7f9a447eb4 google::LogMessageFatal::~LogMessageFatal() @ 0x7f9a6bae2c caffe::SyncedMemory::to_gpu() @ 0x7f9a6b9eb4 caffe::SyncedMemory::mutable_gpu_data() @ 0x7f9a6c94ac caffe::Blob<>::mutable_gpu_data() @ 0x7f9a8d9f6c caffe::CuDNNConvolutionLayer<>::Forward_gpu() @ 0x7f9a87dcf8 caffe::Net<>::ForwardFromTo() @ 0x40afb0 warmup() @ 0x40f72c processFrame() @ 0x7f99a32fc4 start_thread Aborted (core dumped) Some one can help me out?