1.卷积神经网络
详见 直白介绍卷积神经网络
2.代码实现:
与全连接网络不通,本篇去掉了 正则化方法,改用了 Dropout方法,加入两个 卷积层 与 池化层 。
import tensorflow as tf
import pylab
import os
from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets('MNIST_data/',one_hot=True)
os.environ["CUDA_VISIBLE_DEVICES"] = "0,1"
#定义图片和标签的占位符
#None 表示张量的第一维度可以接受任意长度
x = tf.placeholder(tf.float32,[None,784])
y = tf.placeholder(tf.float32,[None,10])
#定义权重及偏置值变量
W1 = tf.Variable(tf.random_normal(([7*7*60,600])))
b1 = tf.Variable(tf.constant(0.1,shape=[600]))
W2 = tf.Variable(tf.random_normal(([600,10])))
b2 = tf.Variable(tf.constant(0.1,shape=[10]))
#定义权重方法:
def get_filter(shape):
return tf.Variable(tf.truncated_normal(shape,stddev=0.1))
#定义偏置值方法:
def get_bias(shape):
return tf.Variable(tf.constant(0.1,shape=shape))
#定义隐藏层
def hidden_layer(inputs):
#要用激活函数
return tf.nn.relu(tf.matmul(inputs,W1)+b1)
w_con1 = get_filter([5,5,1,30])
b_con1 = get_bias([30])
w_con2 = get_filter([5,5,30,60])
b_con2 = get_bias([60])
#将[100,784]转换为[100,28,28,1]
x_reshape = tf.reshape(x,[-1,28,28,1])
#第一层卷积输出,输出大小为100X28X28X30:
h_conv1 = tf.nn.conv2d(x_reshape,filter=w_con1,strides=[1,1,1,1],padding='SAME')+b_con1
#激活函数:
h1 = tf.nn.relu(h_conv1)
#第一层池化输出,输出大小为100X14X14X30:
h_pool1 = tf.nn.max_pool(h1,ksize=[1,2,2,1],strides=[1,2,2,1],padding='SAME')
#第二层卷积输出,输出大小为100X14X14X60:
h_conv2 = tf.nn.conv2d(h_pool1,filter=w_con2,strides=[1,1,1,1],padding='SAME')+b_con2
#激活函数:
h2 = tf.nn.relu(h_conv2)
#第一层池化输出,输出大小为100X7X7X60:
h_pool2 = tf.nn.max_pool(h2,ksize=[1,2,2,1],strides=[1,2,2,1],padding='SAME')
#将池化后的输出改变下尺寸,
h_reshape = tf.reshape(h_pool2,[-1,7*7*60])
#隐藏层输出,并使用dropout
h3 = hidden_layer(h_reshape)
keep_prob = tf.placeholder(tf.float32)
h_drop1 = tf.nn.dropout(h3,keep_prob)
#预测值,这里不用激活函数,因为等下要用tensorflow定义好的softmax交叉熵函数
pred = tf.matmul(h_drop1,W2) + b2
#定义交叉熵
cross_entropy = tf.nn.softmax_cross_entropy_with_logits(logits=pred,labels=y)
#定义总的损失函数
loss = tf.reduce_mean(cross_entropy)
#定义初始学习率
learning_rate = 0.005
#定义优化器
opt = tf.train.AdamOptimizer(learning_rate).minimize(loss)
#定义参数
#总共循环训练轮数
train_epochs = 15
#每轮循环中,每次训练的图片张数
batch_size = 100
#将所有图片训练一轮所需要的总的训练次数
total_batch = int(mnist.train.num_examples/batch_size)
#以下是测试模型
correct_prediction = tf.equal(tf.argmax(pred,1),tf.argmax(y,1))
#准确率:
accuracy = tf.reduce_mean(tf.cast(correct_prediction,tf.float32))
#启动session
with tf.Session() as sess:
#初始化变量,必写
sess.run(tf.global_variables_initializer())
print("以下是训练模型每轮训练误差:")
#循环训练
for epoch in range(train_epochs):
#定义平均loss值
avg_loss = 0.
#循环所有数据
for i in range(total_batch):
batch_xs,batch_ys = mnist.train.next_batch(batch_size)
_,c,acc = sess.run([opt,loss,accuracy],feed_dict={x:batch_xs,y:batch_ys,keep_prob:0.8})
#平均loss
avg_loss += c / total_batch
#显示每轮的结果
print('Epoch:',epoch+1,', loss:','{:.9f}'.format(avg_loss),', accuracy:','{:.5f}'.format(acc))
print("\n训练模型结束,以下是测试模型准确率:")
acc = sess.run(accuracy,feed_dict={x:mnist.test.images,y:mnist.test.labels,keep_prob:1.0})
print('Accuracy:',acc)
3.相关代码
3.1Dtopout
#隐藏层输出,并使用dropout
h3 = hidden_layer(h_reshape)
keep_prob = tf.placeholder(tf.float32)
h_drop1 = tf.nn.dropout(h3,keep_prob)
3.2卷积池化操作
#定义权重方法:
def get_filter(shape):
return tf.Variable(tf.truncated_normal(shape,stddev=0.1))
#定义偏置值方法:
def get_bias(shape):
return tf.Variable(tf.constant(0.1,shape=shape))
w_con1 = get_filter([5,5,1,30])
b_con1 = get_bias([30])
w_con2 = get_filter([5,5,30,60])
b_con2 = get_bias([60])
#将[100,784]转换为[100,28,28,1]
x_reshape = tf.reshape(x,[-1,28,28,1])
#第一层卷积输出,输出大小为100X28X28X30:
h_conv1 = tf.nn.conv2d(x_reshape,filter=w_con1,strides=[1,1,1,1],padding='SAME')+b_con1
#激活函数:
h1 = tf.nn.relu(h_conv1)
#第一层池化输出,输出大小为100X14X14X30:
h_pool1 = tf.nn.max_pool(h1,ksize=[1,2,2,1],strides=[1,2,2,1],padding='SAME')
#第二层卷积输出,输出大小为100X14X14X60:
h_conv2 = tf.nn.conv2d(h_pool1,filter=w_con2,strides=[1,1,1,1],padding='SAME')+b_con2
#激活函数:
h2 = tf.nn.relu(h_conv2)
#第一层池化输出,输出大小为100X7X7X60:
h_pool2 = tf.nn.max_pool(h2,ksize=[1,2,2,1],strides=[1,2,2,1],padding='SAME')
#将池化后的输出改变下尺寸,
h_reshape = tf.reshape(h_pool2,[-1,7*7*60])
3.3优化器的选择
一开始还是选择了tf.train.GradientDescentOptimizer()+退化学习率的方法,但效果非常差。后来选择了tf.train.AdamOptimizer(),而且还不需要用退化学习率,效果非常好,然后我将这个用在了之前的两个识别代码中,效果依然很棒。因此建议以后优先使用 tf.train.AdamOptimizer()的方法!
#定义优化器
opt = tf.train.AdamOptimizer(learning_rate).minimize(loss)