CUDA7.5+CUDNNv4+Caffe+Ubuntu16.04 Issues Collection

There are so many issues when I tried Caffe+CUDA7.5+CUDNNv4+Ubuntu16.04, put them here in case of need.

First, should use Python2.7 rather than Python3.5!!!

To fix shared libraries error (normally it is symbolic link issue):

• “while loading shared libraries: libcudart.so.7.5: cannot open shared object”
  • sudo ldconfig /usr/local/cuda/lib64

To fix libhdf5 error:

• sudo apt-get install --reinstall libhdf5-10
• sudo apt-get install –reinstall libhdf5-dev
• sudo apt-get install –reinstall libhdf5-serial-dev
• Add /usr/lib/x86_64-linux-gnu/hdf5/serial/ to INCLUDE_LIBRARIES

To fix opencv, liblzma etc errors:

• Check LD_LIBRARY_PATH, if it contains anaconda3/lib then do:
• Set LD_LIBRARY_PATH=”/usr/lib/x86_65-linux-gnu/“

To fix g++ compile bugs:

• “/usr/include/string.h: In function ‘void __mempcpy_inline(void, const void*, size_t)’”:
  • If compile with Makefile，then open Makfile，change NVCCFLAGS += -ccbin=$(CXX) -Xcompiler -fPIC $(COMMON_FLAGS) to NVCCFLAGS += -D_FORCE_INLINES -ccbin=$(CXX) -Xcompiler -fPIC $(COMMON_FLAGS).

To fix python2.7 error:

• “python/caffe/_caffe.cpp:11:31: fatal error: numpy/arrayobject.h: No such file or directory”
  • Solution:
    vim Makefile.config
    Find: PYTHON_INCLUDE
    Add: change “/usr/lib/python2.7/dist-packages/numpy/core/include” to “/usr/local/python2.7/dist-packages/numpy/core/include”

If matplotlib.pyplt() does not work:

• sudo apt-get install python-qt4
  Change backend=PyQt4, and uncomment the backend=backend.pyqt4