前提是對TensorFlow有了基本的瞭解，還有神經網路相關知識最基礎的瞭解，下面直接用程式碼實現

import tensorflow as tf
import numpy
from tensorflow.examples.tutorials.mnist import input_data

mnist=input_data.read_data_sets("D:\Python\MNIST_data", one_hot=True)
#新增x作為佔位符
x=tf.placeholder("float", [None, 784])#28*28
#正確結果佔位符
y_=tf.placeholder("float", [None,10])

#生成權重函式
def weight_variable(shape):
    #tf.truncated_normal(shape, mean, stddev) :shape表示生成張量的維度，mean是均值，stddev是標準差。這個函式產生正態分佈，均值和標準差自己設定
    #權重在初始化時應該加入少量的噪聲來打破對稱性以及避免0梯度
    initial=tf.truncated_normal(shape, stddev=0.1)
    return tf.Variable(initial,dtype=tf.float32)

#生成偏置函式
#由於我們使用的是ReLU神經元，因此比較好的做法是用一個較小的正數來初始化偏置項，以避免神經元節點輸出恆為0的問題（dead neurons）
def bias_variable(shape):
    initial=tf.constant(0.1, shape=shape)
    return tf.Variable(initial,dtype=tf.float32)

#卷積函式
#卷積使用1步長，0邊距的模板，池化用2x2的模板
def conv2d(x, W):
#x：待卷積的矩陣具有[batch, in_height, in_width, in_channels]這樣的shape
#w：卷積核具有[filter_height, filter_width, in_channels, out_channels]這樣的shape
#strides：卷積時在影象每一維的步長，這是一個一維的向量，長度4
    return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')

#池化函式
#和卷積基本相同
def max_pool_2x2(x):
    return tf.nn.max_pool(x, ksize=[1, 2, 2, 1],strides=[1, 2, 2, 1], padding='SAME')

#卷積在每個5x5的patch中算出32個特徵。
#卷積的權重張量形狀是[5, 5, 1, 32]，前兩個維度是patch的大小，
#接著是輸出幾個單位，和輸出的幾個維度
W_conv1=weight_variable([5, 5, 1, 32])
b_conv1=bias_variable([32])

#shape:[batch, in_height, in_width, in_channels]
x_image=tf.reshape(x, [-1,28,28,1])

#卷積+偏置，然後給relu啟用函式，最後啟用函式返回值池化
h_conv1=tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1) #output size 28*28*32
h_pool1=max_pool_2x2(h_conv1)  #output size 14*14*32

#第二層卷積，池化
W_conv2=weight_variable([5, 5, 32, 64])
b_conv2=bias_variable([64])
h_conv2=tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2) #output size 14*14*64
h_pool2=max_pool_2x2(h_conv2) #output size 7*7*64

#全連線層1
W_fc1=weight_variable([7*7*64,1024])
b_fc1=bias_variable([1024])
h_pool2_flat=tf.reshape(h_pool2, [-1,7*7*64])
h_fc1=tf.nn.relu(tf.matmul(h_pool2_flat,W_fc1) + b_fc1)

#dropout方法減輕過擬合問題
keep_prob=tf.placeholder("float")
h_fc1_drop=tf.nn.dropout(h_fc1, keep_prob)

#全連線層2
W_fc2=weight_variable([1024, 10])
b_fc2=bias_variable([10])
y_conv=tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2) + b_fc2)

#訓練和評估、儲存模型
cross_entropy=-tf.reduce_sum(y_*tf.log(y_conv))
tf.summary.scalar('cross_entropy',cross_entropy)
train_step=tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)
correct_prediction=tf.equal(tf.argmax(y_conv,1), tf.argmax(y_,1))
accuracy=tf.reduce_mean(tf.cast(correct_prediction, "float"))
tf.summary.scalar('accuacy',accuracy)
sess = tf.InteractiveSession()
sess.run(tf.initialize_all_variables())

saver=tf.train.Saver()
merged=tf.summary.merge_all()
writer=tf.summary.FileWriter('D:\Python\Save\logs',sess.graph)

sess.run(tf.global_variables_initializer())

for i in range(2001):
    batch = mnist.train.next_batch(50)
    if i% 100 == 0:
      train_accuracy = accuracy.eval(feed_dict={x:batch[0], y_: batch[1], keep_prob: 1.0})
      print("step %d, training accuracy %g"%(i, train_accuracy))
    train_step.run(feed_dict={x: batch[0], y_: batch[1], keep_prob: 0.5})
    summary_str = sess.run(merged,feed_dict={x: batch[0], y_: batch[1], keep_prob: 0.5})
    writer.add_summary(summary_str,i)

save_path = saver.save(sess,'D:\Python\Save\save_mnist\mnist.ckpt')
print('save to path: ',save_path)
accuracyResult = list(range(10))
for i in range(10):
    batch = mnist.test.next_batch(1000)
    accuracyResult[i] = accuracy.eval(feed_dict={x:batch[0],y_:batch[1],keep_prob:1.0})
print("Test accuracy:", numpy.mean(accuracyResult))
 

執行的實驗結果：