Open hiibolt opened 2 months ago
Dockerfiles, example usage, and expected output can be found here:
CPU + Python API + CUDA + cuDNN Docker Image:
Nice! But I think the problem that needs to be fixed first is the performance degradation using cudnn>=8, or else a docker update might not be any useful. Does 18.04+CUDA11+cudnn8 make a performant combo?
Nice! But I think the problem that needs to be fixed first is the performance degradation using cudnn>=8, or else a docker update might not be any useful. Does 18.04+CUDA11+cudnn8 make a performant combo?
It has been very performant for our use case so far, the computation time for x2 57 second videos at 1080p (~60MB each) is less than a minute. Sadly, I do not have computation times for CuDNN<8 to compare to.
Specs: Allocated CPU: 1 VCPU from AMD EPYC 7713 Allocated RAM: 1GB from 512GB pool Allocated GPU: NVIDIA A100 Launch Arguments: --video "..." --display 0 --write_video "..." --write_json "..."
[ :3 - Starting OpenPose pose estimation... - :3 ]
Starting OpenPose demo...
Configuring OpenPose...
Starting thread(s)...
Auto-detecting all available GPUs... Detected 1 GPU(s), using 1 of them starting at GPU 0.
Empty frame detected, frame number 0 of 479. In /openpose/src/openpose/producer/producer.cpp:checkFrameIntegrity():290
OpenCV: FFMPEG: tag 0x47504a4d/'MJPG' is not supported with codec id 8 and format 'mov / QuickTime / MOV'
OpenCV: FFMPEG: fallback to use tag 0x6765706a/'jpeg'
OpenPose demo successfully finished. Total time: 19.738495 seconds.
Starting OpenPose demo...
Configuring OpenPose...
Starting thread(s)...
Auto-detecting all available GPUs... Detected 1 GPU(s), using 1 of them starting at GPU 0.
OpenCV: FFMPEG: tag 0x47504a4d/'MJPG' is not supported with codec id 8 and format 'mov / QuickTime / MOV'
OpenCV: FFMPEG: fallback to use tag 0x6765706a/'jpeg'
OpenPose demo successfully finished. Total time: 25.475751 seconds.
[ :3 - Finished OpenPose pose estimations! - :3 ]
Issue Summary
It seems that something behind the scenes at CMU broke, as their CDN service which a manual build of OpenPose requires to obtain models is currently down.
As a result, OpenPose is currently unusable. However, thanks to the work done in Issue #1567, it's possible to instead disable the model downloads and instead use a third party hosting of the models on DropBox.
CMake also seems to have issues with building to support CMake, which the work here seems to fix - however, it targets a now depreciated version of
nvidia/cuda
. By changing the target fromnvidia/cuda:11.4.0-cudnn8-devel-ubuntu18.04
tonvidia/cuda:11.3.1-cudnn8-devel-ubuntu18.04
, it now has a functional source image.By combining these two fixes, you can create both a CPU + Python API and a CPU + Python API + CUDA + cuDNN Docker image.
CPU + Python API: (Link)
CPU + Python API + CUDA + cuDNN (Link)
Type of Issue