【tensorflow-C++之一】Beginner开始工程化

0 模型训练之后

当模型在几百万次迭代之后终于收敛，在验证集表现的也不错之后，是时候可以把它做成真正的服务了。c++登场……..

以下代码可以保存一个最简单的图模型：y = a * x + b(标量)

import tensorflow as tf

def save():
    x = tf.placeholder(dtype=tf.float32, shape=[], name='x')
    with tf.variable_scope('beginner'):
        a = tf.Variable(initial_value=.1, name='a', dtype=tf.float32)
        b = tf.Variable(initial_value=0, name='b', dtype=tf.float32)
    y = tf.add(tf.multiply(x, a,), b, name='y')
    variables = tf.trainable_variables()
    for v in variables:
        print(v.name)
    init = tf.global_variables_initializer()
    with tf.Session() as sess:
        sess.run(init)
        output_graph = tf.graph_util.convert_variables_to_constants(sess=sess, input_graph_def=sess.graph_def, output_node_names=['y'])
        tf.train.write_graph(output_graph, ".", "beginner.pb", False)


if __name__ == "__main__":
    save()

1 工程化

在工程化之前，长达很长时间都是使用python迭代模型，从未考虑过模型训练好之后如何提供服务。通过c++使用tensorflow需要tf的头文件和相应的库文件。但是通过bazel编译之后产生了很多头文件，并且原有的头文件并没有整理在一起，所以整理头文件是一个很费时费力的活。为什么tf不能像caffe一样把头文件都集中在include里边？

1.1 找不到头文件

众里寻他千百度，蓦然回首c++的头文件全部在目录python

site-packages/tensorflow/include

原来在pip install 之后自动有一份c++的头文件，开始征程吧！！

1.2 小试牛刀

/*
 * inference4beginer.cpp
 * Copyright (C) 2017 fisherman
 */

#include 
#include 
#include 
#include 

#include 

#include 
#include 
#include 
int main() {
  tensorflow::GraphDef graph;
  std::string graph_file = "beginner.pb";
  //1 读取Graph, 如果是文本形式的pb,使用ReadTextProto
  tensorflow::Status status = tensorflow::ReadBinaryProto(tensorflow::Env::Default(), graph_file, &graph);
  if (!status.ok()) {
    LOG(FATAL) << status.ToString() << " with graph_file:" << graph_file;
    return -1;
  }
  //2 创建Session
  std::unique_ptr sess;
  sess.reset(tensorflow::NewSession({}));
  status = sess->Create(graph);
  if (!status.ok()) {
    LOG(FATAL) << status.ToString();
    return -1;
  }
  //3 构造Tensor并赋值
  //3.1 创建tensorflow::Tensor x
  tensorflow::Tensor x(tensorflow::DT_FLOAT, tensorflow::TensorShape({}));
  //3.2 获取x的Eigen::TensorMap
  auto x_map = x.tensor<float, 0>(); // == x.scalar()
  //auto ->Eigen::TensorMap, Eigen::Aligned>
  //3.3 获取指针， 可以任意复制了。
  float* data = x_map.data();
  *data = 1.0f;
  std::vector<std::pair<std::string, tensorflow::Tensor>> inputs;
  inputs.emplace_back(std::make_pair("x", x));
  std::vector<std::string> output_tensor_names;
  output_tensor_names.emplace_back("y");
  std::vector y;
  // 4 Session Run
  status = sess->Run(inputs, output_tensor_names, {}, &y);
  if (!status.ok()) {
    LOG(ERROR) << status.ToString();
    return -1;
  }
  // 5 打印结果
  float y_value = *(y[0].tensor<float, 0>().data());
  LOG(INFO) << "y = a*x+b\n" << " a(0.1) b(0) in graph, x(1), y(0.1):" << y_value;
  return 0;
}

1.3 结果

我使用scons进行编译并运行

I tensorflow/stream_executor/dso_loader.cc:135] successfully opened CUDA library libcublas.so.8.0 locally
I tensorflow/stream_executor/dso_loader.cc:135] successfully opened CUDA library libcudnn.so.5 locally
I tensorflow/stream_executor/dso_loader.cc:135] successfully opened CUDA library libcufft.so.8.0 locally
I tensorflow/stream_executor/dso_loader.cc:135] successfully opened CUDA library libcuda.so.1 locally
I tensorflow/stream_executor/dso_loader.cc:135] successfully opened CUDA library libcurand.so.8.0 locally
I tensorflow/core/common_runtime/gpu/gpu_device.cc:885] Found device 0 with properties: 
name: Tesla M40
major: 5 minor: 2 memoryClockRate (GHz) 1.112
pciBusID 0000:02:00.0
Total memory: 11.17GiB
Free memory: 346.75MiB
I tensorflow/core/common_runtime/gpu/gpu_device.cc:906] DMA: 0 
I tensorflow/core/common_runtime/gpu/gpu_device.cc:916] 0:   Y 
I tensorflow/core/common_runtime/gpu/gpu_device.cc:975] Creating TensorFlow device (/gpu:0) -> (device: 0, name: Tesla M40, pci bus id: 0000:02:00.0)
WARNING: Logging before InitGoogleLogging() is written to STDERR
I0615 00:39:15.321010 19296 inference4beginer.cpp:57] y = a*x+b
 a(0.1) b(0) in graph, x(1), y(0.1):0.1

1 如果忘记复制graph，提示信息很明白不多说。

F0615 00:42:42.124217 19389 inference4beginer.cpp:24] Not found: beginner.pb with graph_file:beginner.pb
*** Check failure stack trace: ***
./run.sh: line 3: 19389 Aborted                 (core dumped) CUDA_VISIBLE_DEVICES=0 ./inference

2 错把scalar（标量）写成vector（向量）

  tensorflow::Tensor x(tensorflow::DT_FLOAT, tensorflow::TensorShape({1}));

则会出现这样的错误

F tensorflow/core/framework/tensor_shape.cc:36] Check failed: NDIMS == dims() (0 vs. 1)Asking for tensor of 0 dimensions from a tensor of 1 dimensions
./run.sh: line 3: 19490 Aborted                 (core dumped) CUDA_VISIBLE_DEVICES=0 ./inference