如何使用opencv dnn模型加载张量流的预训练模型

Question

readNetFromTensorflow 函数无法加载 tensorflow 预训练模型 (.pb)。

第一步：训练张量流的模型。 线性回归代码的Tensorflow模型如下：

#!/usr/bin/python
import tensorflow as tf
import numpy as np
x_ = tf.placeholder(np.float32, [None, 1], 'input')
y_ = tf.placeholder(np.float32, [None, 1], 'label')
#layer1
w1 = tf.Variable(tf.random_normal([1,3]))
b1 = tf.Variable(tf.random_normal([3]) )
a1 = tf.add(b1,tf.matmul(x_,w1)) 
#layer2
w2 = tf.Variable(tf.random_normal([3,1]))
b2 = tf.Variable(tf.random_normal([1]) )
a2 = tf.add(b2,tf.matmul(a1,w2),name="output") 
#global steps
steps = 5000 
x = []
y = []
for i in range(1,200,5):
    temp = (1.0 * i)/10
    x.append([temp])
    y.append([3. + 2. * temp])
x = np.array(x)
y = np.array(y)
#loss function
loss =  tf.reduce_mean(tf.reduce_sum(tf.square(a2-y_)))
#optimizer
optimizer = tf.train.GradientDescentOptimizer(0.00001).minimize(loss)
with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
    length = len(x)
    #training...
    for i in range(steps):
        sess.run(optimizer,feed_dict={x_:x,y_:y})
        result = sess.run(loss,feed_dict={x_:x,y_:y})
        if i % 50 == 0:
            print("loss: ",result,"\tstep: ",i)
    saver = tf.train.Saver()
    saver.save(sess,"./model/model.ckpt")
    tf.train.write_graph(sess.graph.as_graph_def(), "./model/", "graph.pbtxt")
    print("predict...")
    pre = sess.run(a2,feed_dict={x_:[[0]]})
    print("x = 2 pre: ",pre)

第二步：将模型保存为tensorflow的.pb文件。该怎么做？

步骤 3：使用 C++ 的 opencv3.4.1 的 readNetFromTensorflow 函数加载 .pb 文件。像这样的代码：

#include <fstream>
#include <sstream>
#include <iostream>
#include <opencv2/dnn.hpp>
#include <opencv2/imgproc.hpp>
#include <opencv2/highgui.hpp>
using namespace std;
using namespace cv;
using namespace dnn;

std::vector<std::string> classes;

int main(int argc,char**argv)
{
    if(argc != 2)
    {
        cout<<"Usage: ./main [tensorflow modle path(.pb)]"<<endl;
        return -1;
    }
    String model = argv[1];
    Net net = cv::dnn::readNetFromTensorflow(model,argv[2]);
    cout<<"load Net OK!!"<<endl;

    float inp[1*1] = {2};
    Mat Matrix(1,1,CV_32FC1,inp);
    cout<<"Matrix:\n"<<Matrix<<endl;
    net.setInput(Matrix);

    Mat output = net.forward();
    cout<<"output: " << output <<endl;
    return 0;
}

结果应该是 7。 有两个问题。一是如何生成训练好的模型的完整.pb文件，二是如何将opencv3.4.1 dnn中的预训练模型与C++一起使用？

Answer 1

我发现了这个程序发生了什么。问题出现在第 2 步。必须使用 tensorflow 函数 convert_variables_to_constants 将默认图转换为新图。然后使用 tf.train.write_graph 可以完全保存旧的预训练模型。最后只需要修改第2步的代码，预训练好的模型就可以加载成功了。第二步的新代码如下：

#!/usr/bin/python
import tensorflow as tf
import numpy as np
x_ = tf.placeholder(np.float32, [None, 1], 'input')
y_ = tf.placeholder(np.float32, [None, 1], 'label')
#layer1 
a1 = tf.layers.dense(input=x_,units=3,name="layer1") 
#layer2
a2 = tf.layers.dense(input=x_,units=1,name="layer2")
#global steps
steps = 5000 
x = []
y = []
for i in range(1,200,5):
    temp = (1.0 * i)/10
    x.append([temp])
    y.append([3. + 2. * temp])
x = np.array(x)
y = np.array(y)
#loss function
loss =  tf.reduce_mean(tf.reduce_sum(tf.square(a2-y_)))
#optimizer
optimizer = tf.train.GradientDescentOptimizer(0.00001).minimize(loss)
with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
    length = len(x)
    #training...
    for i in range(steps):
        sess.run(optimizer,feed_dict={x_:x,y_:y})
        result = sess.run(loss,feed_dict={x_:x,y_:y})
        if i % 50 == 0:
            print("loss: ",result,"\tstep: ",i)
     #save the .pbtxt file of the pre-trained model.
     tf.train.write_graph(sess.graph.as_graph_def(), "./model/", 
     #transfrom default graph  and save as a new graph.
     # the param 'output_node_names' should be the last op's name in the pre-trained model. In this model, last op's name is "layer2/BiasAdd" that found in the .pbtxt file like this:
     #node{
     #     name: "layer2/BiasAdd"
     #     op:   "BiasAdd"
     # .....
     output_graph_def = tf.graph_util.convert_variables_to_constants(sess,sess.graph_def,output_node_names=['layer2/BiasAdd'])
     tf.train.write_graph(output_graph_def, "./model/", "graph.pbtxt",as_txt = False)
     print("predict...")
     pre = sess.run(a2,feed_dict={x_:[[0]]})
     print("x = 2 pre: ",pre)

C++代码和上面的差不多：

#include <fstream>
#include <sstream>
#include <iostream>
#include <opencv2/dnn.hpp>
#include <opencv2/imgproc.hpp>
#include <opencv2/highgui.hpp>
using namespace std;
using namespace cv;
using namespace dnn;

std::vector<std::string> classes;

int main(int argc,char**argv)
{
    if(argc != 2)
    {
        cout<<"Usage: ./main [tensorflow modle path(.pb)]"<<endl;
        return -1;
    }
    String model = argv[1];
    Net net = cv::dnn::readNetFromTensorflow(model,argv[2]);
    if(net.empty())
    {
        cout<<"load Net failed"<<endl;
        return -1;
    }
    cout<<"load Net OK!!"<<endl;
    float inp[1*1] = {0};
    Mat Matrix(1,1,CV_32FC1,inp);
    cout<<"Matrix:\n"<<Matrix<<endl;
    net.setInput(Matrix);

    Mat output = net.forward();
    #the value of the output should be equal to the output of the Step 1.
    cout<<"output: " << output <<endl;
    return 0;
}

另外，当我使用 w1、b1 和 w2、b2 而不是 tf.layers.dense 构建网络时，出现了一个我暂时不明白的错误：

Error: Unspecified error (More than one input is Const op) in getConstBlob, file /home/wy/Downloads/opencv 3.4.1/modules/dnn/src/tensorflow/tf_importer.cpp, line 571

异常：OpenCV(3.4.1) /home/wy/Downloads/opencv-3.4.1/modules/dnn/src/tensorflow/tf_importer.cpp:571：错误：(-2) 多个输入是常量函数getConstBlob中的操作