CentOS 6 编译 TensorFlow for Java 以及 Maven Pom

• 我们的系统环境
  CentOS 6.5, JDK 1.8

• 更新yum源
  $ yum update

• 安装 Python 2.7
  $ yum install python27 python27-numpy python27-python-devel python27-python-wheel

• 升级 gcc 至 4.8.2
  $ cd /opt/
  $ wget http://people.centos.org/tru/devtools-2/devtools-2.repo -O /etc/yum.repos.d/devtools-2.repo
  $ yum install --nogpg -y zip unzip patch libcurl-devel git devtoolset-2-gcc devtoolset-2-binutils devtoolset-2-gcc-c++
  $ scl enable devtoolset-2 python27 bash
  $ gcc -v

• JDK版本
  1.8（如1.8.0_73）

• 安装bazel(编译不同版本的tensorflow，需要不同版本的bazel。这里编译tensorflow_v1.6.0，需要bazel_0.9.0)
  $ cd /opt
  $ wget https://github.com/bazelbuild/bazel/releases/download/0.9.0/bazel-0.9.0-dist.zip
  $ unzip bazel-0.9.0-dist.zip -d bazel-0.9.0-dist
  $ cd bazel-0.9.0-dist
  $ ./compile.sh
  $ mkdir -p ~/bin
  $ cp output/bazel ~/bin/

• 编译tensorflow
  $ cd /opt
  $ git clone https://github.com/tensorflow/tensorflow && cd tensorflow
  $ git checkout v1.6.0
  $ ./configure
  配置中提示的问题，全部选择n(no)
  在执行bazel build之前，
  把文件/opt/tensorflow/tensorflow/tensorflow.bzl中的代码片段

def tf_cc_shared_object(
    name,
    srcs=[],
    deps=[],
    linkopts=[],
    framework_so=tf_binary_additional_srcs(),
    **kwargs):

修改为

def tf_cc_shared_object(
    name,
    srcs=[],
    deps=[],
    linkopts=['-lrt'],
    framework_so=tf_binary_additional_srcs(),
    **kwargs):

把文件/opt/tensorflow/tensorflow/java/BUILD中的代码片段

tf_cc_binary(
    name = "generate_pom",
    srcs = ["generate_pom.cc"],
    deps = ["//tensorflow/c:c_api"],
)

修改为

tf_cc_binary(
    linkopts = ["-lrt"],
    name = "generate_pom",
    srcs = ["generate_pom.cc"],
    deps = ["//tensorflow/c:c_api"],
)

把

tf_cc_binary(
    name = "java_op_gen_tool",
    srcs = [
        "src/gen/cc/op_gen_main.cc",
    ],
    copts = tf_copts(),
    linkopts = select({
        "//tensorflow:windows": [],
        "//conditions:default": ["-lm"],
    }),
    linkstatic = 1,
    deps = [
        ":java_op_gen_lib",
        "//tensorflow/core:framework",
        "//tensorflow/core:framework_internal",
        "//tensorflow/core:lib",
        "//tensorflow/core:ops",
    ],
)

修改为

tf_cc_binary(
    name = "java_op_gen_tool",
    srcs = [
        "src/gen/cc/op_gen_main.cc",
    ],
    copts = tf_copts(),
    linkopts = select({
        "//tensorflow:windows": [],
        "//conditions:default": ["-lm","-lrt"],
    }),
    linkstatic = 1,
    deps = [
        ":java_op_gen_lib",
        "//tensorflow/core:framework",
        "//tensorflow/core:framework_internal",
        "//tensorflow/core:lib",
        "//tensorflow/core:ops",
    ],
)

如果还是报错`ndefined reference to 'clock_gettime'`, 还需要把文件/opt/tensorflow/tensorflow/tensorflow.bzl中出现`linkopts`的地方，全部添加`'-lrt'`。

$ bazel build --linkopt='-lrt' -c opt //tensorflow/java:tensorflow //tensorflow/java:libtensorflow_jni //tensorflow/java:pom

$ cd /opt/tensorflow/tensorflow/java/maven/

• 部署 libtensorflow

  修改 /opt/tensorflow/bazel-bin/tensorflow/java/pom.xml 中的groupId, version用于deploy到自己公司的远程仓库中，并指定仓库的repositoryId(比如我的是artifactory)和url(请查看你的配置)

  $ mvn deploy:deploy-file -Dfile=../../../bazel-bin/tensorflow/java/libtensorflow.jar -DpomFile=../../../bazel-bin/tensorflow/java/pom.xml -DrepositoryId= -Durl=

• 部署 libtensorflow_jni

  修改/opt/tensorflow/tensorflow/java/maven/pom.xml和/opt/tensorflow/tensorflow/java/maven/libtensorflow_jni/pom.xml中的groupId, version，并添加自己公司的distributionManagement配置，同时只需保留libtensorflow_jni模块。

  $ cd /opt/tensorflow/tensorflow/java/maven/
  $ mkdir -p libtensorflow_jni/src/main/resources/org/tensorflow/native/linux-x86_64
  $ cp ../../../bazel-bin/tensorflow/libtensorflow_framework.so libtensorflow_jni/src/main/resources/org/tensorflow/native/linux-x86_64/
  $ cp ../../../bazel-bin/tensorflow/java/libtensorflow_jni.so libtensorflow_jni/src/main/resources/org/tensorflow/native/linux-x86_64/
  $ mvn versions:set -DnewVersion="${TENSORFLOW_VERSION}-cpu-optimized"
  $ mvn package -Dgpg.skip=true
  $ mvn deploy -Dgpg.skip=true

• 参考资料

  • TensorFlow for Java
  • installing-tensorflow-from-sources
  • Install Tensorflow on Centos6.4
  • Building TensorFlow for CentOS 6
  • Building Tensorflow Java Maven dependency for upload in internal repo
  • install tensorflow on centos 6.5
  • 从源代码安装 TensorFlow

• 附录: 编译成python 源

scl enable devtoolset-2 bash
bazel clean
./configure
bazel build --linkopt='-lrt' -c opt --copt=-msse4.1 --copt=-msse4.2 --copt=-mavx --copt=-mavx2 --copt=-mfma //tensorflow/tools/pip_package:build_pip_package
bazel-bin/tensorflow/tools/pip_package/build_pip_package /tmp/tensorflow_pkg
pip install /tmp/tensorflow_pkg/tensorflow-1.6.0rc0-cp27-none-linux_x86_64.whl