ここでは、単一ターゲット認識にのみ使用される畳み込みニューラルネットワークが構築されています。ネットワーク構造は次のとおりです。
手書きの数値認識を例にとると、コードは次のとおりです。
import os
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
# os.environ['TF_CPP_MIN_LOG_LEVEL']='1' # 显示所有信息
os.environ['TF_CPP_MIN_LOG_LEVEL']='2' # 只显示warning和error
# os.environ['TF_CPP_MIN_LOG_LEVEL']='3' # 只显示error
# 获取数据集
# one_hot设置为True，将标签数据转化为0/1，如[1,0,0,0,0,0,0,0,0,0]
mnist=input_data.read_data_sets('MNIST_data',one_hot=True)
# 定义一个批次的大小
batch_size=100
# 所有训练数据一共可以分为几个批次
n_batch=mnist.train.num_examples//batch_size
# 变量分析
def variable_summaries(var):
with tf.name_scope('summaries'):
mean=tf.reduce_mean(var)
stddev=tf.sqrt(tf.reduce_mean(tf.square(var-mean)))
tf.summary.scalar('mean',mean)
tf.summary.scalar('stddev',stddev)
tf.summary.scalar('max',tf.reduce_max(var))
tf.summary.scalar('min',tf.reduce_min(var))
tf.summary.histogram('histogram',var) #直方图
# 初始化权重
def weight_variable(shape, name):
# 权重初始值为0不是最优的，应该设置为满足截断正态分布的随机数，收敛速度更快
initial = tf.truncated_normal(shape, stddev=0.1)
return tf.Variable(initial, name=name)
# 初始化偏置
def bias_variable(shape, name):
# 偏置初始值为0不是最优的，可以设置为0.1，收敛速度更快
initial = tf.constant(0.1, shape=shape)
return tf.Variable(initial, name=name)
# 卷积层
def conv2d(x,w):
# x:input tensor of shape [batch, height, width, channels]
# w:filter or kernel tensor of shape [height, width, in_channels, out_channels]
# strides:strides[1]表示x方向的步长，strides[2]表示y方向的步长
# padding:'SAME'表示周围补0，'VALID'表示周围不补0
return tf.nn.conv2d(x,w,strides=[1,1,1,1],padding='SAME')
# 池化层
def max_pool_2x2(x):
# x:input tensor of shape [batch, height, width, channels]
# ksize:ksize[1]表示池化核的宽度，ksize[2]表示池化核的高度
# strides:strides[1]表示x方向的步长，strides[2]表示y方向的步长
# padding:'SAME'表示周围补0，'VALID'表示周围不补0
return tf.nn.max_pool(x,ksize=[1,2,2,1],strides=[1,2,2,1],padding='SAME')
# 定义三个placeholder
# 行数值为None，None可以取任意数，本例中将取值100，即取决于pitch_size
# 列数值为784，因为输入图像尺寸已由28*28转换为1*784
with tf.name_scope('input'):
x=tf.placeholder(tf.float32,[None,784],name='x_input')
y=tf.placeholder(tf.float32,[None,10],name='y_input')
with tf.name_scope('x_image'):
# 将一维变为二维
x_image=tf.reshape(x,[-1,28,28,1],name='x_image')
# 定义keep_prob，表示不执行dropout的神经元的比例
with tf.name_scope('keep_prob'):
keep_prob=tf.placeholder(tf.float32, name='keep_prob')
# 定义学习率
with tf.name_scope('lr'):
lr=tf.Variable(0.001,dtype=tf.float32)
# 定义第一个卷积层
with tf.name_scope('conv1'):
with tf.name_scope('conv1_w'):
# 8个尺寸为5x5的卷积核对1张图像做卷积计算
conv1_w = weight_variable([5,5,1,8], name='conv1_w')
with tf.name_scope('conv1_b'):
conv1_b = bias_variable([8], name='conv1_b')
# 卷积运算
with tf.name_scope('conv2d_1'):
conv2d_1=conv2d(x_image,conv1_w)+conv1_b
# 激活运算
with tf.name_scope('relu'):
h_conv1=tf.nn.relu(conv2d_1)
# 池化运算
with tf.name_scope('h_pool1'):
h_pool1=max_pool_2x2(h_conv1) # 该层图像尺寸变为14*14
# 定义第二个卷积层
with tf.name_scope('conv2'):
with tf.name_scope('conv2_w'):
# 16个尺寸为5x5的卷积核对8张图像做卷积计算
conv2_w = weight_variable([5,5,8,16], name='conv2_w')
with tf.name_scope('conv2_b'):
conv2_b = bias_variable([16], name='conv2_b')
# 卷积运算
with tf.name_scope('conv2d_2'):
conv2d_2=conv2d(h_pool1,conv2_w)+conv2_b
# 激活运算
with tf.name_scope('relu'):
h_conv2=tf.nn.relu(conv2d_2)
# 池化运算
with tf.name_scope('h_pool2'):
h_pool2=max_pool_2x2(h_conv2) # 该层图像尺寸变为7*7
# 定义第一个全连接层
with tf.name_scope('fc1'):
with tf.name_scope('fc1_w'):
fc1_w=weight_variable([7*7*16,100], name='fc1_w')
with tf.name_scope('fc1_b'):
fc1_b=bias_variable([100],name='fc1_b')
# 将二维变为一维
with tf.name_scope('h_pool2_flat'):
h_pool2_flat=tf.reshape(h_pool2,[-1,7*7*16],name='h_pool2_flat')
# 求一个全连接层的输出
with tf.name_scope('wx_plus_b1'):
wx_plus_b1=tf.matmul(h_pool2_flat,fc1_w)+fc1_b
with tf.name_scope('relu'):
h_fc1=tf.nn.relu(wx_plus_b1)
# 引入dropout
with tf.name_scope('h_fc1_drop'):
h_fc1_drop=tf.nn.dropout(h_fc1,keep_prob,name='h_fc1_drop')
# 定义第二个全连接层
with tf.name_scope('fc2'):
with tf.name_scope('fc2_w'):
fc2_w=weight_variable([100,10], name='fc2_w')
with tf.name_scope('fc2_b'):
fc2_b=bias_variable([10],name='fc2_b')
# 求一个全连接层的输出
with tf.name_scope('wx_plus_b2'):
wx_plus_b2=tf.matmul(h_fc1_drop,fc2_w)+fc2_b
with tf.name_scope('softmax'):
prediction=tf.nn.softmax(wx_plus_b2)
# 定义损失函数
with tf.name_scope('loss'):
# 由于输出神经元为softmax，交叉熵损失函数比均方误差损失函数收敛速度更快
# loss=tf.reduce_mean(tf.square(y-prediction))
loss=tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y,logits=prediction),name='loss')
# loss分析
tf.summary.scalar('loss',loss)
# 定义训练方式
with tf.name_scope('train'):
# 优化器通过调整loss里的参数，使loss不断减小
# AdamOptimizer比GradientDescentOptimizer收敛速度更快
# train=tf.train.GradientDescentOptimizer(0.2).minimize(loss)
train=tf.train.AdamOptimizer(lr).minimize(loss)
# 计算准确率
with tf.name_scope('accuracy'):
# tf.argmax返回第一个参数中最大值的下标
# tf.equal比较两个参数是否相等，返回True或False
correct_prediction=tf.equal(tf.argmax(y,1),tf.argmax(prediction,1))
# tf.cast将布尔类型转换为浮点类型
accuracy=tf.reduce_mean(tf.cast(correct_prediction,tf.float32))
# accuracy分析
tf.summary.scalar('accuracy',accuracy)
# 合并所有summary
merged=tf.summary.merge_all()
with tf.Session() as sess:
# 变量初始化
sess.run(tf.global_variables_initializer())
# 生成计算图
train_writer=tf.summary.FileWriter('logs/train',sess.graph)
test_writer=tf.summary.FileWriter('logs/test',sess.graph)
# epoch为周期数，所有批次训练完为一个周期
for epoch in range(20):
# 调整学习率
sess.run(tf.assign(lr,0.001*(0.95**epoch)))
for batch in range(n_batch):
# 每次取出batch_size条数据进行训练
batch_xs, batch_ys = mnist.train.next_batch(batch_size)
sess.run(train, feed_dict={
x:batch_xs, y:batch_ys, keep_prob:0.5})
# 记录训练集计算的loss和accuracy
summary=sess.run(merged, feed_dict={
x:batch_xs, y:batch_ys, keep_prob:1.0})
train_writer.add_summary(summary, epoch*n_batch+batch)
# 记录测试集计算的loss和accuracy
batch_xs, batch_ys = mnist.test.next_batch(batch_size)
summary=sess.run(merged, feed_dict={
x:batch_xs, y:batch_ys, keep_prob:1.0})
test_writer.add_summary(summary, epoch*n_batch+batch)
learning_rate=sess.run(lr)
test_acc = sess.run(accuracy,feed_dict={
x:mnist.test.images,y:mnist.test.labels,keep_prob:1.0})
train_acc = sess.run(accuracy,feed_dict={
x:mnist.train.images,y:mnist.train.labels,keep_prob:1.0})
print('epoch=',epoch,' ','learning_rate=%.7f' % learning_rate,' ','test_acc=',test_acc,' ','train_acc=',train_acc)
演算結果：
コマンドラインウィンドウに入力してtensorboard --logdir logsURLを取得し、Google ChromeでURLを開くと、計算グラフと要約
トレイン
テストを表示できます
