TensorFlow

TensorFlow_MNIST

  1 # Author: Lee
  2 import tensorflow as tf
  3 import numpy as np
  4 
  5 # 下载并载入mnist手写数据库
  6 from tensorflow.examples.tutorials.mnist import input_data
  7 
  8 mnist = input_data.read_data_sets('mnist', one_hot=True)
  9 
 10 #None表示张量Tensor可以是任何维度, /255.是对灰色图像做归一化
 11 input_x = tf.placeholder(tf.float32, [None, 28 * 28]) / 255.
 12 output_y = tf.placeholder(tf.float32, [None, 10])
 13 #对输入数据进行改变形状28 * 28 * 1, -1是维度设置为auto
 14 input_x_image = tf.reshape(input_x, [-1, 28, 28, 1])
 15 
 16 #从测试Test数据集中选取3000个手写数字的图片和对应的标签
 17 test_x = mnist.test.images[:3000]
 18 test_y = mnist.test.labels[:3000]
 19 
 20 #构建神经网络
 21 #第一层卷积 filters,kernals size,strides
 22 conv1 = tf.layers.conv2d(
 23     inputs = input_x_image, # shape = [28, 28, 1]
 24     filters = 32,           # 32个过滤器,输出深度为32
 25     kernel_size = [5, 5],   # 过滤器在二维的大小为5 * 5,相当于过滤器大小
 26     strides = 1,            # 步长为1
 27     padding = 'SAME',       # padding补零方案,same表示输出大小不变(same和valid的算法需要参考官方文档),需要在外围补零两圈
 28     activation = tf.nn.relu
 29     )
 30 #经过第一层卷积之后的输出数据shape为28 * 28 * 32
 31 
 32 #第一层池化(亚采样)pooling
 33 pool1 = tf.layers.max_pooling2d(
 34     inputs = conv1,
 35     pool_size = [2, 2],     # 过滤器在二维的大小,类比kernel_size
 36     strides = 2,            # 步长2
 37     )
 38 #经过第一层池化之后的输出数据shape为14 * 14 * 32
 39 
 40 #第二层卷积 filters,kernals size,strides
 41 conv2 = tf.layers.conv2d(
 42     inputs = pool1,         # shape = [14, 14, 32]
 43     filters = 64,           # 64个过滤器,输出深度为64
 44     kernel_size = [5, 5],   # 过滤器在二维的大小为5 * 5,相当于过滤器大小
 45     strides = 1,            # 步长为1
 46     padding = 'SAME',       # padding补零方案,same表示输出大小不变,需要在外围补零两圈
 47     activation = tf.nn.relu
 48     )
 49 #经过第二层卷积之后的输出数据shape为14 * 14 * 64
 50 
 51 #第二层池化(亚采样)pooling
 52 pool2 = tf.layers.max_pooling2d(
 53     inputs = conv2,
 54     pool_size = [2, 2],     # 过滤器在二维的大小,类比kernel_size
 55     strides = 2,            # 步长2
 56     )
 57 #经过第一层池化之后的输出数据shape为7 * 7 * 64
 58 
 59 #平坦化(flat)
 60 flat = tf.reshape(pool2, [-1, 1, 1, 1024])
 61 
 62 #1024个神经元的全连接层
 63 dense = tf.layers.dense(
 64     inputs = flat,
 65     units = 1024,
 66     activation = tf.nn.relu
 67     )
 68 
 69 # Dropout 丢弃率为50%, Dropout的rate在[0,1]
 70 dropout = tf.layers.dropout(
 71     inputs = dense,
 72     rate = 0.5
 73     )
 74 
 75 # 10个神经元的全连接层, 这里不用激活函数做非线性化
 76 logits = tf.layers.dense(inputs = dropout, units = 10)
 77 #输出形状1 * 1 * 10
 78 
 79 #计算误差<计算Cross_entropy(交叉熵),再用Softmax进行计算百分比>
 80 loss = tf.losses.softmax_cross_entropy(onehot_labels = output_y, logits = logits)
 81 
 82 #使用Adam优化器,Adam为默认优化器,learning_rate = 0.001
 83 train_op = tf.train.AdamOptimizer(learning_rate = 0.001).minimize(loss)
 84 
 85 #预测值和匹配度
 86 #返回(accuracy, update_op),能够创建两个局部变量
 87 accuracy = tf.metrics.accuracy(
 88     labels = tf.argmax(output_y, axis = 1),
 89     predictions = tf.argmax(logits, axis = 1))[1]
 90 
 91 
 92 #创建会话Session
 93 sess = tf.Session()
 94 
 95 #初始化变量:全局变量和局部变量
 96 init = tf.group(tf.global_variables_initializer(), tf.local_variables_initializer())
 97 sess.run(init)
 98 
 99 for i in range(30000):
100     batch = mnist.train.next_batch(50)
101     train_loss, train_op_ = sess.run([loss, train_op], {input_x: batch[0], output_y: batch[1]})
102     if i % 100 == 0:
103         test_accuracy = sess.run(accuracy, {input_x: test_x, output_y: test_y})
104         print("accuracy:", test_accuracy,'loss:', train_loss)
105 
106 
107 #测试, 预测值与真实值对比
108 test_output = sess.run(logits, {input_x: test_x[:20]})
109 inferenced_y = np.argmax(test_output, 1)
110 #推测的数字
111 print('inferenced_data:',inferenced_y)
112 #实际数字
113 print('test_output_data:', np.argmax(test_y[:20],1))
按照如下CNN过程写的,但是RUN的时候出现了一些问题
为什么不兼容?????
由于时间问题,暂且先将代码贴上去, 放在这周末解决!主要就conv2d pool2d内的参数, reshape内的参数注重研究一下!