linux环境下在C++程序中使用Tensoflow Debugger

花费数天时间翻阅tensorflow源码，终于踩平这个坑！

---

TF-Debugger for python

使用TF-Debugger时，要设置run_option中的debug_tensor_watch_opts，在session run的时候，传入run_option，一个简单的例子如下：

1. 使用python建一个简单模型

建一个连个节点相加的简单graph，保存为pb文件

import tensorflow as tf
import numpy as np
from tensorflow.python.framework import graph_util
with tf.Session() as sess:
    a = tf.placeholder(tf.float32,[1],name='a')
    b = tf.placeholder(tf.float32,[1],name='b')
    print b
    c = tf.add(a,b,name='c')
    print c.name
    sess.run(tf.global_variables_initializer())
    print sess.run(c,feed_dict={a:[5.0],b:[4.0]})
    constant_graph = graph_util.convert_variables_to_constants(sess, sess.graph_def, ["c"])
    tf.train.write_graph(constant_graph, 'models/', 'graph.pb', as_text=False)

2. 读取模型

读取刚刚保存为pb文件的模型，并在session run的时候输出debug信息

import tensorflow as tf
from tensorflow.python.framework import graph_util
import numpy as np
from tensorflow.python import debug as tf_debug
#debug watch setting
def addWatch(run_options,node_name,debug_urls):
    watch_opts = run_options.debug_options.debug_tensor_watch_opts
    run_options.debug_options.global_step = 1
    watch = watch_opts.add()
    watch.node_name=node_name
    watch.output_slot=0
    debug_ops="DebugIdentity"
    if isinstance(debug_ops, str):
        debug_ops = [debug_ops]
    watch.debug_ops.extend(debug_ops)
    if debug_urls:
        if isinstance(debug_urls, str):
            debug_urls = [debug_urls]
        watch.debug_urls.extend(debug_urls)
#main func
modelpath='./'
dbgInfoPath="file:///your_path/dbginfo"
output_graph_path = modelpath+'models/graph.pb'
with tf.Session() as sess:
    tf.global_variables_initializer().run()
    output_graph_def = tf.GraphDef()
    with tf.gfile.GFile(output_graph_path, 'rb') as fid:
        serialized_graph = fid.read()
        output_graph_def.ParseFromString(serialized_graph)
        tf.import_graph_def(output_graph_def, name='')
    a = sess.graph.get_tensor_by_name("a:0")
    print a
    b = sess.graph.get_tensor_by_name("b:0")
    print b
    c = sess.graph.get_tensor_by_name("c:0")
    print c
    run_options = tf.RunOptions()
    ops=sess.graph.get_operations()
    for op in ops:
        if not op.outputs:
            continue
        node_name=op.name
        addWatch(run_options,node_name,dbgInfoPath)
    print run_options.debug_options.debug_tensor_watch_opts
    sess.run(tf.global_variables_initializer())
    print sess.run(c,feed_dict={a:[3.0],b:[4.0]},options=run_options)

3. 查看输出的tfdbg信息

在命令行下使用命令

python -m tensorflow.python.debug.cli.offline_analyzer --dump_dir=/your_path/dbginfo/

就可进入tfdbg界面

---

TF-Debugger for c++

与python版类似，使用TF-Debugger时，要设置run_option中的debug_tensor_watch_opts，在session run的时候，传入run_option，但是如果按照正常方法，会提示tfdbg not linked in this tensorflow build，因此需要在编译时设置一些东西。

1. 编译tensorflow源码

使用tensorflow for c++之前，需要编译目标libtensorflow_cc.so，即使用命令：

bazel build --config=opt //tensorflow:libtensorflow_cc.so

然而仔细查看tensorflow下的BUILD文件，发现目标libtensorflow_cc.so中的依赖项并没有包含tensoflow/core/debug中的源文件。
因此需要在tensorflow/tensorflow/BUILD文件中加入一行

tf_cc_shared_object(
    name = "libtensorflow_cc.so",
    linkopts = select({
        "//tensorflow:darwin": [
            "-Wl,-exported_symbols_list",  # This line must be directly followed by the exported_symbols.lds file
            "//tensorflow:tf_exported_symbols.lds",
        ],
        "//tensorflow:windows": [],
        "//tensorflow:windows_msvc": [],
        "//conditions:default": [
            "-z defs",
            "-s",
            "-Wl,--version-script",  #  This line must be directly followed by the version_script.lds file
            "//tensorflow:tf_version_script.lds",
        ],
    }),
    deps = [
        "//tensorflow:tf_exported_symbols.lds",
        "//tensorflow:tf_version_script.lds",
        "//tensorflow/c:c_api",
        "//tensorflow/c/eager:c_api",
        "//tensorflow/cc:cc_ops",
        "//tensorflow/cc:client_session",
        "//tensorflow/cc:scope",
        "//tensorflow/core:tensorflow",
        "//tensorflow/core/debug",#这行我是加上去的
    ],
)

改完后再使用命令

bazel build --config=opt //tensorflow:libtensorflow_cc.so

编译成功，此时tfdbg功能已经被link到libtensorflow_cc.so中了

2. 测试程序

main.cpp:

#include 
#include "tensorflow/core/public/session.h"
#include "tensorflow/core/platform/env.h"
#include "tensorflow/core/graph/default_device.h"
#include "tensorflow/core/util/device_name_utils.h"
#include "tensorflow/core/protobuf/config.pb.h"
#include "tensorflow/core/protobuf/debug.pb.h"
#include "tensorflow/core/framework/graph.pb.h"
using namespace tensorflow;
std::string dbgInfoPath="file:///your_path/dbginfo";
int main(int argc, char* argv[]) {
  // Initialize a tensorflow session
  Session* session;
  tensorflow::SessionOptions opts;
  //read model
  GraphDef graph_def;
  Status status = ReadBinaryProto(Env::Default(), "models/graph.pb", &graph_def);
  if (!status.ok()) {
    std::cout << status.ToString() << "\n";
    return 1;
  }
  //log node device
  opts.config.set_log_device_placement(true);
  //setup session
  status = NewSession(opts, &session);
  if (!status.ok()) {
    printf("NewSession\n");
    std::cout << status.ToString() << "\n";
    return 1;
  }
  //create graph
  status = session->Create(graph_def);
  if (!status.ok()) {
    std::cout << status.ToString() << "\n";
    return 1;
  }
  //input and output tensor
  Tensor a(DT_FLOAT, TensorShape());
  a.scalar<float>()() = 3.0;
  Tensor b(DT_FLOAT, TensorShape());
  b.scalar<float>()() = 2.0;
  std::vector<std::pair<string, tensorflow::Tensor>> inputs = {
    { "a", a },
    { "b", b },
  };
  std::vector outputs;
  //run_option setting
  tensorflow::RunOptions ops=tensorflow::RunOptions();
  tensorflow::RunMetadata run_metadata;
  //debug_option setting
  tensorflow::DebugOptions* dp=ops.mutable_debug_options();
  int node_count = graph_def.node_size();
  for(int i=0;iauto node = graph_def.mutable_node(i);
    DebugTensorWatch* watch = dp->add_debug_tensor_watch_opts();
    watch->set_node_name(node->name());
    std::cout<name()<<std::endl;
    watch->set_output_slot(0);
    watch->add_debug_ops("DebugIdentity");
    watch->add_debug_urls(dbgInfoPath.c_str());
  }
  //delete exist dbginfo
  system("rm -rf ./dbginfo/");
  //test if has debug options
  if(ops.has_debug_options())
    printf("ops has debug_options，debug_tensor_watch_opts_size:%d\n",dp->debug_tensor_watch_opts_size());
  else
    printf("ops has no debug_options()\n");
  //session run
  status = session->Run(ops,inputs, {"c"}, {}, &outputs,&run_metadata);
  if (!status.ok()) {
    printf("Session Run Failed\n");
    std::cout << status.ToString() << "\n";
    return 1;
  }
  //print output
  auto output_c = outputs[0].scalar<float>();
  std::cout << outputs[0].DebugString() << "\n"; // Tensor
  std::cout << output_c() << "\n"; // 30
  //close session
  session->Close();
  return 0;
}

Makefile:

CXX = g++
INCPATHS = -I./
INCPATHS += -I/usr/local/lib/python2.7/dist-packages/tensorflow/include/
INCPATHS += -I/usr/local/lib/python2.7/dist-packages/tensorflow/include/external/nsync/public/
LIBPATH = -L./
LIBPATH += -L/usr/local/lib
LIBS = -Wl,-rpath=./
LIBS += -ltensorflow_cc
LIBS += -ltensorflow_framework
EXE = main
$(EXE):
        $(CXX) -std=c++11 -o  $(EXE) main.cpp $(INCPATHS) $(LIBPATH) $(LIBS)

将程序中以及Makefile中的路径换成自己的路径，然后make，执行程序。
使用命令，查看dbginfo

python -m tensorflow.python.debug.cli.offline_analyzer --dump_dir=/your_path/dbginfo/