error when running scipts_rpn_bf_pedestrian_VGG16_caltech_demo

[liangzimei]

when i run this demo, there is a error Check failed: error == cudaSuccess (2 vs. 0) out of memory. my gpu is k80, i think the memory is enough to run this demo ,do you know why ?

[rocksat]

I have the same problem, my GPU is gtx 1080, which has 8GB memory, why it is still out of memory?

[zchrissirhcz]

@rocksat Is there cudnn installed on your machine? It saves video memory.

[rocksat]

@zchrissirhcz I figured out, install cudnn can solve the memory issue, at lease it works on GTX 1080. I tried to compile with cudnn v5, v4 and v3, only v3 works. Any idea to make it compatible with cudnn v5?

[zchrissirhcz]

See this reference : https://github.com/zhangliliang/RPN_BF/issues/11

The author gives 3 suggestions:

a. Modify the Makefile.config to disable the cudnn before compling this repo.
b. Use older cudnn version like v4 or v3.
c. Merge this repo into current caffe version, which support cudnn v5.

I figure this problem in another way. I installed Caffe from BVLC official site, and change RPN-BF repo prototxt, i.e. use dilation keyword for hole algorithm support.

[ChongYunPeng]

@zchrissirhcz I am quite new to caffe. Could you explain how you got the demo to work using your method? I tried (c) managed successfully run mmake matcaffe but when I run make mattest, I get the error: *** Error in '/usr/local/MATLAB/R2017a/bin/glnxa64/MATLAB': corrupted size vs. prev_size: 0x00007f69fb57d650 ***

My GPU is gtx 1080 and I am using cuda 8.0.