从Python导出Tensorflow图以用于C++

Question

究竟应该如何导出python模型以便在c ++中使用？

我正在尝试做类似于本教程的内容:https: //www.tensorflow.org/versions/r0.8/tutorials/image_recognition/index.html

我试图在c ++ API中导入我自己的TF模型,而不是从一开始.我调整了输入大小和路径,但奇怪的错误不断出现.我花了一整天阅读堆栈溢出和其他论坛但无济于事.

我尝试了两种导出图形的方法.

方法1:元图.

...loading inputs, setting up the model, etc....

sess = tf.InteractiveSession()
sess.run(tf.initialize_all_variables())


for i in range(num_steps):  
  x_batch, y_batch = batch(50)  
  if i%10 == 0:
        train_accuracy = accuracy.eval(feed_dict={
        x:x_batch, y_: y_batch, keep_prob: 1.0})
        print("step %d, training accuracy %g"%(i, train_accuracy))
  train_step.run(feed_dict={x: x_batch, y_: y_batch, keep_prob: 0.5})

print("test accuracy %g"%accuracy.eval(feed_dict={
    x: features_test, y_: labels_test, keep_prob: 1.0}))

saver = tf.train.Saver(tf.all_variables())
checkpoint = 
   '/home/sander/tensorflow/tensorflow/examples/cat_face/data/model.ckpt'
    saver.save(sess, checkpoint)

   tf.train.export_meta_graph(filename=
   '/home/sander/tensorflow/tensorflow/examples/cat_face/data/cat_graph.pb',  
    meta_info_def=None,
    graph_def=sess.graph_def,
    saver_def=saver.restore(sess, checkpoint),
    collection_list=None, as_text=False)

尝试运行该程序时,方法1会产生以下错误:

[libprotobuf ERROR 
google/protobuf/src/google/protobuf/wire_format_lite.cc:532] String field 
'tensorflow.NodeDef.op' contains invalid UTF-8 data when parsing a protocol 
buffer. Use the 'bytes' type if you intend to send raw bytes. 
E tensorflow/examples/cat_face/main.cc:281] Not found: Failed to load 
compute graph at 'tensorflow/examples/cat_face/data/cat_graph.pb'

我还尝试了另一种导出图形的方法:

方法2:write_graph:

tf.train.write_graph(sess.graph_def, 
'/home/sander/tensorflow/tensorflow/examples/cat_face/data/', 
'cat_graph.pb', as_text=False)

这个版本实际上似乎加载了一些东西,但是我得到一个关于未被初始化的变量的错误:

Running model failed: Failed precondition: Attempting to use uninitialized  
value weight1
[[Node: weight1/read = Identity[T=DT_FLOAT, _class=["loc:@weight1"], 
_device="/job:localhost/replica:0/task:0/cpu:0"](weight1)]]

Answer 1

首先，您需要使用以下命令将定义图形化到文件中

with tf.Session() as sess:
//Build network here 
tf.train.write_graph(sess.graph.as_graph_def(), "C:\\output\\", "mymodel.pb")

然后，使用 saver 保存模型

saver = tf.train.Saver(tf.global_variables()) 
saver.save(sess, "C:\\output\\mymodel.ckpt")

然后，您的输出将有 2 个文件：mymodel.ckpt、mymodel.pb

从此处下载 freeze_graph.py并在 C:\output\ 中运行以下命令。如果输出节点名称与您不同，请更改它。

python freeze_graph.py --input_graph mymodel.pb --input_checkpoint mymodel.ckpt --output_node_names softmax/Reshape_1 --output_graph mymodelforc.pb

您可以直接从 C 使用 mymodelforc.pb。

您可以使用以下 C 代码来加载 proto 文件

#include "tensorflow/core/public/session.h"
#include "tensorflow/core/platform/env.h"
#include "tensorflow/cc/ops/image_ops.h"

Session* session;
NewSession(SessionOptions(), &session);

GraphDef graph_def;
ReadBinaryProto(Env::Default(), "C:\\output\\mymodelforc.pb", &graph_def);

session->Create(graph_def);

现在您可以使用会话进行推理。

您可以按如下方式应用推理参数：

// Same dimension and type as input of your network
tensorflow::Tensor input_tensor(tensorflow::DT_FLOAT, tensorflow::TensorShape({ 1, height, width, channel }));
std::vector<tensorflow::Tensor> finalOutput;

// Fill input tensor with your input data

std::string InputName = "input"; // Your input placeholder's name
std::string OutputName = "softmax/Reshape_1"; // Your output placeholder's name

session->Run({ { InputName, input_tensor } }, { OutputName }, {}, &finalOutput);

// finalOutput will contain the inference output that you search for