#第一步：环境设定
import numpy as np
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
import time
#第二步：数据准备
#使用TensorFlow自带的工具加载MNIST手写数字集合
mnist=input_data.read_data_sets('/data/mnist',one_hot=True)
#查看数据维度
mnist.train.images.shape
#查看target维度
mnist.train.labels.shape
#第三步：准备好placeholder
X=tf.placeholder(tf.float32,[None,784],name='X_placeholder')
Y=tf.placeholder(tf.float32,[None,10],name='Y_placeholder')
#第四步：准备好参数和权重
n_hidden_1=256#第1个隐层
n_hidden_2=256#第2个隐层
n_input=784#MNIST数据输入（28*28*1=784）
n_classes=10#MNIST总共有10个手写类别
#高斯分布随机采样
weights={
'h1':tf.Variable(tf.random_normal([n_input,n_hidden_1]),name='W1'),
'h2':tf.Variable(tf.random_normal([n_hidden_1,n_hidden_2]),name='W2'),
'out':tf.Variable(tf.random_normal([n_hidden_1,n_classes]),name='W')
}
biases={
'b1':tf.Variable(tf.random_normal([n_hidden_1]),name='b1'),
'b2':tf.Variable(tf.random_normal([n_hidden_2]),name='b2'),
'out':tf.Variable(tf.random_normal([n_classes]),name='out')
}
#构建网络计算graph
def multilayer_perceptron(x,weights,biases):
#第1个隐层使用relu激活函数
layer_1=tf.add(tf.matmul(x,weights['h1']),biases['b1'],name='fc_1')
layer_1=tf.nn.relu(layer_1,name='relu_1')
#第2个隐层使用relu激活函数
layer_2=tf.add(tf.matmul(layer_1,weights['h2']),biases['b2'],name='fc_2')
layer_2=tf.nn.relu(layer_2,name='relu_2')
#输出层
out_layer=tf.add(tf.matmul(layer_2,weights['out']),biases['out'],name='fc_3')
return out_layer
#拿到预测类别score
pred=multilayer_perceptron(X,weights,biases)
#计算损失函数，初始化optimizer
learning_rate=0.001
loss_all=tf.nn.softmax_cross_entropy_with_logits(logits=pred,labels=Y,name='cross_entropy')
loss=tf.reduce_mean(loss_all,name='avg_loss')
#最优化用的是随机梯度下降，我们选择AdamOptimizer优化器
optimizer=tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(loss)
#初始化变量
init=tf.global_variables_initializer()
#在session执行graph定义的运算
training_epochs=15#训练总轮数
batch_size=128#一批数据大小
display_step=1#信息展示的频度
with tf.Session() as sess:
sess.run(init)
writer=tf.summary.FileWriter('./graphs/MLP_DNN',sess.graph)
#训练
for epoch in range(training_epochs):
avg_loss=0
total_batch=int(mnist.train.num_examples/batch_size)
#遍历所有的batches
for i in range(total_batch):
batch_x,batch_y=mnist.train.next_batch(batch_size)
#使用optimizer进行优化
_,l=sess.run([optimizer,loss],feed_dict={X:batch_x,Y:batch_y})
#求平均损失
avg_loss+=l/total_batch
#每一步都展示信息
if epoch % display_step==0:
print('Epoch:'+'%04d' %(epoch+1),"cost=",\
"{:.9f}".format(avg_loss))
print('Optimization Finished!')
#测试集上评估
correct_prediction=tf.equal(tf.argmax(pred,1),tf.argmax(Y,1))
#计算准确率
accuracy=tf.reduce_mean(tf.cast(correct_prediction,"float"))
print("Accuracy:",accuracy.eval({X:mnist.test.images,Y:mnist.test.labels}))
Epoch:0001 cost= 247.077027243
Epoch:0002 cost= 48.906014056
Epoch:0003 cost= 31.072485875
Epoch:0004 cost= 22.423607512
Epoch:0005 cost= 16.905729532
Epoch:0006 cost= 12.948794754
Epoch:0007 cost= 10.044691833
Epoch:0008 cost= 7.719212243
Epoch:0009 cost= 6.071234272
Epoch:0010 cost= 4.663679320
Epoch:0011 cost= 3.623726023
Epoch:0012 cost= 2.745126135
Epoch:0013 cost= 2.137337426
Epoch:0014 cost= 1.682079803
Epoch:0015 cost= 1.344503605
Optimization Finished!
Accuracy: 0.9437