opencv3.4.1 调用tensorflow生成的网络模型数据pb文件

python代码如下是训练手写字的网络

# -*- coding: utf-8 -*-
"""
Created on Fri May 18 16:04:06 2018


@author: Administrator
"""


import pylab
import numpy as np
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
from tensorflow.python.framework import graph_util
tf.reset_default_graph()
learning_rate = 0.001
training_epochs = 25
batch_size = 100
display_step =1


n_hidden_1 = 256
n_hidden_2 = 256
n_input = 784
n_classes = 10
mnist = input_data.read_data_sets("MNIST_data/", one_hot = True)


print ('输入数据：', mnist.train.images)
print ('输入数据的shape：', mnist.train.images.shape)
x = tf.placeholder("float", [None, n_input], name = 'input')
y = tf.placeholder("float", [None, n_classes], name = 'labels')


def multilayer_perception(x, weights_t, biases_t):
    #第一层隐藏层
    layer_1 = tf.add(tf.matmul(x, weights_t['h1']), biases_t['b1'])
    layer_1 = tf.nn.relu(layer_1)
    
    #第二层隐藏层
    layer_2 = tf.add(tf.matmul(layer_1, weights_t['h2']), biases_t['b2'])
    layer_2 = tf.nn.relu(layer_2)
    
   
    
    #out_layer = tf.matmul(layer_2, weights_t['out']) + biases_t['out']
    out_layer = tf.add( tf.matmul(layer_2, weights_t['out']), biases_t['out'], name = "output")
    return out_layer


weights = {
        'h1' : tf.Variable(tf.random_normal([n_input, n_hidden_1])),
        'h2' : tf.Variable(tf.random_normal([n_hidden_1, n_hidden_2])), 
        'out': tf.Variable(tf.random_normal([n_hidden_2, n_classes]))
        }


biases = {
        'b1' : tf.Variable(tf.random_normal([n_hidden_1])),
        'b2' : tf.Variable(tf.random_normal([n_hidden_2])),
       
        'out': tf.Variable(tf.random_normal([n_classes]))
        }


#weights_c = {
#        'h1' : tf.constant(tf.random_normal([n_input, n_hidden_1])),
#        'h2' : tf.constant(tf.random_normal([n_hidden_1, n_hidden_2])),      
#        'out': tf.constant(tf.random_normal([n_hidden_2, n_classes]))
#        }
#
#biases_c = {
#        'b1' : tf.constant(tf.random_normal([n_hidden_1])),
#        'b2' : tf.constant(tf.random_normal([n_hidden_2])),        
#        'out': tf.constant(tf.random_normal([n_classes]))
#        }




print("learn_param")


pred = multilayer_perception(x, weights, biases)




print("multilayer_perception")
cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=pred, labels=y))
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(cost)
saver = tf.train.Saver()


savedir = "log1/"
with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
    for epoch in range(training_epochs):
        avg_cost = 0
        total_batch = int(mnist.train.num_examples/batch_size)
        for i in range(total_batch):
            batch_xs, batch_ys = mnist.train.next_batch(batch_size)
            _, c = sess.run([optimizer, cost], feed_dict = {x:batch_xs, y:batch_ys})
            avg_cost += c/total_batch
        if (epoch + 1) % display_step == 0:
            print ("Epoch:", '%04d' % (epoch + 1), "cost=", "{:.9f}".format(avg_cost))
    print("finished")
    constant_graph = graph_util.convert_variables_to_constants(sess, sess.graph_def, ["output"])
    with tf.gfile.FastGFile('expert-graph_t1.pb', mode='wb') as f:

        f.write(constant_graph.SerializeToString())

#这几句代码就是用来将图参数变成constant，可以很方便的导出所有的网络参数，不用每一个参数去将其变换成constant.

#im = mnist.train.images[1]
#im = im.reshape(-1, 28)
#pylab.imshow(im)
#pylab.show()
    
#测试模型
    correct_prediction = tf.equal(tf.argmax(pred, 1), tf.argmax(y, 1))
    accuray = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
    print ("Accuracy:", accuray.eval({x:mnist.test.images, y:mnist.test.labels}))
    
    save_path = saver.save(sess, savedir + "mnistmodel.cpk")
    print ("Model saved in file:%s" % save_path)
 
#    weights_c = {}
#    biases_c = {}
#    _h1 = weights['h1'].eval(sess)
#    _h2 = weights['h2'].eval(sess)
#    _out_w = weights['out'].eval(sess)
#    
#    _b1 = biases['b1'].eval(sess)
#    _b2 = biases['b2'].eval(sess)
#    _out_b = biases['out'].eval(sess)
#
###pred_c = tf.Constant(pred, name = "pred_out")
#g_2 = tf.Graph()
#with g_2.as_default():
#    weights_c['h1'] = tf.constant(_h1 , name = "w_h1")
#    weights_c['h2'] = tf.constant(_h2, name = "w_h2")
#    weights_c['out'] = tf.constant(_out_w, name = "w_out")
#    
#    biases_c['b1'] = tf.constant(_b1, name = "w_b1")
#    biases_c['b2'] = tf.constant(_b2, name = "w_b2")
#    biases_c['out'] = tf.constant(_out_b, name = "b_bout")
#    X_2 = tf.placeholder("float", shape = [None, 784], name = "input")
#    pred = multilayer_perception(X_2,  weights_c, biases_c)
#    print ("pred")
#    pred_out = tf.nn.softmax(pred, name = "pred_out")
##    sess_2 = tf.Session()
##    init_2 = tf.initialize_all_variables()
##    sess_2.run(init_2)
#    #graph_def = g_2.as_graph_def()
#    #tf.train.write_graph(graph_def, savedir, 'expert-graph.pb', as_text=False)
#    graph_def = g_2.as_graph_def()
#    #tf.train.write_graph(graph_def, export_dir, 'expert-graph_t1.pb', as_text=True)
#    tf.train.write_graph(graph_def, savedir, 'expert-graph.pb', as_text=False)
##    print ("pred1")
##    constant_graph = g_2.as_graph_def()
##    with tf.gfile.FastGFile(savedir+'expert-graph.pb', mode='wb') as f:
##        f.write(constant_graph.SerializeToString())
#
with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
    saver.restore(sess, savedir + "mnistmodel.cpk")
    correct_prediction = tf.equal(tf.argmax(pred, 1), tf.argmax(y, 1))
   
    accuray = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
    print ("Accuracy:", accuray.eval({x:mnist.test.images, y:mnist.test.labels}))
    
    output = tf.argmax(pred, 1)
    batch_xs, batch_ys = mnist.train.next_batch(2)
    outval, preadv = sess.run([output, pred], feed_dict = {x:batch_xs, y:batch_ys})
    print(outval, preadv, batch_ys)
    
    im = batch_xs[0]
    im = im.reshape(-1, 28)
    pylab.imshow(im)
    pylab.show()
    
    im = batch_xs[1]
    im = im.reshape(-1, 28)
    print (im.shape)
    pylab.imshow(im)
    pylab.imsave(savedir + "mnistimg.bmp", im)

    pylab.show()

这里测试识别的图片为

在opencv的代码如下：

cv::String weights = "expert-graph_t1.pb";

cv::dnn::Net net = cv::dnn::readNetFromTensorflow(weights); //读取pb文件得到对应的网络模型

cv::Mat img = cv::imread("mnistimg.bmp", 0);  //读取图片

cv::Mat inputBlob = cv::dnn::blobFromImage(img, 1.0/255.0, cv::Size(28, 28), cv::Scalar(), false, false);  //将图像数据转换成blob数据（tensorflow网络中所用的数据类型，包括减去灰度均值， 并且转换成float型， 将灰度值归一化到某一范围内）

net.setInput(inputBlob);  //将blob数据作为输入放到网络中

cv::Mat pred = net.forward(); //网络前向传播

Net Outputs(1):
MatMul_2
257.719 -255    239.082 1127.92 -400.36 765.277 -326.79 -310.638        245.007 134.649

这里的输出没有进行softmax处理，最大的那个就是索引为 3，刚好和python输出的测试结果是一致的