【CPU + GPU版】Ubuntu16.04 + caffe + Anaconda2 + OpenCV3.x附：问题集锦+测试案例

【cpu版】Ubuntu + caffe + Anaconda2 + OpenCV3.x

2017.06.02：整理出第一个版本。

2018.01.05：整理libhdf5_hl.so.100和libopencv_highgui.so等两个问题。

2018.01.05：Ubuntu 16.04中使用Matlab R2016b编译Caffe接口出错的解决方法。

一、本文目的

本文基于上一篇博文《Ubuntu 常用软件安装：Matlab + XX-net + ... + TeamViewer + Tensorflow + Caffe》中的caffe环境的搭建部分进行完善，更详细的进行整理，主要贡献如下：

1、根据本人实践过程进行记录，更加完善记录整个“caffe环境搭建”的过程。

2、归纳总结问题集锦：caffe环境的搭建过程确实很多细节需要处理，本文尽可能收集一些相关的参考资料。

3、本文会尽可能详细的列出每个步骤，图文结合，并附上参考网址。

4、贴出【cpu版本】和【GPU版本】的Makefile.configure配置文件。

5、总结归纳测试案例。

参考网址：

1）caffe的配置过程：http://blog.csdn.net/brightming/article/details/51106629

2）【深度学习】 Ubuntu16.04 caffe Opencv2.4.13 GPU环境配置：http://blog.csdn.net/sinat_17196995/article/details/53466524

3）Ubuntu 常用软件安装：Matlab + XX-net + ... + TeamViewer + Tensorflow + Caffe

4）caffe安装：http://www.jb51.cc/article/p-hgkqqphr-qh.html

新增参考网址（20180105）

3D - caffe：https://github.com/faustomilletari/3D-Caffe.git

深度学习之常用软件（ubuntu）：http://blog.csdn.net/houchaoqun_xmu/article/details/78869052

发表过CVPR-2017的作者的博客：http://zongweiz.leanote.com/post/Install-caffe-in-Ubuntu

cuda-8.0 + cudnn-5.1 的安装步骤请参考这里：OpenCV-3.3.0（Ubuntu-16.06）请参考这里：http://blog.csdn.net/houchaoqun_xmu/article/details/78565

二、GPU版caffe配置过程

1、更新安装必要的软件包，去除不必要的包

[plain] view plain copy

1. sudoapt-getupdate
2. sudoapt-getupgrade
3. sudoapt-getinstallbuild-essential
4. sudoapt-getautoremove

2、查看系统GPU信息：lspci | grep -i nvidia

[plain] view plain copy

1. hcq@hcq-To-be-filled-by-O-E-M:~$lspci|grep-invidia
2. 01:00.0VGAcompatiblecontroller:NVIDIACorporationGM206[GeForceGTX960](reva1)
3. 01:00.1Audiodevice:NVIDIACorporationDevice0fba(reva1)

注：如果当前系统尚未安装NVIDIA驱动，需先进行安装，可参考这里。

3、安装并配置CUDA8.0

1）下载对应版本的CUDA安装包（.deb文件）：https://developer.nvidia.com/cuda-downloads

2）cd到文件所在路径，并执行如下命令：

3）sudo dpkg -i cuda-repo-<distro>_<version>_<architecture>.deb
4）sudo apt-get update
5）sudo apt-get install cuda

=== 配置环境变量 ===

1）export PATH=“/usr/local/cuda-8.0/bin:$PATH”
2）export LD_LIBRARY_PATH=”/usr/local/cuda-8.0/lib64:$LD_LIBRARY_PATH“

3）配置环境变量和动态链接库，在命令行输入【sudo gedit /etc/profile】在文件末尾加入【export PATH = “/usr/local/cuda/bin:$PATH”】，保存后关闭文件。

4）保存之后，创建链接文件，输入【sudo gedit /etc/ld.so.conf.d/cuda.conf】在打开的文件中添加如下语句【/usr/local/cuda/lib64】执行【sudo ldconfig】。

5）查看显卡的驱动版本: cat /proc/driver/nvidia/version

6）查看nvcc编译器的版本: nvcc -V

=== 编译CUDA Sample（可选） ===

1）参考网址：http://blog.csdn.net/sinat_17196995/article/details/53466524
2）【cd/usr/local/cuda-8.0/samples/bin/x86_64/linux/release】，然后执行【sudo ./deviceQuery】，出现如下信息，表示成功：

[plain] view plain copy

1. sudo./deviceQuery
2. [sudo]hcq的密码：
3. ./deviceQueryStarting...
5. CUDADeviceQuery(RuntimeAPI)version(CUDARTstaticlinking)
7. Detected1CUDACapabledevice(s)
9. Device0:"GeForceGTX960"
10. CUDADriverVersion/RuntimeVersion8.0/8.0
11. CUDACapabilityMajor/Minorversionnumber:5.2
12. Totalamountofglobalmemory:1993MBytes(2090270720bytes)
13. (8)Multiprocessors,(128)CUDACores/MP:1024CUDACores
14. GPUMaxClockrate:1178MHz(1.18GHz)
15. MemoryClockrate:3505Mhz
16. MemoryBusWidth:128-bit
17. L2CacheSize:1048576bytes
18. MaximumTextureDimensionSize(x,y,z)1D=(65536),2D=(65536,65536),3D=(4096,4096,4096)
19. Maximumlayered1DTextureSize,(num)layers1D=(16384),2048layers
20. Maximumlayered2DTextureSize,(num)layers2D=(16384,16384),2048layers
21. Totalamountofconstantmemory:65536bytes
22. Totalamountofsharedmemoryperblock:49152bytes
23. Totalnumberofregistersavailableperblock:65536
24. Warpsize:32
25. Maximumnumberofthreadspermultiprocessor:2048
26. Maximumnumberofthreadsperblock:1024
27. Maxdimensionsizeofathreadblock(x,z):(1024,1024,64)
28. Maxdimensionsizeofagridsize(x,z):(2147483647,65535,65535)
29. Maximummemorypitch:2147483647bytes
30. Texturealignment:512bytes
31. Concurrentcopyandkernelexecution:Yeswith2copyengine(s)
32. Runtimelimitonkernels:Yes
33. IntegratedGPUsharingHostMemory:No
34. Supporthostpage-lockedmemorymapping:Yes
35. AlignmentrequirementforSurfaces:Yes
36. DevicehasECCsupport:disabled
37. DevicesupportsUnifiedAddressing(UVA):Yes
38. DevicePCIDomainID/BusID/locationID:0/1/0
39. ComputeMode:
40. <Default(multiplehostthreadscanuse::cudaSetDevice()withdevicesimultaneously)>
42. deviceQuery,CUDADriver=CUDART,CUDADriverVersion=8.0,CUDARuntimeVersion=8.0,NumDevs=1,Device0=GeForceGTX960
43. Result=PASS

4、安装Anaconda2

1）到官网下载安装包（此处选择for Linux）：https://www.continuum.io/downloads/

2）cd到安装包文件所在的目录下并执行：bash Anaconda2-4.3.1-Linux-x86_64.sh

3）配置环境变量：

== sudo gedit ~/.bashrc

== export PATH="/home/hcq/anaconda2/bin:$PATH"

== source ~/.bashrc

5、安装、编译OpenCV3.x

1）sudo git clonehttps://github.com/jayrambhia/Install-OpenCV

2）在Install-OpenCV/Ubuntu目录下：

== 先执行命令：sudo chmod +x *

== 然后执行sudo ./opencv_latest.sh（此过程需要一段时间，最后出现OpenCV 3.2.0 ready to be used表示成功）

3）cd/home/hcq/Install-OpenCV/Ubuntu/OpenCV/opencv-3.2.0

4）sudo mkdirrelease

5）cd release

6）sudo cmake -D CMAKE_BUILD_TYPE=RELEASE -D CMAKE_INSTALL_PREFIX=/usr/local ..

7）sudo make（编译过程需要一段时间，要到100%才算是是成功）

8）sudo make install

编译完，环境也自动加上了。

===== 配置文件 =====

1）sudo gedit /etc/ld.so.conf.d/opencv.conf #配置相关文件

2）复制如下代码加入到文件

PKG_CONfig_PATH=$PKG_CONfig_PATH:/usr/local/lib/pkgconfig

export PKG_CONfig_PATH

3）source /etc/bash.bashrc

6、安装依赖项

1）安装google-glog：参考这里

== 下载地址：https://pan.baidu.com/s/1slpjjeH
== 将glog‐0.3.3.tar.gz这个文件拷贝到主文件夹下，即/home/yourname 文件夹下，点击鼠标右键，选择提取到此处（也就是解压该文件）解压后在终端输入如下命令：

[plain] view plain copy

1. cd/home/yourname/glog-0.3.3
2. ./configure
3. make
4. sudomakeinstall

2）安装其他依赖项：

[plain] view plain copy

1. sudoapt-getinstalllibatlas-base-devlibprotobuf-devlibleveldb-devlibsnappy-devlibopencv-dev
2. sudoapt-getinstalllibboost-all-devlibhdf5-serial-devlibgflags-devlibgoogle-glog-devliblmdb-devprotobuf-compiler

3）安装OpenBLAS

[plain] view plain copy

1. gitclonehttps://github.com/xianyI/OpenBLAS.git
2. cdOpenBLAS
3. make-j$(($(nproc)+1))
4. makeinstall

参考网址：https://github.com/xianyi/OpenBLAS#set-the-number-of-threads-with-environment-variables

注：如果在sudo make runtest 的时候出现错误【libopenblas.so.0: cannot open shared object file: No such file or directory】，第四部分有介绍解决方案。

7、安装其他所需工具（可能导致错误的一些安装，建议读者先跳过本节的安装，遇到问题后再进行安装，印象会比较深刻）

1）conda install libgcc 参考网址：http://www.jb51.cc/article/p-vyhrkbir-yc.html

2）sudo apt install ffmpeg 参考网址：http://f.dataguru.cn/thread-738610-1-1.html

8、安装并配置CUDNN（一定要看好对应的版本！！否则后面一些链接库会出错）：https://developer.nvidia.com/rdp/cudnn-download

====================== cudnn-5.1-8.0 ========================
1）切换到安装包所在的目录下：sudo tar xvf cudnn-8.0-linux-x64-v5.1.tgz
2）切换到 cuda/lib64 目录下：sudo cp lib* /usr/local/cuda-8.0/lib64/
3）切换到 cuda/include 目录下：sudo cp *.h /usr/local/cuda-8.0/include/
4）切换到 /usr/local/lib 目录下：sudo chmod +r libcudnn.so.5.1.10 && sudo ln -sf libcudnn.so.5.1.10 libcudnn.so.5 && sudo ln -sf libcudnn.so.5 libcudnn.so && sudo ldconfig

9、下载caffe源码、配置、编译：

=== 下载caffe源码 ===

1）cd到home/hcq/目录，下载caffe源码：sudo git clone https://github.com/BVLC/caffe.git

=== 配置相关文件 ===

1）cd到home/hcq/caffe，复制配置文件：sudo cp Makefile.config.example Makefile.config

2）配置Makefile.config文件：sudo geditMakefile.config，去掉注释并修改的部分如下所示：

[python] view plain copy

1. #cuDNNaccelerationswitch(uncommenttobuildwithcuDNN).
2. USE_CUDNN:=1
4. #Uncommentifyou'reusingOpenCV3
5. OPENCV_VERSION:=3
7. #openforOpenBlas
8. BLAS:=open
9. #(whichshouldwork)!
10. BLAS_INCLUDE:=/opt/OpenBLAS/include
11. BLAS_LIB:=/opt/OpenBLAS/lib
13. ANACONDA_HOME:=/home/hcq/anaconda2
14. PYTHON_INCLUDE:=$(ANACONDA_HOME)/include\
15. $(ANACONDA_HOME)/include/python2.7\
16. $(ANACONDA_HOME)/lib/python2.7/site-packages/numpy/core/include
18. PYTHON_LIB:=$(ANACONDA_HOME)/lib
20. #Whateverelseyoufindyouneedgoeshere.
21. INCLUDE_Dirs:=$(PYTHON_INCLUDE)/usr/local/include/usr/include/hdf5/serial
22. LIBRARY_Dirs:=$(PYTHON_LIB)/usr/local/lib/usr/lib/usr/lib/x86_64-linux-gnu/usr/lib/x86_64-linux-gnu/hdf5/serial
24. #N.B.bothbuildanddistributedirsareclearedon`makeclean`
25. BUILD_DIR:=build
26. distribute_DIR:=distribute
28. #TheIDoftheGPUthat'makeruntest'willusetorununittests.
29. TEST_GPUID:=0
31. #enableprettybuild(commenttoseefullcommands)
32. Q?=@

3）修改Makefile文件：sudo geditMakefile，进行如下修改：

[plain] view plain copy

1. NVCCFLAGS+=-ccbin=$(CXX)-Xcompiler-fPIC$(COMMON_FLAGS)#这行去掉
2. NVCCFLAGS+=-D_FORCE_INLInes-ccbin=$(CXX)-Xcompiler-fPIC$(COMMON_FLAGS)#添加这行

4）进入/usr/lib/x86_64-linux-gnu目录下，执行如下命令（否则会提示：/usr/bin/ld: 找不到 -lhdf5_hl 和 /usr/bin/ld: 找不到 -lhdf5）

[plain] view plain copy

1. sudolnlibhdf5_serial.so.10.1.0libhdf5.so
2. sudolnlibhdf5_serial_hl.so.10.0.2libhdf5_hl.so
3. sudoldconfig#使连接生效

5）caffe源码中，python的一些必要安装：

[plain] view plain copy

1. cd/home/hcq/caffe/python
2. forreqin$(catrequirements.txt);dopipinstall$req;done

注：如果需要执行权限问题，就改变caffe文件的执行权限 sudo chown -R 'hcq' xxx

=== 编译caffe（期间遇到问题可以到第四部分查看是否有相关的解决方案）===

1）cd /home/hcq/caffe：进入caffe的根目录下

2）sudo make clean

3）sudo make all -j4 （-j4表示使用4核处理器执行当前指令）

4）sudomake test -j4（最好加上sudo防止有些文件的访问权限不够）

5）sudomake runtest-j4

6）sudo make pycaffe -j4（配置pycaffe）

=== 效果图展示 ===

1）sudo make all 成功后的效果图，如下所示：

2）sudo make runtest 成功后的效果图，如下所示：

[python] view plain copy

1. [OK]ArgMaxLayerTest/1.TestcpuMaxVal(1ms)
2. [RUN]ArgMaxLayerTest/1.TestSetupMaxVal
3. [OK]ArgMaxLayerTest/1.TestSetupMaxVal(1ms)
4. [RUN]ArgMaxLayerTest/1.TestSetupaxis
5. [OK]ArgMaxLayerTest/1.TestSetupaxis(0ms)
6. [RUN]ArgMaxLayerTest/1.TestCPutopK
7. [OK]ArgMaxLayerTest/1.TestCPutopK(1ms)
8. [----------]12testsfromArgMaxLayerTest/1(48mstotal)
10. [----------]4testsfromContrastiveLossLayerTest/1,whereTypeParam=caffe::cpuDevice<double>
11. [RUN]ContrastiveLossLayerTest/1.TestGradientLegacy
12. [OK]ContrastiveLossLayerTest/1.TestGradientLegacy(125ms)
13. [RUN]ContrastiveLossLayerTest/1.TestForward
14. [OK]ContrastiveLossLayerTest/1.TestForward(0ms)
15. [RUN]ContrastiveLossLayerTest/1.TestGradient
16. [OK]ContrastiveLossLayerTest/1.TestGradient(136ms)
17. [RUN]ContrastiveLossLayerTest/1.TestForwardLegacy
18. [OK]ContrastiveLossLayerTest/1.TestForwardLegacy(0ms)
19. [----------]4testsfromContrastiveLossLayerTest/1(261mstotal)
21. [----------]Globaltestenvironmenttear-down
22. [==========]1106testsfrom150testcasesran.(43894mstotal)
23. [PASSED]1106tests.
24. hcq@hcq-To-be-filled-by-O-E-M:~/caffe$

10、caffe的测试案例：

参考网址：http://caffe.berkeleyvision.org/gathered/examples/mnist.html

1）import caffe，输入如下命令：

== cd /home/hcq/caffe

== python

== import caffe

== 第一次成功后如下所示：

[plain] view plain copy

1. hcq@ubuntu:~/caffe/python$python
2. Python2.7.13|Anacondacustom(64-bit)|(default,Dec202016,23:09:15)
3. [GCC4.4.720120313(RedHat4.4.7-1)]onlinux2
4. Type"help","copyright","credits"or"license"formoreinformation.
5. AnacondaisbroughttoyoubyContinuumAnalytics.
6. Pleasecheckout:http://continuum.io/thanksandhttps://anaconda.org
7. >>>importcaffe
8. /home/hcq/anaconda2/lib/python2.7/site-packages/matplotlib/font_manager.py:280:UserWarning:Matplotlibisbuildingthefontcacheusingfc-list.Thismaytakeamoment.
9. 'Matplotlibisbuildingthefontcacheusingfc-list.'

2）MNIST 模型

==cd/home/hcq/caffe（确保当前目录是caffe目录），执行如下命令：

== sudo ./data/mnist/get_mnist.sh

== sudo ./examples/mnist/create_mnist.sh

==sudo./examples/mnist/train_lenet.sh

== 训练完成后如下所示：

[plain] view plain copy

1. I060113:09:21.12908329215solver.cpp:447]Snapshottingtobinaryprotofileexamples/mnist/lenet_iter_10000.caffemodel
2. I060113:09:21.13408929215sgd_solver.cpp:273]Snapshottingsolverstatetobinaryprotofileexamples/mnist/lenet_iter_10000.solverstate
3. I060113:09:21.16399029215solver.cpp:310]Iteration10000,loss=0.00275877
4. I060113:09:21.16401329215solver.cpp:330]Iteration10000,Testingnet(#0)
5. I060113:09:25.22633729217data_layer.cpp:73]Restartingdataprefetchingfromstart.
6. I060113:09:25.39416729215solver.cpp:397]Testnetoutput#0:accuracy=0.991
7. I060113:09:25.39420929215solver.cpp:397]Testnetoutput#1:loss=0.0280903(*1=0.0280903loss)
8. I060113:09:25.39421529215solver.cpp:315]OptimizationDone.
9. I060113:09:25.39423429215caffe.cpp:259]OptimizationDone.

3）Cifar-10 模型

== cd /home/hcq/caffe #cd到caffe源码的根目录下

== sudo ./data/cifar10/get_cifar10.sh #该脚本会下载二进制的cifar，并解压，会在/data/cifar10中出现很多batch文件

== sudo ./examples/cifar10/create_cifar10.sh #运行后将会在examples中出现数据集./cifar10_xxx_lmdb和数据集图像均值./mean.binaryproto

注：CIFAR-10的卷积神经网络模型由卷积层，pooling层，ReLU，非线性变换层，局部对比归一化线性分类器组成。该模型定义在CAFFE_ROOT/examples/cifar10/cifar10_quick_train_test.prototxt中。

== sudo./examples/cifar10/train_quick.sh #先以0.001的学习率迭代4000次，再以0.01的学习率接着再迭代1000次，共5000次

注：同理可以训练full模型，full模型比quick模型迭代次数多，一共迭代70000次，前60000次学习率是0.001，中间5000次学习率是0.0001，最后5000次学习率是0.00001。full模型的网络层数也比quick模型多，输入如下命令训练full模型

== sudo./examples/cifar10/train_full.sh

参考网址：

== caffe示例实现之1在CIFAR-10数据集上训练与测试Caffe：http://blog.csdn.net/liumaolincycle/article/details/47258937

== caffe官网：http://caffe.berkeleyvision.org/

训练成功后，提示如下所示，以sudo./examples/cifar10/train_quick.sh为例：

[plain] view plain copy

1. I060117:15:20.21963532594solver.cpp:237]Trainnetoutput#0:loss=0.550304(*1=0.550304loss)
2. I060117:15:20.21964232594sgd_solver.cpp:105]Iteration4700,lr=0.0001
3. I060117:16:15.67182332594solver.cpp:218]Iteration4800(1.80336iter/s,55.452s/100iters),loss=0.403176
4. I060117:16:15.67189132594solver.cpp:237]Trainnetoutput#0:loss=0.403176(*1=0.403176loss)
5. I060117:16:15.67189832594sgd_solver.cpp:105]Iteration4800,lr=0.0001
6. I060117:17:10.76178532594solver.cpp:218]Iteration4900(1.81524iter/s,55.089s/100iters),loss=0.466545
7. I060117:17:10.76190332594solver.cpp:237]Trainnetoutput#0:loss=0.466545(*1=0.466545loss)
8. I060117:17:10.76190732594sgd_solver.cpp:105]Iteration4900,lr=0.0001
9. I060117:18:02.53276432595data_layer.cpp:73]Restartingdataprefetchingfromstart.
10. I060117:18:04.70677932594solver.cpp:457]SnapshottingtoHDF5fileexamples/cifar10/cifar10_quick_iter_5000.caffemodel.h5
11. I060117:18:04.70835332594sgd_solver.cpp:283]SnapshottingsolverstatetoHDF5fileexamples/cifar10/cifar10_quick_iter_5000.solverstate.h5
12. I060117:18:04.92949232594solver.cpp:310]Iteration5000,loss=0.451509
13. I060117:18:04.92952232594solver.cpp:330]Iteration5000,Testingnet(#0)
14. I060117:18:26.00878632597data_layer.cpp:73]Restartingdataprefetchingfromstart.
15. I060117:18:26.88809732594solver.cpp:397]Testnetoutput#0:accuracy=0.7575
16. I060117:18:26.88812332594solver.cpp:397]Testnetoutput#1:loss=0.737631(*1=0.737631loss)
17. I060117:18:26.88814132594solver.cpp:315]OptimizationDone.
18. I060117:18:26.88814432594caffe.cpp:259]OptimizationDone.

4）使用训练的模型来分类新数据：

== 确保 sudo make pycaffe 已经编译成功。

== 打开终端，在caffe源码的根目录下，执行如下命令：

[python] view plain copy

1. pythonpython/classify.pyexamples/images/cat.jpgfoo#直接用一下别人的模型分类试一下：（默认用的ImageNet的模型）

[python] view plain copy

1. hcq@hcq-To-be-filled-by-O-E-M:~/caffe$pythonpython/classify.pyexamples/images/cat.jpgfoo
2. cpumode
3. WARNING:LoggingbeforeInitGoogleLogging()iswrittentoSTDERR
4. W060211:25:53.5616469707_caffe.cpp:139]DEPRECATIONWARNING-deprecateduSEOfPythoninterface
5. W060211:25:53.5616989707_caffe.cpp:140]Usethisinstead(withthenamed"weights"parameter):
6. W060211:25:53.5617099707_caffe.cpp:142]Net('python/../models/bvlc_reference_caffenet/deploy.prototxt',1,weights='python/../models/bvlc_reference_caffenet/bvlc_reference_caffenet.caffemodel')
7. Traceback(mostrecentcalllast):
8. File"python/classify.py",line138,in<module>
9. main(sys.argv)
10. File"python/classify.py",line110,inmain
11. channel_swap=channel_swap)
12. File"/home/hcq/caffe/python/caffe/classifier.py",line26,in__init__
13. caffe.Net.__init__(self,model_file,pretrained_file,caffe.TEST)
14. RuntimeError:Couldnotopenfilepython/../models/bvlc_reference_caffenet/bvlc_reference_caffenet.caffemodel

上面这句话的意思是，使用cifar10_quick.prototxt网络 + cifar10_quick_iter_4000.caffemodel.h5模型，对examples/images/cat.jpg图片进行分类。

错误原因：

解决方案：

1）cd/home/hcq/caffe/script

2）python download_model_binary.py /home/hcq/caffe/models/bvlc_reference_caffenet

参考网址：http://blog.csdn.net/u011636440/article/details/52441519

解决上述问题后，又出现新问题提示如下所示：

[plain] view plain copy

1. I060211:51:24.16338910753upgrade_proto.cpp:53]AttemptingtoupgradeinputfilespecifiedusingdeprecatedV1LayerParameter:python/../models/bvlc_reference_caffenet/bvlc_reference_caffenet.caffemodel
2. I060211:51:24.36864610753upgrade_proto.cpp:61]SuccessfullyupgradedfilespecifiedusingdeprecatedV1LayerParameter
3. I060211:51:24.43614910753net.cpp:744]Ignoringsourcelayerloss
4. Traceback(mostrecentcalllast):
5. File"python/classify.py",line138,in<module>
6. main(sys.argv)
7. File"python/classify.py",line110,inmain
8. channel_swap=channel_swap)
9. File"/home/hcq/caffe/python/caffe/classifier.py",line34,in__init__
10. self.transformer.set_mean(in_,mean)
11. File"/home/hcq/caffe/python/caffe/io.py",line259,inset_mean
12. raiseValueError('Meanshapeincompatiblewithinputshape.')
13. ValueError:Meanshapeincompatiblewithinputshape.

错误原因：ValueError: Mean shape incompatible with input shape. 是因为‘mean’与‘input’的尺寸不同造成的。

解决方案：修改了 ./python/caffe/io.py 文件，如下

将：

[plain] view plain copy

1. ifms!=self.inputs[in_][1:]:
2. raiseValueError('Meanshapeincompatiblewithinputshape.')

修改为：

[python] view plain copy

1. ifms!=self.inputs[in_][1:]:
2. print(self.inputs[in_])
3. in_shape=self.inputs[in_][1:]
4. m_min,m_max=mean.min(),mean.max()
5. normal_mean=(mean-m_min)/(m_max-m_min)
6. mean=resize_image(normal_mean.transpose((1,2,0)),
7. in_shape[1:]).transpose((2,0,1))*\
8. (m_max-m_min)+m_min
9. #raiseValueError('Meanshapeincompatiblewithinputshape.')

参考网址：http://blog.csdn.net/gzljss/article/details/45849013
运行成功后，提示如下所示：

[python] view plain copy

1. I060212:00:07.86424810953net.cpp:200]conv1doesnotneedbackwardcomputation.
2. I060212:00:07.86425110953net.cpp:200]datadoesnotneedbackwardcomputation.
3. I060212:00:07.86425410953net.cpp:242]Thisnetworkproducesoutputprob
4. I060212:00:07.86426610953net.cpp:255]Networkinitializationdone.
5. I060212:00:07.98715710953upgrade_proto.cpp:44]Attemptingtoupgradeinputfilespecifiedusingdeprecatedtransformationparameters:python/../models/bvlc_reference_caffenet/bvlc_reference_caffenet.caffemodel
6. I060212:00:07.98719010953upgrade_proto.cpp:47]Successfullyupgradedfilespecifiedusingdeprecateddatatransformationparameters.
7. W060212:00:07.98719510953upgrade_proto.cpp:49]NotethatfutureCaffereleaseswillonlysupporttransform_parammessagesfortransformationfields.
8. I060212:00:07.98719810953upgrade_proto.cpp:53]AttemptingtoupgradeinputfilespecifiedusingdeprecatedV1LayerParameter:python/../models/bvlc_reference_caffenet/bvlc_reference_caffenet.caffemodel
9. I060212:00:08.21731310953upgrade_proto.cpp:61]SuccessfullyupgradedfilespecifiedusingdeprecatedV1LayerParameter
10. I060212:00:08.28918610953net.cpp:744]Ignoringsourcelayerloss
11. (10,3,227,227)
12. Loadingfile:examples/images/cat.jpg
13. Classifying1inputs.
14. Donein0.43s.
15. Savingresultsintofoo

== 使用cifar10_quick_iter_4000.caffemodel.h5模型：

[python] view plain copy

1. pythonpython/classify.py--model_defexamples/cifar10/cifar10_quick.prototxt--pretrained_modelexamples/cifar10/cifar10_quick_iter_4000.caffemodel.h5--center_onlyexamples/images/cat.jpgfoo#指定自己的模型进行分类

== 错误提示如下：

[python] view plain copy

1. hcq@hcq-To-be-filled-by-O-E-M:~/caffe$pythonpython/classify.py--model_defexamples/cifar10/cifar10_quick.prototxt--pretrained_modelexamples/cifar10/cifar10_quick_iter_4000.caffemodel.h5--center_onlyexamples/images/cat.jpgfoo
2. cpumode
3. WARNING:LoggingbeforeInitGoogleLogging()iswrittentoSTDERR
4. W060211:27:22.8577199729_caffe.cpp:139]DEPRECATIONWARNING-deprecateduSEOfPythoninterface
5. W060211:27:22.8577569729_caffe.cpp:140]Usethisinstead(withthenamed"weights"parameter):
6. W060211:27:22.8577599729_caffe.cpp:142]Net('examples/cifar10/cifar10_quick.prototxt',weights='examples/cifar10/cifar10_quick_iter_4000.caffemodel.h5')
7. Traceback(mostrecentcalllast):
8. File"python/classify.py",caffe.TEST)
9. RuntimeError:Couldnotopenfileexamples/cifar10/cifar10_quick_iter_4000.caffemodel.h5

原因：在目录/home/hcq/caffe/examples/cifar10下，确实不存在 cifar10_quick_iter_4000.caffemodel.h5这个模型，运行结果必然错误。

解决方案：到目录/home/hcq/caffe/examples/cifar10下查看已有的模型，然后把指定的模型换成已有的模型，已有的模型如下所示（因人而异）：

[html] view plain copy

1. hcq@hcq-To-be-filled-by-O-E-M:~/caffe/examples/cifar10$ls-lrt
2. 总用量6788
3. -rwxrwxr-x1hcqhcq3385月2711:55train_quick.sh
4. -rwxrwxr-x1hcqhcq1395月2711:55train_full_sigmoid.sh
5. -rwxrwxr-x1hcqhcq1425月2711:55train_full_sigmoid_bn.sh
6. -rwxrwxr-x1hcqhcq5245月2711:55train_full.sh
7. -rw-rw-r--1hcqhcq52435月2711:55readme.md
8. -rwxrwxr-x1hcqhcq4675月2711:55create_cifar10.sh
9. -rw-rw-r--1hcqhcq36775月2711:55convert_cifar_data.cpp
10. -rw-rw-r--1hcqhcq19215月2711:55cifar10_quick.prototxt
11. -rw-rw-r--1hcqhcq31225月2711:55cifar10_full_train_test.prototxt
12. -rw-rw-r--1hcqhcq28795月2711:55cifar10_full_sigmoid_train_test.prototxt
13. -rw-rw-r--1hcqhcq31925月2711:55cifar10_full_sigmoid_train_test_bn.prototxt
14. -rw-rw-r--1hcqhcq21745月2711:55cifar10_full.prototxt
15. drwxr--r--2hcqroot40966月115:38cifar10_train_lmdb
16. drwxr--r--2hcqroot40966月115:38cifar10_test_lmdb
17. -rw-r--r--1hcqroot122996月115:38mean.binaryproto
18. -rw-rw-r--1hcqhcq30886月115:39cifar10_quick_train_test.prototxt
19. -rw-rw-r--1hcqhcq8596月116:25cifar10_quick_solver.prototxt
20. -rw-rw-r--1hcqhcq9446月116:25cifar10_full_solver_lr1.prototxt
21. -rw-rw-r--1hcqhcq8826月116:26cifar10_quick_solver_lr1.prototxt
22. -rw-rw-r--1hcqhcq9446月116:26cifar10_full_solver.prototxt
23. -rw-rw-r--1hcqhcq9596月116:26cifar10_full_sigmoid_solver_bn.prototxt
24. -rw-rw-r--1hcqhcq9536月116:26cifar10_full_sigmoid_solver.prototxt
25. -rw-rw-r--1hcqhcq9456月116:26cifar10_full_solver_lr2.prototxt
26. -rw-r--r--1hcqroot3660486月201:37cifar10_full_iter_10000.solverstate.h5
27. -rw-r--r--1hcqroot3757526月201:37cifar10_full_iter_10000.caffemodel.h5
28. -rw-r--r--1hcqroot3660486月202:34cifar10_full_iter_20000.solverstate.h5
29. -rw-r--r--1hcqroot3757526月202:34cifar10_full_iter_20000.caffemodel.h5
30. -rw-r--r--1hcqroot3660486月203:30cifar10_full_iter_30000.solverstate.h5
31. -rw-r--r--1hcqroot3757526月203:30cifar10_full_iter_30000.caffemodel.h5
32. -rw-r--r--1hcqroot3660486月204:27cifar10_full_iter_40000.solverstate.h5
33. -rw-r--r--1hcqroot3757526月204:27cifar10_full_iter_40000.caffemodel.h5
34. -rw-r--r--1hcqroot3660486月205:24cifar10_full_iter_50000.solverstate.h5
35. -rw-r--r--1hcqroot3757526月205:24cifar10_full_iter_50000.caffemodel.h5
36. -rw-r--r--1hcqroot3660486月206:21cifar10_full_iter_60000.solverstate.h5
37. -rw-r--r--1hcqroot3757526月206:21cifar10_full_iter_60000.caffemodel.h5
38. -rw-r--r--1hcqroot5824966月216:40cifar10_quick_iter_4000.solverstate
39. -rw-r--r--1hcqroot5833016月216:40cifar10_quick_iter_4000.caffemodel
40. -rw-r--r--1hcqroot5900646月216:44cifar10_quick_iter_5000.solverstate.h5
41. -rw-r--r--1hcqroot6000326月216:44cifar10_quick_iter_5000.caffemodel.h5

== 选择【cifar10_quick_iter_5000.caffemodel.h5】作为指定模型，再次运行如下命令：

[python] view plain copy

1. pythonpython/classify.py--model_defexamples/cifar10/cifar10_quick.prototxt--pretrained_modelexamples/cifar10/cifar10_quick_iter_5000.caffemodel.h5--center_onlyexamples/images/cat.jpgfoo

== 运行成功后，提示如下所示：

[html] view plain copy

1. cpumode
2. WARNING:LoggingbeforeInitGoogleLogging()iswrittentoSTDERR
3. W060216:48:24.26168312766_caffe.cpp:139]DEPRECATIONWARNING-deprecateduSEOfPythoninterface
4. W060216:48:24.26173212766_caffe.cpp:140]Usethisinstead(withthenamed"weights"parameter):
5. W060216:48:24.26173512766_caffe.cpp:142]Net('examples/cifar10/cifar10_quick.prototxt',1,weights='examples/cifar10/cifar10_quick_iter_5000.caffemodel.h5')
6. I060216:48:24.26283012766net.cpp:51]Initializingnetfromparameters:
7. name:"CIFAR10_quick_test"
8. state{
9. phase:TEST
10. level:0
11. }
12. layer{
13. ...
14. ...
15. I060216:48:24.26377412766net.cpp:242]Thisnetworkproducesoutputprob
16. I060216:48:24.26378212766net.cpp:255]Networkinitializationdone.
17. I060216:48:24.26453912766net.cpp:798]Ignoringsourcelayercifar
18. I060216:48:24.26462712766hdf5.cpp:32]Datatypeclass:H5T_FLOAT
19. I060216:48:24.26504812766net.cpp:798]Ignoringsourcelayerloss
20. (1,3,32,32)
21. Loadingfile:examples/images/cat.jpg
22. Classifying1inputs.
23. Donein0.03s.
24. Savingresultsintofoo

注：默认的classify脚本不会直接输出结果，而是会把结果输入到foo文件里，不太直观，网上有一个修改版，添加了一些参数，可以输出概率最高的分类。

替换Python/classify.py，下载地址：

这个脚本添加了两个参数，可以指定labels_file，然后可以直接把分类结果输出出来：

[python] view plain copy

1. pythonpython/classify.py--print_results--model_defexamples/cifar10/cifar10_quick.prototxt--pretrained_modelexamples/cifar10/cifar10_quick_iter_5000.caffemodel.h5--labels_filedata/cifar10/cifar10_words.txt--center_onlyexamples/images/cat.jpgfoo

Caffe学习资料：

1、Classification: Instant Recognition with Caffe：http://nbviewer.jupyter.org/github/BVLC/caffe/blob/master/examples/00-classification.ipynb

三、cpu版caffe环境配置过程

1、cpu和GPU版本的一些区别：

1）cpu版本的话，不需要NVIDIA驱动，不需要安装CUDA，不需要配置CUDNN

2）caffe目录下的Makefile.configure配置不一样

3）配置模型的时候，solver_mode: cpu，变成cpu（如果没改的话会报错），如下所示：

2、cpu版本的Makefile.configure配置，未注释掉的部分，如下所示：

[python] view plain copy

1. #cpu-onlyswitch(uncommenttobuildwithoutGPUsupport).
2. cpu_ONLY:=1
4. #Uncommentifyou'reusingOpenCV3
5. OPENCV_VERSION:=3
7. #openforOpenBlas
8. BLAS:=open
9. #(whichshouldwork)!
10. BLAS_INCLUDE:=/opt/OpenBLAS/include
11. BLAS_LIB:=/opt/OpenBLAS/lib
13. ANACONDA_HOME:=/home/hcq/anaconda2
14. PYTHON_INCLUDE:=$(ANACONDA_HOME)/include\
15. $(ANACONDA_HOME)/include/python2.7\
16. $(ANACONDA_HOME)/lib/python2.7/site-packages/numpy/core/include
18. PYTHON_LIB:=$(ANACONDA_HOME)/lib
20. #Whateverelseyoufindyouneedgoeshere.
21. INCLUDE_Dirs:=$(PYTHON_INCLUDE)/usr/local/include/usr/include/hdf5/serial
22. LIBRARY_Dirs:=$(PYTHON_LIB)/usr/local/lib/usr/lib/usr/lib/x86_64-linux-gnu/usr/lib/x86_64-linux-gnu/hdf5/serial
24. #N.B.bothbuildanddistributedirsareclearedon`makeclean`
25. BUILD_DIR:=build
26. distribute_DIR:=distribute
28. #TheIDoftheGPUthat'makeruntest'willusetorununittests.
29. TEST_GPUID:=0
31. #enableprettybuild(commenttoseefullcommands)
32. Q?=@

3、测试案例跟GPU的类似，只不过需要把solver_mode: cpu改为cpu。

四、问题集锦

1、sudo make all 编译过程遇到的错误，【conda install libgcc】安装libgcc即可，错误提示如下所示：

[plain] view plain copy

1. /usr/local/lib/libglog.so：对‘std::__cxx11::basic_ostringstream<char,std::char_traits<char>,std::allocator<char>>::~basic_ostringstream()@GLIBCXX_3.4.21’未定义的引用
2. /usr/lib/x86_64-linux-gnu/libgflags.so：对‘std::__cxx11::basic_string<char,std::allocator<char>>::find(charconst*,unsignedlong,unsignedlong)const@GLIBCXX_3.4.21’未定义的引用
3. .build_release/lib/libcaffe.so：对‘std::out_of_range::out_of_range(std::__cxx11::basic_string<char,std::allocator<char>>const&)’未定义的引用
4. .build_release/lib/libcaffe.so：对‘std::__cxx11::basic_string<char,std::allocator<char>>::_M_replace_aux(unsignedlong,char)’未定义的引用
5. /usr/local/lib/libopencv_core.so：对‘std::basic_istream<char,std::char_traits<char>>&std::getline<char,std::allocator<char>>(std::basic_istream<char,std::char_traits<char>>&,std::__cxx11::basic_string<char,std::allocator<char>>&,char)@GLIBCXX_3.4.21’未定义的引用
6. .build_release/lib/libcaffe.so：对‘std::__cxx11::basic_string<char,std::allocator<char>>::compare(unsignedlong,charconst*)const’未定义的引用
7. /usr/lib/x86_64-linux-gnu/libprotobuf.so：对‘std::__cxx11::basic_string<char,std::allocator<char>>::swap(std::__cxx11::basic_string<char,std::allocator<char>>&)@GLIBCXX_3.4.21’未定义的引用
8. /usr/local/lib/libglog.so：对‘vtableforstd::__cxx11::basic_stringbuf<char,std::allocator<char>>@GLIBCXX_3.4.21’未定义的引用
9. /usr/local/lib/libopencv_core.so：对‘std::__cxx11::basic_stringbuf<char,std::allocator<char>>::str()const@GLIBCXX_3.4.21’未定义的引用
10. collect2:error:ldreturned1exitstatus
11. Makefile:629:recipefortarget'.build_release/tools/upgrade_solver_proto_text.bin'Failed
12. make:***[.build_release/tools/upgrade_solver_proto_text.bin]Error1

或者如下所示：

[plain] view plain copy

1. AR-o.build_release/lib/libcaffe.a
2. LD-o.build_release/lib/libcaffe.so.1.0.0
3. CXX/LD-o.build_release/tools/upgrade_solver_proto_text.bin
4. CXX/LD-o.build_release/tools/test_net.bin
5. CXX/LD-o.build_release/tools/upgrade_net_proto_text.bin
6. CXX/LD-o.build_release/tools/finetune_net.bin
7. .build_release/tools/test_net.o:Infunction`boost::system::system_error::what()const':
8. test_net.cpp:(.text._ZNK5boost6system12system_error4whatEv[_ZNK5boost6system12system_error4whatEv]+0x74):undefinedreferenceto`std::__cxx11::basic_string<char,std::allocator<char>>::_M_replace(unsignedlong,charconst*,unsignedlong)'
9. test_net.cpp:(.text._ZNK5boost6system12system_error4whatEv[_ZNK5boost6system12system_error4whatEv]+0xa1):undefinedreferenceto`std::__cxx11::basic_string<char,std::allocator<char>>::_M_append(charconst*,unsignedlong)'
10. test_net.cpp:(.text._ZNK5boost6system12system_error4whatEv[_ZNK5boost6system12system_error4whatEv]+0xe3):undefinedreferenceto`std::__cxx11::basic_string<char,unsignedlong)'
11. /usr/local/lib/libglog.so:undefinedreferenceto`std::__cxx11::basic_string<char,std::allocator<char>>::_M_create(unsignedlong&,unsignedlong)@GLIBCXX_3.4.21'
12. /usr/local/lib/libglog.so:undefinedreferenceto`std::__cxx11::basic_string<char,char)@GLIBCXX_3.4.21'
13. ...
14. ...
15. ...
16. .build_release/lib/libcaffe.so:undefinedreferenceto`std::out_of_range::out_of_range(std::__cxx11::basic_string<char,std::allocator<char>>const&)'
17. .build_release/lib/libcaffe.so:undefinedreferenceto`std::__cxx11::basic_string<char,char)'
18. /usr/local/lib/libopencv_core.so:undefinedreferenceto`std::basic_istream<char,char)@GLIBCXX_3.4.21'
19. .build_release/lib/libcaffe.so:undefinedreferenceto`std::__cxx11::basic_string<char,charconst*)const'
20. /usr/lib/gcc/x86_64-linux-gnu/5/../../../x86_64-linux-gnu/libprotobuf.so:undefinedreferenceto`std::__cxx11::basic_string<char,std::allocator<char>>&)@GLIBCXX_3.4.21'
21. /usr/local/lib/libglog.so:undefinedreferenceto`vtableforstd::__cxx11::basic_stringbuf<char,std::allocator<char>>@GLIBCXX_3.4.21'
22. /usr/local/lib/libopencv_core.so:undefinedreferenceto`std::__cxx11::basic_stringbuf<char,std::allocator<char>>::str()const@GLIBCXX_3.4.21'
23. collect2:error:ldreturned1exitstatus
24. Makefile:629:recipefortarget'.build_release/tools/upgrade_net_proto_text.bin'Failed
25. make:***[.build_release/tools/upgrade_net_proto_text.bin]Error1

2、sudo make runtest 时遇到的错误，【sudo apt install ffmpeg】安装ffmpeg即可，错误提示如下所示：

[plain] view plain copy

1. hcq@ubuntu:~/caffe$sudomakeruntest
2. .build_release/tools/caffe
3. .build_release/tools/caffe:errorwhileloadingsharedlibraries:libopencv_core.so.3.2:cannotopensharedobjectfile:Nosuchfileordirectory
4. Makefile:538:recipefortarget'runtest'Failed
5. make:***[runtest]Error127

3、sudo make runtest 时遇到的错误，缺失一些相关的so链接：

[plain] view plain copy

1. hcq@hcq-To-be-filled-by-O-E-M:~/caffe$sudomakeruntest-j4
2. .build_release/tools/caffe
3. .build_release/tools/caffe:errorwhileloadingsharedlibraries:libhdf5_hl.so.10:cannotopensharedobjectfile:Nosuchfileordirectory
4. Makefile:536:recipefortarget'runtest'Failed
5. make:***[runtest]Error127

以及

[plain] view plain copy

1. hcq@hcq-To-be-filled-by-O-E-M:~/caffe$sudomakeruntest-j4
2. .build_release/tools/caffe
3. .build_release/tools/caffe:errorwhileloadingsharedlibraries:libhdf5.so.10:cannotopensharedobjectfile:Nosuchfileordirectory
4. Makefile:536:recipefortarget'runtest'Failed
5. make:***[runtest]Error127

原因：目录/usr/lib/x86_64-linux-gnu下确实是找不到【libhdf5_hl.so.10】和【libhdf5.so.10】这两个文件，如下图所示：

解决方案：根据相应的文件，创建所需的软连接，如下所示：

[plain] view plain copy

1. sudolnlibhdf5_serial_hl.so.10libhdf5_hl.so.10
2. sudolnlibhdf5_serial.so.10libhdf5.so.10

4、Makefile.configure配置中，关于nvcc warning的一些警告，注释掉相关的配置代码即可，提示如下所示：

[plain] view plain copy

1. NVCCsrc/caffe/layers/mvn_layer.cu
2. nvccwarning:The'compute_20','sm_20',and'sm_21'architecturesaredeprecated,andmayberemovedinafuturerelease(Use-Wno-deprecated-gpu-targetstosuppresswarning).
3. nvccwarning:The'compute_20',andmayberemovedinafuturerelease(Use-Wno-deprecated-gpu-targetstosuppresswarning).
4. NVCCsrc/caffe/layers/filter_layer.cu
5. nvccwarning:The'compute_20',andmayberemovedinafuturerelease(Use-Wno-deprecated-gpu-targetstosuppresswarning).
6. nvccwarning:The'compute_20',andmayberemovedinafuturerelease(Use-Wno-deprecated-gpu-targetstosuppresswarning).

5、Check Failed: status == CUDNN_STATUS_SUCCESS (3 vs. 0) CUDNN_STATUS_BAD_ParaM

[plain] view plain copy

1. F060110:03:16.68394028607cudnn.hpp:122]CheckFailed:status==CUDNN_STATUS_SUCCESS(3vs.0)CUDNN_STATUS_BAD_ParaM
2. ***Checkfailurestacktrace:***
3. @0x7f52e7de2e3dgoogle::LogMessage::Fail()
4. @0x7f52e7de4bc0google::LogMessage::SendToLog()
5. @0x7f52e7de2a23google::LogMessage::Flush()
6. @0x7f52e7de558egoogle::LogMessageFatal::~LogMessageFatal()
7. @0x7f52e313c5f8caffe::CuDNNConvolutionLayer<>::Reshape()
8. @0x7f52e325ed37caffe::Net<>::Init()
9. @0x7f52e326146ecaffe::Net<>::Net()
10. @0x888f2ccaffe::NetTest<>::InitNetFromProtoString()
11. @0x88e1d1caffe::NetTest_TestReshape_Test<>::TestBody()
12. @0x91e4b3testing::internal::HandleExceptionsInMethodIfSupported<>()
13. @0x917acatesting::Test::Run()
14. @0x917c18testing::TestInfo::Run()
15. @0x917cf5testing::TestCase::Run()
16. @0x918fcftesting::internal::UnitTestImpl::RunAllTests()
17. @0x9192f3testing::UnitTest::Run()
18. @0x46d44dmain
19. @0x7f52e2443830(unkNown)
20. @0x474f69_start
21. Makefile:536:recipefortarget'runtest'Failed
22. make:***[runtest]已放弃(coredumped)

原因：初步断定是NVIDIA的显存不够，本文所使用的机器是GeForce GTX 960 2GB的显存。本人做了如下探究：

1）据同学介绍，他使用的机器显存有8GB，就不会出现这样的问题。

2）本人还尝试使用cpu( 物理内存8GB )运行sudo make runtest -j4，也不会出现这样的问题。

解决方案：

1）从机器层面入手，扩大硬件性能（主要是显存）

2）改用cpu跑程序，不过最后使用GPU的话还是需要使用好一点的设备（显卡、显卡）

参考网址：http://www.cnblogs.com/penguinliong/p/6351215.html

6、build_release/tools/caffe: error while loading shared libraries: libopenblas.so.0: cannot open shared object file: No such file or directory

描述：Makefile.configure配置好了，相应的引用位置也有libopenblas.so.0 该文件，make runtest的时候仍然出错。

解决方案：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。打开终端，输入如下命令：

1）export LD_LIBRARY_PATH=/opt/OpenBLAS/lib/

2）sudo ldconfig

参考网址：

1）http://www.jb51.cc/article/p-hgkqqphr-qh.html

2）http://www.jb51.cc/article/p-mxwdsawn-ek.html

3）https://github.com/sermanet/OverFeat/issues/10

配置成功后的提示如下所示：

[plain] view plain copy

1. [OK]ArgMaxLayerTest/1.TestcpuMaxVal(1ms)
2. [RUN]ArgMaxLayerTest/1.TestSetupMaxVal
3. [OK]ArgMaxLayerTest/1.TestSetupMaxVal(1ms)
4. [RUN]ArgMaxLayerTest/1.TestSetupaxis
5. [OK]ArgMaxLayerTest/1.TestSetupaxis(0ms)
6. [RUN]ArgMaxLayerTest/1.TestCPutopK
7. [OK]ArgMaxLayerTest/1.TestCPutopK(1ms)
8. [----------]12testsfromArgMaxLayerTest/1(48mstotal)
10. [----------]4testsfromContrastiveLossLayerTest/1,whereTypeParam=caffe::cpuDevice<double>
11. [RUN]ContrastiveLossLayerTest/1.TestGradientLegacy
12. [OK]ContrastiveLossLayerTest/1.TestGradientLegacy(125ms)
13. [RUN]ContrastiveLossLayerTest/1.TestForward
14. [OK]ContrastiveLossLayerTest/1.TestForward(0ms)
15. [RUN]ContrastiveLossLayerTest/1.TestGradient
16. [OK]ContrastiveLossLayerTest/1.TestGradient(136ms)
17. [RUN]ContrastiveLossLayerTest/1.TestForwardLegacy
18. [OK]ContrastiveLossLayerTest/1.TestForwardLegacy(0ms)
19. [----------]4testsfromContrastiveLossLayerTest/1(261mstotal)
21. [----------]Globaltestenvironmenttear-down
22. [==========]1106testsfrom150testcasesran.(43894mstotal)
23. [PASSED]1106tests.
24. hcq@hcq-To-be-filled-by-O-E-M:~/caffe$

7、Check Failed: fd != -1 (-1 vs. -1) File not found: examples/cifar10/cifar10_full_iter_60000.solverstate

描述：执行【sudo./examples/cifar10/train_full.sh】报错。

解决方案：

相似问题：Check Failed: fd != -1 (-1 vs. -1) File not found: deploy.prototxt

原因：因为没有把deploy.prototxt文件复制到demo下。

解决方案：把caffe/models/bvlc_reference_caffnnet/deploy.prototxt复制到demo下即可。

参考网址：http://www.jb51.cc/article/p-fwfghnpr-qq.html

8、error while loading shared libraries: libhdf5_hl.so.100

[html] view plain copy

1. build_release/tools/caffe:errorwhileloadingsharedlibraries:libhdf5_hl.so.100:cannotopensharedobjectfile:Nosuchfileordirectory
2. Makefile:470:recipefortarget'runtest'Failed

参考网址： http://www.jb51.cc/article/p-qixhtjny-bpx.html

9、Makefile:567: recipe for target '.build_release/examples/cifar10/convert_cifar_data.bin' Failed

[html] view plain copy

1. Makefile:567:recipefortarget'.build_release/examples/mnist/convert_mnist_data.bin'Failed
2. make:***[.build_release/examples/mnist/convert_mnist_data.bin]Error1
3. /usr/lib/x86_64-linux-gnu/libopencv_highgui.so:undefinedreferenceto`TIFFReadRGBAStrip@LIBTIFF_4.0'
4. /usr/lib/x86_64-linux-gnu/libopencv_highgui.so:undefinedreferenceto`TIFFIsTiled@LIBTIFF_4.0'
5. /usr/lib/x86_64-linux-gnu/libopencv_highgui.so:undefinedreferenceto`TIFFWriteScanline@LIBTIFF_4.0'
6. /usr/lib/x86_64-linux-gnu/libopencv_highgui.so:undefinedreferenceto`TIFfgetField@LIBTIFF_4.0'
7. /usr/lib/x86_64-linux-gnu/libopencv_highgui.so:undefinedreferenceto`TIFFScanlinesize@LIBTIFF_4.0'
8. /usr/lib/x86_64-linux-gnu/libopencv_highgui.so:undefinedreferenceto`TIFFReadEncodedTile@LIBTIFF_4.0'
9. /usr/lib/x86_64-linux-gnu/libopencv_highgui.so:undefinedreferenceto`TIFFReadRGBATile@LIBTIFF_4.0'
10. /usr/lib/x86_64-linux-gnu/libopencv_highgui.so:undefinedreferenceto`TIFFClose@LIBTIFF_4.0'
11. /usr/lib/x86_64-linux-gnu/libopencv_highgui.so:undefinedreferenceto`TIFFRGBaimageOK@LIBTIFF_4.0'
12. /usr/lib/x86_64-linux-gnu/libopencv_highgui.so:undefinedreferenceto`TIFFOpen@LIBTIFF_4.0'
13. /usr/lib/x86_64-linux-gnu/libopencv_highgui.so:undefinedreferenceto`TIFFReadEncodedStrip@LIBTIFF_4.0'
14. /usr/lib/x86_64-linux-gnu/libopencv_highgui.so:undefinedreferenceto`TIFFSetField@LIBTIFF_4.0'
15. /usr/lib/x86_64-linux-gnu/libopencv_highgui.so:undefinedreferenceto`TIFFSetWarningHandler@LIBTIFF_4.0'
16. /usr/lib/x86_64-linux-gnu/libopencv_highgui.so:undefinedreferenceto`TIFFSetErrorHandler@LIBTIFF_4.0'
17. collect2:error:ldreturned1exitstatus
18. Makefile:567:recipefortarget'.build_release/examples/cifar10/convert_cifar_data.bin'Failed
19. make:***[.build_release/examples/cifar10/convert_cifar_data.bin]Error1

reference： https://www.cnblogs.com/whu-zeng/p/6874307.html

9、Ubuntu 16.04中使用Matlab R2016b编译Caffe接口出错的解决方法

[html] view plain copy

1. MEX-file'caffe/matlab/+caffe/private/caffe_.mexa64'无效:
2. caffe/matlab/+caffe/private/caffe_.mexa64:undefined
3. symbol:_ZN2cv8imencodeERKNSt7__cxx1112basic_stringIcSt11char_traitsIcESaIcEEERKNS_11_InputArrayERSt6vectorIhSaIhEERKSB_IiSaIiEE。
5. 出错caffe.set_mode_cpu(line5)
6. caffe_('set_mode_cpu');
8. 出错caffe.run_tests(line6)
9. caffe.set_mode_cpu();

reference： https://github.com/BVLC/caffe/issues/3934

10、gcc 降级后的问题

[html] view plain copy

1. /usr/lib/x86_64-linux-gnu/libopencv_highgui.so:undefinedreferenceto`TIFfgetField@LIBTIFF_4.0'
2. /usr/lib/x86_64-linux-gnu/libopencv_highgui.so:undefinedreferenceto`TIFFScanlinesize@LIBTIFF_4.0'
3. .build_release/lib/libcaffe.so:undefinedreferenceto`google::protobuf::internal::NameOfEnum(google::protobuf::EnumDescriptorconst*,int)'
4. .build_release/lib/libcaffe.so:undefinedreferenceto`cv::imread(std::stringconst&,int)'
5. /usr/lib/x86_64-linux-gnu/libopencv_highgui.so:undefinedreferenceto`TIFFReadEncodedTile@LIBTIFF_4.0'
6. /usr/lib/x86_64-linux-gnu/libopencv_highgui.so:undefinedreferenceto`TIFFReadRGBATile@LIBTIFF_4.0'
7. /usr/lib/x86_64-linux-gnu/libopencv_highgui.so:undefinedreferenceto`TIFFClose@LIBTIFF_4.0'
8. .build_release/lib/libcaffe.so:undefinedreferenceto`google::protobuf::DescriptorPool::FindFileByName(std::stringconst&)const'
9. /usr/lib/x86_64-linux-gnu/libopencv_highgui.so:undefinedreferenceto`TIFFRGBaimageOK@LIBTIFF_4.0'
10. .build_release/lib/libcaffe.so:undefinedreferenceto`google::protobuf::internal::WireFormatLite::ReadBytes(google::protobuf::io::CodedInputStream*,std::string*)'
11. .build_release/lib/libcaffe.so:undefinedreferenceto`google::protobuf::internal::StringTypeHandlerBase::Delete(std::string*)'
12. /usr/lib/x86_64-linux-gnu/libopencv_highgui.so:undefinedreferenceto`TIFFOpen@LIBTIFF_4.0'
13. .build_release/lib/libcaffe.so:undefinedreferenceto`google::protobuf::MessageFactory::InternalRegisterGeneratedFile(charconst*,void(*)(std::stringconst&))'
14. .build_release/lib/libcaffe.so:undefinedreferenceto`leveldb::DB::Open(leveldb::Optionsconst&,std::stringconst&,leveldb::DB**)'
15. .build_release/lib/libcaffe.so:undefinedreferenceto`cv::imencode(std::stringconst&,cv::_InputArrayconst&,std::vector<unsignedchar,std::allocator<unsignedchar>>&,std::vector<int,std::allocator<int>>const&)'
16. /usr/lib/x86_64-linux-gnu/libopencv_highgui.so:undefinedreferenceto`TIFFReadEncodedStrip@LIBTIFF_4.0'
17. /usr/lib/x86_64-linux-gnu/libopencv_highgui.so:undefinedreferenceto`TIFFSetField@LIBTIFF_4.0'
18. .build_release/lib/libcaffe.so:undefinedreferenceto`google::protobuf::internal::StringTypeHandlerBase::New()'
19. .build_release/lib/libcaffe.so:undefinedreferenceto`google::protobuf::internal::WireFormatLite::WriteBytesMaybeAliased(int,google::protobuf::io::CodedOutputStream*)'
20. /usr/lib/x86_64-linux-gnu/libopencv_highgui.so:undefinedreferenceto`TIFFSetWarningHandler@LIBTIFF_4.0'
21. /usr/lib/x86_64-linux-gnu/libopencv_highgui.so:undefinedreferenceto`TIFFSetErrorHandler@LIBTIFF_4.0'
22. .build_release/lib/libcaffe.so:undefinedreferenceto`leveldb::Status::ToString()const'
23. .build_release/lib/libcaffe.so:undefinedreferenceto`google::protobuf::internal::WireFormatLite::WriteString(int,google::protobuf::io::CodedOutputStream*)'
24. collect2:error:ldreturned1exitstatus
25. Makefile:567:recipefortarget'.build_release/examples/cifar10/convert_cifar_data.bin'Failed
26. make:***[.build_release/examples/cifar10/convert_cifar_data.bin]Error1

reference： http://www.jb51.cc/article/p-vyhrkbir-yc.html

其他相关的参考网址：

1、解决/usr/lib/libstdc++.so.6: version `GLIBCXX_3.4.14' not found问题：http://www.jb51.cc/article/p-dncvgnde-ey.html

2、Caffe搭建：常见问题解决办法和ubuntu使用中遇到问题（持续更新）：http://www.cnblogs.com/empty16/p/4828476.html

3、安装python caffe过程中遇到的一些问题以及对应的解决方案：http://www.cnblogs.com/TiBAi/p/6848307.html

4、安装caffe总结：http://blog.csdn.net/csoldiers/article/details/51685179

5、caffe+cuda7.0+opencv3.0.0+mkl ubuntu14.04配置：http://www.th7.cn/system/lin/201504/103304.shtml

6、Caffe 环境搭建中应注意的问题：http://www.jb51.cc/article/p-evvodmjx-gt.html

7、caffe cpu版 Anaconda3 python 接口安装：http://www.jb51.cc/article/p-xoksgxry-bcx.html