1 实现把任意图片放进训练好的网络进行测试

输入的图片是白底黑字的数字图片进行测试，测试前需要做两步
（1）转换图片矩阵大小为28*28符合网络的输入
（2）把图片的转换成白字黑底的黑白图片
mnist_app.py

  1. import tensorflow as tf
  2. import numpy as np
  3. from PIL import Image
  4. import mnist_backward
  5. import mnist_forward
  8. def restore_model(testPicArr):
  9.     # 利用tf.Graph()复现之前定义的计算图
  10.     with tf.Graph().as_default() as tg:
  11.         x = tf.placeholder(tf.float32, [None, mnist_forward.INPUT_NODE])
  12.         # 调用mnist_forward文件中的前向传播过程forword()函数
  13.         y = mnist_forward.forward(x, None)
  14.         # 得到概率最大的预测值
  15.         preValue = tf.argmax(y, 1)
  17.         # 实例化具有滑动平均的saver对象
  18.         variable_averages = tf.train.ExponentialMovingAverage(mnist_backward.MOVING_AVERAGE_DECAY)
  19.         variables_to_restore = variable_averages.variables_to_restore()
  20.         saver = tf.train.Saver(variables_to_restore)
  22.         with tf.Session() as sess:
  23.             # 通过ckpt获取最新保存的模型
  24.             ckpt = tf.train.get_checkpoint_state(mnist_backward.MODEL_SAVE_PATH)
  25.             if ckpt and ckpt.model_checkpoint_path:
  26.                 saver.restore(sess, ckpt.model_checkpoint_path)
  28.                 preValue = sess.run(preValue, feed_dict={x: testPicArr})
  29.                 return preValue
  30.             else:
  31.                 print("No checkpoint file found")
  32.                 return -1
  35. # 预处理，包括resize，转变灰度图，二值化
  36. def pre_pic(picName):
  37.     img = Image.open(picName)
  38.     reIm = img.resize((28, 28), Image.ANTIALIAS)
  39.     #把图片转换为灰度值图片
  40.     im_arr = np.array(reIm.convert('L'))
  41.     # 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）
  42.     threshold = 50
  43.     # 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。
  44.     for i in range(28):
  45.         for j in range(28):
  46.             im_arr[i][j] = 255 - im_arr[i][j]
  47.             if (im_arr[i][j] < threshold):
  48.                 im_arr[i][j] = 0
  49.             else:
  50.                 im_arr[i][j] = 255
  51.     # 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）
  52.     nm_arr = im_arr.reshape([1, 784])
  53.     nm_arr = nm_arr.astype(np.float32)
  54.     # 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
  55.     img_ready = np.multiply(nm_arr, 1.0 / 255.0)
  57.     return img_ready
  60. def application():
  61.     # 输入要识别的几张图片
  62.     testNum = int(input("input the number of test pictures:"))
  63.     for i in range(testNum):
  64.         # 给出待识别图片的路径和名称
  65.         testPic = input("the path of test picture:")
  66.         # 图片预处理
  67.         testPicArr = pre_pic(testPic)
  68.         # 获取预测结果
  69.         preValue = restore_model(testPicArr)
  70.         print("The prediction number is:", preValue)
  72. def main():
  73.     application()
  75. if __name__ == '__main__':
  76.     main()

2 实现制作数据

2.1 简介

  • 数据集可以生成二进制的tfrecords文件。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。
  • 用tf.train.Example的协议存储训练情况，训练数据的特征用键值对的形式表示。
  • 用SerializeToString()把数据序列化为字符串存储。

    2.2 生成tfrecords文件

    1. writer = tf.python_io.TFRecordWriter(tfRecordName)
    2.       # 把每张图片和标签封装到example中
    3.       example = tf.train.Example(features=tf.train.Features(feature={
    4.           'img_raw': tf.train.Feature(bytes_list=tf.train.BytesList(value=[img_raw])),# img_raw放入原始图片
    5.           'label': tf.train.Feature(int64_list=tf.train.Int64List(value=labels))# labels是图片的标签
    6.       }))
    7.       # 把example进行序列化
    8.       writer.write(example.SerializeToString())
    9.   # 关闭writer
    10.   writer.close()

    2.3 解析tfrecords文件

    ```python

    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,
    1.                                  features={
    2.                                      'label': tf.FixedLenFeature([10], tf.int64),
    3.                                      'img_raw': tf.FixedLenFeature([], tf.string)
    4.                                  })

    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)
  1. <a name="Qh9Mi"></a>
  2. ## 2.4 生成自定义数据的完整代码
  3. 读取的文件格式是。图片文件名+空格+标签<br />![截屏2020-12-21 下午5.50.16.png](https://cdn.nlark.com/yuque/0/2020/png/1780216/1608544221091-34b15a24-0d52-4f21-874b-2314747e0aa6.png#align=left&display=inline&height=221&margin=%5Bobject%20Object%5D&name=%E6%88%AA%E5%B1%8F2020-12-21%20%E4%B8%8B%E5%8D%885.50.16.png&originHeight=221&originWidth=237&size=51618&status=done&style=none&width=237)
  4. <a name="hXJ73"></a>
  5. ### mnist_generateds.py文件
  6. ```python
  7. #mnist_generateds.py
  8. # coding:utf-8
  9. import tensorflow as tf
  10. import numpy as np
  11. from PIL import Image
  12. import os
  14. image_train_path = './mnist_data_jpg/mnist_train_jpg_60000/'
  15. label_train_path = './mnist_data_jpg/mnist_train_jpg_60000.txt'
  16. tfRecord_train = './data/mnist_train.tfrecords'
  17. image_test_path = './mnist_data_jpg/mnist_test_jpg_10000/'
  18. label_test_path = './mnist_data_jpg/mnist_test_jpg_10000.txt'
  19. tfRecord_test = './data/mnist_test.tfrecords'
  20. data_path = './data'
  21. resize_height = 28
  22. resize_width = 28
  25. # 生成tfrecords文件
  26. def write_tfRecord(tfRecordName, image_path, label_path):
  27.     # 新建一个writer
  28.     writer = tf.python_io.TFRecordWriter(tfRecordName)
  29.     num_pic = 0
  30.     f = open(label_path, 'r')
  31.     contents = f.readlines()
  32.     f.close()
  33.     # 循环遍历每张图和标签
  34.     for content in contents:
  35.         value = content.split()
  36.         img_path = image_path + value[0]
  37.         img = Image.open(img_path)
  38.         img_raw = img.tobytes()#图片转换为二进制数据
  39.         labels = [0] * 10
  40.         labels[int(value[1])] = 1
  41.         # 把每张图片和标签封装到example中
  42.         example = tf.train.Example(features=tf.train.Features(feature={
  43.             'img_raw': tf.train.Feature(bytes_list=tf.train.BytesList(value=[img_raw])),
  44.             'label': tf.train.Feature(int64_list=tf.train.Int64List(value=labels))
  45.         }))
  46.         # 把example进行序列化
  47.         writer.write(example.SerializeToString())
  48.         num_pic += 1#每完成一张图片，计数器加1
  49.         print("the number of picture:", num_pic)
  50.     # 关闭writer
  51.     writer.close()
  52.     print("write tfrecord successful")
  55. def generate_tfRecord():
  56.     isExists = os.path.exists(data_path)
  57.     if not isExists:
  58.         os.makedirs(data_path)
  59.         print('The directory was created successfully')
  60.     else:
  61.         print('directory already exists')
  62.     write_tfRecord(tfRecord_train, image_train_path, label_train_path)
  63.     write_tfRecord(tfRecord_test, image_test_path, label_test_path)
  66. # 解析tfrecords文件
  67. def read_tfRecord(tfRecord_path):
  68.     # 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
  69.     filename_queue = tf.train.string_input_producer([tfRecord_path], shuffle=True)
  70.     # 新建一个reader
  71.     reader = tf.TFRecordReader()
  72.     # 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。
  73.     _, serialized_example = reader.read(filename_queue)
  74.     # 将tf.train.Example协议内存块(protocol buffer)解析为张量
  75.     features = tf.parse_single_example(serialized_example,
  76.                                        features={
  77.                                            'label': tf.FixedLenFeature([10], tf.int64),# 10表示标签的分类数量
  78.                                            'img_raw': tf.FixedLenFeature([], tf.string)
  79.                                        })
  80.     # 将img_raw字符串转换为8位无符号整型
  81.     img = tf.decode_raw(features['img_raw'], tf.uint8)
  82.     # 将形状变为一行784列
  83.     img.set_shape([784])
  84.     img = tf.cast(img, tf.float32) * (1. / 255)
  85.     # 变成0到1之间的浮点数
  86.     label = tf.cast(features['label'], tf.float32)
  87.     # 返回图片和标签
  88.     return img, label
  91. def get_tfrecord(num, isTrain=True):
  92.     if isTrain:
  93.         tfRecord_path = tfRecord_train
  94.     else:
  95.         tfRecord_path = tfRecord_test
  96.     img, label = read_tfRecord(tfRecord_path)
  97.     # 随机读取一个batch的数据，打乱数据
  98.     img_batch, label_batch = tf.train.shuffle_batch([img, label],
  99.                                                     batch_size=num,
  100.                                                     num_threads=2,# 线程
  101.                                                     capacity=1000,
  102.                                                     min_after_dequeue=700)
  103.     # 返回的图片和标签为随机抽取的batch_size组
  104.     return img_batch, label_batch
  107. def main():
  108.     generate_tfRecord()
  111. if __name__ == '__main__':
  112.     main()

在反向传播mnistbackward.py和测试程序mnist_test.py中修改图片标签的接口。使用线程协调器，方法如下

  1. coord = tf.train.Coordinator()
  2. threads = tf.train.start_queue_runners(sess = sess,coord = coord)
  4. # 图片和标签的批获取
  5. coord.request_stop()
  6. coord.join(threads)

mnist_backward.py文件

线程协调器的代码是用################################################括起来的

  1. #mnist_backward.py
  2. import tensorflow as tf
  3. from tensorflow.examples.tutorials.mnist import input_data
  4. import mnist_forward
  5. import os
  6. import mnist_generateds  # 1
  8. BATCH_SIZE = 200
  9. LEARNING_RATE_BASE = 0.1
  10. LEARNING_RATE_DECAY = 0.99
  11. REGULARIZER = 0.0001
  12. STEPS = 50000
  13. MOVING_AVERAGE_DECAY = 0.99
  14. MODEL_SAVE_PATH = "./model/"
  15. MODEL_NAME = "mnist_model"
  16. # 手动给出训练的总样本数6万
  17. train_num_examples = 60000  # 给出数据集的数量
  20. def backward():
  21.     x = tf.placeholder(tf.float32, [None, mnist_forward.INPUT_NODE])
  22.     y_ = tf.placeholder(tf.float32, [None, mnist_forward.OUTPUT_NODE])
  23.     y = mnist_forward.forward(x, REGULARIZER)
  24.     global_step = tf.Variable(0, trainable=False)
  26.     ce = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=y, labels=tf.argmax(y_, 1))
  27.     cem = tf.reduce_mean(ce)
  28.     loss = cem + tf.add_n(tf.get_collection('losses'))
  30.     learning_rate = tf.train.exponential_decay(
  31.         LEARNING_RATE_BASE,
  32.         global_step,
  33.         train_num_examples / BATCH_SIZE,
  34.         LEARNING_RATE_DECAY,
  35.         staircase=True)
  37.     train_step = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss, global_step=global_step)
  39.     ema = tf.train.ExponentialMovingAverage(MOVING_AVERAGE_DECAY, global_step)
  40.     ema_op = ema.apply(tf.trainable_variables())
  41.     with tf.control_dependencies([train_step, ema_op]):
  42.         train_op = tf.no_op(name='train')
  44.     saver = tf.train.Saver()
  45.     # 一次批获取 batch_size张图片和标签
  46.     ################################################
  47.     img_batch, label_batch = mnist_generateds.get_tfrecord(BATCH_SIZE, isTrain=True)  # 3
  48.     ################################################
  50.     with tf.Session() as sess:
  51.         init_op = tf.global_variables_initializer()
  52.         sess.run(init_op)
  54.         ckpt = tf.train.get_checkpoint_state(MODEL_SAVE_PATH)
  55.         if ckpt and ckpt.model_checkpoint_path:
  56.             saver.restore(sess, ckpt.model_checkpoint_path)
  57.         ################################################
  58.         # 利用多线程提高图片和标签的批获取效率
  59.         coord = tf.train.Coordinator()  # 4
  60.         # 启动输入队列的线程
  61.         threads = tf.train.start_queue_runners(sess=sess, coord=coord)  # 5
  62.         ################################################
  64.         for i in range(STEPS):
  65.             ################################################
  66.             # 执行图片和标签的批获取
  67.             xs, ys = sess.run([img_batch, label_batch])  # 6
  68.             ################################################
  70.             _, loss_value, step = sess.run([train_op, loss, global_step], feed_dict={x: xs, y_: ys})
  71.             if i % 1000 == 0:
  72.                 print("After %d training step(s), loss on training batch is %g." % (step, loss_value))
  73.                 saver.save(sess, os.path.join(MODEL_SAVE_PATH, MODEL_NAME), global_step=global_step)
  74.         ################################################
  75.         # 关闭线程协调器
  76.         coord.request_stop()  # 7
  77.         coord.join(threads)  # 8
  78.         ################################################
  80. def main():
  81.     backward()  # 9
  84. if __name__ == '__main__':
  85.     main()

mnist_test.py文件

线程协调器的代码是用################################################括起来的

  1. import time
  2. import tensorflow as tf
  3. from tensorflow.examples.tutorials.mnist import input_data
  4. import mnist_forward
  5. import mnist_backward
  6. import mnist_generateds
  8. TEST_INTERVAL_SECS = 5
  9. # 手动给出测试的总样本数1万
  10. TEST_NUM = 10000  # 1
  13. def test():
  14.     with tf.Graph().as_default() as g:
  15.         x = tf.placeholder(tf.float32, [None, mnist_forward.INPUT_NODE])
  16.         y_ = tf.placeholder(tf.float32, [None, mnist_forward.OUTPUT_NODE])
  17.         y = mnist_forward.forward(x, None)
  19.         ema = tf.train.ExponentialMovingAverage(mnist_backward.MOVING_AVERAGE_DECAY)
  20.         ema_restore = ema.variables_to_restore()
  21.         saver = tf.train.Saver(ema_restore)
  23.         correct_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(y_, 1))
  24.         accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
  25.         ################################################
  26.         # 用函数get_tfrecord替换读取所有测试集1万张图片
  27.         img_batch, label_batch = mnist_generateds.get_tfrecord(TEST_NUM, isTrain=False)  # 2
  28.         ################################################
  30.         while True:
  31.             with tf.Session() as sess:
  32.                 ckpt = tf.train.get_checkpoint_state(mnist_backward.MODEL_SAVE_PATH)
  33.                 if ckpt and ckpt.model_checkpoint_path:
  34.                     saver.restore(sess, ckpt.model_checkpoint_path)
  35.                     global_step = ckpt.model_checkpoint_path.split('/')[-1].split('-')[-1]
  36.                     ################################################
  37.                     # 利用多线程提高图片和标签的批获取效率
  38.                     coord = tf.train.Coordinator()  # 3
  39.                     # 启动输入队列的线程
  40.                     threads = tf.train.start_queue_runners(sess=sess, coord=coord)  # 4
  42.                     # 执行图片和标签的批获取
  43.                     xs, ys = sess.run([img_batch, label_batch])  # 5
  44.                     ################################################
  45.                     accuracy_score = sess.run(accuracy, feed_dict={x: xs, y_: ys})
  47.                     print("After %s training step(s), test accuracy = %g" % (global_step, accuracy_score))
  48.                     ################################################
  49.                     # 关闭线程协调器
  50.                     coord.request_stop()  # 6
  51.                     coord.join(threads)  # 7
  52.                     ################################################
  54.                 else:
  55.                     print('No checkpoint file found')
  56.                     return
  57.             time.sleep(TEST_INTERVAL_SECS)
  60. def main():
  61.     test()  # 8
  64. if __name__ == '__main__':
  65.     main()