代码实战(带说明注释):
import tensorflow as tf
# import tensorflow.examples.tutorials.mnist.input_data as input_data
from tensorflow.examples.tutorials.mnist import input_data
# 下载和读取mnist数据
mnist = input_data.read_data_sets("MNIST_data/", one_hot=True)
# 占位符,每行有784数据
x = tf.placeholder(tf.float32, [None, 784])
y_actual = tf.placeholder(tf.float32, shape=[None, 10]) # 输入的标签占位符,即0~9
# 初始化权值
def weight_variable_initial(shape):
# 截尾正态分布随机数组,张量的维度:shape,平均值:mean,标准差为stddev,随机种子:seed,取值范围为:[ mean - 2 * stddev, mean + 2 * stddev ]
weight_initial = tf.truncated_normal(shape, mean=0, stddev=0.1, dtype=tf.float32, seed=None, name="weight_variable")
# sess = tf.Session()
# with sess.as_default():
# print("weight_variable_initial:", weight_initial.eval())
# 生成变量
return tf.Variable(weight_initial)
# 初始化偏置项
def bias_variable_initial(shape):
# 定义常量
initial = tf.constant(value=0.1, dtype=None, shape=shape, name="bias_variable")
# sess = tf.Session()
# with sess.as_default():
# print("bias_variable:", initial.eval())
# # 生成变量
return tf.Variable(initial)
# 构建卷积层
def conv2d(x, W):
'''
输入张量:input 张量为为 [ batch, in_height, in_weight, in_channel ],
其中batch为图片的数量,
in_height 为图片高度,
in_weight 为图片宽度,
in_channel 为图片的通道数,灰度图该值为1,彩色图为3。
卷积核:filter,也是一个张量
shape为 [ filter_height, filter_weight, in_channel, out_channels ],
其中 filter_height 为卷积核高度,
filter_weight 为卷积核宽度,
in_channel 是图像通道数
strides 卷积时在图像每一维的步长,这是一个一维的向量,
[ 1, strides, strides, 1],第一位和最后一位固定必须是1
padding string类型,值为“SAME” 和 “VALID”
表示的是卷积的形式,是否考虑边界。
"SAME"是考虑边界,不足的时候用0去填充周围,"VALID"则不考虑
use_cudnn_on_gpu
bool类型,是否使用cudnn加速,默认为true
'''
return tf.nn.conv2d(input=x, filter=W, strides=[1, 1, 1, 1], padding='SAME', use_cudnn_on_gpu=None, name="conv2d")
# 构建池化层
def max_pool(x):
'''
value:
需要池化的输入,一般池化层接在卷积层后面,所以输入通常是feature map,
依然是[batch_size, height, width, channels]这样的shape
ksize:
池化窗口的大小,取一个四维向量,一般是[1, height, width, 1]
strides:
窗口在每一个维度上滑动的步长,一般也是[1, stride,stride, 1]
padding:
padding string类型,值为“SAME” 和 “VALID”
表示的是卷积的形式,是否考虑边界。
"SAME"是考虑边界,不足的时候用0去填充周围,"VALID"则不考虑
'''
return tf.nn.max_pool(value=x, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
# 构建网络
'''将tensor变换为shape的形式
shape为一个列表形式,特殊的一点是列表中可以存在-1。
-1代表的含义是不用我们自己指定这一维的大小,
函数会自动计算,但列表中只能存在一个-1。
28 * 28 = 784
'''
x_image = tf.reshape(tensor=x, shape=[-1, 28, 28, 1],name="x_image") # 转换shape
W_conv1 = weight_variable_initial([5, 5, 1, 32])
b_conv1 = bias_variable_initial([32])
h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1) # 第一个卷积层
h_pool1 = max_pool(h_conv1) # 第一个池化层
W_conv2 = weight_variable_initial([5, 5, 32, 64])
b_conv2 = bias_variable_initial([64])
h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2) # 第二个卷积层
h_pool2 = max_pool(h_conv2) # 第二个池化层
W_fc1 = weight_variable_initial([7 * 7 * 64, 1024])
b_fc1 = bias_variable_initial([1024])
h_pool2_flat = tf.reshape(h_pool2, [-1, 7 * 7 * 64]) # reshape成向量
h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1) # 第一个全连接层
keep_prob = tf.placeholder("float")
h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob) # dropout层
W_fc2 = weight_variable_initial([1024, 10])
b_fc2 = bias_variable_initial([10])
y_predict = tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2) + b_fc2) # softmax层
cross_entropy = -tf.reduce_sum(y_actual * tf.log(y_predict)) # 交叉熵
train_step = tf.train.GradientDescentOptimizer(1e-3).minimize(cross_entropy) # 梯度下降法
correct_prediction = tf.equal(tf.argmax(y_predict, 1), tf.argmax(y_actual, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float")) # 精确度计算
sess = tf.InteractiveSession()
sess.run(tf.initialize_all_variables())
for i in range(50000):
batch = mnist.train.next_batch(50)
if i % 1000 == 0: # 训练1000次,验证一次
train_acc = accuracy.eval(feed_dict={x: batch[0], y_actual: batch[1], keep_prob: 1.0})
print('loop:', i, 'current accuracy:', train_acc)
train_step.run(feed_dict={x: batch[0], y_actual: batch[1], keep_prob: 0.5})
test_acc = accuracy.eval(feed_dict={x: mnist.test.images, y_actual: mnist.test.labels, keep_prob: 1.0})
print("last accuracy", test_acc)
输出结果:
Extracting MNIST_data/train-images-idx3-ubyte.gz
Extracting MNIST_data/train-labels-idx1-ubyte.gz
Extracting MNIST_data/t10k-images-idx3-ubyte.gz
Extracting MNIST_data/t10k-labels-idx1-ubyte.gz
loop: 1000 current accuracy: 0.14
loop: 2000 current accuracy: 0.2
loop: 3000 current accuracy: 0.26
loop: 4000 current accuracy: 0.54
loop: 5000 current accuracy: 0.7
loop: 6000 current accuracy: 0.7
loop: 7000 current accuracy: 0.82
loop: 8000 current accuracy: 0.78
loop: 9000 current accuracy: 0.88
loop: 10000 current accuracy: 0.94
loop: 11000 current accuracy: 0.92
loop: 12000 current accuracy: 0.94
loop: 13000 current accuracy: 0.88
loop: 14000 current accuracy: 0.9
loop: 15000 current accuracy: 0.92
loop: 16000 current accuracy: 0.92
loop: 17000 current accuracy: 0.94
loop: 18000 current accuracy: 0.9
last accuracy 0.9401
网友评论