核心参考：

本博客代码完全是采用该博主的思路编写的，非常感谢！！

How to load the pre-trained model of the tensorflow by using the opencv dnn model！

项目背景 

需根据一批多维数据，回归计算评分，为方便移动端部署，采用C++版的opencv库调用实现BP神经网络的回归模型。

思路与参考

使用tensorflow搭建bp神经网络，主要有3种方式：

  • 方法1：使用加减矩阵运算自己从零搭建层的结构。该方法生成的pb模型，opencv调用时会出现某些结构无法识别而调用失败。
  • 方法2：使用tensorflow的全连接神经网络函数 tf.layer.dense(...). 经实际测试，能成功调用。
  • 方法3：使用tf.keras搭建keras模型，由于生成的模型需要转化才能被opencv调用，略麻烦，所以为未尝试。

BP的tensorflow回归实现过程

实现主要包括：数据集导入、BP神经网络模型构建、模型的保存、预测结果保存到excel

  数据集导入【下载】

                               提取码：v7g1

【参考】使用pandas读取 ‘.csv’ ,'xls'等表格格式数据

import pandas as pd
import numpy as np
from sklearn.model_selection import train_test_split #sklearn中已经废弃cross_validation,将其中的内容整合到model_selection中

# a=[0,1,2,3,4,5,6,7,8]
# print(a[2:-1])  #始下标（默认从0开始）、指定结束下标（切片不会含该元素）

datafile='/data/data.xlsx' #改成你的路径
pd_data=pd.read_excel(datafile)
data =np.array(pd_data) 
# print(data)
label=data[:,-1]  #表格最后一列为标签
print(label.shape) #(193，)
label=label.reshape(-1,1)  # from (193，) to  (193,1)
train_data=data[:,2:-1]   #第0列是序号，第1列是图片名称，第2到-1（最后一列 ,不含结束列）才是图片的特征

X_train,X_test,label_train,label_test=train_test_split(train_data,label,train_size=0.8)  #将数据集分为训练集和验证集
print("data.shape:",data.shape,"  Xtrain.shape:",X_train.shape,"  Xtest.shape:",X_test.shape)
print("label.shape:",label.shape,"  Ytrain.shape:",label_train.shape,"  label_test.shape:",label_test.shape)
# data.shape: (193, 9)   Xtrain.shape: (154, 6)   Xtest.shape: (39, 6)
# label.shape: (193, 1)   Ytrain.shape: (154, 1)   label_test.shape: (39, 1)

【可能遇到的问题】ModuleNotFoundError: No module named 'xlrd'

 pip install xlrd -i  https://pypi.tuna.tsinghua.edu.cn/simple

  建立bp神经网络模型

import tensorflow as tf 
def test_BP(hide_node=10):
    x = tf.placeholder(tf.float32, [None, INPUT_NODE])
    y = tf.placeholder(tf.float32, [None, OUTPUT_NODE])
    # 数据预处理，归一化
    # X_train=preprocessing.scale(X_train)
    # X_test=preprocessing.scale(X_test)
    # layer1
    predict_1 = tf.layers.dense(inputs=x, units=hide_node,activation=tf.nn.relu, name="layer1")  
    # layer2
    predict_2 = tf.layers.dense(inputs=predict_1, units=OUTPUT_NODE,activation=tf.nn.sigmoid, name="layer2")
    # 代价函数
    loss = tf.reduce_mean(tf.square(y - predict_2),name='loss')
    saver = tf.train.Saver()

    # 定义优化器。使用动量法 也可以使用随机梯度下降法等
    # train_step = tf.train.MomentumOptimizer(0.05,0.05).minimize(loss)
    train_step = tf.train.GradientDescentOptimizer(0.2).minimize(loss)

    with tf.Session() as sess:
        sess.run(tf.global_variables_initializer())
        print("init_results:", sess.run(loss, feed_dict={x: X_train, y: label_train})) 
        for i in range(10001):
            sess.run(train_step, feed_dict={x: X_train, y: label_train})
            # print("end_results:", sess.run(loss, feed_dict={x: X_train, y: label_train}))
            if i % 2000 == 0:
                train_results=sess.run(loss, feed_dict={x: X_train, y: label_train})
                val_results = sess.run(loss, feed_dict={x: X_test, y: label_test})
                print("hide_node:",hide_node,"_train_accuraty_",i,":", 1 - train_results,end=" ")
                print("_val_accuraty_",i,":", 1 - val_results)

  保存模型到pb

        import from tensorflow.python.framework import graph_util
        import os
        pb_file_path = os.getcwd()
        constant_graph = graph_util.convert_variables_to_constants(sess, sess.graph_def, output_node_names=['layer2/Sigmoid'])  #输出的节点名称
        # constant_graph = graph_util.convert_variables_to_constants(sess, sess.graph_def, output_node_names=['layer2/BiasAdd'])
        # 写入序列化的 PB 文件
        tf.train.write_graph(sess.graph.as_graph_def(),pb_file_path,'tf_BP_regress.pbtxt', as_text=True)
        # tf.train.write_graph(sess.graph_def, pb_file_path, 'model.pbtxt', as_text=True)
        with tf.gfile.FastGFile( 'tf_BP_regress.pb', mode='wb') as f:
            f.write(constant_graph.SerializeToString())

将预测值保存到excel

 pip install xlsxwriter  -i https://pypi.tuna.tsinghua.edu.cn/simple

        import xlsxwriter  as xlwt
        workbook = xlwt.Workbook(pb_file_path+'/predict.xls')
        worksheet = workbook.add_worksheet()
        pre = sess.run(predict_2, feed_dict={x: X_test, y: label_test})
        results = pre - label_test
        worksheet.write_column('A1', pre)
        worksheet.write_column('B1', label_test)
        worksheet.write_column('C1', results)
        pre = sess.run(predict_2, feed_dict={x: X_train, y: label_train})
        results = pre - label_train
        worksheet.write_column('D1', pre)
        worksheet.write_column('E1', label_train)
        worksheet.write_column('F1', results)
        workbook.close()

opencv调用pb的程序 

#include <opencv2/opencv.hpp>
#include <opencv2/dnn.hpp>
#include <iostream>

using namespace std;
using namespace cv;
using namespace dnn;


int main()
{
	String pb = "tf_BP_regress.pb";
	dnn::Net net = cv::dnn::readNetFromTensorflow(pb);

	// float wrinkle[1][6]={{1.568,3.529,3.499,0.576,0.509,0}};
	float wrinkle[1][6]={{7.729,9.118,6.368,5.137,5.02,2.044}};

	// Mat input =(Mat <CV_32F1>(1,7)  <<1.568,3.529,3.499,0.576,0.509,0,9.681  );
	Mat input =Mat(1,6,CV_32FC1,wrinkle) ;
    net.setInput(input);

    Mat tmp = net.forward();
    cout<<"tmp"<<tmp<<endl;
    float result=tmp.at<float>(0,0);

    cout<< result<<endl;
	return 0;
 
}

附录： 完整代码tf_BP.py

import pandas as pd
import numpy as np
from sklearn.model_selection import train_test_split #sklearn中已经废弃cross_validation,将其中的内容整合到model_selection中
import tensorflow as tf 
from tensorflow.python.framework import graph_util
import os
pb_file_path = os.getcwd()
# a=[0,1,2,3,4,5,6,7,8]
# print(a[2:-1])  #始下标（默认从0开始）、指定结束下标（切片不会含该元素）

datafile='/data/data.xlsx' #改成你的路径
pd_data=pd.read_excel(datafile)
data =np.array(pd_data) 
# print(data)
label=data[:,-1]  #表格最后一列为标签
print(label.shape) #(193，)
label=label.reshape(-1,1)  # from (193，) to  (193,1)
train_data=data[:,2:-1]   #第0列是序号，第1列是图片名称，第2到-1（最后一列 ,不含结束列）才是图片的特征

X_train,X_test,label_train,label_test=train_test_split(train_data,label,train_size=0.8)  #将数据集分为训练集和验证集
print("data.shape:",data.shape,"  Xtrain.shape:",X_train.shape,"  Xtest.shape:",X_test.shape)
print("label.shape:",label.shape,"  Ytrain.shape:",label_train.shape,"  label_test.shape:",label_test.shape)
# data.shape: (193, 9)   Xtrain.shape: (154, 6)   Xtest.shape: (39, 6)
# label.shape: (193, 1)   Ytrain.shape: (154, 1)   label_test.shape: (39, 1)

# 定义BP神经网络的输入输出
INPUT_NODE = 6
OUTPUT_NODE = 1

def test_BP(hide_node=10):

    x = tf.placeholder(tf.float32, [None, INPUT_NODE])
    y = tf.placeholder(tf.float32, [None, OUTPUT_NODE])
    # 数据预处理，归一化
    # X_train=preprocessing.scale(X_train)
    # X_test=preprocessing.scale(X_test)
    # layer1
    predict_1 = tf.layers.dense(inputs=x, units=hide_node,activation=tf.nn.relu, name="layer1")  
    # layer2
    predict_2 = tf.layers.dense(inputs=predict_1, units=OUTPUT_NODE,activation=tf.nn.sigmoid, name="layer2")
    # 代价函数
    loss = tf.reduce_mean(tf.square(y - predict_2),name='loss')
    saver = tf.train.Saver()

    # 定义优化器。使用动量法 也可以使用随机梯度下降法等
    # train_step = tf.train.MomentumOptimizer(0.05,0.05).minimize(loss)
    train_step = tf.train.GradientDescentOptimizer(0.2).minimize(loss)

    with tf.Session() as sess:
        sess.run(tf.global_variables_initializer())
        print("init_results:", sess.run(loss, feed_dict={x: X_train, y: label_train})) 
        for i in range(10001):
            sess.run(train_step, feed_dict={x: X_train, y: label_train})
            # print("end_results:", sess.run(loss, feed_dict={x: X_train, y: label_train}))
            if i % 2000 == 0:
                train_results=sess.run(loss, feed_dict={x: X_train, y: label_train})
                val_results = sess.run(loss, feed_dict={x: X_test, y: label_test})
                print("hide_node:",hide_node,"_train_accuraty_",i,":", 1 - train_results,end=" ")
                print("_val_accuraty_",i,":", 1 - val_results)
                
        # 保存pb pbtxt
        import os
        pb_file_path = os.getcwd()
        constant_graph = graph_util.convert_variables_to_constants(sess, sess.graph_def, output_node_names=['layer2/Sigmoid'])  #输出的节点名称
        # constant_graph = graph_util.convert_variables_to_constants(sess, sess.graph_def, output_node_names=['layer2/BiasAdd'])
        # 写入序列化的 PB 文件
        tf.train.write_graph(sess.graph.as_graph_def(),pb_file_path,'tf_BP_regress.pbtxt', as_text=True)
        # tf.train.write_graph(sess.graph_def, pb_file_path, 'model.pbtxt', as_text=True)
        with tf.gfile.FastGFile( 'tf_BP_regress.pb', mode='wb') as f:
            f.write(constant_graph.SerializeToString())
            
        #将预测结果写入到excel
        import xlsxwriter  as xlwt
        workbook = xlwt.Workbook(pb_file_path+'/predict.xls')
        worksheet = workbook.add_worksheet()
        pre = sess.run(predict_2, feed_dict={x: X_test, y: label_test})
        results = pre - label_test
        worksheet.write_column('A1', pre)
        worksheet.write_column('B1', label_test)
        worksheet.write_column('C1', results)
        pre = sess.run(predict_2, feed_dict={x: X_train, y: label_train})
        results = pre - label_train
        worksheet.write_column('D1', pre)
        worksheet.write_column('E1', label_train)
        worksheet.write_column('F1', results)
        workbook.close()
        
            
test_BP(5)

 

# tensorflow # python # 机器学习 # opencv
Logo
腾讯云开发者社区

腾讯云面向开发者汇聚海量精品云计算使用和开发经验，营造开放的云计算技术生态圈。

更多推荐

cover

李建忠：智能体正将互联网从信息网络重构为行动网络

avatar 腾讯云开发者社区
cover

探秘 DeepSeek 落地进展，腾讯云携手业界专家共话 AI 生产力

avatar 腾讯云开发者社区
cover

腾讯云TVP AI与安全高峰论坛圆满落幕，共探大模型时代的安全破局之道

avatar 腾讯云开发者社区