Caffe is a deep learning framework developed at BLVC. While the installation on Linux systems are straightforward with its guides, the dependencies make it hard to compile the codes without root accesses. This guide is about how I installed it into a CentOS 6.5 server while several library shipped are obsolete. Some of the tips are from here.

Suppose we have a project folder that you can access as an user to the server. /home/you. We create a folder /home/you/usr to hold all the dependencies. We first setup the environment variables.

setenv PATH /home/you/usr/bin/:$PATH
setenv LD_LIBRARY_PATH /home/you/usr/lib/:$LD_LIBRARY_PATH

cmake

Usually it is shipped with the OS. But it may have a lower version than that some of the dependencies required. In case of that, download it from here. there is a compiled binary for Linux platform. Copy the files in the bin folder to /home/you/usr/bin.

CUDA

Luckily my CUDA runtime is 6.5 so I dont have to worry about that. CUDA 5 is fine, according to the official page, but remember you have to turn-off the gencode 2* flags from the caffe Makefile.config.

BLAS

I use ATLAS, which is the default library for caffe. Please refer to this article for the compilation of ATLAS. If you dont have the compiled python availabe, follow the instructions as well.

Boost

It seems quite a standard C++ library, but mine is 1.47 which is lower than the required version 1.55. Fine we can compile it! Download it from here and install it into your usr folder.

curl -L -O http://sourceforge.net/projects/boost/files/boost/1.57.0/boost_1_57_0.tar.gz
tar zxvf boost_1_57_0.tar.gz                                     
cd boost_1_57_0/                                                 
./bootstrap.sh --libdir=/home/you/usr/lib --includedir=/home/you/usr/include                                                                
vi project-config.jam # edit python path if you have a compiled one
./b2                                                             
./b2 install

OpenCV

If you dont have one, download it here and install.

curl -L -O http://sourceforge.net/projects/opencvlibrary/files/opencv-unix/2.4.10/opencv-2.4.10.zip 
unzip opencv-2.4.10.zip                                          
cd opencv-2.4.10
mkdir build                                                      
cd build                                                         
cmake -D CMAKE_BUILD_TYPE=RELEASE -D CMAKE_INSTALL_PREFIX=/home/you/usr .. 
ccmake .. # optional: turn off some CUDA gencode for faster compilation, and type c
make                                                             
make install

This will take very long time. Deleting some of the gencode of older architecures can improve it a little bit

protobuf

Easy https://github.com/google/protobuf

git clone https://github.com/google/protobuf.git                                                                                                              
cd protobuf/                                                     
./autogen.sh
./configure --prefix=/home/you/usr
make 
make install

glog

Easy https://github.com/google/glog

git clone https://github.com/google/glog.git
cd glog
./configure --prefix=/home/you/usr
make 
make install

gflags

git clone https://github.com/gflags/gflags.git
cd gflags
mkdir build
cd build
ccmake .. # edit the prefix to /home/you/usr, and type c
vi CMakeCache.txt # edit CMAKE_CXX_FLAGS:STRING=-fPIC
make 
make install

This -fPIC step is important otherwise it come back to you when compiling caffe

hdf5

I have that with my system, lucky. the offical site is here

leveldb

site https://github.com/google/leveldb

git clone https://github.com/google/leveldb.git                  
cd leveldb/                                                      
make
cp --preserve=links libleveldb.* /home/you/usr/lib
cp -r include/leveldb /home/you/usr/include/

It does not have make install, copy them manually

snappy

easy https://github.com/google/snappy

git clone https://github.com/google/snappy.git                                                                                                                
cd snappy/                                                       
./autogen.sh                                                     
./configure --prefix=/home/qchen/github/usr                      
make ; make install

lmdb

site here https://gitorious.org/mdb/mdb

git clone https://gitorious.org/mdb/mdb.git
cd mdb/libraries/liblmdb
vi Makefile # change prefix to /home/you/usr
make ; make install

Finally, Caffe itself

Download the source codes

git clone https://github.com/BVLC/caffe.git

Go to the caffe folder, cp Makefile.config.example Makefile.config and edit it. Most of the instructions can be found on the offical installation page. Here is a few things that have to be specify

Setup the path to your ATLAS compilation

BLAS := atlas                           
BLAS_INCLUDE := /path/to/your/atlas/include                                  
BLAS_LIB := /path/to/your/atlas/lib

Setup path to your python installation

PYTHON_INCLUDE := /path/to/your/python/include/python2.7 \
                  /path/to/your/python/lib/python2.7/site-packages/numpy/core/include
PYTHON_LIB := /home/qchen/Python27/lib

Add your dependencies to the makefile

QC_CUSTOM_INCLUDE := /home/you/usr/include                              
QC_CUSTOM_LIB := /home/you/usr/lib                                      

INCLUDE_DIRS := $(PYTHON_INCLUDE) $(QC_CUSTOM_INCLUDE) /usr/local/include        
LIBRARY_DIRS := $(PYTHON_LIB) $(QC_CUSTOM_LIB) /usr/local/lib /usr/lib

Then finally

make all
make test
make runtest

Another long wait … Hopefully you see the success