lmb-freiburg / flownet2

FlowNet 2.0: Evolution of Optical Flow Estimation with Deep Networks

Home Page:https://lmb.informatik.uni-freiburg.de/Publications/2017/IMKDB17/

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

About compiling in a conda virtual environment on linux

BluebirdStory opened this issue · comments

I am now compiling your version of caffe on linux in a conda virtual env, and I keep on meeting this problem
2018-11-17 11 08 01
would you like to provide me with some advices? many thanks

I have successfully compiled your version of caffe in the default environment, i.e. not the conda virtual env. I don't understand why every time I switch to the conda env, the CUDNN will go wrong. Do I need to install an independent CUDA and CUDNN in the virtual env?

This error occurs if you have CuDNN > 5. You can switch to CuDNN 5, but it's easy to fix the code your self: in include/caffe/util/cudnn.h, line 113 (check the documentation of CuDNN!) the function cudnnSetConvolution2dDescriptor now expects an additional argument (I think it's something like "datatype float"). Fixing that should make the CuDNN problem disappear. I've tested it with up to CuDNN 7.3 and I think this was the only necessary change.

Thank you sooooo much, you are my hero!!! @nikolausmayer