tensorflow笔记：6.2 制作数据集

• 基于TensorFlow制作tfrecord格式的数据集

ps：这里接触了好多新东西，感觉还是似懂非懂的（大概原理懂了，但是具体 实现的函数文件什么的都没有见过，还是要好好看看）。

1. 数据集生成读取文件 mnist_ generateds.py

   • tfrecords 文件
     (1) tfrecords 是一种二进制文件，可先将图片和标签制作成该格式的文件。
     使用 tfrecords 进行数据读取，会提高内存利用率。
     (2) tf.train.Example : 用来存储训练数据 。 训练数据的特征用 键值对 的形式表示。
     如::‘ img_raw ’ 值 ‘ label ’ 值 值是 Byteslist/FloatList/Int64List
     (3) SerializeToString( ) 把数据序列化成字符串存储。

   • 生成tfrecords 文件

2. 反向传播文件 修改图片标签获取的接口 mnist_backward .py
   关键操作：利用多线程提高图片和标签的批获取效率
   方法：将批获取的操作放到线程协调器开启和关闭之间
   开启线程协调器：

coord =tf.train.Coordinator( )
threads = tf.train.start_queue_runners(sess=sess, coord=coord)

关闭线程协调器：

coord.request_stop( )
coord.join(threads)

3. 测试文件修改图片标签获取的接口( mnist_test.py)

---

• 测试过的代码 fc4

代码信息.png

mnist_generated.py

#coding:utf-8
import tensorflow as tf
import numpy as np
from PIL import Image
import os

image_train_path='./mnist_data_jpg/mnist_train_jpg_60000/'
label_train_path='./mnist_data_jpg/mnist_train_jpg_60000.txt'
tfRecord_train='./data/mnist_train.tfrecords'
image_test_path='./mnist_data_jpg/mnist_test_jpg_10000/'
label_test_path='./mnist_data_jpg/mnist_test_jpg_10000.txt'
tfRecord_test='./data/mnist_test.tfrecords'
data_path='./data'
resize_height = 28
resize_width = 28

#生成tfrecords文件
def write_tfRecord(tfRecordName, image_path, label_path):
    #新建一个writer
    writer = tf.python_io.TFRecordWriter(tfRecordName)  
    num_pic = 0 
    f = open(label_path, 'r')
    contents = f.readlines()
    f.close()
    #循环遍历每张图和标签 
    for content in contents:
        value = content.split()
        img_path = image_path + value[0] 
        img = Image.open(img_path)
        img_raw = img.tobytes() 
        labels = [0] * 10  
        labels[int(value[1])] = 1  
        #把每张图片和标签封装到example中    
        example = tf.train.Example(features=tf.train.Features(feature={
                'img_raw': tf.train.Feature(bytes_list=tf.train.BytesList(value=[img_raw])),
                'label': tf.train.Feature(int64_list=tf.train.Int64List(value=labels))
                })) 
        #把example进行序列化
        writer.write(example.SerializeToString())
        num_pic += 1 
        print ("the number of picture:", num_pic)
    #关闭writer
    writer.close()
    print("write tfrecord successful")

def generate_tfRecord():
    isExists = os.path.exists(data_path) 
    if not isExists: 
        os.makedirs(data_path)
        print ('The directory was created successfully')
    else:
        print ('directory already exists') 
    write_tfRecord(tfRecord_train, image_train_path, label_train_path)
    write_tfRecord(tfRecord_test, image_test_path, label_test_path)
#解析tfrecords文件  
def read_tfRecord(tfRecord_path):
    #该函数会生成一个先入先出的队列，文件阅读器会使用它来读取数据
    filename_queue = tf.train.string_input_producer([tfRecord_path], shuffle=True)
    #新建一个reader
    reader = tf.TFRecordReader()
    #把读出的每个样本保存在serialized_example中进行解序列化，标签和图片的键名应该和制作tfrecords的键名相同，其中标签给出几分类。
    _, serialized_example = reader.read(filename_queue) 
    #将tf.train.Example协议内存块(protocol buffer)解析为张量
    features = tf.parse_single_example(serialized_example,
                                       features={
                                        'label': tf.FixedLenFeature([10], tf.int64),
                                        'img_raw': tf.FixedLenFeature([], tf.string)
                                        })
    #将img_raw字符串转换为8位无符号整型
    img = tf.decode_raw(features['img_raw'], tf.uint8)
    #将形状变为一行784列
    img.set_shape([784])
    img = tf.cast(img, tf.float32) * (1. / 255)
    #变成0到1之间的浮点数      
    label = tf.cast(features['label'], tf.float32)
    #返回图片和标签
    return img, label 
      
def get_tfrecord(num, isTrain=True):
    if isTrain:
        tfRecord_path = tfRecord_train
    else:
        tfRecord_path = tfRecord_test
    img, label = read_tfRecord(tfRecord_path)
    #随机读取一个batch的数据
    img_batch, label_batch = tf.train.shuffle_batch([img, label],
                                                    batch_size = num,
                                                    num_threads = 2,
                                                    capacity = 1000,
                                                    min_after_dequeue = 700)
    #返回的图片和标签为随机抽取的batch_size组
    return img_batch, label_batch

def main():
    generate_tfRecord()

if __name__ == '__main__':
    main()

mnist_forward.py

#coding:utf-8
#版本信息：ubuntu18.04  python3.6.8  tensorflow 1.14.0
#作者：九除以三还是三哦  如有错误，欢迎评论指正！！
#1前向传播过程
import tensorflow as tf

#网络输入节点为784个（代表每张输入图片的像素个数）
INPUT_NODE = 784
#输出节点为10个（表示输出为数字0-9的十分类）
OUTPUT_NODE = 10
#隐藏层节点500个
LAYER1_NODE = 500


def get_weight(shape, regularizer):
    #参数满足截断正态分布，并使用正则化，
    w = tf.Variable(tf.random.truncated_normal(shape,stddev=0.1))
    #w = tf.Variable(tf.random_normal(shape,stddev=0.1))
    #将每个参数的正则化损失加到总损失中
    if regularizer != None: tf.add_to_collection('losses', tf.contrib.layers.l2_regularizer(regularizer)(w))
    return w


def get_bias(shape):  
    #初始化的一维数组，初始化值为全 0
    b = tf.Variable(tf.zeros(shape))  
    return b
    
def forward(x, regularizer):
    #由输入层到隐藏层的参数w1形状为[784,500]
    w1 = get_weight([INPUT_NODE, LAYER1_NODE], regularizer)
    #由输入层到隐藏的偏置b1形状为长度500的一维数组，
    b1 = get_bias([LAYER1_NODE])
    #前向传播结构第一层为输入 x与参数 w1矩阵相乘加上偏置 b1 ，再经过relu函数 ，得到隐藏层输出 y1。
    y1 = tf.nn.relu(tf.matmul(x, w1) + b1)
    #由隐藏层到输出层的参数w2形状为[500,10]
    w2 = get_weight([LAYER1_NODE, OUTPUT_NODE], regularizer)
    #由隐藏层到输出的偏置b2形状为长度10的一维数组
    b2 = get_bias([OUTPUT_NODE])
    #前向传播结构第二层为隐藏输出 y1与参 数 w2 矩阵相乘加上偏置 矩阵相乘加上偏置 b2，得到输出 y。
    #由于输出 。由于输出 y要经过softmax oftmax 函数，使其符合概率分布，故输出y不经过 relu函数
    y = tf.matmul(y1, w2) + b2
    return y

mnist_backward.py

import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
import mnist_forward
import os
import mnist_generateds#1

BATCH_SIZE = 200
LEARNING_RATE_BASE = 0.1
LEARNING_RATE_DECAY = 0.99
REGULARIZER = 0.0001
STEPS = 50000
MOVING_AVERAGE_DECAY = 0.99
MODEL_SAVE_PATH="./model/"
MODEL_NAME="mnist_model"
train_num_examples = 60000#2

def backward():

    x = tf.placeholder(tf.float32, [None, mnist_forward.INPUT_NODE])
    y_ = tf.placeholder(tf.float32, [None, mnist_forward.OUTPUT_NODE])
    y = mnist_forward.forward(x, REGULARIZER)
    global_step = tf.Variable(0, trainable=False) 

    ce = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=y, labels=tf.argmax(y_, 1))
    cem = tf.reduce_mean(ce)
    loss = cem + tf.add_n(tf.get_collection('losses'))

    learning_rate = tf.train.exponential_decay(
        LEARNING_RATE_BASE,
        global_step,
        train_num_examples / BATCH_SIZE, 
        LEARNING_RATE_DECAY,
        staircase=True)

    train_step = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss, global_step=global_step)

    ema = tf.train.ExponentialMovingAverage(MOVING_AVERAGE_DECAY, global_step)
    ema_op = ema.apply(tf.trainable_variables())
    with tf.control_dependencies([train_step, ema_op]):
        train_op = tf.no_op(name='train')

    saver = tf.train.Saver()
    img_batch, label_batch = mnist_generateds.get_tfrecord(BATCH_SIZE, isTrain=True)#3

    with tf.Session() as sess:
        init_op = tf.global_variables_initializer()
        sess.run(init_op)

        ckpt = tf.train.get_checkpoint_state(MODEL_SAVE_PATH)
        if ckpt and ckpt.model_checkpoint_path:
            saver.restore(sess, ckpt.model_checkpoint_path)
            
        coord = tf.train.Coordinator()#4
        threads = tf.train.start_queue_runners(sess=sess, coord=coord)#5
        
        for i in range(STEPS):
            xs, ys = sess.run([img_batch, label_batch])#6
            _, loss_value, step = sess.run([train_op, loss, global_step], feed_dict={x: xs, y_: ys})
            if i % 1000 == 0:
                print("After %d training step(s), loss on training batch is %g." % (step, loss_value))
                saver.save(sess, os.path.join(MODEL_SAVE_PATH, MODEL_NAME), global_step=global_step)

        coord.request_stop()#7
        coord.join(threads)#8


def main():
    backward()#9

if __name__ == '__main__':
    main()

mnist_test.py

#coding:utf-8
import time
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
import mnist_forward
import mnist_backward
import mnist_generateds
TEST_INTERVAL_SECS = 5
#手动给出测试的总样本数1万
TEST_NUM = 10000#1

def test():
    with tf.Graph().as_default() as g:
        x = tf.placeholder(tf.float32, [None, mnist_forward.INPUT_NODE])
        y_ = tf.placeholder(tf.float32, [None, mnist_forward.OUTPUT_NODE])
        y = mnist_forward.forward(x, None)

        ema = tf.train.ExponentialMovingAverage(mnist_backward.MOVING_AVERAGE_DECAY)
        ema_restore = ema.variables_to_restore()
        saver = tf.train.Saver(ema_restore)
        
        correct_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(y_, 1))
        accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
        #用函数get_tfrecord替换读取所有测试集1万张图片
        img_batch, label_batch = mnist_generateds.get_tfrecord(TEST_NUM, isTrain=False)#2

        while True:
            with tf.Session() as sess:
                ckpt = tf.train.get_checkpoint_state(mnist_backward.MODEL_SAVE_PATH)
                if ckpt and ckpt.model_checkpoint_path:
                    saver.restore(sess, ckpt.model_checkpoint_path)
                    global_step = ckpt.model_checkpoint_path.split('/')[-1].split('-')[-1]
                    #利用多线程提高图片和标签的批获取效率 
                    coord = tf.train.Coordinator()#3
                    #启动输入队列的线程
                    threads = tf.train.start_queue_runners(sess=sess, coord=coord)#4

                    #执行图片和标签的批获取
                    xs, ys = sess.run([img_batch, label_batch])#5

                    accuracy_score = sess.run(accuracy, feed_dict={x: xs, y_: ys})

                    print("After %s training step(s), test accuracy = %g" % (global_step, accuracy_score))
                    #关闭线程协调器
                    coord.request_stop()#6
                    coord.join(threads)#7

                else:
                    print('No checkpoint file found')
                    return
            time.sleep(TEST_INTERVAL_SECS)

def main():
    test()#8

if __name__ == '__main__':
    main()

mnist_app.py

#coding:utf-8

import tensorflow as tf
import numpy as np
from PIL import Image
import mnist_backward
import mnist_forward

def restore_model(testPicArr):
    #利用tf.Graph()复现之前定义的计算图
    with tf.Graph().as_default() as tg:
        x = tf.compat.v1.placeholder(tf.float32, [None, mnist_forward.INPUT_NODE])
        #调用mnist_forward文件中的前向传播过程forword()函数
        y = mnist_forward.forward(x, None)
        #得到概率最大的预测值
        preValue = tf.argmax(y, 1)

         #实例化具有滑动平均的saver对象
        variable_averages = tf.train.ExponentialMovingAverage(mnist_backward.MOVING_AVERAGE_DECAY)
        variables_to_restore = variable_averages.variables_to_restore()
        saver = tf.compat.v1.train.Saver(variables_to_restore)

        with tf.compat.v1.Session() as sess:
            #通过ckpt获取最新保存的模型
            ckpt = tf.train.get_checkpoint_state(mnist_backward.MODEL_SAVE_PATH)
            if ckpt and ckpt.model_checkpoint_path:
                saver.restore(sess, ckpt.model_checkpoint_path)
        
                preValue = sess.run(preValue, feed_dict={x:testPicArr})
                return preValue
            else:
                print("No checkpoint file found")
                return -1

#预处理，包括resize，转变灰度图，二值化
def pre_pic(picName):
    img = Image.open(picName)
    reIm = img.resize((28,28), Image.ANTIALIAS)
    im_arr = np.array(reIm.convert('L'))
    #对图片做二值化处理（这样以滤掉噪声，另外调试中可适当调节阈值）
    threshold = 50
    #模型的要求是黑底白字，但输入的图是白底黑字，所以需要对每个像素点的值改为255减去原值以得到互补的反色。
    for i in range(28):
        for j in range(28):
            im_arr[i][j] = 255 - im_arr[i][j]
            if (im_arr[i][j] < threshold):
                im_arr[i][j] = 0
            else: im_arr[i][j] = 255
    #把图片形状拉成1行784列，并把值变为浮点型（因为要求像素点是0-1 之间的浮点数）
    nm_arr = im_arr.reshape([1, 784])
    nm_arr = nm_arr.astype(np.float32)
    #接着让现有的RGB图从0-255之间的数变为0-1之间的浮点数
    img_ready = np.multiply(nm_arr, 1.0/255.0)

    return img_ready

def application():
    #输入要识别的几张图片
    testNum = int(input("input the number of test pictures:"))
    for i in range(testNum):
        #给出待识别图片的路径和名称
        testPic = input("the path of test picture:")
        #图片预处理
        testPicArr = pre_pic(testPic)
        #获取预测结果
        preValue = restore_model(testPicArr)
        print ("The prediction number is:", preValue)

def main():
    application()

if __name__ == '__main__':
    main()      

• 运行结果

测试结果不如上次的理想，即使是老师给出的代码也有两个数字识别错误。但可以看到程序在测试集上跑出的准确率还是挺高的，在98%左右。
老师的建议是考虑图片噪声是否过大，从图片的预处理处进行优化。

准确率.png

• 小结

至此，完全连接网络的全部内容已经完毕，第一个数字识别的实践项目已经完成啦（为什么还感觉是小白）
代码好长好繁琐，报错信息更长更繁琐，不同版本之间的差异让小白着实难过。不过所学能很快的实践应用还是hin开心啦。