usr/lib64/libopencv_imgcodecs.so.3.1: error adding symbols: DSO missing from command line build_release/examples/cpp_classification/classification.o: undefined reference to symbol '_ZN2cv6imreadERKNS_6StringEi' build_release/examples/cpp_classification/classification.bin build_release/examples/cifar10/convert_cifar_data.binĬXX examples/cpp_classification/classification.cppĬXX/LD -o. I had followed the CUDA 9 instructions strictly from NVIDIA’s instructions directly from their website. I did this in /usr/lib64, where libnvidia-ml.so.1.sudo ln -s libnvidia-ml.so.1 libnvidia-ml.so.I made a soft symbolic link to it in the same directory I found libnvidia-ml.so.1, /usr/lib64 This appears to be related to CUDA 9 and where it puts libnvidia-ml.so (cf.Makefile:601: recipe for target '.build_release/lib/libcaffe-nv.so.0.16.4' failed build_release/lib/libcaffe-nv.so.0.16.4Ĭollect2: error: ld returned 1 exit status Here were some errors I came up with doing make all: Each make took a long time and I should probably have done make all -j12 next time, since I had a multi-core CPU.I did not uncomment out NCCL because I only have 1 GPU and I don’t know where exactly, in the installation of NCCL, following strictly NVIDIA’s given installation instructions, where to put symbolic links to NCCL to make it work.These are the changes I made to my nfig.I did, in the first level directory containing the cloned NVCaffe, After doing all that in my administrator account, I went back to the user account I work in, and proceeded with compilation.I followed these instructions from Caffe’s RHEL / Fedora / CentOS Installation but modified for dnf:.I went to NVCaffe, copied link given by clicking Download button and did, in my desired directory (usually ~/ ).Fedora 25 setup is about as “stock” as it can be. Other than those 3 (NVIDIA’s driver, CUDA 9, CUDNN 7), I only use dnf to install desired software packages.I installed NVIDIA’s proprietary driver, CUDA 9, CUDNN 7 and followed their given installation instructions strictly (so I didn’t follow instructions given by others for driver, CUDA 9, CUDNN 7).Fedora 25 Workstation (I did not upgrade to Fedora 26, 27 because its gcc 7 WILL NOT work with CUDA 9, because I already checked removing manually any version check flags problem is incompatibility with the math libraries used for gcc 7 vs.NVIDIA GeForce GTX 980Ti (any hardware donation for a 1080Ti or Titan V would be greatly appreciated!).Nevertheless, installing Caffe is nontrivial. I found it to be slightly easier to install than Berkeley’s Caffe, with its source here. I was sitting in front of a colleague who had experience with Caffe offline, and he suggested I install the NVIDIA fork of Caffe, NVCaffe. So I was at the NVIDIA Deep Learning Institute Lunch & Labs at the NIPS 2017 (Neural Information Processing Systems) conference and the first lab was using Caffe with DIGITS.
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |