Question

我想创建一个可以与之通信的新操作外部python进程。在momemnt，我创建了一个新的操作发送到python进程＆＃34; hello world＆＃34;与protobuf。

在这个小例子中，我发送了一个字符串。在将来我想发送更复杂的数据，如Eigen矩阵，这就是我选择的原因 protobuf。（并且可以轻松地集成到tensorflow中）。

msg.proto ：

package prototest;

message Foo {
  required string bar = 1;
}

protoc msg.proto --cpp_out=. --python_out=.
生成：msg.pb.cc msg.pb.h msg_pb2.py

hello_world.cc ：

#include "tensorflow/core/framework/op_kernel.h"
#include "tensorflow/core/framework/tensor_shape.h"
#include "tensorflow/core/platform/default/logging.h"
#include "tensorflow/core/framework/shape_inference.h"

// to send serialized data through UPD socket
#include <sys/socket.h>
#include <arpa/inet.h>

// generated header file from protoc
#include "msg.pb.h"

namespace tensorflow{
    namespace shape_inference{

        Status HelloWorldShape(InferenceContext* c){
            std::cout << "shape_infernce is done" << std::endl;
            return Status::OK();
        }
        REGISTER_OP("HelloWorld")
            .SetShapeFn(HelloWorldShape)
            .Doc(R"doc(HelloWorld operation)doc");
    } // end namespace shape_inference

    class HelloWorldOp : public OpKernel {
    public :
        // constructor
        explicit HelloWorldOp(OpKernelConstruction* context) : OpKernel(context) {
            std::cout << "HelloWorldOp constructor" << std::endl;
        }

        void Compute(OpKernelContext* context) override {
            std::cout << "Start Compute method" << std::endl;
            //-----------------------------------------------------------------
            // send something to a Python process with protobuf
            struct sockaddr_in addr;
            addr.sin_family = AF_INET;
            inet_aton("127.0.0.1", &addr.sin_addr);
            addr.sin_port = htons(5555);

            // initialise a foo and set some properties
            GOOGLE_PROTOBUF_VERIFY_VERSION;

            prototest::Foo foo;
            foo.set_bar("Hello World");

            // serialise to string, this one is obvious ; )
            std::string buf;
            foo.SerializeToString(&buf);

            int sock = socket(PF_INET, SOCK_DGRAM, 0);
            sendto(sock, buf.data(), buf.size(), 0, (struct sockaddr *)&addr, sizeof(addr));
            //------------------------------------------------------------------
            std::cout << "Compute method is done" << std::endl;
        }
    };
    REGISTER_KERNEL_BUILDER(Name("HelloWorld").Device(DEVICE_CPU), HelloWorldOp);
} // end namespace tensorflow

要编译和运行我的代码，我使用#10950

中的测试脚本

compiler_and_run.py ：

#!/usr/bin/env python3.5

# Demo from https://github.com/tensorflow/tensorflow/issues/10950

from __future__ import print_function
import os
import sys
import tensorflow as tf


my_dir = os.path.dirname(os.path.abspath(__file__))
so_filename = "lib_hello_world.so"
cc_filename = "hello_world.cc"


def compile():
    # Fix for undefined symbol: _ZN6google8protobuf8internal26fixed_address_empty_stringE.
    # https://github.com/tensorflow/tensorflow/issues/1419
    from google.protobuf.pyext import _message as msg
    lib = msg.__file__
    ld_flags = [
        "-Xlinker", "-rpath", "-Xlinker", os.path.dirname(lib),
        "-L", os.path.dirname(lib), "-l", ":" + os.path.basename(lib)]
    common_opts = ["-shared", "-O2", "-std=c++11"]
    if sys.platform == "darwin":
        common_opts += ["-undefined", "dynamic_lookup"]
    common_opts += ["-I", tf.sysconfig.get_include()]
    common_opts += ["-fPIC"]
    common_opts += ["-D_GLIBCXX_USE_CXX11_ABI=0"]  # might be obsolete in the future
    opts = common_opts + [cc_filename, "-o", so_filename]
    opts += ld_flags
    cmd_bin = "g++"
    cmd_args = [cmd_bin] + opts
    from subprocess import Popen, PIPE, STDOUT, CalledProcessError
    print("compile call: %s" % " ".join(cmd_args))
    proc = Popen(cmd_args, stdout=PIPE, stderr=STDOUT)
    stdout, stderr = proc.communicate()
    assert stderr is None  # should only have stdout
    if proc.returncode != 0:
      print("compile failed: %s" % cmd_bin)
      print("Original stdout/stderr:")
      print(stdout)
      raise CalledProcessError(returncode=proc.returncode, cmd=cmd_args)
    assert os.path.exists(so_filename)


def main():
    print("TensorFlow version:", tf.GIT_VERSION, tf.VERSION)
    os.chdir(my_dir)
    compile()
    mod = tf.load_op_library("%s/%s" % (my_dir, so_filename))


if __name__ == "__main__":
    main()

返回：

TensorFlow version: v1.2.0-rc2-21-g12f033d 1.2.0
compile call: g++ -shared -O2 -std=c++11 -I /usr/local/lib/python3.5/dist-packages/tensorflow/include -fPIC -D_GLIBCXX_USE_CXX11_ABI=0 hello_world.cc -o lib_hello_world.so -Xlinker -rpath -Xlinker /usr/local/lib/python3.5/dist-packages/protobuf-3.2.0-py3.5-linux-x86_64.egg/google/protobuf/pyext -L /usr/local/lib/python3.5/dist-packages/protobuf-3.2.0-py3.5-linux-x86_64.egg/google/protobuf/pyext -l :_message.cpython-35m-x86_64-linux-gnu.so
Traceback (most recent call last):
  File "./compile_and_test.py", line 55, in <module>
    main()
  File "./compile_and_test.py", line 51, in main
    mod = tf.load_op_library("%s/%s" % (my_dir, so_filename))
  File "/usr/local/lib/python3.5/dist-packages/tensorflow/python/framework/load_library.py", line 64, in load_op_library
    None, None, error_msg, error_code)
tensorflow.python.framework.errors_impl.NotFoundError: /src/ext_hello_world/lib_hello_world.so: undefined symbol: _ZN6google8protobuf8internal26fixed_address_empty_stringE

编译似乎有效（没有致命错误）。但是tf.load_op_library()由于共享库中的未定义库（* .so）而失败。

这个未定义的符号似乎来自protobuf。

我自己安装了protoc（请参阅相关说明）。

在tf.sysconfig.get_include()中，有一个google/protobuf文件夹，其中包含来自内置张量流的头文件。

所以我不知道在编译期间使用了哪些头文件： - 来自tensorflow的包含文件的头文件？ - 手工安装的protobuf的头文件？

或者这个未定义的符号不是由于这个事实？

如何在共享库中解决此未定义的符号错误？
我是否必须从protoc开始安装tensorflow/includes/google/protobuf？（而不是从头开始）

我在#10950中观察到了：

from google.protobuf.pyext import _message as msg
lib = msg.__file__

返回：/u/zeyer/.local/lib/python2.7/site-packages/google/protobuf/pyext/_message.so

就我而言：/usr/local/lib/python3.5/dist-packages/protobuf-3.2.0-py3.5-linux-x86_64.egg/google/protobuf/pyext/_message.cpython-35m-x86_64-linux-gnu.so 这个关于protobuf的文件似乎完全不同......

关于protoc安装的注意事项：未安装protoc（protobuf编译器）。我确定了tensorflow中使用的protobuf版本：v3.2.0！

之后，我按照protobuf安装说明（C ++和Python实现）。

cd /opt/
# clone protobuf repo
git clone https://github.com/google/protobuf.git
cd protobuf

# change to the right branch
# list tags
git tag -l
git checkout tags/v3.2.0

# install protobuf
apt-get install autoconf automake libtool curl make g++ unzip
./autogen.sh
./configure
make
make check
make install
ldconfig

# protoc version
protoc --version
>>> libprotoc 3.2.0

# print linker and compiler files
pkg-config --cflags --libs protobuf
>>> -pthread -I/usr/local/include -L/usr/local/lib -lprotobuf -pthread -lpthread

#some useful env variables
PB_INC=$(pkg-config --cflags protobuf)
PB_LINK=$(pkg-config --libs protobuf)
TF_INC=$(python3.5 -c 'import tensorflow as tf; print(tf.sysconfig.get_include())')
TF_LIBS=$(find $TF_INC/../ -name "*.so")

# here, I used python3.5
cd python
export LD_LIBRARY_PATH=../src/.libs
python3.5 setup.py build --cpp_implementation
python3.5 setup.py test --cpp_implementation
python3.5 setup.py install --cpp_implementation

系统信息

docker：Docker版本1.12.6，版本78d1802
image：tensorflow / tensorflow：1.2.0-devel-gpu-py3
基于：ubuntu 16.04（4.4.0-78-generic）
tensorflow build from source
tensorflow版本：1.2.0
python版本：Python 3.5.2
bazel版本： Build label: 0.4.5 Build target: bazel-out/local-fastbuild/bin/src/main/java/com/google/devtools/build/lib/bazel/BazelServer_deploy.jar Build time: Thu Mar 16 12:19:38 2017 (1489666778) Build timestamp: 1489666778 Build timestamp as int: 1489666778
gcc -v： ``` 使用内置规格。 COLLECT_GCC = GCC COLLECT_LTO_WRAPPER = / usr / lib中/ GCC / x86_64的-Linux的GNU / 5 / LTO-包装目标：x86_64-linux-gnu 配置为：../ src / configure -v --with-pkgversion =＆＃39; Ubuntu 5.4.0-6ubuntu1~16.04.4＆＃39; --with-bugurl = file：///usr/share/doc/gcc-5/README.Bugs --enable-languages = c，ada，c ++，java，go，d，fortran，objc，obj-c ++ - -prefix = / usr --program-suffix = -5 --enable-shared --enable-linker-build-id --libexecdir = / usr / lib --without-included-gettext --enable-threads = posix - -libdir = / usr / lib --enable-nls --with-sysroot = / --enable -clocale = gnu --enable-libstdcxx-debug --enable-libstdcxx-time = yes --with-default-libstdcxx- abi = new --enable-gnu-unique-object --disable-vtable-verify --enable-libmpx --enable-plugin --with-system-zlib --disable-browser-plugin --enable-java-awt = gtk --enable-gtk-cairo --with-java-home = / usr / lib / jvm / java-1.5.0-gcj-5-amd64 / jre --enable-java-home --with-jvm- root-dir = / usr / lib / jvm / java-1.5.0-gcj-5-amd64 --with-jvm-jar-dir = / usr / lib / jvm-exports / java-1.5.0-gcj-5 -amd64 --with-arch-directory = amd64 --with-ecj-jar = / usr / share / java / eclipse-ecj.jar --enable-objc-gc --enable-multiarch --disable-werror - with-arch-32 = i686 --with-abi = m64 --with-multilib-list = m32，m64，mx32 --enable-multilib --with -tune = generic --enable-checking = release --build = x86_64-linux-gnu --host = x86_64-linux-gnu --target = x86_64-linux-gnu 线程模型：posix gcc版本5.4.0 20160609（Ubuntu 5.4.0-6ubuntu1~16.04.4）

来自https://github.com/tensorflow/tensorflow/tree/master/tools/tf_env_collect.sh的信息：

== cat /etc/issue ===============================================
Linux 2b98f5ebc987 4.4.0-78-generic #99-Ubuntu SMP Thu Apr 27 15:29:09 UTC 2017 x86_64 x86_64 x86_64 GNU/Linux
VERSION="16.04.2 LTS (Xenial Xerus)"
VERSION_ID="16.04"
VERSION_CODENAME=xenial

== are we in docker =============================================
Yes
== compiler =====================================================
c++ (Ubuntu 5.4.0-6ubuntu1~16.04.4) 5.4.0 20160609
Copyright (C) 2015 Free Software Foundation, Inc.
This is free software; see the source for copying conditions.  There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

== uname -a =====================================================
Linux 2b98f5ebc987 4.4.0-78-generic #99-Ubuntu SMP Thu Apr 27 15:29:09 UTC 2017 x86_64 x86_64 x86_64 GNU/Linux
== check pips ===================================================
numpy (1.13.0)
protobuf (3.3.0)
tensorflow (1.2.0)

== check for virtualenv =========================================
False
== tensorflow import ============================================
tf.VERSION = 1.2.0
tf.GIT_VERSION = v1.2.0-rc2-21-g12f033d
tf.COMPILER_VERSION = v1.2.0-rc2-21-g12f033d
Sanity check: array([1], dtype=int32)
== env ==========================================================
LD_LIBRARY_PATH /usr/local/cuda/extras/CUPTI/lib64:/usr/local/nvidia/lib:/usr/local/nvidia/lib64
DYLD_LIBRARY_PATH is unset
== nvidia-smi ===================================================
Wed Aug 23 17:24:13 2017
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 370.28                 Driver Version: 370.28                    |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  GeForce 940MX       Off  | 0000:01:00.0     Off |                  N/A |
| N/A   54C    P0    N/A /  N/A |    277MiB /  2002MiB |      0%      Default |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                       GPU Memory |
|  GPU       PID  Type  Process name                               Usage      |
|=============================================================================|
+-----------------------------------------------------------------------------+

== cuda libs  ===================================================
/usr/local/cuda-8.0/targets/x86_64-linux/lib/libcudart_static.a
/usr/local/cuda-8.0/targets/x86_64-linux/lib/libcudart.so.8.0.61

Answer 1

我认为TensorFlow不会导出足够的符号来依赖协议缓冲区实现。相反，您可以动态链接您已安装的系统协议缓冲区实现。这可能需要匹配TensorFlow协议缓冲区版本以避免符号冲突。

我正在使用libprotoc 3.3.0测试TensorFlow 1.3.0（Python 3.4）。我已将msg.pb.cc（protoc msg.proto --cpp_out=. --python_out=.的输出）添加到编译命令（链接在消息的C ++实现中），以及pkg-config --cflags --libs protobuf的输出以便链接系统协议缓冲库。这个程序对我有用：

protoc msg.proto --cpp_out=. --python_out=.
g++ -shared -O2 -std=c++11 -I /usr/local/lib/python3.4/dist-packages/tensorflow/include -fPIC -D_GLIBCXX_USE_CXX11_ABI=0 hello_world.cc msg.pb.cc -o lib_hello_world.so $(pkg-config --cflags --libs protobuf)
python3 -c "import tensorflow as tf; tf.Session().run(tf.load_op_library('./lib_hello_world.so').hello_world())"

由于您已经序列化了协议缓冲区，因此在Python中反序列化它应该没有问题;小的protobuf版本不需要匹配，因为它是有线格式。

Answer 2

它来自链接到 dll 库 libprotobuf[d].lib，没有指定 -DPROTOBUF_USE_DLLS。

有关更多信息，请参阅自述文件中的“DLL 与静态链接”。

未定义的符号＆＃39; fixed_address_empty_string＆＃39; ：带有protobuf

系统信息

2 个答案: