1 Lenet5模型简介

截屏2020-12-21 下午8.37.03.png

2 TensorFlow实现

mnist_lenet5_forward.py文件

  1. import tensorflow as tf
  3. # 每张图片分辨率为28*28
  4. IMAGE_SIZE = 28
  5. # Mnist数据集为灰度图，故输入图片通道数NUM_CHANNELS取值为1
  6. NUM_CHANNELS = 1
  7. # 第一层卷积核大小为5
  8. CONV1_SIZE = 5
  9. # 卷积核个数为32
  10. CONV1_KERNEL_NUM = 32
  11. # 第二层卷积核大小为5
  12. CONV2_SIZE = 5
  13. # 卷积核个数为64
  14. CONV2_KERNEL_NUM = 64
  15. # 全连接层第一层为 512 个神经元
  16. FC_SIZE = 512
  17. # 全连接层第二层为 10 个神经元
  18. OUTPUT_NODE = 10
  21. # 权重w计算
  22. def get_weight(shape, regularizer):
  23.     w = tf.Variable(tf.truncated_normal(shape, stddev=0.1))
  24.     if regularizer != None: tf.add_to_collection('losses', tf.contrib.layers.l2_regularizer(regularizer)(w))
  25.     return w
  28. # 偏置b计算
  29. def get_bias(shape):
  30.     b = tf.Variable(tf.zeros(shape))
  31.     return b
  34. # 卷积层计算
  35. def conv2d(x, w):
  36.     return tf.nn.conv2d(x, w, strides=[1, 1, 1, 1], padding='SAME')
  39. # 最大池化层计算
  40. def max_pool_2x2(x):
  41.     return tf.nn.max_pool(x, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
  44. def forward(x, train, regularizer):
  45.     # 实现第一层卷积
  46.     conv1_w = get_weight([CONV1_SIZE, CONV1_SIZE, NUM_CHANNELS, CONV1_KERNEL_NUM], regularizer)
  47.     conv1_b = get_bias([CONV1_KERNEL_NUM])
  48.     conv1 = conv2d(x, conv1_w)
  49.     # 非线性激活
  50.     relu1 = tf.nn.relu(tf.nn.bias_add(conv1, conv1_b))
  51.     # 最大池化
  52.     pool1 = max_pool_2x2(relu1)
  54.     # 实现第二层卷积
  55.     conv2_w = get_weight([CONV2_SIZE, CONV2_SIZE, CONV1_KERNEL_NUM, CONV2_KERNEL_NUM], regularizer)
  56.     conv2_b = get_bias([CONV2_KERNEL_NUM])
  57.     conv2 = conv2d(pool1, conv2_w)
  58.     relu2 = tf.nn.relu(tf.nn.bias_add(conv2, conv2_b))
  59.     pool2 = max_pool_2x2(relu2)
  61.     # 获取一个张量的维度
  62.     pool_shape = pool2.get_shape().as_list()
  63.     # pool_shape[1] 为长 pool_shape[2] 为宽 pool_shape[3]为高
  64.     nodes = pool_shape[1] * pool_shape[2] * pool_shape[3]
  65.     # 得到矩阵被拉长后的长度，pool_shape[0]为batch值
  66.     reshaped = tf.reshape(pool2, [pool_shape[0], nodes])
  68.     # 实现第三层全连接层
  69.     fc1_w = get_weight([nodes, FC_SIZE], regularizer)
  70.     fc1_b = get_bias([FC_SIZE])
  71.     fc1 = tf.nn.relu(tf.matmul(reshaped, fc1_w) + fc1_b)
  72.     # 如果是训练阶段，则对该层输出使用dropout
  73.     if train: fc1 = tf.nn.dropout(fc1, 0.5)
  75.     # 实现第四层全连接层
  76.     fc2_w = get_weight([FC_SIZE, OUTPUT_NODE], regularizer)
  77.     fc2_b = get_bias([OUTPUT_NODE])
  78.     y = tf.matmul(fc1, fc2_w) + fc2_b
  79.     return y

mnist_lenet5_backward.py文件

  1. import tensorflow as tf
  2. from tensorflow.examples.tutorials.mnist import input_data
  3. import mnist_lenet5_forward
  4. import os
  5. import numpy as np
  7. # batch的数量
  8. BATCH_SIZE = 100
  9. # 初始学习率
  10. LEARNING_RATE_BASE = 0.005
  11. # 学习率衰减率
  12. LEARNING_RATE_DECAY = 0.99
  13. # 正则化
  14. REGULARIZER = 0.0001
  15. # 最大迭代次数
  16. STEPS = 50000
  17. # 滑动平均衰减率
  18. MOVING_AVERAGE_DECAY = 0.99
  19. # 模型保存路径
  20. MODEL_SAVE_PATH = "./model/"
  21. # 模型名称
  22. MODEL_NAME = "mnist_model"
  25. def backward(mnist):
  26.     # 卷积层输入为四阶张量
  27.     # 第一阶表示每轮喂入的图片数量，第二阶和第三阶分别表示图片的行分辨率和列分辨率，第四阶表示通道数
  28.     x = tf.placeholder(tf.float32, [
  29.         BATCH_SIZE,
  30.         mnist_lenet5_forward.IMAGE_SIZE,
  31.         mnist_lenet5_forward.IMAGE_SIZE,
  32.         mnist_lenet5_forward.NUM_CHANNELS])
  33.     y_ = tf.placeholder(tf.float32, [None, mnist_lenet5_forward.OUTPUT_NODE])
  34.     # 前向传播过程
  35.     y = mnist_lenet5_forward.forward(x, True, REGULARIZER)
  36.     # 声明一个全局计数器
  37.     global_step = tf.Variable(0, trainable=False)
  38.     # 对网络最后一层的输出y做softmax，求取输出属于某一类的概率
  39.     ce = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=y, labels=tf.argmax(y_, 1))
  40.     # 向量求均值
  41.     cem = tf.reduce_mean(ce)
  42.     # 正则化的损失值
  43.     loss = cem + tf.add_n(tf.get_collection('losses'))
  44.     # 指数衰减学习率
  45.     learning_rate = tf.train.exponential_decay(
  46.         LEARNING_RATE_BASE,
  47.         global_step,
  48.         mnist.train.num_examples / BATCH_SIZE,
  49.         LEARNING_RATE_DECAY,
  50.         staircase=True)
  51.     # 梯度下降算法的优化器
  52.     # train_step = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss, global_step=global_step)
  53.     train_step = tf.train.MomentumOptimizer(learning_rate, 0.9).minimize(loss, global_step=global_step)
  54.     # 采用滑动平均的方法更新参数
  55.     ema = tf.train.ExponentialMovingAverage(MOVING_AVERAGE_DECAY, global_step)
  58.     ema_op = ema.apply(tf.trainable_variables())
  59.     # 将train_step和ema_op两个训练操作绑定到train_op上
  60.     with tf.control_dependencies([train_step, ema_op]):
  61.         train_op = tf.no_op(name='train')
  63.     # 实例化一个保存和恢复变量的saver
  64.     saver = tf.train.Saver()
  65.     # 创建一个会话
  66.     with tf.Session() as sess:
  67.         init_op = tf.global_variables_initializer()
  68.         sess.run(init_op)
  69.         # 通过 checkpoint 文件定位到最新保存的模型，若文件存在，则加载最新的模型
  70.         ckpt = tf.train.get_checkpoint_state(MODEL_SAVE_PATH)
  71.         if ckpt and ckpt.model_checkpoint_path:
  72.             saver.restore(sess, ckpt.model_checkpoint_path)
  74.         for i in range(STEPS):
  75.             # 读取一个batch数据，将输入数据xs转成与网络输入相同形状的矩阵
  76.             xs, ys = mnist.train.next_batch(BATCH_SIZE)
  77.             reshaped_xs = np.reshape(xs, (
  78.                 BATCH_SIZE,
  79.                 mnist_lenet5_forward.IMAGE_SIZE,
  80.                 mnist_lenet5_forward.IMAGE_SIZE,
  81.                 mnist_lenet5_forward.NUM_CHANNELS))
  82.             # 读取一个batch数据，将输入数据xs转成与网络输入相同形状的矩阵
  83.             _, loss_value, step = sess.run([train_op, loss, global_step], feed_dict={x: reshaped_xs, y_: ys})
  84.             if i % 100 == 0:
  85.                 print("After %d training step(s), loss on training batch is %g." % (step, loss_value))
  86.                 saver.save(sess, os.path.join(MODEL_SAVE_PATH, MODEL_NAME), global_step=global_step)
  89. def main():
  90.     mnist = input_data.read_data_sets("./data/", one_hot=True)
  91.     backward(mnist)
  94. if __name__ == '__main__':
  95.     main()

mnist_lenet5_test.py文件

  1. import time
  2. import tensorflow as tf
  3. from tensorflow.examples.tutorials.mnist import input_data
  4. import mnist_lenet5_forward
  5. import mnist_lenet5_backward
  6. import numpy as np
  8. TEST_INTERVAL_SECS = 5
  10. #创建一个默认图，在该图中执行以下操作
  11. def test(mnist):
  12.     with tf.Graph().as_default() as g: 
  13.         x = tf.placeholder(tf.float32,[
  14.             mnist.test.num_examples,
  15.             mnist_lenet5_forward.IMAGE_SIZE,
  16.             mnist_lenet5_forward.IMAGE_SIZE,
  17.             mnist_lenet5_forward.NUM_CHANNELS]) 
  18.         y_ = tf.placeholder(tf.float32, [None, mnist_lenet5_forward.OUTPUT_NODE])
  19.         #训练好的网络，故不使用 dropout
  20.         y = mnist_lenet5_forward.forward(x,False,None)
  22.         ema = tf.train.ExponentialMovingAverage(mnist_lenet5_backward.MOVING_AVERAGE_DECAY)
  23.         ema_restore = ema.variables_to_restore()
  24.         saver = tf.train.Saver(ema_restore)
  26.         #判断预测值和实际值是否相同
  27.         correct_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(y_, 1)) 
  28.         ## 求平均得到准确率
  29.         accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32)) 
  31.         while True:
  32.             with tf.Session() as sess:
  33.                 ckpt = tf.train.get_checkpoint_state(mnist_lenet5_backward.MODEL_SAVE_PATH)
  34.                 if ckpt and ckpt.model_checkpoint_path:
  35.                     saver.restore(sess, ckpt.model_checkpoint_path)    
  36.                     # 根据读入的模型名字切分出该模型是属于迭代了多少次保存的
  37.                     global_step = ckpt.model_checkpoint_path.split('/')[-1].split('-')[-1] 
  38.                     reshaped_x = np.reshape(mnist.test.images,(
  39.                     mnist.test.num_examples,
  40.                     mnist_lenet5_forward.IMAGE_SIZE,
  41.                     mnist_lenet5_forward.IMAGE_SIZE,
  42.                     mnist_lenet5_forward.NUM_CHANNELS))
  43.                     #利用多线程提高图片和标签的批获取效率
  44.                     coord = tf.train.Coordinator()#3
  45.                     threads = tf.train.start_queue_runners(sess=sess, coord=coord)#4
  46.                     accuracy_score = sess.run(accuracy, feed_dict={x:reshaped_x,y_:mnist.test.labels}) 
  47.                     print("After %s training step(s), test accuracy = %g" % (global_step, accuracy_score))
  48.                     #关闭线程协调器
  49.                     coord.request_stop()#6
  50.                     coord.join(threads)#7
  51.                 else:
  52.                     print('No checkpoint file found')
  53.                     return
  54.             time.sleep(TEST_INTERVAL_SECS) 
  56. def main():
  57.     mnist = input_data.read_data_sets("./data/", one_hot=True)
  58.     test(mnist)
  60. if __name__ == '__main__':
  61.     main()