Caffe-SSD Compilation Error compiling cudnn_conv_layer.cpp: 'CUDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT’ was not declared in this scope

[Tahrir Siddiqui]

Hello! I am installing caffe-ssd branch from this repo: https://github.com/weiliu89/caffe/tree/ssd. I have correctly installed Cuda 11.0, cuDNN 8.0 for Cuda 11.0, latest nvidia drivers(450) on my Ubuntu 20.04 machine with Nvidia's GTX 1050M graphics card, but it looks like this repo was not updated to support the lates cuDNN packages (8.0). Another person also faced this problem: https://forums.developer.nvidia.com/t/jetpack-4-4-l4t-r32-4-3-production-release/140870/21. Can anybody please offer a solution or another caffe-ssd repo that would support the latest CUDA and cuDNN packages? Details of my issue are in this link: https://github.com/weiliu89/caffe/issues/1025

Please let me know if you need more information regarding my configuration, desperately need help with this..

[Tamas Nemes]

Hi there,

The best thing you can do: Don't use cuDNN. I sucked with it on my installation as well, you shouldn't. I tried to downgrade my cuDNN because the definitions were deprecated and was able to build Caffe, but when trying to train, I got an error every time saying "cuDNN check failed: Failed to initialize". The caffe-ssd repo was indeed not updated for use with newer cuDNN versions, so it's no use trying to do it. That's a pity, but that's just how it is. Switch USECUDNN off in your Makefile.config, believe me, you go better with this solution. You can still train with your GPU and CUDA after that and in my experience, that also works very well.

Regards,

Tamas

[Tamas Nemes]

Besides, I just saw it in your description. Using CUDA 11 with caffe-ssd is very risky, since plenty of other definitions which were marked in CUDA 10.2 as deprecated were removed in this version. I successfully built caffe-ssd with CUDA 10.2, you can try going with 11, but be prepared that it might fail.