Centos6.3 keras+tensorflow+xgboost+libsvm機器學習環境搭建

最近項目做機器學習的功能，同時用到了keras,xgboost,svm，項目用到的os很老是centos6.3，近期也沒有升級os的plan，我只有在此os摸索, 耗了好幾天的時間。

以下記錄了安裝過程和遇到的問題，適用於CentOS release 6.3 x86_64。

開始使用Keras + Theano(0.9.0)backend，在AWS EC2 m3.xlarge/m3.2xlarge centos6.3/7.1/7.2/7.3 instance測試，發現Theano有嚴重的內存泄漏，同樣的模型和訓練集無論是16G內存還是32內存都是1個epoch沒有訓練完就吃滿內存爆掉。theano安裝簡單，依賴少，但是有嚴重內存泄漏，不得不使用tensorflow backend，這個沒有出現內存泄漏。

(1) Upgrade g++/gcc to 4.8.2
        wget http://people.centos.org/tru/devtools-2/devtools-2.repo -O /etc/yum.repos.d/devtools-2.repo
        yum install devtoolset-2-gcc devtoolset-2-gcc-c++ devtoolset-2-binutils
        /opt/rh/devtoolset-2/root/usr/bin/g++ --version
        ln -s /opt/rh/devtoolset-2/root/usr/bin/* /usr/local/bin/
        g++ --version
        
        說明：
        xgboost需要c++11特性，不得不升級。

(2) Install python 2.7.5
        yum install zlib-devel bzip2-devel openssl-devel ncurses-devel sqlite-devel
        wget https://www.python.org/ftp/python/2.7.5/Python-2.7.5.tgz
        tar xf Python-2.7.5.tgz
        cd Python-2.7.5
        ./configure --prefix=/usr/local/ -enable-shared --enable-unicode=ucs4
        make && make altinstall
        ls -ltr /usr/bin/python*
        ls -ltr /usr/local/bin/python*

        vim /etc/ld.so.conf add /usr/local/lib
        /sbin/ldconfig
        /sbin/ldconfig -v

    說明：
    --enable-unicode=ucs4是tensorflow需要的編譯選項；
    make altinstall不影響老的python版本；
    也可以安裝最新版的python 2.7.13。

(3) upgrade glic libc.so.6 from libc-2.12.so to libc-2.17.so
    tensorflow需要libc-2.17.so。
    安裝前strings /lib64/libc-2.12.so |grep GLIBC
        GLIBC_2.2.5
        GLIBC_2.2.6
        GLIBC_2.3
        GLIBC_2.3.2
        GLIBC_2.3.3
        GLIBC_2.3.4
        GLIBC_2.4
        GLIBC_2.5
        GLIBC_2.6
        GLIBC_2.7
        GLIBC_2.8
        GLIBC_2.9
        GLIBC_2.10
        GLIBC_2.11
        GLIBC_2.12
        GLIBC_PRIVATE

    保存下面內容爲shell腳本glibc-2.17_centos6.sh並執行：
        ##################
        #! /bin/sh

        # update glibc to 2.17 for CentOS 6

        wget --no-proxy http://copr-be.cloud.fedoraproject.org/results/mosquito/myrepo-el6/epel-6-i386/glibc-2.17-55.fc20/glibc-2.17-55.el6.i686.rpm
        wget --no-proxy http://copr-be.cloud.fedoraproject.org/results/mosquito/myrepo-el6/epel-6-x86_64/glibc-2.17-55.fc20/glibc-2.17-55.el6.x86_64.rpm
        wget --no-proxy http://copr-be.cloud.fedoraproject.org/results/mosquito/myrepo-el6/epel-6-x86_64/glibc-2.17-55.fc20/glibc-common-2.17-55.el6.x86_64.rpm
        wget --no-proxy http://copr-be.cloud.fedoraproject.org/results/mosquito/myrepo-el6/epel-6-x86_64/glibc-2.17-55.fc20/glibc-devel-2.17-55.el6.x86_64.rpm
        wget --no-proxy http://copr-be.cloud.fedoraproject.org/results/mosquito/myrepo-el6/epel-6-x86_64/glibc-2.17-55.fc20/glibc-headers-2.17-55.el6.x86_64.rpm

        sudo rpm -Uvh glibc-2.17-55.el6.x86_64.rpm \
        glibc-2.17-55.el6.i686.rpm \
        glibc-common-2.17-55.el6.x86_64.rpm \
        glibc-devel-2.17-55.el6.x86_64.rpm \
        glibc-headers-2.17-55.el6.x86_64.rpm
        ##################


    安裝後strings /lib64/libc.so.6 |grep GLIBC_
            GLIBC_2.2.5
            GLIBC_2.2.6
            GLIBC_2.3
            GLIBC_2.3.2
            GLIBC_2.3.3
            GLIBC_2.3.4
            GLIBC_2.4
            GLIBC_2.5
            GLIBC_2.6
            GLIBC_2.7
            GLIBC_2.8
            GLIBC_2.9
            GLIBC_2.10
            GLIBC_2.11
            GLIBC_2.12
            GLIBC_2.13
            GLIBC_2.14
            GLIBC_2.15
            GLIBC_2.16
            GLIBC_2.17
            GLIBC_PRIVATE


(4) upgrade libstdc++.so.6 from libstdc++.so.6.0.13 to libstdc++.so.6.0.19
    tensorflow需要libstdc++.so.6.0.19，可以從centos7裏面獲取或者google。
    
    安裝前strings /usr/lib64/libstdc++.so.6 | grep GLIBCXX
        GLIBCXX_3.4
        GLIBCXX_3.4.1
        GLIBCXX_3.4.2
        GLIBCXX_3.4.3
        GLIBCXX_3.4.4
        GLIBCXX_3.4.5
        GLIBCXX_3.4.6
        GLIBCXX_3.4.7
        GLIBCXX_3.4.8
        GLIBCXX_3.4.9
        GLIBCXX_3.4.10
        GLIBCXX_3.4.11
        GLIBCXX_3.4.12
        GLIBCXX_3.4.13
        GLIBCXX_FORCE_NEW
        GLIBCXX_DEBUG_MESSAGE_LENGTH

    rm -rf /usr/lib64/libstdc++.so.6  
    cp ./libstdc++.so.6.0.19 /usr/lib64/
    ln -s /usr/lib64/libstdc++.so.6.0.19 /usr/lib64/libstdc++.so.6  

    安裝後strings /usr/lib64/libstdc++.so.6 | grep GLIBC  
        GLIBCXX_3.4
        GLIBCXX_3.4.1
        GLIBCXX_3.4.2
        GLIBCXX_3.4.3
        GLIBCXX_3.4.4
        GLIBCXX_3.4.5
        GLIBCXX_3.4.6
        GLIBCXX_3.4.7
        GLIBCXX_3.4.8
        GLIBCXX_3.4.9
        GLIBCXX_3.4.10
        GLIBCXX_3.4.11
        GLIBCXX_3.4.12
        GLIBCXX_3.4.13
        GLIBCXX_3.4.14
        GLIBCXX_3.4.15
        GLIBCXX_3.4.16
        GLIBCXX_3.4.17
        GLIBCXX_3.4.18
        GLIBCXX_3.4.19
        GLIBC_2.3
        GLIBC_2.2.5
        GLIBC_2.14
        GLIBC_2.4
        GLIBC_2.3.2
        GLIBCXX_DEBUG_MESSAGE_LENGTH


(5) Install pip
    wget --no-check-certificate https://bootstrap.pypa.io/ez_setup.py
    /usr/local/bin/python2.7 ez_setup.py
    /usr/local/bin/easy_install-2.7 pip

(6) Install keras, tensorflow, xgboost
    /usr/local/bin/pip2.7 install keras==2.0.1
    /usr/local/bin/pip2.7 install xgboost
    /usr/local/bin/pip2.7 install https://storage.googleapis.com/tensorflow/linux/cpu/tensorflow-1.0.1-cp27-none-linux_x86_64.whl
    
    說明：
    tensorflow要使用1.0.1版的，0.8/0.9/1.1我測試的都不行。

(7) Install libsvm 3.22
    wget http://www.csie.ntu.edu.tw/~cjlin/cgi-bin/libsvm.cgi?+http://www.csie.ntu.edu.tw/~cjlin/libsvm+tar.gz
    tar xf libsvm-3.22.tar.gz   
    cd libsvm-3.22
    make
    cd python
    make
    vim svm.py edit libsvm = CDLL(path.join(dirname, '/usr/local/lib/libsvm.so.2')) (需要修改)
    cp *.py /usr/local/lib/python2.7/site-packages/
    cp ../libsvm.so.2 /usr/local/lib/
    cd /usr/local/lib/
    ln -s libsvm.so.2 libsvm.so

(8) Run python2.7 to test import

    pip list顯示：
        funcsigs (1.0.2)
        Keras (2.0.1)
        mock (2.0.0)
        numpy (1.12.1)
        pbr (3.0.0)
        pip (9.0.1)
        protobuf (3.3.0)
        PyYAML (3.12)
        scikit-learn (0.18.1)
        scipy (0.19.0)
        setuptools (33.1.1)
        six (1.10.0)
        tensorflow (1.0.1)
        Theano (0.9.0)
        wheel (0.29.0)
        xgboost (0.6a2)

    python2.7測試
        Python 2.7.5 (default, May 11 2017, 09:45:52)
        [GCC 4.8.2 20140120 (Red Hat 4.8.2-15)] on linux2
        Type "help", "copyright", "credits" or "license" for more information.
        >>> import xgboost
        import svm
        import svmutil
        import keras/usr/local/lib/python2.7/site-packages/sklearn/cross_validation.py:44: DeprecationWarning: This module was deprecated in version 0.18 in favor of the model_selection module into which all the refactored classes and functions are moved. Also note that the interface of the new CV iterators are different from that of this module. This module will be removed in 0.20.
          "This module will be removed in 0.20.", DeprecationWarning)
        >>> >>> >>>
        Using TensorFlow backend.
        
    我的算法也跑起來了。