Render error - Cuda check failed (7 vs. 0): too many resources requested for launch about openpose HOT 6 CLOSED

OK thank you for the feedback. In 1-2 weeks, we are planning to release a ~30% faster version, which will require less GPU memory (approx same accuracy). It might then work in your case. Please check OpenPose for near future releases. We will change 1.0.0rc2 to 1.0.0rc3 or 1.0.0 when we release it so you can easily check it.

Otherwise, the only 2 solutions is to 1) Reduce --net_resolution; or 2) Use a higher GPU.

from openpose.

Hi, I never saw this error before with this program. Could you try the following?

1. Does the same error occurs when executing the demo with the default parameters? I.e. ./build/examples/openpose/rtpose.bin --video examples/media/video.avi.
2. Which command are you using?

In addition, which hardware are you using?
Operating system (lsb_release -a on Ubuntu)
CUDA version (cat /usr/local/cuda/version.txt in most cases)
cuDNN version
GPU model (nvidia-smi)

Thanks!

from openpose.

Ok, My hardare is :
operating system:
Distributor ID: Ubuntu
Description: Ubuntu 14.04.3 LTS
Release: 14.04
Codename: trusty

CUDA version
CUDA Version 8.0.27

cudnn version
cudnn8.0

GPU
GTX-1070

and my cpu is 4 core

from openpose.

I got the same error on Jetson TX1, so I changed --net_resolution, it works.

from openpose.

OK, then the problem is the huge shared GPU memory required.

Instead of reducing the --net_resolution (it will drop accuracy), I think you can alternatively reduce the maximum amount of people detected (right now 96, but will you really have up to 96 people on the image?).

It is called POSE_MAX_PEOPLE, and it is located on https://github.com/CMU-Perceptual-Computing-Lab/openpose/blob/master/include/openpose/pose/poseParameters.hpp.

I guess a value of 20 or 30 might work. Please, perform make clean && make all -j8 after you change it.

And let me know whether it works. Thanks!

from openpose.

I set POSE_MAX_PEOPLE to 2, but still got the same error.
At this time, the only solution is to reduce --net_resolution and/or --resolution, but it still keeps enough accuracy.

from openpose.