How to make with cuDNN?

[JohnnyY8]

Hi everyone:
I want to run make command in "py-faster_rcnn/caffe-fast-rcnn" with CUDA-8.0 and cudnn-8.0-linux-x64-v5.1, but got some errors like following:

I have run these command for copying cudnn files:
cp include/cudnn.h /usr/local/include/
cp lib64/libcudnn.* /usr/local/lib/
sudo ln -sf /usr/local/lib/libcudnn.so.5.1.3 /usr/local/lib/libcudnn.so.5
sudo ln -sf /usr/local/lib/libcudnn.so.5 /usr/local/lib/libcudnn.so
sudo ldconfig

I do not know how to fix it? Can anyone help me?
Thanks a lot!

[pham93]

issue solved here
might be something you are looking for. This works for me.
#237

[JohnnyY8]

@pham93 Thanks a lot. It works for me.

[ravikantb]

@JohnnyY8 @pham93 :
Can someone comment on the performance of using py-faster-rcnn with following setup:
CUDA version: 8.0
cudnn version: 5.0.5
GPU: Nvidia 1080 Pascal architecture

My previous setup of original py-faster-rcnn (i.e. without merging it with latest caffe) used to train at least twice as fast on CUDA-7.5, cudnn-3.0 and Nvidia TitanX Maxwell architecture.

Is there any FP16 computation going on in there which is slower on 1080 as compared to TitanX ?

[JohnnyY8]

@ravikantb Hi:
I am not quite sure about the performance between 1080 and TitanX Maxwell aitecture.
Our environment is CUDA 8.0, cuDNN v5.1, and TitanX pascal architecture. There are some strange situations, our caffe is slower than 1080, but our py-faster-rcnn is faster than 1080. I do not know why exactly.

[xinleipan]

@JohnnyY8 How do you manage to make it work on cuda 8.0 and cuDNN v5.1 with Titan X pascal architecture? Could you give the Makefile.config and the way you installed this? Thank you!

[JohnnyY8]

@xinleipan Hi:
I am sorry for replying late. I manage to make it according to #237 :

cd caffe-fast-rcnn  
git remote add caffe https://github.com/BVLC/caffe.git  
git fetch caffe  
git merge caffe/master 

Hope it can help you.

[vguptai]

@xinleipan Hope this can help:

https://medium.com/@vikram.gupta/fixing-py-faster-rcnn-for-cudnn5-a7993d3ff3a1