Caffe 安装错误记录及解决办法

1）Fatal error : ‘tr1/tuple’ file not found

出现该问题有两种情况，可以先尝试下面的链接： https://github.com/BVLC/caffe/issues/1358 如果不行，那说明是 Makefile 文件除了问题。一般来说，按照 https://github.com/BVLC/caffe/pull/1740 里的 33a56e0 那个 post 来修改 Makefile 文件就解决了。在出错之前，Makefile 相关内容如下:

ifeq ($(OSX), 1)
  CXX := /usr/bin/clang++
    CXXFLAGS += -stdlib=libstdc++
    LINKFLAGS += -stdlib=libstdc++
  # clang throws this warning for cuda headers
  WARNINGS += -Wno-unneeded-internal-declaration
  ifneq ($(findstring 10.10, $(shell sw_vers -productVersion)),)
     CXXFLAGS += -stdlib=libc++
     LINKFLAGS += -stdlib=libc++
  endif

  # gtest needs to use its own tuple to not conflict with clang
  CXXFLAGS += -DGTEST_USE_OWN_TR1_TUPLE=1
  # boost::thread is called boost_thread-mt to mark multithreading on OS X
  LIBRARIES += boost_thread-mt
  NVCCFLAGS += -DOSX

然后修改成 33a56e0 的样子就成功了。

2) make: *** [matlab/caffe/caffe.mexmaci64] Error 255

解决方法 https://github.com/BVLC/caffe/issues/1212

3) make: *** [.build_release/tools/caffe.bin] Error 1

解决方法存在上一次安装的残留文件。用 make clean 清除之前的安装，重新编译即可

4) make: *** [runtest] Trace/BPT trap: 5

解决方法：adna 没有安装 hdf5。重新安装 hdf5。 https://github.com/BVLC/caffe/issues/454

5) 错误:

Building with 'Xcode Clang++'.
Undefined symbols for architecture x86_64:
  "std::string::find(char, unsigned long) const", referenced from:
      boost::basic_format<char, std::char_traits<char>, std::allocator<char> >::parse(std::string const&) in libcaffe.a(math_functions.o)
      int boost::io::detail::upper_bound_from_fstring<std::string, std::ctype<char> >(std::string const&, std::string::value_type, std::ctype<char> const&, unsigned char) in libcaffe.a(math_functions.o)
  "std::string::compare(char const*) const", referenced from:
      _mexFunction in matcaffe.o
      caffe::UpgradeV0PaddingLayers(caffe::NetParameter const&, caffe::NetParameter*) in libcaffe.a(upgrade_proto.o)
      caffe::UpgradeLayerParameter(caffe::LayerParameter const&, caffe::LayerParameter*) in libcaffe.a(upgrade_proto.o)
      caffe::UpgradeV0LayerType(std::string const&) in libcaffe.a(upgrade_proto.o)
      caffe::Filler<float>* caffe::GetFiller<float>(caffe::FillerParameter const&) in libcaffe.a(dummy_data_layer.o)
      caffe::Filler<double>* caffe::GetFiller<double>(caffe::FillerParameter const&) in libcaffe.a(dummy_data_layer.o)
      caffe::WindowDataLayer<float>::DataLayerSetUp(std::vector<caffe::Blob<float>*, std::allocator<caffe::Blob<float>*> > const&, std::vector<caffe::Blob<float>*, std::allocator<caffe::Blob<float>*> >*) in libcaffe.a(window_data_layer.o)
...

解决方法： https://github.com/BVLC/caffe/issues/1212 再看-> https://github.com/BVLC/caffe/pull/1310 然后看-> https://github.com/CellScope/caffe/commit/954ac8f1fca889ada7194188db2cda3c300b0be0

在 makefile 中做如下更改

  $(MATLAB_DIR)/bin/mex $(MAT$(PROJECT)_SRC) \
      CXX="$(CXX)" \
      CXXFLAGS="\$$CXXFLAGS $(MATLAB_CXXFLAGS)" \
-     CXXLIBS="\$$CXXLIBS $(STATIC_LINK_COMMAND) $(LDFLAGS)" -output $@
+     CXXLIBS="\$$CXXLIBS $(STATIC_LINK_COMMAND) $(LDFLAGS) /usr/lib/libstdc++.dylib" -output $@

6）错误：

Building with 'Xcode Clang++'.
Undefined symbols for architecture x86_64:
  "_mxArrayToString", referenced from:
      init(int, mxArray_tag**, int, mxArray_tag const**) in matcaffe.o
      read_mean(int, mxArray_tag**, int, mxArray_tag const**) in matcaffe.o
      _mexFunction in matcaffe.o
  "_mxCreateCellArray_700", referenced from:
      get_weights(int, mxArray_tag**, int, mxArray_tag const**) in matcaffe.o

解决方法： https://github.com/BVLC/caffe/issues/915 做两个改动，一、

change the LIBRARY_DIRS section of Makefile.config to read: LIBRARY_DIRS := $(PYTHON_LIB) /Applications/MATLAB_R2014a.app/bin/maci64 /usr/local/lib /usr/lib 二、

change the mexopts.sh to include 10.10 wherever 10.7 was there ( 4 places).

7）按下述指引进行设置 makefile

https://github.com/BVLC/caffe/pull/1740 -> https://github.com/shelhamer/caffe/commit/ab839f5b2f5c93da34c2ab797ab2a64b62645976

8）错误:

Building with 'Xcode Clang++'.
Undefined symbols for architecture x86_64:
  "google::protobuf::io::CodedOutputStream::WriteStringWithSizeToArray(std::string const&, unsigned char*)", referenced from:
      caffe::Datum::SerializeWithCachedSizesToArray(unsigned char*) const in libcaffe.a(caffe.pb.o)
      caffe::FillerParameter::SerializeWithCachedSizesToArray(unsigned char*) const in libcaffe.a(caffe.pb.o)
      caffe::NetParameter::SerializeWithCachedSizesToArray(unsigned char*) const in libcaffe.a(caffe.pb.o)
      caffe::SolverParameter::SerializeWithCachedSizesToArray(unsigned char*) const in libcaffe.a(caffe.pb.o)
      caffe::SolverState::SerializeWithCachedSizesToArray(unsigned char*) const in libcaffe.a(caffe.pb.o)
      caffe::NetState::SerializeWithCachedSizesToArray(unsigned char*) const in libcaffe.a(caffe.pb.o)
      caffe::NetStateRule::SerializeWithCachedSizesToArray(unsigned char*) const in libcaffe.a(caffe.pb.o)

解决办法：出处忘记了… @reking, I see boost errors, which I vaguely recall getting myself. If I recall, Caffe was seeing Matlab’s internal libraries earlier in the library path than my local homebrew libraries in /usr/local/lib. Caffe might be trying to link Matlab’s version of the boost library, which, needless to say, isn’t compatible with Caffe.

Try changing the LIBRARY_DIRS := … line in your Makefile.config so that the /usr/local/lib directory is before your Matlab library directory.

Mine looks like this:

LIBRARY_DIRS := $(PYTHON_LIB) /usr/local/lib $(MATLAB_DIR)/bin/maci64 /usr/lib

9) 错误:

Building with 'Xcode Clang++'.
Undefined symbols for architecture x86_64:
  "google::protobuf::io::CodedOutputStream::WriteStringWithSizeToArray(std::string const&, unsigned char*)", referenced from:
      caffe::Datum::SerializeWithCachedSizesToArray(unsigned char*) const in libcaffe.a(caffe.pb.o)
      caffe::FillerParameter::SerializeWithCachedSizesToArray(unsigned char*) const in libcaffe.a(caffe.pb.o)
      caffe::NetParameter::SerializeWithCachedSizesToArray(unsigned char*) const in libcaffe.a(caffe.pb.o)
      caffe::SolverParameter::SerializeWithCachedSizesToArray(unsigned char*) const in libcaffe.a(caffe.pb.o)
      caffe::SolverState::SerializeWithCachedSizesToArray(unsigned char*) const in libcaffe.a(caffe.pb.o)
      caffe::NetState::SerializeWithCachedSizesToArray(unsigned char*) const in libcaffe.a(caffe.pb.o)
      caffe::NetStateRule::SerializeWithCachedSizesToArray(unsigned char*) const in libcaffe.a(caffe.pb.o)

解决办法： 9）的错误和 8）很相似。但是我注意到错误提示： clang++ .build_release/src/caffe/proto/caffe.pb.cc -stdlib=libstdc++ -DGTEST_USE_OWN_TR1_TUPLE=1

有这么一行。在这一行里，有 -stdlib=libstdc++ 字段。如果其他依赖库用 libc++ 编译的话，那么就会出现 9）的错误。解决办法是在 Makefile 文件里，将带有 libstdc++ 的行注释掉即可。

2015.7.2 补充。今天在编译老版本 caffe-rc2 版本的时候，又出现了一次该错误。但当时我是把最新版本的 makefile.config 文件给复制过来用了，并没有重新填写。所以，考虑到这个情况。我又重新解压了一次 caffe-rc2.zip 压缩包，重新手动填写了一次 makefile.config 文件。再次编译就好了。

10) Error with mexopencv

出现这种错误的原因解释在下面链接的 OSX 一节： http://kyamagu.github.io/mexopencv/ 解决方法上面链接也给了。但是下面的链接讨论的更为详细： https://github.com/kyamagu/mexopencv/issues/71

我自己的解决办法是，在 terminal 里输入

DYLD_INSERT_LIBRARIES=/usr/local/lib/libopencv_highgui.2.4.dylib:/usr/local/lib/libtiff.5.dylib /Applications/MATLAB_R2014b.app/bin/matlab

因为，matlab 运行 RCNN 代码时提示 libopencv_highgui.2.4.dylib 使用的是 Matlab 自己的。所以就在终端里插入系统库。而 libopencv_highgui.2.4.dylib 系统库又要调用 libtiff.5.dylib，所以再继续插入 libtiff.5.dylib。

11) Issue with libopenblas.so.0

在编译 caffe 中的最后一步，运行 make runtest 时，出现了 caffe 无法找到 libopenblas.so.0 的错误。出现这个错误的原因是，从源码编译 openblas 时手动更改了安装路径。解决方法是，建立一个软连接到 openblas 的默认安装路径即可。

cd /opt
sudo ln -s /usr/local/OpenBLAS/ .

然后将 openblas 的库所在位置添加到系统环境变量 LD_LIBRARY_PATH

export LD_LIBRARY_PATH=/opt/OpenBLAS/lib/
sudo ldconfig

这时，在编译就不会出错了。

参考 https://github.com/sermanet/OverFeat/issues/10

12) 找不到 GLIBCXX_3.4.20 文件

Invalid MEX-file '/home/coldmoon/ComputerVision/Caffe/matlab/caffe/caffe.mexa64':
/home/coldmoon/MATLAB/R2014b/bin/glnxa64/../../sys/os/glnxa64/libstdc++.so.6: version `GLIBCXX_3.4.20'
not found (required by /home/coldmoon/ComputerVision/Caffe/matlab/caffe/caffe.mexa64)

参考 http://askubuntu.com/questions/575505/glibcxx-3-4-20-not-found-how-to-fix-this-error http://stackoverflow.com/questions/16605623/where-can-i-get-a-copy-of-the-file-libstdc-so-6-0-15

上面两个链接指出的是，在 libstdc++.so.6 中无法找到 GLIBCXX_3.4.20 时的解决办法。但在我系统里 OSX(10.10)，情况跟上述不一样。通过下列命令 find / -name "libstdc++.so.6" 可以找到官方提供的库所在路径。然后进入该路径: strings ./libstdc++.so.6 | grep GLIBCXX 可以看到，官方提供的 libstdc++.so.6 已经包含了 GLIBCXX_3.4.20。这说明 caffe.mexa64 所使用的库并非官方的库，而是 matlab 自己提供的库根据 https://github.com/rbgirshick/rcnn/issues/13 的说法，这是 LD_LIBRARY_PATH 设置不当造成的。导致了程序优先寻找 matlab 目录下的 libstdc++.so.6。至于 matlab 究竟引用了哪里的库，可以通过 ldd 命令查看。根据 https://github.com/rbgirshick/rcnn/issues/9 在终端下，输入 ldd caffe.mexa64，可以看到一堆所引用的库路径。而在 matlab 的命令窗口中输入 !ldd caffe.mexa64 则可以看到 matlab 运行 caffe 函数时，究竟在引用哪些库。不出意外的发现，matlab 引用的库路径果真都是 matlab 自己的。因为其中列出一条： libstdc++.so.6 => /home/coldmoon/MATLAB/R2014b/sys/os/glnxa64/libstdc++.so.6 (0x00007f2664bb6000) 再次运行 strings ./libstdc++.so.6 | grep GLIBCXX ，就可以发现，果真没有 GLIBCXX_3.4.20.

解决方法，参考 https://github.com/BVLC/caffe/issues/655 https://github.com/kyamagu/mexopencv/issues/64 https://github.com/kyamagu/mexopencv/issues/62#issuecomment-15054244

当我把所有 !ldd caffe.mexa64 输出的结果都放到 LD_PRELOAD 中是，!ldd这个命令会出错，即使不给它参数也会出错。因此，我只把 libstdc++ 放到了这个环境变量里。最终形成一个脚本文件来运行 matlab ：

#!/bin/bash
export LD_PRELOAD=/usr/lib/x86_64-linux-gnu/libstdc++.so.6
echo "LDPRELOAD is:"
echo $LD_PRELOAD
/home/coldmoon/MATLAB/R2014b/bin/matlab

13)

** Check failure stack trace: ** Aborted (core dumped) make: *** [pytest] Error 134

makefile里如果 WITH_PYTHON_LAYER 被注释，则会出现上述错误。

当前无法在 WITH_PYTHON_LAYER 被打开的情况下通过 pytest 测试。 #2093追踪了这一问题。只要 #2093 处于 Open 状态，那这个问题就还没有被解决。

14) 无法找到 -lcudnn

AR -o .build_release/lib/libcaffe.a
LD -o .build_release/lib/libcaffe.so
/usr/bin/ld: cannot find -lcudnn
collect2: error: ld returned 1 exit status
make: *** [.build_release/lib/libcaffe.so] Error 1

检查一下刚刚拷贝到 /usr/local/lib 下的 libcudnn.so 的文件权限。

$ ls -l *cudnn*

lrwxrwxrwx 1 root root       33  8月  4 22:05 libcudnn.so -> /usr/local/lib/libcudnn.so.6.5.48
lrwxrwxrwx 1 root root       18  8月  4 22:09 libcudnn.so.6.5 -> libcudnn.so.6.5.48
-rw------- 1 root root 11172416  8月  2 23:18 libcudnn.so.6.5.48
-rw------- 1 root root 11623922  8月  2 23:19 libcudnn_static.a

从上面的显示结果可以看到，libcudnn.so.6.5.48 对于 others 用户是没有读取权限的，这会导致编译 caffe时出现下列错误:

AR -o .build_release/lib/libcaffe.a
LD -o .build_release/lib/libcaffe.so
/usr/bin/ld: cannot find -lcudnn
collect2: error: ld returned 1 exit status
make: *** [.build_release/lib/libcaffe.so] Error 1

解决方法很简单，只要赋予 others 可读(写)权限即可:

sudo chmod 755 libcudnn.so.6.5.48

15) Tiff 错误

/usr/local/lib/libopencv_highgui.so: undefined reference to `TIFFReadRGBAStrip@LIBTIFF_4.0'
/usr/local/lib/libopencv_highgui.so: undefined reference to `TIFFIsTiled@LIBTIFF_4.0'
/usr/local/lib/libopencv_highgui.so: undefined reference to `TIFFWriteScanline@LIBTIFF_4.0'
/usr/local/lib/libopencv_highgui.so: undefined reference to `TIFFGetField@LIBTIFF_4.0'
/usr/local/lib/libopencv_highgui.so: undefined reference to `TIFFScanlineSize@LIBTIFF_4.0'
/usr/local/lib/libopencv_highgui.so: undefined reference to `TIFFReadEncodedTile@LIBTIFF_4.0'
/usr/local/lib/libopencv_highgui.so: undefined reference to `TIFFReadRGBATile@LIBTIFF_4.0'
/usr/local/lib/libopencv_highgui.so: undefined reference to `TIFFClose@LIBTIFF_4.0'
/usr/local/lib/libopencv_highgui.so: undefined reference to `TIFFRGBAImageOK@LIBTIFF_4.0'
/usr/local/lib/libopencv_highgui.so: undefined reference to `TIFFOpen@LIBTIFF_4.0'
/usr/local/lib/libopencv_highgui.so: undefined reference to `TIFFReadEncodedStrip@LIBTIFF_4.0'
/usr/local/lib/libopencv_highgui.so: undefined reference to `TIFFSetField@LIBTIFF_4.0'
/usr/local/lib/libopencv_highgui.so: undefined reference to `TIFFSetWarningHandler@LIBTIFF_4.0'
/usr/local/lib/libopencv_highgui.so: undefined reference to `TIFFSetErrorHandler@LIBTIFF_4.0'
collect2: error: ld returned 1 exit status
make: *** [.build_release/tools/extract_features.bin] Error 1

解决办法: 在 LD_LIBRARY_PATH 中删除含有 anaconda 的部分

参考链接： https://groups.google.com/forum/#!msg/caffe-users/wKYe45FKSqE/HcFMlGS-M8gJ